text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
abstract: |
Low resolution spectra have been used to measure individual metal abundances of RR Lyrae stars in NGC6441, a Galactic globular cluster known to have very unusual horizontal branch morphology and periods of the RR Lyrae stars for its high metallicity. We find an average metal abundance of \[Fe/H\]=$-0.69 \pm 0.06$ (r.m.s.=0.33 dex) and \[Fe/H\]=$-0.41 \pm 0.06$ (r.m.s.=0.36 dex) on Zinn & West and Carretta & Gratton metallicity scales, respectively, consistent with the cluster metal abundance derived by Armandroff & Zinn. Most of the metallicities were extrapolated from calibration relations defined for \[Fe/H\]$\leq -1$; however, they are clearly high and contrast with the rather long periods of the NGC6441 variables, thus confirming that the cluster does not fit in the general Oosterhoff classification scheme. The r.m.s. scatter of the average is larger than observational errors (0.15-0.16 dex) possibly indicating some spread in metallicity. However, even the metal poor variables, if confirmed to be cluster members, are still more metal rich than those commonly found in the Oosterhoff type II globular clusters.\
author:
- 'Gisella Clementini, Raffaele G. Gratton, Angela Bragaglia, Vincenzo Ripepi, Aldo F. Martinez Fiorenzano$^{,}$ Enrico V. Held, Eugenio Carretta'
title: Metal abundances of RR Lyrae stars in the metal rich globular cluster NGC 6441
---
Introduction
============
NGC6441, as well as its twin NGC6388, are metal rich, massive globular clusters (\[Fe/H\]=$-0.53 \pm 0.11$, and $-0.60 \pm 0.15$, respectively: @armandroff88) with very unusual horizontal branches (HBs) extending from stubby red, as expected for their high metallicities, to extremely blue, and with the red HBs sloping upward as one moves blueward in the $V,B-V$ color magnitude diagram (@rich97; @pritzl01 [@pritzl02; @pritzl03]). Given their high metallicities, we would expect them to have very few RR Lyrae stars with the short periods typical of the Oosterhoff type I systems [@oosterhoff39]. Rather unexpectedly, large numbers of RR Lyrae stars with unusually long periods, even longer than those commonly observed in the Oosterhoff type II systems, have been discovered in both clusters (@layden99 [@pritzl01; @pritzl02; @pritzl03]). Indeed, NGC6441 and NGC6388 seem to violate the trend of decreasing period with increasing metallicity followed by the Galactic globular clusters (GCs), and stand apart in the mean period vs. \[Fe/H\] diagram (@pritzl01). The clusters have been suggested to be a further and extreme manifestation of the so called second-parameter effect, meaning that metallicity is not the only factor governing the morphology of the HB, but other parameters such as age, helium or CNO abundances, core rotation, or dynamical interactions are at work. Some of these possibilities (e.g. a high helium abundance, higher interaction rates, etc.) are not supported by observations for NGC6441 and NGC6388 (@rich97 [@layden99; @raimondo02]; Moehler, Sweigart, & Catelan 1999). Thus, we still lack a satisfactory explanation for the cluster’s peculiar HBs and for the unusual properties of their RR Lyrae stars.
A metallicity spread, as first argued by @piotto97 from the intrinsic spread in color of the Red Giant Branch, with the RR Lyrae and the blue HB stars being at the metal-poor tail (\[Fe/H\]$\lesssim -1.6$) of the cluster’s metallicity distributions, could in principle explain their anomalous HBs (@sweigart02 [@ree02]). In this scenario NGC6441 and NGC6388, the two most massive Galactic GCs after $\omega$ Cen and M54, have been suggested to be the relics of disrupted dwarf galaxies [@ree02], similarly to $\omega$ Cen which has a metallicity spread, and to M54 which is considered the nucleus of the Sagittarius galaxy [@layden00]. However, even the metallicity spread does not completely explain the unusual nature of NGC6441 and NGC6388 RR Lyrae stars (see discussions in @pritzl01 [@pritzl02]). On the other hand, if the RR Lyrae in NGC6388 and NGC6441 are metal rich, they would form a new, distinct subclass of long-period, metal rich RR Lyrae stars [@layden99], that has no counterpart among the field and cluster RR Lyrae stars known so far, except perhaps V9 in 47 Tuc (Carney, Storm, & Williams 1993).
No direct measure of the metal abundance of the RR Lyrae stars in either cluster existed so far. @pritzl01 [@pritzl02] derived metallicities for some of the [*ab-*]{}type RR Lyrae stars in the two clusters using the parameters of the Fourier decomposition of the light curve and the Jurcsik-Kovács method [@jurcsik96; @kovacs97]. They estimated average metal abundances of \[Fe/H\]=$-0.99$ and $-1.21$ for NGC 6441 and NGC6388, respectively, that according to @jurcsik95 correspond to $-1.3$ and $-1.4$ on the Zinn & West (1984, hereafter ZW84) metallicity scale. These metallicities are much lower than the cluster metal abundances derived by @armandroff88, but are close to the metallicity of @sweigart02 models which yield a best-fitting model of NGC6441 HB for a \[Fe/H\]$\sim -1.4$ and an $\alpha$ enhancement of \[$\alpha$/Fe\]=+0.3. However, given the uncertainty of the Jurcsik-Kovács method (see discussions in @difabrizio05; @gratton04, hereafter G04; and @clementini05, hereafter C05), and the unusual nature of the NGC 6441 and NGC6388 RR Lyrae stars, there is some question as to the validity of these metallicity determinations.
In this Letter we present the first direct measure of metallicity for RR Lyrae stars in NGC 6441 through spectroscopy, and provide the first quantitative assessment that the cluster variables are indeed metal rich, with a few outliers possibly suggesting some spread in metallicity.
The data
========
Spectra of 12 RR Lyrae stars in NGC 6441 were obtained in July 2003, in the course of a spectroscopic survey of RR Lyrae stars in the Sculptor dwarf spheroidal galaxy (C05). Observations were performed using the FORS2 spectrograph at the ESO Very Large Telescope (VLT, Paranal, Chile). We observed two $6.8\arcmin \times 6.8 \arcmin$ FORS2 subfields of NGC 6441 centered at $\alpha_{2000}$=$17^h 50^m 19.9^s$, $\delta_{2000}$=$-37^\circ 05^\prime 21.9^{\prime\prime}$; $\alpha_{2000}$=$17^h 50^m 24.5^s$, $\delta_{2000}$=$-37^\circ 00^\prime
06.3^{\prime\prime}$, and comprising 5 and 7 RR Lyrae stars, respectively. Spectra were collected using slits 1$^{\prime\prime}$ wide, and about 14$^{\prime\prime}$ long to allow for sky subtraction. With this configuration, each pixel corresponds to 0.75Å. Our wavelength range contains both the Ca [ii]{} K and the hydrogen Balmer lines up to H$\beta$. Exposure time was of 300 sec, as an optimal compromise between S/N and time resolution of the light curve. Details on the observations and data reduction procedures can be found in C05.
Time series photometry for all the target stars and classification in types has been published by @layden99 and @pritzl01. Accordingly, our sample includes 5 [*ab-*]{} and 7 [*c-*]{}type RR Lyrae stars; their identification is provided in Table \[t:NGC6441\]. Figure \[f:fig1\] shows examples of spectra for some of the variable stars in the cluster.
We estimated radial velocities from our spectra; they are given in Column 12 of Table \[t:NGC6441\]. According to C05 typical errors of these radial velocity determinations are of about 15 km s$^{-1}$. Our radial velocity estimates do not exclude the cluster membership for any of the RR Lyrae stars we have analyzed. The 12 stars have $\langle v_{r} \rangle =-$1 km s$^{-1}$ (r.m.s. = 12 km s$^{-1}$, and zero point error of $\pm$7 km s$^{-1}$). Our average radial velocity differs somewhat from the value of +16.4 km s$^{-1}$ (@harris96, on line catalogue, available at http://www.physics.mcmaster.ca/Globular.html). A reason for part of this discrepancy is that our mean value includes the phase-dependent contributions due to the star pulsations. Further residual differences may be due to systematic offsets possibly caused by offcentering of the cluster variables on the slits.
Measure of the metal abundances
===============================
Metal abundances for the RR Lyrae stars in NGC 6441 were derived using a modified (and improved) version of the $\Delta$S method [@preston59]. Our technique is fully described in G04, and is based on the definition of Hydrogen and Ca [ii]{} spectral indices, $\langle H\rangle$ and $K$, for each variable star by directly integrating the instrumental fluxes in spectral bands centered on the H$\delta$, H$\gamma$, H$\beta$, and Ca [ii]{} K lines. These spectral indices are then used to measure metallicities by comparison to the same quantities for variable stars in a number of globular clusters of known metal abundance. A summary of the method and an update of the metallicity calibration procedure can be found in C05. An advantage of our technique is that we do not need to know the phases of our spectra. On the other hand, the accuracy of our \[Fe/H\] may be a function of phase, as represented by the strength of the H lines. According to figure 12 of G04, most accurate metallicity determinations are obtained for values of $\langle H\rangle <$ 0.20, and 0.25 for [*ab-*]{} and [*c-*]{}type RR Lyrae stars, respectively. Outside these ranges metal abundances determinations may be more uncertain, depending on the actual value of $\langle H\rangle $. However, metallicities of individual stars in the calibrating clusters (see Tables 3, 4 in G04, and Tables 5, 6, and 7 in C05) show this effect to be small, if present. This point is particularly relevant for the NGC6441 RR Lyrae stars for which, based on the available photometric data [@layden99; @pritzl01] it may be difficult to reliably define the pulsation phase at the epochs of the observations. Line indices measured for the RR Lyrae stars in NGC 6441 are provided in Columns from 4 to 8 of Table \[t:NGC6441\]. The stellar K values were not corrected for the interstellar K lines contribution since we estimated it to be much less than 0.1 dex in \[Fe/H\]. In fact, given the small absolute value of the cluster radial velocity, interstellar lines are expected to lie in the core of the stellar K-lines where there is almost no flux, and then can subtract only a negligible fraction of flux[^1]. This would not be the case for halo stars where, due to the high velocities, the interstellar absorptions occur outside the line cores.
The calibration of the line indices of the variable stars in terms of metal abundances \[Fe/H\] was obtained using RR Lyrae stars observed in the clusters M15, M2 and NGC6171 (C05), and M68, NGC1851 and NGC3201 (G04). For all these clusters, precise metal abundances are available on both the ZW84 and the Carretta & Gratton (1997, hereafter CG97) metallicity scales. Figure \[f:fig2\] shows the correlation between $K$ and $\langle H\rangle$ indices for the calibrating clusters and, in the bottom-right panel, the position on the $K$ vs $\langle H\rangle$ plane of the NGC6441 variable stars. The figure shows that there are three objects lying near the ridge line of NGC1851 at an average metallicity of \[Fe/H\]$\simeq -1.3$ on the ZW84 scale, while all the remaining stars define a tight correlation at higher metallicity. However, one of the deviating objects, star V41, is probably blended with a cooler companion, as suggested by its spectrum (see Fig. 1). Thus its metallicity is uncertain and we will drop it from any further consideration. V49 was observed in the safe range where metal abundance determinations are most reliable. The spectrum of this star is shown in Fig. \[f:fig1\], along with spectra of all the variables having $\langle H\rangle < $0.25, and clearly shows that the star has a shallower $K$ line. Finally, V74 was observed at $\langle H\rangle$=0.311, thus its metallicity may be slightly more uncertain.
Metallicity indices, $M.I.$’s, for the program stars were derived from their $K$ and $\langle H\rangle$ values using equations (3), (4), and (5) of G04. They are listed in Column 9 of Table \[t:NGC6441\]. Metal abundances \[Fe/H\] were then deduced from the $M.I.$ values using the metallicity calibrations defined by C05. These are described by linear regressions, namely equations (4) and (5) of C05, between the average $\langle M.I.\rangle$ values of RR Lyrae stars in the calibrating clusters M15, M2, NGC6171, M68, NGC1851 and NGC3201 and the cluster’s metal abundances on the ZW84 and CG97 metallicity scales, respectively. In Figure \[f:cal02\] we show the calibration relation of the metallicity indices in the ZW84 metallicity scale. We note that the calibrating clusters have metallicities that do not extend higher than \[Fe/H\]$\sim -1$, thus the metal abundance we derive for NGC6441 is an extrapolation of the calibration equations. However, the cluster is found to fall well on the extrapolation to higher metallicities of the linear relation defined by the calibrating clusters. This results is in agreement with the original Preston’s $\Delta$S method, which appears to have a linear calibration up to about solar metallicity. Individual metal abundances derived with this procedure are given in Columns 10 and 11 of Table \[t:NGC6441\]. According to C05 we attach internal errors of 0.15 and 0.16 dex to individual abundance determinations in ZW84 and CG97 scale, respectively. However, to take into account any additional uncertainty which might affect the spectra corresponding to $\langle H\rangle $ values outside the safest range, in the following we will divide the stars in three groups corresponding to $\langle H\rangle \leq $0.224, 0.257 $ \leq \langle H\rangle \leq $ 0.271, and 0.311 $ \leq \langle H\rangle
\leq $ 0.329, and all averages will be computed giving different weights to the 3 groups of stars (i.e. 1, 0.75 and 0.5).
For comparison, in the last column of Table \[t:NGC6441\] we report the metal abundances derived by @pritzl01 from the Fourier decomposition of the light curve for 4 of the [*ab-*]{}type RR Lyrae stars analyzed here. The average of @pritzl01 values for the 4 stars is \[Fe/H\]=$-1.02 ~~(r.m.s.=0.15)$, in @jurcsik95 metallicity scale. According to equation (4) in @jurcsik95 this corresponds to \[Fe/H\]=$-1.33$, on the ZW84 scale. The average of our metal abundances for these 4 stars is \[Fe/H\]=$-0.67 ~~(r.m.s.=0.11)$, on the ZW84 scale, i.e. much more metal rich. This result seems to suggest that the Jurcsik-Kovács method may not be reliable when applied to the anomalous RR Lyrae stars in NGC6441.
Adopting the averaging scheme described above, the mean metal abundance of our RR Lyrae sample in NGC6441 is \[Fe/H\]=$-0.69 \pm 0.06$ (r.m.s.=0.33 dex, 11 stars), and $-0.41 \pm 0.06$ (r.m.s.=0.36 dex) in ZW84 and CG97 scale, respectively. The scatter of the average is larger than expected from measurement errors alone ($<$0.2 dex) thus suggesting that, if the 11 variables are all cluster members, there is some spread in metallicity with two more metal poor objects at an average metal abundance around \[Fe/H\]$\sim -1.3$, and the remaining 9 more metal rich stars around \[Fe/H\]$\sim -0.6$ dex (on the ZW84 scale). If the two metal poor stars are disregarded, the mean metal abundances become \[Fe/H\]=$-0.57\pm 0.04$ (r.m.s.=0.19 dex, 9 stars), and $-0.28 \pm 0.04$ (r.m.s.=0.21 dex). The former value is in excellent agreement with both ZW84 and @armandroff88 metallicity estimates for the cluster.
Summary and conclusions
=======================
Metal abundances from low resolution spectroscopy obtained with FORS2 at the ESO VLT have revealed that the NGC6441 RR Lyrae stars are metal rich, with an average metal abundance of \[Fe/H\]=$-0.69 \pm 0.06$, on ZW84 scale. The spectroscopic analysis also reveals that there are two variables (out of eleven) having metallicities around \[Fe/H\]$\sim -1.3$. However, the metal poor stars, if confirmed cluster members, are a minor component of our sample. Even allowing for measurement errors, we do not find in this cluster RR Lyrae stars as metal poor as the variables commonly found in the Galactic Oosterhoff type II GCs, as instead one would expect from the extraordinarily long periods and the position near the Oosterhoff II line of the NGC6441 variables in the period-amplitude diagram (see figure 6 of @pritzl03). Clearly, metal abundances and memberships for a larger number of variable stars are needed to better assess the metallicity distribution of the NGC6441 stars and the relevance of the metal poor component, if any. Nevertheless, the existence of extremely long period RR Lyrae stars with extraordinarily high metal abundances, as some of the variables in our sample, demonstrates that the NGC6441 variable stars are different from the RR Lyrae stars known so far both in the Milky Way GCs and in the field, and confirms that this cluster does not conform to the Oosterhoff dichotomy described by the other Galactic GCs.
Which mechanism may be able to produce such metal rich variables with pulsation characteristics similar to the Oosterhoff type II ones remains unexplained. For instance, @pritzl02 show that it is difficult to model NGC6441 as an Oosterhoff II system under the hypothesis that its variables are evolved from a position on the blue zero-age HB, as a result of the small number of progenitors on the blue HB. As suggested by @layden99 and @pritzl01, the theoretical reproduction of the observed light curves with pulsation models may shed some light on the physical properties responsible for the anomalous properties of the NGC6441 RR Lyrae stars. Such modeling for the variables analyzed in the present paper is currently under way (Clementini & Marconi 2005, in preparation). This research was funded by MIUR, under the scientific project: 2003029437, “Continuity and Discontinuity in the Milky Way Formation” (P.I.: Raffaele Gratton).
Armandroff, T.E., & Zinn, R. 1988, AJ, 96, 92
Carney, B.W., Storm, J., & Williams, C. 1993, PASP, 105, 294
Carretta, E., & Gratton, R.G. 1997, A&AS, 121, 95 (CG97)
Clement, et al. 2001, AJ, 122, 2587
Clementini, G., Ripepi, V., Bragaglia, A., Martinez Fiorenzano, A., Held, E.V., Gratton, R.G. 2005, MNRAS, in press (astro-ph/0506206; C05)
Di Fabrizio, L. Clementini, G., Maio, M., Bragaglia, A., Carretta, E., Gratton, R., Montegriffo, P., Zoccali, M. 2005, A&A, 430, 603
Gratton, R.G., Bragaglia A., Clementini, G., Carretta, E., Di Fabrizio, L., Maio, M., Taribello, E. 2004, A&A, 421, 937 (G04)
Harris, W.E. 1996, , 112, 1487,
Jurcsik, J. 1995, Acta Astron., 45, 653
Jurcsik, J., & Kovács, G. 1996, A&A, 312, 111
Kovács, G., & Jurcsik, J., 1997, A&A, 322, 218
Layden, A.C., Ritter, L.A., Welch, D.L., Webb, T.M.A. 1999, , 117, 1313
Layden, A.C., & Sarajedini, A. 2000, , 119, 1760
Moehler, S., Sweigart, A.V., & Catelan, M. 1999, A&A, 351, 519
Oosterhoff, P.Th. 1939, Observatory, 62, 104
Piotto, G. et al. 1997, in Advances in Stellar Evolution, ed. R.T. Rood & A. Renzini (Cambridge:Cambridge Univ. Press), 84
Preston, G.W., 1959, , 130, 507
Pritzl, B.J., Smith, H.A., Catelan, M., Sweigart, A.V. 2001, , 122, 2600 (P01)
Pritzl, B.J., Smith, H.A., Catelan, M., Sweigart, A.V. 2002, 124, 949
Pritzl, B.J., Smith, H.A., Stetson, P.B., Catelan, M., Sweigart, A.V., Layden A.C., Rich R.M. 2003, , 126, 1381
Raimondo, G., Castellani, V., Cassisi, S., Brocato, E., & Piotto, G. 2002, , 569, 975
Ree, C.H., Yoon, S.-J., Lee, S.-C., Lee, Y.-W. 2002, in ASP Conf. Ser. 265, Omega Centauri: A Unique Window Into Astrophysics, ed. F. van Leeuwen, J.D. Hughes, & G. Piotto, (San Francisco: ASP), 101
Rich, R.M. et al. 1997, , 484, L25
Sweigart, A.V. 2002, in Highlights of Astronomy, ed. H. Rickman (San Francisco: ASP), 12, 292
Zinn, R., West, M.J. 1984, , 55, 45 (ZW84)
[^1]: We checked this point on the spectra of a few NGC6441 giants we recently observed with UVES@VLT. We estimated that, according to the definition of the spectral indices used in the present study, the interstellar K line contribution may cause a systematic offset $<$ 0.02 dex in the metal abundance of the NGC6441 variable stars.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this article we solve the problem of maximizing the expected utility of future consumption and terminal wealth to determine the optimal pension or life-cycle fund strategy for a cohort of pension fund investors. The setup is strongly related to a DC pension plan where additionally (individual) consumption is taken into account. The consumption rate is subject to a time-varying minimum level and terminal wealth is subject to a terminal floor. Moreover, the preference between consumption and terminal wealth as well as the intertemporal coefficient of risk aversion are time-varying and therefore depend on the age of the considered pension cohort. The optimal consumption and investment policies are calculated in the case of a Black-Scholes financial market framework and hyperbolic absolute risk aversion (HARA) utility functions. We generalize Ye (2008) (2008 American Control Conference, 356-362) by adding an age-dependent coefficient of risk aversion and extend Steffensen (2011) (Journal of Economic Dynamics and Control, 35(5), 659-667), Hentschel (2016) (Doctoral dissertation, Ulm University) and Aase (2017) (Stochastics, 89(1), 115-141) by considering consumption in combination with terminal wealth and allowing for consumption and terminal wealth floors via an application of HARA utility functions. A case study on fitting several models to realistic, time-dependent life-cycle consumption and relative investment profiles shows that only our extended model with time-varying preference parameters provides sufficient flexibility for an adequate fit. This is of particular interest to life-cycle products for (private) pension investments or pension insurance in general.'
address:
- 'Department of Mathematics, Technical University of Munich, Munich, Germany'
- 'Department of Actuarial Studies and Business Analytics, Macquarie University, Sydney, Australia'
author:
- Andreas Lichtenstern
- 'Pavel V. Shevchenko'
- Rudi Zagst
bibliography:
- 'Bibliography.bib'
title: 'Optimal life-cycle consumption and investment decisions under age-dependent risk preferences[^1]'
---
Pension investments ,optimal life-cycle consumption and investment ,age-dependent risk aversion ,HARA utility function ,martingale method
G11 ,G22 ,C61 ,D14
Introduction {#sec:Introduction}
============
A suitable management of pensions needs to consider earnings/contributions and investment, but should also account for the required consumption during the accumulation and/or decumulation phase. For this sake, in this paper we consider the finite horizon portfolio problem of maximizing expected utility of future consumption and terminal wealth to determine the optimal pension or life-cycle fund strategy for a cohort of pension fund investors. The setup is strongly related to a DC pension plan where additionally (individual) consumption is taken into account. Within this framework, [@LaknerNygren2006] describe the trade-off the investor faces as a compromise between ‘living well’ (consumption) and ‘becoming rich’ (terminal wealth). Classical consumption-investment problems consider constant risk aversion in the intertemporal utility functions for consumption besides a personal discount rate or impatience factor, see [@Merton1969] or [@Merton1971]. Within classical models (where constant relative risk aversion (CRRA) utilities are applied), optimal portfolio policies turn out to be constant over the life-cycle, meaning time and wealth independent. According to [@Aase2017] this is ‘against empirical evidence, and against the typical recommendations of portfolio managers’. Furthermore, [@Aase2017] and [@YangFang2014] argue that the tendency of stocks to outperform bonds over long horizons in the past is one of the reasons why people at a younger age are advised to allocate a higher proportion of wealth to equities compared to older people. Evidence for changing risk aversion over the life-cycle is reported in the literature, although there is no broad agreement on its behavior: [@Morin1983], [@Bakshi1994], [@Palsson1996], [@Bellante2004], [@AlAjmi2008], [@Ho2009], [@Yao2011] and [@AlbertDuffy2012] observe increasing risk aversion by age, [@Bellante1986] and [@Wang1997] find risk aversion decreasing by age and [@Riley1992] detect different behavior between the pre- and post-retirement phase. Age-depending risk preferences can economically be motivated by the observed behavior of people to stepwise reduce their investment risk the closer to retirement. This behavior is reflected in many life-cycle fund allocation policies, see for instance [@Milliman2010] or [@TIAACREF_LifecycleFunds2015]. An important economic reasoning behind is that the older the person, the less time to retirement entrance is left and therefore the less likely it is for her to overcome a potential market crash, strongly connected to the fear of having an insufficient wealth left for retirement. Moreover, it is reasonable that the closer to retirement time, the more satisfaction is connected with savings, i.e. with a lower consumption surplus, which yields a higher initial wealth for the decumulation phase. Based on these economic reasons, it is meaningful to consider age-varying preference parameters (dependent on the age of the pension cohort or the individual investor) in form of a coefficient of risk aversion, later called $b(t)$, and a weighting factor, later referred to as $a(t)$, that governs the relative importance of consumption at different points in time. The latter has no impact on risk aversion but can control for the varying preference between consumption and terminal wealth over time. In an analysis of the optimal controls in Section \[sec:NCS\] we show that our proposed model can explain and describe people’s observed behavior of reducing relative risky investments by time while simultaneously targeting a certain function for the consumption rate on average. In opposite, we find that the previously described existing models are not able to capture this behavior. Therefore, particulary Section \[sec:NCS\] shows that it is economically important to have separate functions or parameters for risk aversion and preference of consumption over terminal wealth, $a(t)$ and $b(t)$.
In addition, consumption and wealth floors are introduced which have an economic meaning as minimum required levels of consumption and wealth. This motivates the development of a dynamic life-cycle model with time-varying risk preferences such as coefficient of risk aversion and consumption and wealth floors which can capture age-depending consumption and investment behavior of investors.
Related literature to this topic consider stochastic income and unemployment risks, see [@BodieMertonSamuelson1992], [@Koo1998], [@Munk2000], [@Viceira2001], [@Huang2008], [@Jang2013], [@Bensoussan2016], [@Wang2016] or [@Chen2018]. Setups where the investor faces uncertain lifetime, mortality and optimal life insurance are considered in [@Yaari1965], [@PliskaYe2007], [@MenoncinRegis2017], [@ZouCadenillas2014], [@KronborgSteffensen2015], [@ShenWei2016], [@Duarte2011], [@Huang2012], [@KronborgSteffensen2015], [@ShenWei2016] and [@Ye2008]; optimal consumption and investment under insurer default risk is studied by [@JangKooPark2019]. [@KraftMunk2011], [@KraftMunkWagner2018], [@AndreassonShevchenkoNovikov2017], [@CuocoLiu2000] and [@DamgaardFuglsbjergMunk2003] analyze optimal housing as a durable good. Constraints in the optimization problem are considered in [@Cuoco1997], [@Elie2008] and [@Grandits2015]. Moreover, [@Akian1996], [@Altarovici2017] and [@Dai2009] analyze the portfolio problem under transaction costs. The application of HARA utility functions in a life-cycle context can be found in [@Huang2008], [@Ye2008], [@ChangRong2014], [@ChangChang2017] and [@Wang2017]. Moreover, [@BackLiuTeguia2019] study a life-cycle consumption problem for HARA utility with time-independent, increasing risk aversion and examine the relation between age and portfolio risk by using Monte Carlo analysis. [@TangPurcalZhang2018] study an optimal consumption-investment problem under CRRA utility function with age-independent risk aversion, but examine the impact of hyperbolic discounting, where the rate of time preference is a function of time. We generalize this approach by considering general $a(t)$ or ${e^{- \beta t} a(t)}$, respectively, and by introducing age-varying risk aversion.
In this paper we apply HARA utility functions on both the consumption and terminal wealth and consider time-varying preferences: an age-depending preference between consumption and terminal wealth and an age-depending coefficient of risk aversion in the intertemporal consumption utility. For simplicity, income is treated as a deterministic process. Furthermore, we do not model mortality and consider a fixed time horizon $T$ that corresponds to a retirement age, thus we assume the agent to survive up to the age of retirement. A positive, fixed floor in the terminal utility ensures a minimum liquid asset wealth level at the age of retirement, which is meaningful as the retiree needs wealth to live from and could possibly afford housing from this wealth. In addition, a positive, time-varying floor in the consumption utility guarantees a minimum (time-dependent) consumption rate. This is essential during the accumulation phase as for instance living expenses, rental payments when home is rented or mortgage payments and maintenance costs when home is bought and financed by debt or only maintenance costs when the agent already fully owns a house (e.g. inherited) need to be covered. Therefore, the economic demand for both a positive minimum level of consumption and terminal wealth can be motivated.
Most related to our work are [@Ye2008], [@Steffensen2011], [@Hentschel2016] and [@Aase2017]. The difference of our approach to these papers is as follows. [@Ye2008] considers income, mortality and HARA utilities for both consumption and terminal wealth under a constant coefficient of risk aversion, i.e. constant $b(t)$, but where the age-dependent preference between consumption and wealth $a(t)$ is incorporated. We generalize the results by introducing a time-dependent coefficient of risk aversion $b(t)$. [@Steffensen2011] provides a first insight into the optimal policy when the utility parameters of the intertemporal utility, which is of a CRRA type, are time-varying; thus $a(t)$ and $b(t)$ are captured. But the model disregards terminal wealth, consumption floor and labor income. In a similar fashion, [@Hentschel2016] studies the consumption problem for CRRA utility with habit formation and considers $a(t)$ and $b(t)$. Similar to [@Steffensen2011], neither terminal wealth nor consumption floor nor income are included in their model. Finally, [@Aase2017] uses the martingale method (that allows to reformulate the optimal stochastic control problem to a simpler maximization problem with constraint) to determine optimal consumption and investment under mortality and a CRRA utility with age-depending risk aversion $b(t)$. But the model does not consider terminal wealth, consumption floor, income or time-varying preference $a(t)$.
The main contributions and innovations of this paper can be summarized as follows: we consider all the ‘ingredients’ of the models in the above mentioned papers ($a(t)$, $b(t)$, terminal wealth, floors for consumption and terminal wealth via HARA utilities, income process) that leads to a novel, very flexible and more realistic dynamic life-cycle model framework. We extend or generalize [@Ye2008] by adding an age-dependent coefficient of risk aversion $b(t)$ and [@Steffensen2011], [@Hentschel2016] and [@Aase2017] by considering terminal wealth and allowing for consumption and terminal wealth floors via an application of HARA utility functions. The corresponding consumption-investment problem is solved analytically and interpretation is provided. In a case study, where we fit realistic predetermined target policies for consumption and relative allocation to several models, we realize that only our proposed and most general model is sufficiently flexible to describe human preferences on consumption and investment in a suitable fashion. This implies that modeling the agent’s preferences in an age-depending fashion is inevitable.
To solve the respective portfolio problem, we follow a separation approach similar to the ones developed by [@KaratzasShreve1998] and [@LaknerNygren2006]. It divides the original consumption-terminal wealth optimization problem into two sub-problems, the corresponding consumption problem and the terminal wealth problem. These separate problems are to be solved individually. Due to time-dependent preference parameters we apply the martingale method in line with [@Aase2017] to solve the individual problems in closed form. Afterwards, we show how the individual solutions have to be glued together in order to obtain the general solution to the original consumption-terminal wealth problem.
The remainder of this paper is organized as follows. Section \[sec:FinancialMarketModel\] introduces the financial market and the portfolio problem of interest, Section \[sec:SeparationTechnique\] shows the separation approach and the solution to the problem. A fit of the analytic strategy to suitable consumption and investment curves is conducted in Section \[sec:NCS\], followed by an investigation of the optimal controls and corresponding wealth process. Section \[sec:Conclusion\] concludes. \[app:Proofs\] summarizes all proofs of the claimed statements: the proofs for Section \[sec:ConsumptionProblem:y\] on the consumption problem can be found in \[app:ProofsConsumptionProblem\], the proofs related to Section \[sec:TerminalWealthProblem\] on the terminal wealth problem in \[app:ProofsTerminalWealthProblem\], and for the proofs associated with Section \[sec:OptimalMerging\] on merging both individual solutions, see \[app:ProofsOptimalMerging\].
The financial market model and consumption-investment problem {#sec:FinancialMarketModel}
=============================================================
We consider a frictionless financial market $M$ which consists of $N+1$ continuously traded assets, one risk-free asset and $N$ risky assets. Let $[0,T]$ represent the fixed and finite investment horizon. Uncertainty in the continuous-time financial market is modeled by a complete, filtered probability space $(\Omega, \CMcal{F}, \left(\CMcal{F}_{t}\right)_{t \in [0,T]}, {\mathbb{P}})$, where $\Omega$ is the sample space, ${\mathbb{P}}$ the real-world probability measure, $\CMcal{F}_{t}$ is the natural filtration generated by $W(s)$, ${0 \le s \le t}$, augmented by all the null sets, and ${W = \left(W(t)\right)_{t \in [0,T]}}$, $W(t) = (W_{1}(t), \hdots, W_{N}(t))'$, $N \in \mathbb{N}$, is a standard $N$-dimensional Brownian motion. The price of the risk-free asset at time $t$ is denoted by $P_{0}(t)$ and is subject to the equation $$\begin{aligned}
dP_{0}(t) = r P_{0}(t)dt,\ P_{0}(0) = 1,\end{aligned}$$
with constant risk-free interest rate $r > 0$. The remaining $N$ assets in the market are risky assets with price $P_{i}(t)$, $i = 1, \hdots, N$, at time $t$ subject to the stochastic differential equations
$$\begin{aligned}
dP_{i}(t) = {} & P_{i}(t) \left(\mu_{i} dt + \sigma_{i} dW(t)\right) = P_{i}(t) \left(\mu_{i} dt + \sum_{j = 1}^{N} \sigma_{ij} dW_{j}(t)\right),\ P_{i}(0) = p_{i} > 0,\end{aligned}$$
with constant drift ${\mu = \left(\mu_{1},\hdots,\mu_{N}\right)' \in \mathbb{R}_{+}^{N}}$, ${\mu - r \mathbf{1} > \mathbf{0}}$, and constant volatility vector $\sigma_{i} = (\sigma_{i1}, \hdots, \sigma_{iN}) \in \mathbb{R}_{+}^{1 \times N}$. The volatility matrix is defined by ${\sigma = \left(\sigma_{ij}\right)_{i,j = 1,\hdots,N}}$, the covariance matrix of the log-returns is ${\Sigma = \sigma \sigma'}$ which is assumed to be strongly positive definite, i.e. there exists $K > 0$ such that ${\mathbb{P}}$-a.s. it holds ${x' \Sigma x \ge K x' x}$, ${\forall x \in {\mathbb{R}}^{N}}$. Furthermore, let ${\gamma = \sigma^{-1} (\mu - r \mathbf{1})}$ denote the market price of risk. In this case of Black-Scholes market dynamics, according to [@KaratzasShreve1998], there exists a unique risk-neutral probability measure ${{\mathbb{Q}}\sim {\mathbb{P}}}$ defined by ${\frac{d{\mathbb{Q}}}{d{\mathbb{P}}} | _{\CMcal{F}_{t}} := e^{-\frac{1}{2} \|\gamma\|^{2} t - \gamma' W(t)}}$ and the market is complete (that allows to value payment streams under the measure ${\mathbb{Q}}$ as expected discounted values, meaning that the cost of a portfolio replicating the contract is given by its expected discounted value under ${\mathbb{Q}}$). The corresponding pricing kernel or state price deflator, denoted by $\tilde{Z}(t)$, is defined as $$\begin{aligned}
\label{eq:PricingKernel}
\tilde{Z}(t) := e^{- \left(r + \frac{1}{2}\|\gamma\|^{2}\right) t - \gamma' W(t)}\end{aligned}$$ and can be used for the valuation of payment streams under the real-world probability measure. Its dynamics are subject to the stochastic differential equation $$\begin{aligned}
d \tilde{Z}(t) = - \tilde{Z}(t) \left(r dt + \gamma' dW(t)\right),\ \tilde{Z}(0) = 1.\end{aligned}$$ We consider $\CMcal{F}_{t}$-progressively measurable trading strategies $\varphi = (\varphi_{0}, \hat{\varphi})'$, ${\hat{\varphi} = (\varphi_{1}, \hdots, \varphi_{N})'}$, such that ${\mathbb{P}}$-a.s. it holds $\int_{0}^{T} |\varphi_{0}(t)| dt < \infty$ and $\int_{0}^{T} \varphi_{i}(t)^{2} dt < \infty$. $\varphi_{i}(t)$ represents the number of individual shares of asset $i$ held by the investor at time $t$. The corresponding relative portfolio process is denoted by ${\pi = (\pi_{0}, \hat{\pi}')'}$ with risky relative investment ${\hat{\pi} = (\pi_{1}, \hdots, \pi_{N})'}$ and risk-free relative investment ${\pi_{0}(t) = 1-\hat{\pi}(t)'\mathbf{1}}$, where $\pi_{i}(t)$ denotes the fraction of wealth allocated to asset $i$ at time $t$. It is to satisfy $\int_{0}^{T} \pi_{i}(t)^{2} dt < \infty$, ${\mathbb{P}}$-a.s.. Moreover, let $(c(t))_{t \in [0,T]}$ denote a non-negative, progressively measurable, real-valued stochastic consumption rate process with $\int_{0}^{T} c(t) dt < \infty$, ${\mathbb{P}}$-a.s., and $(y(t))_{t \in [0,T]}$ a non-negative, deterministic income-rate process with $\int_{0}^{T} y(t) dt < \infty$. Those technical conditions are assumed to get a solution for the subsequently formulated stochastic problem. The dynamics of the investor’s wealth process $V = \left(V(t)\right)_{t \in [0,T]}$ under the strategy $(\pi,c)$ to initial wealth $V(0) = v_{0} > 0$, including liquid assets, consumption and income, is then given by $$\begin{aligned}
\label{eq:SDE:V:y}
dV(t) = {} & V(t) \left[\left(r + \hat{\pi}(t)' \left(\mu - r \mathbf{1}\right)\right) dt + \hat{\pi}(t)'\sigma dW(t)\right] - c(t) dt + y(t) dt.\end{aligned}$$ The relative investment in the risk-free asset is ${\pi_{0}(t) = 1 - \hat{\pi}(t)' \mathbf{1}}$. We consider the objective of maximizing expected utility of future terminal wealth and consumption, starting at time $0$ and ending at $T$. Hence the objective function to be maximized is $$\begin{aligned}
\label{eq:ObjectiveFunction}
J(\pi,c;v_{0}) = {\mathbb{E}}\left[\int_{0}^{T} U_{1}(t,c(t)) dt + U_{2}(V(T))\right],\end{aligned}$$ where $v_{0} > 0$ denotes the initial endowment of the investor. All expectations in this paper are with respect to the real-world measure ${\mathbb{P}}$. The general portfolio optimization problem with initial wealth ${V(0) = v_{0} > 0}$ to be solved is then given by $$\begin{aligned}
\label{eq:OptimizationProblem}
\mathcal{V}(v_{0}) = \sup_{(\pi,c) \in \Lambda} J(\pi,c;v_{0})\end{aligned}$$ subject to . $\mathcal{V}(v_{0})$ is the value function of the problem. $\Lambda$ denotes the set of admissible strategies $(\pi,c)$ such that ${V(t) + \int_{t}^{T} e^{- r (s-t)} y(s) ds \ge 0}$, ${\mathbb{P}}$-a.s., $\forall t \in [0,T]$, and which admit a unique solution to while satisfying the integrability condition ${{\mathbb{E}}\left[\int_{0}^{T} |U_{1}(t,c(t))| dt\right] < \infty}$. The so-called budget constraint reads $$\begin{aligned}
\label{eq:BudgetConstraint:y}
{\mathbb{E}}\left[\int_{0}^{T} \tilde{Z}(t) c(t) dt + \tilde{Z}(T) V(T)\right] \le v_{0} + {\mathbb{E}}\left[\int_{0}^{T} \tilde{Z}(t) y(t) dt\right] = v_{0} + \int_{0}^{T} e^{-r t} y(t) dt.\end{aligned}$$ It describes the requirement that today’s value of future consumption and terminal wealth, less income, must not exceed the initial endowment. It can be shown that for the optimal $(\hat{\pi}^{\star}, c^{\star})$ to Problem , Equation holds with equality. We consider a preference utility model given by the utility functions $$\begin{aligned}
\begin{split} \label{eq:utilitymodel:new}
U_{1}(t,c) = {} & \left(e^{- \beta t} a(t)\right) \frac{1-b(t)}{b(t)} \left(\frac{1}{1-b(t)} \left(c - \bar{c}(t)\right)\right)^{b(t)}, \\
U_{2}(v) = {} & e^{- \beta T} \hat{a} \frac{1-\hat{b}}{\hat{b}} \left(\frac{1}{1-\hat{b}} (v-F)\right)^{\hat{b}},
\end{split}\end{aligned}$$ for $\beta \ge 0$, $b : [0,T] \to (- \infty, 1)\backslash \{0\}$ continuous, $\hat{b} < 1$, $\hat{b} \neq 0$, $a(t) > 0$, $\hat{a} > 0$, $c(t) > \bar{c}(t)$, $\bar{c}(t) \ge 0$ deterministic, and $v > F$ with $F \ge 0$. $U_{2}$ is a continuously differentiable and strictly concave terminal utility function, $U_{1}$ denotes a continuous (intertemporal consumption) utility function which is continuously differentiable and strictly concave in the second argument. This utility model accounts for several desired aspects: minimum liquid asset wealth level $F \ge 0$ at the age of retirement $T$, minimum consumption rate $\bar{c}(t) \ge 0$ and time-varying preference of consumption over terminal wealth in terms of $a(t)$. Moreover, the coefficient of risk aversion $b(t)$ in the consumption utility is now a continuous function in time.
Notice that the associated *Arrow-Pratt measure* ${\mathcal{A}(v) := - \frac{U''(v)}{U'(v)} = - \frac{\partial}{\partial v} \ln U'(v)}$ of absolute risk aversion, developed by [@Pratt1964] and [@Arrow1970], admits the following hyperbolic representation $$\begin{aligned}
\mathcal{A}_{1}(t,c) = \frac{1 - b(t)}{c - \bar{c}(t)},\ \mathcal{A}_{2}(v) = \frac{1 - \hat{b}}{v - F}.\end{aligned}$$ For this reason, we use the notation of an increasing $b(t)$ as a synonym for a decreasing coefficient of risk aversion and vice versa. Further note that $a(t)$ does not appear in $\mathcal{A}_{1}(t,c)$. Therefore we have two input functions $a(t)$ and $b(t)$ where $a(t)$ has no influence on risk aversion, but $b(t)$ determines it; hence a very flexible model.
Since we have $c(t) > \bar{c}(t)$ and $V(T) > F$ by definition of the utility functions in , we restrict $$\begin{aligned}
\label{eq:Condition:v0:minimalrequirement}
v_{0} > \int_{0}^{T} e^{- r s} \left(\bar{c}(s) - y(s)\right) ds + e^{- r T} F =: F(0)\end{aligned}$$ on the initial endowment in . It is useful to define $$\begin{aligned}
F(t) = {} & {\mathbb{E}}\left[\int_{t}^{T} \frac{\tilde{Z}(s)}{\tilde{Z}(t)} \bar{c}(s) ds + \frac{\tilde{Z}(T)}{\tilde{Z}(t)} F - \int_{t}^{T} \frac{\tilde{Z}(s)}{\tilde{Z}(t)} y(s) ds \Big| \mathcal{F}_{t}\right] \nonumber \\
= {} & \int_{t}^{T} {\mathbb{E}}\left[\frac{\tilde{Z}(s)}{\tilde{Z}(t)} \Big| \mathcal{F}_{t}\right] \left(\bar{c}(s) - y(s)\right) ds + F {\mathbb{E}}\left[\frac{\tilde{Z}(T)}{\tilde{Z}(t)} \Big| \mathcal{F}_{t}\right] = \int_{t}^{T} e^{- r (s-t)} \left(\bar{c}(s) - y(s)\right) ds + e^{- r (T-t)} F. \label{eq:def:F(t)}\end{aligned}$$ $F(t)$ can be interpreted as the time $t$ value of all future minimal liabilities less income. $F(t)$ equals the sum of the time $t$ wealth necessarily required to meet all the future minimum living expenses and expenditures $\bar{c}(s)$, ${s \in [t,T]}$ during the remaining time and the time $t$ value of the minimum desired terminal wealth level $F$; future salary income is subtracted as it reduces the time $t$ value of the minimum required capital.
Solution: Separation technique {#sec:SeparationTechnique}
==============================
In the sequel we follow the separation technique approach by [@KaratzasShreve1998] and [@LaknerNygren2006] for solving the consumption-terminal wealth problem as defined by . We split the problem into two sub-problems: the consumption-only and terminal wealth-only problem. Both individual problems are separately solved via the martingale method, similar to the approach by [@Aase2017]. The individual problem solutions are optimally merged at the end. For this sake, let us consider the two individual problems first.
The consumption problem {#sec:ConsumptionProblem:y}
-----------------------
The consumption-only problem is $$\begin{aligned}
\begin{split} \label{eq:OptimizationProblem:ConsumptionOnly}
J_{1}(\pi,c;v_{1}) = {} & {\mathbb{E}}\left[\int_{0}^{T} U_{1}(t,c(t)) dt\right], \\
\mathcal{V}_{1}(v_{1}) = {} & \sup_{(\pi,c) \in \Lambda_{1}} J_{1}(\pi,c;v_{1})
\end{split}\end{aligned}$$ subject to the budget constraint $$\begin{aligned}
\label{eq:BudgetConstraint:ConsumptionOnly:y}
{\mathbb{E}}\left[\int_{0}^{T} \tilde{Z}(t) c(t) dt\right] \le v_{1} + {\mathbb{E}}\left[\int_{0}^{T} \tilde{Z}(t) y(t) dt\right] = v_{1} + \int_{0}^{T} e^{-r t} y(t) dt.\end{aligned}$$ $\Lambda_{1}$ denotes the set of admissible strategies $(\pi,c)$ such that ${V(t) + \int_{t}^{T} e^{- r (s-t)} y(s) ds \ge 0}$, ${\mathbb{P}}$-a.s., $\forall t \in [0,T]$, and which admit a unique solution to while satisfying ${{\mathbb{E}}\left[\int_{0}^{T} |U_{1}(t,c(t))| dt\right] < \infty}$.
[@Steffensen2011] provides a proof for CRRA utility functions by solving the associated Hamilton-Jacobi-Bellman (HJB) equation. We follow the approach by [@Aase2017], likewise for a HARA utility function. We extend the findings of [@Aase2017] by introducing a time-varying, deterministic consumption floor $\bar{c}(t)$, a time-varying preference function $a(t)$ of consumption over terminal wealth and an income-rate process $y(t)$.
In order to guarantee the consumption rate floor, note $c(t) > \bar{c}(t)$, let us assume the following lower boundary for $v_{1}$ which equals the integral over the discounted consumption floor rate minus income rate over the whole horizon of interest: $$\begin{aligned}
\begin{split} \label{eq:Condition:v1_&_eq:def:F1(t)}
v_{1} > {} & \int_{0}^{T} e^{- r s} \left(\bar{c}(s) - y(s)\right) ds =: F_{1}(0), \\
F_{1}(t) := {} & {\mathbb{E}}\left[\int_{t}^{T} \frac{\tilde{Z}(s)}{\tilde{Z}(t)} \left(\bar{c}(s) - y(s)\right) ds \Big| \mathcal{F}_{t}\right] = \int_{t}^{T} e^{- r (s-t)} \left(\bar{c}(s) - y(s)\right) ds.
\end{split}\end{aligned}$$ Notice that ${v_{1} < 0}$ is possible since a sufficiently large positive income stream can be high enough to finance consumption. Using the martingale method we solve the problem as summarized by the theorem below.
\[thm:Solution:ConsumptionOnly\] The solution to the optimal stochastic control problem with intertemporal utility function $U_{1}$ in is $$\begin{aligned}
\hat{\pi}_{1}(t ; v_{1}) = {} & \frac{1}{1 - b(\tilde{t}_{1})} \Sigma^{-1} (\mu - r \mathbf{1}) \frac{V_{1}(t ; v_{1}) - F_{1}(t)}{V_{1}(t ; v_{1})}, \\
c_{1}(t;v_{1}) = {} & g(t,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(t)-1}} + \bar{c}(t) = (1-b(t)) \left(\lambda_{1} \frac{e^{\beta t}}{a(t)} \tilde{Z}(t)\right)^{\frac{1}{b(t)-1}} + \bar{c}(t), \\
V_{1}(t ; v_{1}) = {} & \int_{t}^{T} g(s,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(s)-1}} ds + F_{1}(t), \\
V_{1}(T ; v_{1}) = {} & 0,\end{aligned}$$ for all $t \in [0,T]$, where $$\begin{aligned}
g(s,t; v_{1}) = (1-b(s)) \left(\frac{e^{\beta s - b(s) \left(r - \frac{1}{2} \frac{1}{b(s)-1} \|\gamma\|^{2}\right) (s-t)}}{a(s)}\right)^{\frac{1}{b(s)-1}} \lambda_{1}^{\frac{1}{b(s)-1}}.\end{aligned}$$ $\lambda_{1} = \lambda_{1}(v_{1}) > 0$ satisfies the budget constraint uniquely and is subject to the equation $$\begin{aligned}
\label{eq:ConsumptionOnly:lambda}
\int_{0}^{T} g(t,0; v_{1}) dt = v_{1} - F_{1}(0).\end{aligned}$$ $\tilde{t}_{1} = \tilde{t}_{1}(v_{1}) \in (t,T)$ is the solution to the equation $$\begin{aligned}
\label{eq:ConsumptionOnly:tautilde}
\int_{t}^{T} \frac{1}{b(s)-1} g(s,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(s)-1}} ds = \frac{1}{b(\tilde{t}_{1})-1} \int_{t}^{T} g(s,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(s)-1}} ds.\end{aligned}$$ For the optimal $c_{1}(t;v_{1})$, Equation is fulfilled with equality.
We remind the reader that all proofs can be found in \[app:Proofs\]. It is clear that ${c_{1}(t;v_{1}) > \bar{c}(t)}$, a.s.. We now aim to interpret the optimal investment strategy as proportional portfolio insurance (PPI) strategy. The first strategy family corresponds to a constant multiple, the latter one is more general and also covers proportional strategies with time-varying or even state-dependent multiples. [@ZielingMahayniBalder2014] evaluate the performance of such strategies. Theorem \[thm:Solution:ConsumptionOnly\] shows that the optimal investment strategy generally is a PPI strategy with time-varying floor $F_{1}(t)$ at time $t$, equal to the time $t$ value of the accumulated outstanding future consumption floor minus income. Notice that $\tilde{t}_{1}$ can firstly be determined at time $t$, since the value depends on the stochastic $\tilde{Z}(t)$ which is not known before time $t$. Hence, $\tilde{t}_{1}$ is time- and also state-dependent and thus the optimal PPI strategy itself is time- and state-dependent through its PPI multiple. The PPI multiple in summary is time-varying, state-dependent and depends on all future coefficients of risk aversion via $b(\tilde{t}_{1})$.
Furthermore, ${V_{1}(0 ; v_{1}) > F_{1}(0)}$ holds by the assumption in . In addition, $\hat{\pi}_{1}(t ; v_{1})$ converges to $0$ when ${V_{1}(t ; v_{1})}$ approaches $F_{1}(t)$. Thus, ${V_{1}(t ; v_{1}) > F_{1}(t)}$ a.s., which additionally follows directly from the formula for $V_{1}(t ; v_{1})$ in Theorem \[thm:Solution:ConsumptionOnly\]. This further implies that $(\hat{\pi}_{1},c_{1})$ is an admissible pair, i.e. ${(\hat{\pi}_{1},c_{1}) \in \Lambda_{1}}$. The next remark provides the solution under time-independent risk aversion.
When $b(t) \equiv b$, then $$\begin{aligned}
\hat{\pi}_{1}(t ; v_{1}) = \frac{1}{1 - b} \Sigma^{-1} (\mu - r \mathbf{1}) \frac{V_{1}(t ; v_{1}) - F_{1}(t)}{V_{1}(t ; v_{1})}\end{aligned}$$ which is a conventional CPPI strategy with constant multiple. Moreover, if $\bar{c}(t) - y(t) \equiv 0$, i.e. the minimum consumption is eating up the whole income, then $$\begin{aligned}
\hat{\pi}_{1}(t ; v_{1}) = \frac{1}{1 - b} \Sigma^{-1} (\mu - r \mathbf{1}),\end{aligned}$$ which is a constant mix strategy and represents the standard, well-known result for CRRA utility with constant risk aversion parameter.
Some comments on the initial capital $v_{1}$ and the sign of the risky investments come next. As already pointed out, a start with a negative initial capital ${V_{1}(0 ; v_{1}) = v_{1} < 0}$ to Problem is possible and might be reasonable in a sense that accumulated income over the life-cycle is expected to exceed total consumption. Hence, there is no need to require positive capital to this problem. For this reason, ${V_{1}(t ; v_{1}) < 0}$ can happen and might be reasonable, too.
Theorem \[thm:Solution:ConsumptionOnly\] tells that the optimal relative investment strategy is given by $$\begin{aligned}
\hat{\pi}_{1}(t ; v_{1}) = {} & \frac{1}{1 - b(\tilde{t}_{1})} \Sigma^{-1} (\mu - r \mathbf{1}) \frac{V_{1}(t ; v_{1}) - F_{1}(t)}{V_{1}(t ; v_{1})},\end{aligned}$$ where ${V_{1}(t ; v_{1}) > F_{1}(t)}$ a.s.. Let ${\left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} > 0}$ for ${i \in \left\{1, \hdots, N\right\}}$, which for instance is the case when there is only one risky asset ($N = 1$) because then ${\Sigma^{-1} \left(\mu - r\right) = \frac{\mu - r}{\sigma^{2}} > 0}$ since $\mu - r \mathbf{1} > 0$ was assumed. Then $$\begin{aligned}
& \left(\hat{\pi}_{1}(t ; v_{1})\right)_{i} > 0\ \Leftrightarrow\ V_{1}(t ; v_{1}) > 0, \\
& \left(\hat{\pi}_{1}(t ; v_{1})\right)_{i} < 0\ \Leftrightarrow\ V_{1}(t ; v_{1}) < 0.\end{aligned}$$ Even if the first part of the remark argues that ${V_{1}(t ; v_{1}) < 0}$ is a meaningful case, the conclusion ${\left(\hat{\pi}_{1}(t ; v_{1})\right)_{i} < 0}$ under ${\left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} > 0}$ sounds odd at a first glance. But when looking at the optimal exposure to risky asset $i$, one finds that $$\begin{aligned}
\left(\hat{\pi}_{1}(t ; v_{1}) V_{1}(t ; v_{1})\right)_{i} = {} & \frac{1}{1 - b(\tilde{t}_{1})} \left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} \left(V_{1}(t ; v_{1}) - F_{1}(t)\right),\end{aligned}$$ which, under ${\left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} > 0}$, is positive no matter if ${V_{1}(t ; v_{1}) < 0}$ or ${V_{1}(t ; v_{1}) > 0}$. Therefore, the amount of money invested in the risky assets is always positive. The opposite inequalities and conclusions for $\left(\hat{\pi}_{1}(t ; v_{1})\right)_{i}$ and $\left(\hat{\pi}_{1}(t ; v_{1}) V_{1}(t ; v_{1})\right)_{i}$ apply if ${\left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} < 0}$. In summary, the sign of the optimal exposure to the single risky assets is determined by $$\begin{aligned}
\left(\hat{\pi}_{1}(t ; v_{1}) V_{1}(t ; v_{1})\right)_{i} > 0\ \Leftrightarrow\ \left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} > 0.\end{aligned}$$ Thus, ${\left(\hat{\pi}_{1}(t ; v_{1}) V_{1}(t ; v_{1})\right)_{i} > 0}$ is possible although it might be ${\left(\hat{\pi}_{1}(t ; v_{1})\right)_{i} < 0}$.
Finally, let ${\left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} > 0}$ for all ${i \in \left\{1, \hdots, N\right\}}$. When ${V_{1}(t ; v_{1}) < 0}$, the optimal exposure to the risk-free asset is negative because $$\begin{aligned}
\underbrace{V_{1}(t ; v_{1})}_{< 0} \left(1 - \underbrace{\hat{\pi}_{1}(t ; v_{1})' \mathbf{1}}_{< 0}\right) < V_{1}(t ; v_{1}) < 0.\end{aligned}$$ This in turn implies that in case of ${V_{1}(t ; v_{1}) < 0}$, the investor takes leverage by borrowing from the risk-free account to achieve her investment goals. Leverage at this point can make sense as future income provides some security; note that ${V_{1}(t ; v_{1}) < 0}$ immediately implies that the time $t$ value of accumulated future income exceeds the expected value of consumption.
Some more properties of $\hat{\pi}_{1}(t ; v_{1})$ can be found analytically as follows. The first and second derivative of $\left(\hat{\pi}_{1}(t ; v_{1})\right)_{i}$, $i = 1, \hdots, N$, with respect to wealth $V_{1}(t ; v_{1})$ are $$\begin{aligned}
\frac{\partial}{\partial V_{1}(t ; v_{1})} \left(\hat{\pi}_{1}(t ; v_{1})\right)_{i} = {} & \frac{1}{1 - b(\tilde{t}_{1})} \left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} \frac{F_{1}(t)}{V_{1}(t ; v_{1})^{2}}, \\
\frac{\partial^{2}}{\partial V_{1}(t ; v_{1})^{2}} \left(\hat{\pi}_{1}(t ; v_{1})\right)_{i} = {} & - 2 \frac{1}{1 - b(\tilde{t}_{1})} \left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} \frac{F_{1}(t)}{V_{1}(t ; v_{1})^{3}}.\end{aligned}$$ Let ${\left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} > 0}$ for ${i \in \left\{1, \hdots, N\right\}}$, then
1. ${\frac{\partial}{\partial V_{1}(t ; v_{1})} \left(\hat{\pi}_{1}(t ; v_{1})\right)_{i} \stackrel{(>)}{\ge} 0\ \Leftrightarrow\ F_{1}(t) \stackrel{(>)}{\ge} 0}$.
2. ${\frac{\partial^{2}}{\partial V_{1}(t ; v_{1})^{2}} \left(\hat{\pi}_{1}(t ; v_{1})\right)_{i} \stackrel{(<)}{\le} 0\ \Leftrightarrow}$ either ${F_{1}(t) \stackrel{(>)}{\ge} 0}$ and ${V_{1}(t ; v_{1}) > 0}$ or ${F_{1}(t) \stackrel{(<)}{\le} 0}$ and ${V_{1}(t ; v_{1}) < 0}$.
This implies that at time $t$:
1. $\left(\hat{\pi}_{1}(t ; v_{1})\right)_{i}$ is increasing in $V_{1}(t ; v_{1})$ if and only if ${F_{1}(t) \ge 0}$, and decreasing in $V_{1}(t ; v_{1})$ otherwise.
2. $\left(\hat{\pi}_{1}(t ; v_{1})\right)_{i}$ is concave in $V_{1}(t ; v_{1})$ if and only if
1. either ${F_{1}(t) \ge 0}$ and ${V_{1}(t ; v_{1}) > 0}$
2. or ${F_{1}(t) \le 0}$ and ${V_{1}(t ; v_{1}) < 0}$,
and convex in $V_{1}(t ; v_{1})$ otherwise.
The opposite inequalities and conclusions for $\left(\hat{\pi}_{1}(t ; v_{1})\right)_{i}$ and its derivatives apply if ${\left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} < 0}$.
The optimal controls in Theorem \[thm:Solution:ConsumptionOnly\] determine the value function and the value for $\lambda_{1}$ as follows.
\[thm:Solution:ConsumptionOnly:ValueFunction:lambda\] The optimal value function $\mathcal{V}_{1}(v_{1})$ to Problem is strictly increasing and concave in $v_{1}$. Its value and first and second derivative with respect to the initial budget $v_{1}$ are given by $$\begin{aligned}
\mathcal{V}_{1}(v_{1}) = {} & \int_{0}^{T} \frac{1-b(t)}{b(t)} \left(\frac{e^{\left[\beta - b(t) \left(r - \frac{1}{2} \frac{1}{b(t)-1} \|\gamma\|^{2}\right)\right] t}}{a(t)}\right)^{\frac{1}{b(t)-1}} \lambda_{1}^{\frac{b(t)}{b(t)-1}} dt, \\
\mathcal{V}_{1}^{\prime}(v_{1}) = {} & \lambda_{1} > 0, \\
\mathcal{V}_{1}^{\prime\prime}(v_{1}) = {} & \lambda_{1}^{\prime} = - \left(\int_{0}^{T} \left(\frac{e^{\left[\beta - b(t) \left(r - \frac{1}{2} \frac{1}{b(t)-1} \|\gamma\|^{2}\right)\right] t}}{a(t)}\right)^{\frac{1}{b(t)-1}} \lambda_{1}^{- \frac{b(t)-2}{b(t)-1}} dt\right)^{-1} < 0.\end{aligned}$$
The terminal wealth problem {#sec:TerminalWealthProblem}
---------------------------
The terminal wealth-only problem is $$\begin{aligned}
\begin{split} \label{eq:OptimizationProblem:TerminalWealthOnly}
J_{2}(\pi,c;v_{2}) = {} & {\mathbb{E}}\left[U_{2}(V(T))\right], \\
\mathcal{V}_{2}(v_{2}) = {} & \sup_{(\pi,c) \in \Lambda_{2}} J_{2}(\pi,c;v_{2})
\end{split}\end{aligned}$$ subject to the budget constraint $$\begin{aligned}
\label{eq:BudgetConstraint:TerminalWealthOnly}
{\mathbb{E}}\left[\tilde{Z}(T) V(T)\right] \le v_{2},\ v_{2} \ge 0.\end{aligned}$$ $\Lambda_{2}$ denotes the set of admissible strategies $(\pi,c)$ such that ${V(t) \ge 0}$, ${\mathbb{P}}$-a.s., $\forall t \in [0,T]$, and which admit a unique solution to for ${y(t) \equiv 0}$.
In order to guarantee the terminal wealth floor, note $V(T) > F$, let us assume the following lower bound for $v_{2}$ which equals the discounted terminal floor: $$\begin{aligned}
v_{2} > {} & e^{- r T} F =: F_{2}(0),\ F_{2}(t) := {\mathbb{E}}\left[\frac{\tilde{Z}(T)}{\tilde{Z}(t)} F \Big| \mathcal{F}_{t}\right] = e^{- r (T-t)} F \ge 0. \label{eq:Condition:v2:WithoutProbabilityConstraint_&_eq:def:F2(t)}\end{aligned}$$ Applying the martingale approach leads to the solution to the terminal wealth problem according to the upcoming theorem.
\[thm:Solution:TerminalWealthOnly:WithoutProbabilityConstraint\] The solution to Problem with terminal utility function $U_{2}$ in is $$\begin{aligned}
\hat{\pi}_{2}(t ; v_{2}) = {} & \frac{1}{1-\hat{b}} \Sigma^{-1} (\mu - r \mathbf{1}) \frac{V_{2}(t ; v_{2}) - F_{2}(t)}{V_{2}(t ; v_{2})}, \\
c_{2}(t;v_{2}) = {} & 0, \\
V_{2}(t ; v_{2}) = {} & \left(v_{2} - e^{- r T} F\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} + F_{2}(t), \\
V_{2}(T ; v_{2}) = {} & \left(v_{2} - e^{- r T} F\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) T} \tilde{Z}(T)^{\frac{1}{\hat{b}-1}} + F,\end{aligned}$$ for all $t \in [0,T]$. For the optimal $\hat{\pi}_{2}(t ; v_{2})$, Equation is fulfilled with equality.
Theorem \[thm:Solution:TerminalWealthOnly:WithoutProbabilityConstraint\] shows that the optimal fraction of wealth allocated to the risky assets follows a CPPI strategy with floor $F_{2}(t) \ge 0$ at time $t$, with constant multiple. Moreover, ${V_{2}(0 ; v_{2}) > F_{2}(0) = e^{- r T} F}$ by the assumption in . In addition, $\hat{\pi}_{2}(t ; v_{2})$ converges to $0$ when ${V_{2}(t ; v_{1})}$ approaches $F_{2}(t)$. Thus, it follows ${V_{2}(t ; v_{2}) > F_{2}(t)}$ a.s., which additionally yields that $(\hat{\pi}_{2},0)$ is an admissible pair, i.e. ${(\hat{\pi}_{2},0) \in \Lambda_{2}}$. The characteristics ${V_{2}(t ; v_{2}) > F_{2}(t)}$ a.s. also directly follows from the formula for ${V_{2}(t ; v_{2})}$ in Theorem \[thm:Solution:TerminalWealthOnly:WithoutProbabilityConstraint\]. The next remark shows that the optimal proportion allocated to the risky assets is constant over time if one disregards the floor $F$.
When $F = 0$, then $$\begin{aligned}
\hat{\pi}_{2}(t ; v_{2}) = \frac{1}{1 - \hat{b}} \Sigma^{-1} (\mu - r \mathbf{1})\end{aligned}$$ which is a constant mix strategy and equals the standard result for CRRA utility with constant risk aversion parameter, where the optimal fraction of wealth allocated to the single risky assets does not depend on time or wealth.
In what follows we analyze some characteristics of the optimal strategy $\hat{\pi}_{2}(t ; v_{2})$. The first and second derivative of $\left(\hat{\pi}_{2}(t ; v_{2})\right)_{i}$, $i = 1, \hdots, N$, with respect to wealth $V_{2}(t ; v_{2})$ are $$\begin{aligned}
\frac{\partial}{\partial V_{2}(t ; v_{2})} \left(\hat{\pi}_{2}(t ; v_{2})\right)_{i} = {} & \frac{1}{1 - \hat{b}} \left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} \frac{F_{2}(t)}{V_{2}(t ; v_{2})^{2}}, \\
\frac{\partial^{2}}{\partial V_{2}(t ; v_{2})^{2}} \left(\hat{\pi}_{2}(t ; v_{2})\right)_{i} = {} & - 2 \frac{1}{1 - \hat{b}} \left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} \frac{F_{2}(t)}{V_{2}(t ; v_{2})^{3}}.\end{aligned}$$ Let ${\left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} > 0}$ for ${i \in \left\{1, \hdots, N\right\}}$. Then ${\frac{\partial}{\partial V_{2}(t ; v_{2})} \left(\hat{\pi}_{2}(t ; v_{2})\right)_{i} \ge 0}$ and ${\frac{\partial^{2}}{\partial V_{2}(t ; v_{2})^{2}} \left(\hat{\pi}_{2}(t ; v_{2})\right)_{i} \le 0}$, where the inequalities hold strictly when $F > 0$. Hence, $\left(\hat{\pi}_{2}(t ; v_{2})\right)_{i}$ increases and is concave in the wealth $V_{2}(t ; v_{2})$. Otherwise, if ${\left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} < 0}$ for ${i \in \left\{1, \hdots, N\right\}}$, then $\left(\hat{\pi}_{2}(t ; v_{2})\right)_{i}$ decreases and is convex in the wealth $V_{2}(t ; v_{2})$. For the optimal exposure to the risky assets it therefore holds $$\begin{aligned}
\left(\hat{\pi}_{2}(t ; v_{2}) V_{2}(t ; v_{2})\right)_{i} > 0\ \Leftrightarrow\ \left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} > 0.\end{aligned}$$ Thus, either it is ${\left(\hat{\pi}_{2}(t ; v_{2})\right)_{i} > 0}$ and ${\left(\hat{\pi}_{2}(t ; v_{2}) V_{2}(t ; v_{2})\right)_{i} > 0}$ or ${\left(\hat{\pi}_{2}(t ; v_{2})\right)_{i} < 0}$ and ${\left(\hat{\pi}_{2}(t ; v_{2}) V_{2}(t ; v_{2})\right)_{i} < 0}$.
The optimal controls in Theorem \[thm:Solution:TerminalWealthOnly:WithoutProbabilityConstraint\] determine the value function and the value for $\lambda_{2}$.
\[thm:Solution:TerminalWealthOnly:ValueFunction:lambda\] The optimal value function $\mathcal{V}_{2}(v_{2})$ to Problem is strictly increasing and concave in $v_{2}$. Its value and first and second derivative with respect to the initial budget $v_{2}$ are given by $$\begin{aligned}
\mathcal{V}_{2}(v_{2}) = {} & e^{\left[- \beta + \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)\right] T} \frac{\left(1-\hat{b}\right)^{1-\hat{b}}}{\hat{b}} \hat{a} \left(v_{2} - F_{2}(0)\right)^{\hat{b}}, \\
\mathcal{V}_{2}^{\prime}(v_{2}) = {} & \lambda_{2} > 0, \\
\mathcal{V}_{2}^{\prime\prime}(v_{2}) = {} & \lambda_{2}^{\prime} = - e^{\left[- \beta + \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)\right] T} \left(1-\hat{b}\right)^{2-\hat{b}} \hat{a} \left(v_{2} - F_{2}(0)\right)^{\hat{b}-2} < 0.\end{aligned}$$ The Lagrange multiplier is given by as $$\begin{aligned}
\lambda_{2} = e^{-\left[\beta - \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)\right] T} \left(1-\hat{b}\right)^{1-\hat{b}} \hat{a} \left(v_{2} - F_{2}(0)\right)^{\hat{b}-1} > 0.\end{aligned}$$
Optimal merging of the individual solutions {#sec:OptimalMerging}
-------------------------------------------
Let $(\pi_{1}(t;v_{1}), c_{1}(t;v_{1}))$ denote the optimal controls to Problem with optimal wealth process $V_{1}(t;v_{1})$ to the initial wealth ${v_{1} \ge \int_{0}^{T} e^{- r t} \left(\bar{c}(t) - y(t)\right) dt = F_{1}(0)}$ and $(\pi_{2}(t;v_{2}), c_{2}(t;v_{2}))$ the optimal controls to Problem with optimal wealth process $V_{2}(t;v_{2})$ to the initial wealth ${v_{2} \ge e^{- r T} F = F_{2}(0)}$. Then merging the two solutions to solve Problem is based on the following theorem.
\[thm:ConnectionValueFunctions\] The connection between the value functions is $$\begin{aligned}
\mathcal{V}(v_{0}) = \sup_{v_{1} \ge F_{1}(0),\ v_{2} \ge F_{2}(0),\ v_{1} + v_{2} = v_{0}} \left\{\mathcal{V}_{1}(v_{1}) + \mathcal{V}_{2}(v_{2})\right\}.\end{aligned}$$
Notice that ${F(t) = F_{1}(t) + F_{2}(t)}$, hence ensures that ${v_{0} = v_{1} + v_{2} > F_{1}(0) + F_{2}(0)}$ is claimed. When discounted future income exceeds consumption over the considered period, i.e. when the initial budget to the consumption problem is negative (${v_{1} < 0}$), then ${v_{2} > v_{0}}$ and a higher amount of money $v_{2}$ is invested according to the terminal wealth problem at initial time as the initial endowment $v_{0}$ of the investor.
Theorem \[thm:ConnectionValueFunctions\] shows that an optimal allocation to consumption and terminal wealth at $t = 0$ together with the solution to the two separate problems equals the solution to the original optimization problem. The optimal initial budgets are denoted by $v_{1}^{\star}$ and $v_{2}^{\star}$. The next lemma provides a condition for $v_{1}^{\star}$ and $v_{2}^{\star}$.
\[lemma:EqualityConditionValueFunctions\] The optimal $v_{1}^{\star}$ solves $$\begin{aligned}
\label{eq:OptimalMerging:ValueFunction}
\mathcal{V}_{1}^{\prime}(v_{1}) - \mathcal{V}_{2}^{\prime}(v_{0} - v_{1}) = 0\end{aligned}$$ and is subject to ${F_{1}(0) \le v_{1}^{\star} \le v_{0} - F_{2}(0)}$. The optimal $v_{2}^{\star}$ is then given by ${v_{2}^{\star} = v_{0} - v_{1}^{\star}}$.
Within our specified setup, we can address the condition in Lemma \[lemma:EqualityConditionValueFunctions\] in more detail, the result is provided next.
\[lemma:EqualityConditionValueFunctions:Solution\] The optimal $v_{1}^{\star}$ to exists uniquely and satisfies the boundary condition $F_{1}(0) \le v_{1}^{\star} \le v_{0} - F_{2}(0)$. $v_{1}^{\star}$ is the solution to the equation $$\begin{aligned}
\label{eq:Separation:Optimalv1}
v_{1} - \int_{0}^{T} \chi(t) \left(v_{0} - v_{1} - F_{2}(0)\right)^{\frac{\hat{b}-1}{b(t)-1}} dt = F_{1}(0)\end{aligned}$$ with $$\begin{aligned}
\label{eq:definition:chi}
\chi(t) = (1-b(t)) \left(1-\hat{b}\right)^{\frac{1-\hat{b}}{b(t)-1}} \left(\frac{\hat{a}}{a(t)}\right)^{\frac{1}{b(t)-1}} \left(\frac{e^{\left[\beta - b(t) \left(r - \frac{1}{2} \frac{1}{b(t)-1} \|\gamma\|^{2}\right)\right] t}}{e^{\left[\beta - \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)\right] T}}\right)^{\frac{1}{b(t)-1}} > 0.\end{aligned}$$ The optimal $v_{2}^{\star}$ is given by ${v_{2}^{\star} = v_{0} - v_{1}^{\star}}$.
Moreover, the optimal Lagrange multiplier $\lambda_{1}^{\star} = \lambda_{1}(v_{1}^{\star})$ is given by $$\begin{aligned}
\lambda_{1}^{\star} = \left(1-\hat{b}\right)^{1-\hat{b}} \hat{a} e^{-\left[\beta - \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)\right] T} \left(v_{0} - v_{1}^{\star} - F_{2}(0)\right)^{\hat{b}-1}.\end{aligned}$$
For general $a(t)$ and $b(t)$, $v_{1}^{*}$ as the unique solution to Equation can for instance be determined numerically. Denote by $v_{1}^{\star} \ge F_{1}(0),\ v_{2}^{\star} \ge F_{2}(0)$ with $v_{1}^{\star} + v_{2}^{\star} = v_{0}$ the optimal allocation of the initial wealth according to Lemma \[lemma:EqualityConditionValueFunctions:Solution\] in what follows and denote $\lambda_{1}^{\star} = \lambda_{1}(v_{1}^{\star})$ and $\tilde{t}_{1}^{\star} = \tilde{t}_{1}(v_{1}^{\star})$. We use the individual solutions to the two separate Problems and and merge both solutions optimally according to Lemma \[lemma:EqualityConditionValueFunctions:Solution\] to obtain the solution to the original Problem .
\[thm:Solution:Merging:OriginalProblem\] The optimal wealth process is given by ${V^{\star}(t; v_{0}) = V_{1}(t;v_{1}^{\star}) + V_{2}(t;v_{2}^{\star})}$. The optimal controls to Problem are $$\begin{aligned}
c^{\star}(t; v_{0}) = c_{1}(t;v_{1}^{\star}),\ \hat{\pi}^{\star}(t; v_{0}) = \frac{\hat{\pi}_{1}(t;v_{1}^{\star}) V_{1}(t;v_{1}^{\star}) + \hat{\pi}_{2}(t;v_{2}^{\star}) V_{2}(t;v_{2}^{\star})}{V_{1}(t;v_{1}^{\star}) + V_{2}(t;v_{2}^{\star})}.\end{aligned}$$ The optimal controls and the optimal wealth process to Problem under the utility function setup are given by $$\begin{aligned}
\hat{\pi}^{\star}(t ; v_{0}) = {} & \Sigma^{-1} (\mu - r \mathbf{1}) \frac{\frac{1}{1 - b(\tilde{t}_{1}^{\star})} \left(V_{1}(t ; v_{1}^{\star}) - F_{1}(t)\right) + \frac{1}{1-\hat{b}} \left(V_{2}(t ; v_{2}^{\star}) - F_{2}(t)\right)}{V^{\star}(t ; v_{0})}, \\
c^{\star}(t;v_{0}) = {} & g(t,t; v_{1}^{\star}) \tilde{Z}(t)^{\frac{1}{b(t)-1}} + \bar{c}(t) = (1-b(t)) \left(\lambda_{1}^{\star} \frac{e^{\beta t}}{a(t)} \tilde{Z}(t)\right)^{\frac{1}{b(t)-1}} + \bar{c}(t), \\
V^{\star}(t ; v_{0}) = {} & \int_{t}^{T} g(s,t; v_{1}^{\star}) \tilde{Z}(t)^{\frac{1}{b(s)-1}} ds + \left(v_{2}^{\star} - F_{2}(0)\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} + F(t), \\
V^{\star}(T ; v_{0}) = {} & \left(v_{2}^{\star} - F_{2}(0)\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) T} \tilde{Z}(T)^{\frac{1}{\hat{b}-1}} + F, \\
V_{1}(t ; v_{1}^{\star}) = {} & \int_{t}^{T} g(s,t; v_{1}^{\star}) \tilde{Z}(t)^{\frac{1}{b(s)-1}} ds + F_{1}(t), \\
V_{2}(t ; v_{2}^{\star}) = {} & \left(v_{2}^{\star} - F_{2}(0)\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} + F_{2}(t) \text{,\ $\forall t \in [0,T]$, with}\end{aligned}$$ ${g(s,t; v_{1}^{\star}) = \chi(s) e^{\frac{b(s)}{b(s)-1} \left(r - \frac{1}{2} \frac{1}{b(s)-1} \|\gamma\|^{2}\right) t} \left(v_{0} - v_{1}^{\star} - F_{2}(0)\right)^{\frac{\hat{b}-1}{b(s)-1}}}$, and ${\tilde{t}_{1}^{\star} = \tilde{t}_{1}(v_{1}^{\star}) \in (t,T)}$ solves $$\begin{aligned}
b(\tilde{t}_{1}^{\star}) = {} & 1 + \frac{\int_{t}^{T} g(s,t; v_{1}^{\star}) \tilde{Z}(t)^{\frac{1}{b(s)-1}} ds}{\int_{t}^{T} \frac{1}{b(s)-1} g(s,t; v_{1}^{\star}) \tilde{Z}(t)^{\frac{1}{b(s)-1}} ds}.\end{aligned}$$ For the optimal $(\hat{\pi}^{\star}(t ; v_{0}), c^{\star}(t;v_{0}))$, Equation holds with equality.
It follows immediately that ${c_{1}(t;v_{1}) > \bar{c}(t)}$, a.s.. Theorem \[thm:Solution:Merging:OriginalProblem\] furthermore proves that the general optimal relative investment strategy can be written as a mixture of a PPI and a CPPI strategy, but is not necessarily of a PPI or even CPPI type itself. The PPI comes from the consumption-only problem, see Theorem \[thm:Solution:ConsumptionOnly\], the CPPI arises as the solution to the terminal wealth-only problem, see Theorem \[thm:Solution:TerminalWealthOnly:WithoutProbabilityConstraint\]. The way which of the two strategies dominates the overall optimal investment policy is initially determined by the wealth distribution through $v_{1}^{\star}$ and $v_{2}^{\star}$ and later through $V_{1}(t ; v_{1}^{\star})$ and $V_{2}(t ; v_{2}^{\star})$. The special case where the coefficient of risk aversion $b(t)$ from consumption equals the one from terminal wealth $\hat{b}$ at any time is covered by the next remark.
\[remark:Ye:case\] Assume $b(t) \equiv \hat{b}$ constant. Then the optimal controls turn into $$\begin{aligned}
\hat{\pi}^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0}) = {} & \frac{1}{1-\hat{b}} \Sigma^{-1} (\mu - r \mathbf{1}) \frac{V^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0}) - F(t)}{V^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0})}, \\
c^{\star}_{(b(t) \equiv \hat{b})}(t;v_{0}) = {} & \zeta(t) \left(V^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0}) -F(t)\right) + \bar{c}(t),\end{aligned}$$ with $$\begin{aligned}
\zeta(t) = \frac{\chi(t)}{\int_{t}^{T} \chi(s) ds + 1} > 0,\end{aligned}$$ where $$\begin{aligned}
\chi(t) = \left(\frac{\hat{a}}{a(t)}\right)^{\frac{1}{\hat{b}-1}} e^{- \frac{1}{\hat{b}-1} \left[\beta - \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)\right] (T-t)} > 0.\end{aligned}$$ The optimal investment strategy $\hat{\pi}^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0})$ now is a traditional CPPI strategy with floor $F(t)$ and constant multiple vector ${\frac{1}{1-\hat{b}} \Sigma^{-1} (\mu - r \mathbf{1})}$. The optimal consumption rate $c^{\star}_{(b(t) \equiv \hat{b})}(t;v_{0})$ is the sum of the consumption floor $\bar{c}(t)$ and the time-varying proportion $\zeta(t)$ of the cushion ${V^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0}) - F(t)}$ at time $t$. The fraction between the risky exposure (vector) and consumption is time-varying and it holds $$\begin{aligned}
\frac{\hat{\pi}^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0}) V^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0})}{c^{\star}_{(b(t) \equiv \hat{b})}(t;v_{0})} = \frac{1}{1-\hat{b}} \Sigma^{-1} (\mu - r \mathbf{1}) \left(\zeta(t) + \frac{\bar{c}(t)}{V^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0}) - F(t)}\right)^{-1}. \label{eq:fraction:b(t)constant}\end{aligned}$$ Optimal consumption $c^{\star}_{(b(t) \equiv \hat{b})}(t;v_{0})$ as well as, under ${\Sigma^{-1} (\mu - r \mathbf{1}) > \mathbf{0}}$, optimal risky exposure ${\hat{\pi}^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0}) V^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0})}$ linearly increase in the cushion ${V^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0}) - F(t)}$. Hence, the higher the surplus ${V^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0}) - F(t)}$, the more is invested risky and the more is consumed. The formula shows that, under ${\Sigma^{-1} (\mu - r \mathbf{1}) > \mathbf{0}}$, an increase in the cushion ${V^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0}) - F(t)}$ leads to a stronger increase in the risky exposure ${\hat{\pi}^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0}) V^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0})}$ than in consumption $c^{\star}_{(b(t) \equiv \hat{b})}(t;v_{0})$. Therefore, for a larger surplus ${V^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0}) - F(t)}$, also the relative increase in the risky exposure is larger than the relative increase in consumption, thus investing money in stocks is preferred to consuming.
The associated optimal wealth process is given as a function of the pricing kernel $$\begin{aligned}
V^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0}) = \frac{1}{\zeta(t)} \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} \left(v_{0} - F(0)\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} \frac{\chi(t)}{\int_{0}^{T} \chi(t) dt + 1} + F(t).\end{aligned}$$ This special case result coincides with the findings by [@Ye2008], who used the HJB approach, extended by additionally providing the optimal wealth process $V^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0})$.
We aim to interpret the optimal $\hat{\pi}^{\star}(t ; v_{0})$ for time-varying $b(t)$ and particularly to point out the difference to constant $b(t)$ in Remark \[remark:Ye:case\]. Writing ${V_{1}(t ; v_{1}^{\star}) = V^{\star}(t ; v_{0}) - V_{2}(t ; v_{2}^{\star})}$ where $V_{2}(t ; v_{2}^{\star})$ follows the wealth process of a standard CPPI strategy with floor $F_{2}(t)$ at time $t$ to the initial endowment $v_{2}^{\star}$ and constant multiplier vector ${\frac{1}{1-\hat{b}} \Sigma^{-1} (\mu - r \mathbf{1})}$, we obtain the following representation of the optimal investment decision
$$\begin{aligned}
\hat{\pi}^{\star}(t ; v_{0}) = {} & \Sigma^{-1} (\mu - r \mathbf{1}) \frac{\frac{1}{1 - b(\tilde{t}_{1}^{\star})} \left(V^{\star}(t ; v_{0}) - V_{2}(t ; v_{2}^{\star}) - F_{1}(t)\right) + \frac{1}{1-\hat{b}} \left(V_{2}(t ; v_{2}^{\star}) - F_{2}(t)\right)}{V^{\star}(t ; v_{0})} \nonumber \\
= {} & \Sigma^{-1} (\mu - r \mathbf{1}) \left\{\frac{1}{1 - b(\tilde{t}_{1}^{\star})} \frac{V^{\star}(t ; v_{0}) - F_{1}(t)}{V^{\star}(t ; v_{0})} + \frac{\hat{b}- b(\tilde{t}_{1}^{\star})}{(1-\hat{b}) \left(1 - b(\tilde{t}_{1}^{\star})\right)} \frac{V_{2}(t ; v_{2}^{\star}) - \frac{1 - b(\tilde{t}_{1}^{\star})}{\hat{b}- b(\tilde{t}_{1}^{\star})} F_{2}(t)}{V^{\star}(t ; v_{0})}\right\} \nonumber \\
= {} & \Sigma^{-1} (\mu - r \mathbf{1}) \left\{\frac{1}{1 - b(\tilde{t}_{1}^{\star})} \frac{V^{\star}(t ; v_{0}) - F(t)}{V^{\star}(t ; v_{0})} + \frac{\hat{b} - b(\tilde{t}_{1}^{\star})}{(1-\hat{b}) \left(1 - b(\tilde{t}_{1}^{\star})\right)} \frac{V_{2}(t ; v_{2}^{\star}) - F_{2}(t)}{V^{\star}(t ; v_{0})}\right\} \nonumber \\
= {} & \Sigma^{-1} (\mu - r \mathbf{1}) \left\{\frac{1}{1 - b(\tilde{t}_{1}^{\star})} \frac{V^{\star}(t ; v_{0}) - F(t)}{V^{\star}(t ; v_{0})} + \frac{\hat{b} - b(\tilde{t}_{1}^{\star})}{(1-\hat{b}) \left(1 - b(\tilde{t}_{1}^{\star})\right)} \frac{V_{2}(t ; v_{2}^{\star})}{V^{\star}(t ; v_{0})} \frac{V_{2}(t ; v_{2}^{\star}) - F_{2}(t)}{V_{2}(t ; v_{2}^{\star})}\right\}, \label{eq:SeparationTechnique:pihat:2xPPI}\end{aligned}$$
which can be implemented easily; $F(t)$ is defined in . Formula shows that the optimal relative allocation $\hat{\pi}^{\star}(t ; v_{0})$ can be written as a PPI strategy in $V^{\star}(t ; v_{0})$ with floor $F(t)$ plus a PPI strategy in $V_{2}(t ; v_{2}^{\star})$ with floor $F_{2}(t)$. Alternatively, write ${V_{2}(t ; v_{2}^{\star}) = V^{\star}(t ; v_{0}) - V_{1}(t ; v_{1}^{\star})}$, where $V_{1}(t ; v_{1}^{\star})$ is the replicating wealth process of a PPI strategy with floor $F_{1}(t)$ to the initial wealth $v_{1}^{\star}$ and now time- and state-varying multiplier vector $\frac{1}{1 - b(\tilde{t}_{1})} \Sigma^{-1} (\mu - r \mathbf{1})$ and, in contrast to $V_{2}(t ; v_{2}^{\star})$, a non-zero consumption rate process. Then $\hat{\pi}^{\star}(t ; v_{0})$ can be reformulated as $$\begin{aligned}
\hat{\pi}^{\star}(t ; v_{0}) = {} & \Sigma^{-1} (\mu - r \mathbf{1}) \left\{\left(\frac{1}{1 - b(\tilde{t}_{1}^{\star})} - \frac{1}{1-\hat{b}}\right) \frac{V_{1}(t ; v_{1}^{\star}) - F_{1}(t)}{V^{\star}(t ; v_{0})} + \frac{1}{1-\hat{b}} \frac{V^{\star}(t ; v_{0}) - F(t)}{V^{\star}(t ; v_{0})}\right\} \nonumber \\
= {} & \Sigma^{-1} (\mu - r \mathbf{1}) \left\{\frac{1}{1-\hat{b}} \frac{V^{\star}(t ; v_{0}) - F(t)}{V^{\star}(t ; v_{0})} - \frac{\hat{b} - b(\tilde{t}_{1}^{\star})}{(1-\hat{b}) (1 - b(\tilde{t}_{1}^{\star}))} \frac{V_{1}(t ; v_{1}^{\star}) - F_{2}(t)}{V^{\star}(t ; v_{0})}\right\} \nonumber \\
\begin{split} \label{eq:SeparationTechnique:pihat:V1:1xCPPI:1xPPI}
= {} & \Sigma^{-1} (\mu - r \mathbf{1}) \left\{\frac{1}{1-\hat{b}} \frac{V^{\star}(t ; v_{0}) - F(t)}{V^{\star}(t ; v_{0})} - \frac{\hat{b} - b(\tilde{t}_{1}^{\star})}{(1-\hat{b}) (1 - b(\tilde{t}_{1}^{\star}))} \frac{V_{1}(t ; v_{1}^{\star})}{V^{\star}(t ; v_{0})} \frac{V_{1}(t ; v_{1}^{\star}) - F_{1}(t)}{V_{1}(t ; v_{1}^{\star})}\right\}.
\end{split}\end{aligned}$$ This formula shows that the optimal relative investment $\hat{\pi}^{\star}(t ; v_{0})$ is the sum of a conventional CPPI strategy on $V^{\star}(t ; v_{0})$ with floor $F(t)$ and a PPI strategy on $V_{1}(t ; v_{1}^{\star})$ with floor $F_{1}(t)$.
Recall from Remark \[remark:Ye:case\] that $\hat{\pi}^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0})$ for constant $b(t) \equiv \hat{b}$ follows a traditional CPPI strategy ${\frac{1}{1-\hat{b}} \Sigma^{-1} (\mu - r \mathbf{1}) \frac{V^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0}) - F(t)}{V^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0})}}$ to the floor $F(t)$. The formula for $\hat{\pi}^{\star}(t ; v_{0})$ in shows that the optimal strategy $\hat{\pi}^{\star}(t ; v_{0})$ for time-varying $b(t)$ consists of two parts:
1. The first part coincides with $\hat{\pi}^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0})$ and is a traditional CPPI strategy ${\frac{1}{1-\hat{b}} \Sigma^{-1} (\mu - r \mathbf{1}) \frac{V^{\star}(t ; v_{0}) - F(t)}{V^{\star}(t ; v_{0})}}$ in $V^{\star}(t ; v_{0})$ to the floor $F(t)$.
2. The second, additional part is a time- and state-varying term which can be either positive, negative or zero; hence it can reduce or increase risky investments or can leave it unmodified in comparison with $\hat{\pi}^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0})$.
It is the second part which leads to a deviation in $\hat{\pi}^{\star}(t ; v_{0})$ compared to $\hat{\pi}^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0})$. For this sake, we analyze this second piece in what follows. Note that by Theorem \[thm:Solution:ConsumptionOnly\] it holds ${V_{1}(t ; v_{1}) > F_{1}(t)}$ a.s..
1. If ${V^{\star}(t ; v_{0}) > 0}$, for instance this is reasonable for ${v_{0} > 0}$ and an income rate that outweighs or exceeds consumption, then it follows $$\begin{aligned}
\frac{V_{1}(t ; v_{1}^{\star})}{V^{\star}(t ; v_{0})} \frac{V_{1}(t ; v_{1}^{\star}) - F_{1}(t)}{V_{1}(t ; v_{1}^{\star})} = \frac{V_{1}(t ; v_{1}^{\star}) - F_{1}(t)}{V^{\star}(t ; v_{0})} > 0.\end{aligned}$$ This implies for ${i = 1, \hdots, N}$ at time $t$:
1. $\hat{b} > b(\tilde{t}_{1}^{\star})$: $$\begin{aligned}
\left(\hat{\pi}^{\star}(t ; v_{0})\right)_{i} < \frac{1}{1-\hat{b}} \left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} \frac{V^{\star}(t ; v_{0}) - F(t)}{V^{\star}(t ; v_{0})}\ \Leftrightarrow\ \left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} > 0.\end{aligned}$$
2. $\hat{b} = b(\tilde{t}_{1}^{\star})$: $$\begin{aligned}
\hat{\pi}^{\star}(t ; v_{0}) = \frac{1}{1-\hat{b}} \Sigma^{-1} (\mu - r \mathbf{1}) \frac{V^{\star}(t ; v_{0}) - F(t)}{V^{\star}(t ; v_{0})}.\end{aligned}$$
3. $\hat{b} < b(\tilde{t}_{1}^{\star})$: $$\begin{aligned}
\left(\hat{\pi}^{\star}(t ; v_{0})\right)_{i} > \frac{1}{1-\hat{b}} \left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} \frac{V^{\star}(t ; v_{0}) - F(t)}{V^{\star}(t ; v_{0})}\ \Leftrightarrow\ \left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} < 0.\end{aligned}$$
2. If ${V^{\star}(t ; v_{0}) < 0}$, for instance this is reasonable for ${v_{0} < 0}$ and a high demand for consumption in the past, then it follows $$\begin{aligned}
\frac{V_{1}(t ; v_{1}^{\star})}{V^{\star}(t ; v_{0})} \frac{V_{1}(t ; v_{1}^{\star}) - F_{1}(t)}{V_{1}(t ; v_{1}^{\star})} = \frac{V_{1}(t ; v_{1}^{\star}) - F_{1}(t)}{V^{\star}(t ; v_{0})} > 0.\end{aligned}$$ This in turn implies for ${i = 1, \hdots, N}$ at time $t$:
1. $\hat{b} > b(\tilde{t}_{1}^{\star})$: $$\begin{aligned}
\left(\hat{\pi}^{\star}(t ; v_{0})\right)_{i} > \frac{1}{1-\hat{b}} \left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} \frac{V^{\star}(t ; v_{0}) - F(t)}{V^{\star}(t ; v_{0})}\ \Leftrightarrow\ \left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} > 0.\end{aligned}$$
2. $\hat{b} = b(\tilde{t}_{1}^{\star})$: $$\begin{aligned}
\hat{\pi}^{\star}(t ; v_{0}) = \frac{1}{1-\hat{b}} \Sigma^{-1} (\mu - r \mathbf{1}) \frac{V^{\star}(t ; v_{0}) - F(t)}{V^{\star}(t ; v_{0})}.\end{aligned}$$
3. $\hat{b} < b(\tilde{t}_{1}^{\star})$: $$\begin{aligned}
\left(\hat{\pi}^{\star}(t ; v_{0})\right)_{i} < \frac{1}{1-\hat{b}} \left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} \frac{V^{\star}(t ; v_{0}) - F(t)}{V^{\star}(t ; v_{0})}\ \Leftrightarrow\ \left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} < 0.\end{aligned}$$
In particular, consider the situation ${V^{\star}(t ; v_{0}) > 0}$ and let ${\left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} > 0}$ hold for risky asset $i$. Under ${\hat{b} > b(\tilde{t}_{1}^{\star})}$, the optimal relative investment in stock $i$, which is $\left(\hat{\pi}^{\star}(t ; v_{0})\right)_{i}$, is reduced compared to the relative investment decision $\left(\hat{\pi}^{\star}_{(b(t) \equiv \hat{b})}(t ; v_{0})\right)_{i}$ under ${b(t) \equiv \hat{b}}$. Since ${\hat{b} > b(\tilde{t}_{1}^{\star})}$ can be interpreted as higher risk aversion for consumption than terminal wealth, this is meaningful.
In the situation ${V^{\star}(t ; v_{0}) < 0}$ the interpretation seems counterintuitive at first glance. But when looking at risky exposures rather than risky relative investments, analogue conclusions hold. The same approach shall be used when considering ${V^{\star}(t ; v_{0}) = 0}$.
Furthermore, it is worth to mention that $\hat{\pi}^{\star}(t ; v_{0})$ approaches $0$ when $V^{\star}(t ; v_{0})$ approaches $F(t)$, which can be observed in ; the argument is the following: When $V^{\star}(t ; v_{0})$ falls towards $F(t)$, then automatically $V_{1}(t ; v_{1}^{\star})$ approaches $F_{1}(t)$ and $V_{2}(t ; v_{2}^{\star})$ converges towards $F_{2}(t)$ simultaneously, since $V^{\star}(t ; v_{0}) = V_{1}(t ; v_{1}^{\star}) + V_{2}(t ; v_{2}^{\star})$, ${F(t) = F_{1}(t) + F_{2}(t)}$ and ${V_{1}(t ; v_{1}^{\star}) > F_{1}(t)}$, ${V_{2}(t ; v_{2}^{\star}) > F_{2}(t)}$ a.s. which was already shown in Sections \[sec:ConsumptionProblem:y\] and \[sec:TerminalWealthProblem\]. We moreover proved that in this case ${\hat{\pi}_{1}(t;v_{1}^{\star})}$ and ${\hat{\pi}_{2}(t;v_{2}^{\star})}$ approach $0$. By Theorem \[thm:Solution:Merging:OriginalProblem\] it follows that also $\hat{\pi}^{\star}(t; v_{0})$ must converge to $0$. Therefore, as ${v_{0} > F(0)}$ is assumed, it follows that ${V^{\star}(t ; v_{0}) > F(t)}$ a.s., which can additionally be seen in the respective formula in Theorem \[thm:Solution:Merging:OriginalProblem\], and the optimal decision rules provide portfolio insurance over the whole life-cycle. $F(t)$ is called the minimum asset wealth level, it holds $F(T) = F$.
The optimal exposure to the risky assets equals the sum of the optimal risky exposures of the two sub-problems $$\begin{aligned}
\hat{\pi}^{\star}(t; v_{0}) V^{\star}(t ; v_{0}) = \hat{\pi}_{1}(t;v_{1}^{\star}) V_{1}(t;v_{1}^{\star}) + \hat{\pi}_{2}(t;v_{2}^{\star}) V_{2}(t;v_{2}^{\star})\end{aligned}$$ and by the findings in Sections \[sec:ConsumptionProblem:y\] and \[sec:TerminalWealthProblem\] it holds $$\begin{aligned}
\left(\hat{\pi}^{\star}(t ; v_{0}) V^{\star}(t ; v_{0})\right)_{i} > 0\ \Leftrightarrow\ \left(\Sigma^{-1} (\mu - r \mathbf{1})\right)_{i} > 0.\end{aligned}$$ For the ease of exposition we so far assumed that the income process is deterministic. The following remark shows the solution for a stochastic income process.
\[remark:y(t)stochastic\] Let $(y(t))_{t \in [0,T]}$ be a non-negative, stochastic income-rate process with $\int_{0}^{T} y(t) dt < \infty$, ${\mathbb{P}}$-a.s.. The stated results are still valid after replacing integrals of the form ${\int_{t}^{T} e^{-r (s-t)} y(s) ds}$ by the more general conditional expectation ${{\mathbb{E}}\left[\int_{t}^{T} \frac{\tilde{Z}(s)}{\tilde{Z}(t)} y(s) ds \Big| \mathcal{F}_{t}\right] = \int_{t}^{T} {\mathbb{E}}\left[\frac{\tilde{Z}(s)}{\tilde{Z}(t)} y(s) \Big| \mathcal{F}_{t}\right] ds = \int_{t}^{T} e^{-r (s-t)} {\mathbb{E}}_{{\mathbb{Q}}}\left[y(s) \Big| \mathcal{F}_{t}\right] ds}$ by the Bayes formula for arbitrary ${t \in [0,T]}$, in particular in the definition of $F_{1}(t)$ and $F(t)$. If $(y(t))_{t \in [0,T]}$ is supposed to be independent to $\CMcal{F}$, i.e. independent to the market stochastics, then the conditional expectation ${{\mathbb{E}}\left[\int_{t}^{T} \frac{\tilde{Z}(s)}{\tilde{Z}(t)} y(s) ds \Big| \mathcal{F}_{t}\right]}$ can be reduced to ${\int_{t}^{T} e^{-r (s-t)} {\mathbb{E}}_{{\mathbb{Q}}}\left[y(s)\right] ds}$. For the lower bounds of $v_{0}$ and $v_{1}$, and need to be replaced by $$\begin{aligned}
v_{0} > {} & \int_{0}^{T} e^{- r s} \left(\bar{c}(s) - \bar{y}(s)\right) ds + e^{- r T} F, \\
v_{1} > {} & \int_{0}^{T} e^{- r s} \left(\bar{c}(s) - \bar{y}(s)\right) ds,\end{aligned}$$ where ${\bar{y}(s) = \sup \left\{x \ge 0: {\mathbb{P}}(y(s) \ge x) = 1\right\}}$ denotes the minimal level of income; ${\bar{y}(s) > 0}$ is meaningful due to unemployment benefits paid by the government.
Analysis of optimal controls and wealth process: A case study {#sec:NCS}
=============================================================
This section targets to calibrate the life-cycle model to realistic time-dependent structures for consumption and investment observed in practice and outline the difference between our presented solution with age-depending $a(t)$ and $b(t)$ functions and the models with either only $a(t)$ or $b(t)$ time-varying or none. Hence, we not only estimate $\hat{b}$, $a(t)$ and $b(t)$ for our model, but additionally provide the respective estimates when $a(t)$ or $b(t)$, or both, are assumed to be constants. A comparison of the fit of the different models allows for making a statement on the accuracy of the models in describing the agent’s behavior. For notational convenience we call the three benchmark models as follows:
- $M_{a,b(t)}$: $a(t) \equiv a$ constant, $b(t)$ time-varying
- $M_{a(t),b}$: $a(t)$ time-varying, $b(t) \equiv b$ constant
- $M_{a,b}$: $a(t) \equiv a$ and $b(t) \equiv b$ constant
The subscript thus indicates whether $a(t)$ or $b(t)$ are age-varying. Therefore, our model is denoted by $M_{a(t),b(t)}$. As already indicated in Section \[sec:Introduction\], $M_{a,b(t)}$ is (partially) covered by [@Steffensen2011], [@Hentschel2016] and [@Aase2017], $M_{a(t),b}$ and $M_{a,b}$ are covered by [@Ye2008].
In the later Subsection \[sec:ComparisonCRRA\], we additionally analyze the impact of the floors $\bar{c}(t)$ and $F$, where our model $M_{a(t),b(t)}$ is compared to the same model but with CRRA utility functions, i.e. $\bar{c}(t) \equiv 0$ and $F \equiv 0$. The CRRA model is denoted by $M_{a(t),b(t)}^{CRRA}$ and is (partially) considered by [@Steffensen2011], [@Hentschel2016] and [@Aase2017].
Assumptions
-----------
We assume an exemplary agent with average income, liabilities etc. A similar case study can be carried out for a pension cohort, but for simplicity and data availability we consider an individual client. In detail, we make the following (simplifying) assumptions:
Let the market consist of one risk-free and one risky asset ($N = 1$) with parameters ${r = 0.5 \%}$, ${\mu = 5 \%}$, and ${\sigma = 20 \%}$; these values correspond approximately to the EURONIA Overnight Rate and the performance of the DAX 30 Performance Index as an equity index over the $11$ year period from 17 October 2007 to 17 October 2018. The risky asset can coincide with, but is not restricted to a pure equity portfolio. In general it can be any arbitrary given portfolio which consists of risky assets. The price process of the risky asset is assumed to be ${P(t) = p_{1} e^{(\mu - \frac{1}{2} \sigma^2) t + \sigma W(t)} = p_{1} e^{\frac{1}{2} (\mu + r) (1 - \frac{\sigma}{\gamma})} \tilde{Z}(t)^{- \frac{\sigma}{\gamma}}}$ with initial price ${P(0) = p_{1} = 100}$. Furthermore, let ${T = 40}$ years be the time to retirement, $t = 25$ years the current age of the investor and $65$ years the age of retirement. For the net salary function it is assumed ${y(t) = \frac{\tilde{r}}{e^{\tilde{r}} - 1} y_{0} e^{\tilde{r} t}}$ with $y_{0} = 26,200$ EUR and $\tilde{r} = 2.07 \%$. This corresponds to a net annual starting salary approximately equal to the average for a graduate in Germany in 2017 (cf. online portals [@Absolventa2018] or [@Stepstone2017]), with an annual increase equal to the average for a household’s net salary in Germany over years 2011 to 2016 according to [@StatistischesBundesamtEinkommen2018]. Net income accumulated over the first year is ${\int_{0}^{1} y(t)dt = y_{0}}$ and income accumulated within the year from time $s$ to $s+1$ is ${\int_{s}^{s+1} y(t)dt = \frac{\tilde{r}}{e^{\tilde{r}} - 1} y_{0} \frac{e^{\tilde{r} (s+1)} - e^{\tilde{r} s}}{\tilde{r}} = y_{0} e^{\tilde{r} s}}$.
For the agent’s utility functions, let ${\beta = 3 \%}$ (cf. [@Ye2008]) and ${\hat{a} = 1}$. Let the terminal wealth floor be ${F = 435,125}$ EUR which is motivated by the following argument: According to [@StatistischesBundesamtJahrbuch2017], [@aktuare2017] or [@WKO2016] a lifetime around $81$ years can be expected for a currently $25$ year old person in Germany. Thus survival of $81 - 65 = 16$ years are expected after retirement at the age of $65$. We assume that the agent secures the income inflow during retirement to be $75 \%$ of the last wage paid from year $64$ to $65$ (replacement ratio of $75 \%$), which is ${\int_{39}^{40} y(t)dt = y_{0} e^{39 \tilde{r}} = 58,736}$ EUR. Assume that every year, half of this amount is covered by a separate pension account or plan, e.g. provided by the government. In addition, the agent wants to secure against longevity risk, hence considers $16 \times (100 + 30) \% = 20.8$ years instead of $16$ years for the remaining lifetime after the age of retirement. Thus, $F$ as value at time $T$ is chosen to be ${F = \int_{0}^{20.8} \frac{0.75 \times 58,736 \text{ EUR}}{2} e^{-r t} dt = \frac{0.75 \times 58,736 \text{ EUR}}{2} \left(\frac{1 - e^{- 20.8 \times r}}{r}\right) = 435,125 \text{ EUR}}$. Finally, the function for the net consumption floor is supposed to take the form ${\bar{c}(t) = \frac{\bar{r}}{e^{\bar{r}} - 1} \bar{c}_{0} e^{\bar{r} t}}$ with $\bar{c}_{0} = 14,880$ EUR and $\bar{r} = 1.93 \%$. This corresponds to a starting value equal to approximately $50 \%$ of the average household consumption in Germany in 2016 as starting point, with an annual increase equal to the increase in average household consumption in Germany over years 2011 to 2016 (published by [@StatistischesBundesamtEinkommen2018]). Minimum consumption expenses incurred within the first year is ${\int_{0}^{1} \bar{c}(t)dt = \bar{c}_{0}}$, within year $s$ to $s+1$ is ${\int_{s}^{s+1} \bar{c}(t)dt = \bar{c}_{0} e^{\bar{r} s}}$. The assumed income and consumption floor rates are visualized in Figure \[fig:Analysis:cbar:y\].
Fitting / Calibration under exponential preferences and discussion
------------------------------------------------------------------
In what follows we calibrate the remaining utility parameters $\hat{b}$, $a(t)$ and $b(t)$ to suitable curves for consumption and relative allocation. The targeted curves for parameter fitting are summarized by Table \[tab:Analysis:SamplePoints\]. The consumption rate $c^{\star}(t;v_{0})$ is calibrated with respect to the hump-shaped type observed by [@Carroll1997], [@GourinchasParker2002], [@JensenSteffensen2015] and [@TangPurcalZhang2018]. The relative risky investment $\hat{\pi}^{\star}(t ; v_{0})$ is calibrated towards the $(100 - \text{age}) \%$ rule of thumb; a similar structure is frequently applied by financial advisors and asset management companies for life-cycle funds (see [@Malkiel1990], [@BodieCrane1997], [@Shiller2005], [@Minderhoud2011], [@Milliman2010], [@Shafir2013]). Following this popular rule, the client at age $25$ years starts with a $75 \%$ equity investment, linearly decreases it by her age such that she ends with a $35 \%$ investment in equities at the age of retirement with $65$ years. We would like to mention that in particular relative risky investment curves or products provided by asset management companies are to be understood deterministic, i.e. wealth- / state-independent. Therefore, we calibrate the remaining unknown parameters with respect to the expected values for consumption and risky relative investment. In more detail, we fit the expected value for consumption, which is ${{\mathbb{E}}\left[c^{\star}(t;v_{0})\right]}$, to the given consumption curve. For ${{\mathbb{E}}\left[\hat{\pi}^{\star}(t ; v_{0})\right]}$ we apply the following estimate: we estimate the risky exposure ${{\mathbb{E}}\left[\hat{\pi}^{\star}(t ; v_{0}) V^{\star}(t ; v_{0})\right]}$ without any bias and then replace $V^{\star}(t ; v_{0})$ by its unbiased expectation ${{\mathbb{E}}\left[V^{\star}(t ; v_{0})\right]}$ to obtain the estimate ${\frac{{\mathbb{E}}\left[\hat{\pi}^{\star}(t ; v_{0}) V^{\star}(t ; v_{0})\right]}{{\mathbb{E}}\left[V^{\star}(t ; v_{0})\right]}}$ for ${{\mathbb{E}}\left[\hat{\pi}^{\star}(t ; v_{0})\right]}$. By doing this we replace ${{\mathbb{E}}\left[\hat{\pi}^{\star}(t ; v_{0})\right]}$ by ${\frac{{\mathbb{E}}\left[\hat{\pi}^{\star}(t ; v_{0}) V^{\star}(t ; v_{0})\right]}{{\mathbb{E}}\left[V^{\star}(t ; v_{0})\right]}}$ and fit the latter expression to the given linear relative investment curve. For further readings on deterministic investment strategies we refer to [@ChristiansenSteffensen2013] and [@ChristiansenSteffensen2018]. In summary, we have unbiased estimates for the expected values of optimal consumption, risky exposure and wealth process, and a modified estimate for the expectation of the optimal relative risky investment.
Let $a(t)$ and $b(t)$ take the form of an exponential function, i.e. ${a(t) = a_{0} e^{\lambda_{a} t}}$ and ${b(t) = b_{0} e^{\lambda_{b} t}}$. Moreover, let $v_{0} = 250,000$ EUR. The estimation is carried out via the Matlab function *lsqcurvefit* which solves nonlinear curve-fitting (data-fitting) problems in a least-squares sense and minimizes the sum of the squared relative distances. The underlying time points for target consumption and allocation are set weekly on an equidistant grid which yields $2,080$ points in the time interval $[0,T]$ with ${T = 40}$.
Table \[tab:Analysis:ParameterEstimation:Error:Benchmark:RichInvestor\] gives an overview of the estimated utility parameters and provides the sum of squared relative errors as a quality criterion. The errors show that considering age-depending functions $a(t)$ and $b(t)$ simultaneously in model $M_{a(t),b(t)}$ leads to a comparatively huge improvement in accuracy of the fit compared to any of the three benchmark models: model $M_{a(t),b(t)}$ sum of squared relative distances is only $19.38 \%$ of the respective sum for model $M_{a(t),b}$ which provides the second best fit in terms of sum of squared relative residuals.
Figure \[fig:Analysis:InputFunctions:Benchmark:RichInvestor\] visualizes the fitted parameters and preference functions $\hat{b}$, $a(t)$, $b(t)$. The table and figure show that the estimated coefficient of risk aversion $\hat{b}$ for our model $M_{a(t),b(t)}$ is more negative, which means a higher risk aversion, compared to the three benchmark models $M_{a,b(t)}$, $M_{a(t),b}$, $M_{a,b}$. Furthermore, $a(t)$ is decreasing both within model $M_{a(t),b(t)}$ and $M_{a(t),b}$. In contrast, $b(t)$ increases in model $M_{a(t),b(t)}$ over time whereas it decreases in the comparison model $M_{a,b(t)}$. $b(t)$ in models $M_{a,b(t)}$, $M_{a(t),b}$, $M_{a,b}$ stay very close over the whole life-cycle whereas $b(t)$ in $M_{a(t),b(t)}$ starts more negative and ends less negative. In summary, this means that in model $M_{a(t),b(t)}$ the risk aversion decreases through increasing $b(t)$, but preference of the investor between consumption and terminal wealth is shifted more and more to terminal wealth through decreasing $a(t)$.
Figure \[fig:Analysis:FittedConsumptionInvestment:Benchmark:RichInvestor\] illustrates the expected optimal consumption rate and relative risky investment for the fitted parameters in comparison with the given target policies or average profile. In addition to Table \[tab:Analysis:ParameterEstimation:Error:Benchmark:RichInvestor\] the figure illustrates that, under exponential preferences $a(t)$ and $b(t)$, only the most flexible model $M_{a(t),b(t)}$ provides an accurate and precise fit for both consumption rate and risky relative allocation. We realize that the benchmark models $M_{a,b(t)}$, $M_{a(t),b(t)}$, $M_{a,b}$ apparently do not provide enough flexibility to simultaneously describe the predetermined consumption and relative allocation curves. Whereas the fits for the relative investment $\hat{\pi}^{\star}(t ; v_{0})$ look acceptable, all three benchmark models fail in explaining the targeted consumption rate $c^{\star}(t;v_{0})$. We further notice that $c^{\star}(t;v_{0})$ and $\hat{\pi}^{\star}(t ; v_{0})$ for the models $M_{a,b(t)}$ and $M_{a(t),b}$ are very similar (red and black lines in the respective figures).
In summary, Table \[tab:Analysis:ParameterEstimation:Error:Benchmark:RichInvestor\] and Figure \[fig:Analysis:FittedConsumptionInvestment:Benchmark:RichInvestor\] demonstrate that model $M_{a(t),b(t)}$ is the only one among our considered models which provides enough flexibility to model a hump-shaped consumption decision curve besides a linear risky allocation curve. All three benchmark models, which disregard time-dependency of $a(t)$ or $b(t)$ or both, do not lead to a satisfactory fit. In addition, fitting optimal consumption of the four models to the given consumption curve, while ignoring relative investments, shows the same picture. The result is that the sum of the squared distances associated with model $M_{a(t),b(t)}$ is only $21.26 \%$ of the respective sum associated with the second best model $M_{a,b(t)}$. This supports our findings and conclusion that time-varying preference parameters are indeed needed to model the given time-dependent hump-shaped consumption and linear risky allocation in an accurate way.
In addition to the parameter estimation for the expected path, we provide the figures for optimal consumption, risky relative portfolio and wealth process of all four models under two representative scenarios: a mostly upward (see Figure \[fig:Analysis:FittedConsumptionInvestment:IncreasingPath:RichInvestor\]) and a mostly downward (see Figure \[fig:Analysis:FittedConsumptionInvestment:DecreasingPath:RichInvestor\]) moving path for the underlying stock. The corresponding expected paths for the consumption rate, the relative risky investment and the wealth process can be found in Figure \[fig:Analysis:FittedConsumptionInvestment:Benchmark:RichInvestor\].
In the increasing stock price case optimal consumption and risky relative allocation for model $M_{a(t),b(t)}$ stay very close to the targeted curve since the corresponding wealth stays close to its expected path and shows some reverting behavior. For a stronger increasing underlying price process, consumption exceeds the given consumption curve for the expected path. When the stock price decreases, then optimal consumption and risky allocation for model $M_{a(t),b(t)}$ fall below the target curves after approximately $15$ to $20$ years. In particular higher consumption can no longer be afforded due to a poorly performing equity market. This goes hand in hand with a reduction on the relative risky allocation.
At first glance, it seems that there is a big difference in optimal consumption between our model $M_{a(t),b(t)}$ and the three benchmark models $M_{a,b(t)}$, $M_{a(t),b}$ and $M_{a,b}$ while optimal risky investments and wealth paths for all four models remain in a quite narrow area, although deviation of risky investments from its target curve can be high. This is due to different scales for wealth and consumption. Figure \[fig:Analysis:Difference:c:pi:V:exp\] visualizes the differences, denoted by $\Delta$, in the fitted consumption and relative risky investment and the corresponding wealth process for the three benchmark models to our model within the expected path situation. It can be observed that relative risky allocation $\hat{\pi}^{\star}(t;v_{0})$ of model $M_{a(t),b(t)}$ exceeds the ones associated with the three benchmark models in the first half of the considered period of $40$ years by up to eight percentage points, and falls below in the second half. Moreover, the difference looks monotone decreasing in age. Furthermore, the wealth process which corresponds to model $M_{a(t),b(t)}$ outperforms the three benchmark models in the first half, but provides a lower wealth in the second half due to a higher consumption rate from approx. year $8$ to $30$, with a certain recovery in the wealth close to retirement.
The two exemplary scenarios and the expected development situation which was used for fitting show that the benchmark models $M_{a,b(t)}$, $M_{a(t),b}$ and $M_{a,b}$ overestimate the given consumption curve in early and older years (close to $t = 0$ and $t = 40$) and underestimate it in between. For our model $M_{a(t),b(t)}$, the optimal consumption rate stays very close to its target curve until consumption cannot be afforded anymore because of a low wealth as result of a strong market decline. We conclude that especially within phases of poor stock performance, both $c^{\star}(t;v_{0})$ and $\hat{\pi}^{\star}(t ; v_{0})$ can deviate a lot from their given curves.
\
\
\
\
Comparison with CRRA {#sec:ComparisonCRRA}
--------------------
We conclude the case study section by exploring the impact of minimum consumption and wealth floors on calibration and optimal controls. For this sake, we fit the model $M_{a(t),b(t)}$ to the very same parameters and target curves as before, but now enforce $\bar{c}(t) \equiv 0$ and $F \equiv 0$. This CRRA model is referred to as $M_{a(t),b(t)}^{CRRA}$. Table \[tab:Analysis:ParameterEstimation:Error:Benchmark:RichInvestor:CRRA\] provides the estimated parameters and the sum of the squared relative residuals. In terms of this sum, it is clear that model $M_{a(t),b(t)}$ provides a more adequate fit than model $M_{a(t),b(t)}^{CRRA}$, its sum is only $4.82 \%$ of the sum which corresponds to $M_{a(t),b(t)}^{CRRA}$. Going even further, all three benchmark models $M_{a,b(t)}$, $M_{a(t),b}$ and $M_{a,b}$ from the previous subsection, which all consider minimum levels for consumption and wealth, provide a more precise fit than $M_{a(t),b(t)}^{CRRA}$ in view of the sum of squared relative residuals. This shows that the introduction of floors for consumption and wealth in the model is essential.
Figure \[fig:Analysis:InputFunctions:Benchmark:RichInvestor:CRRA\] visualizes the estimated input functions, Figure \[fig:Analysis:FittedConsumptionInvestment:Benchmark:RichInvestor:CRRA\] provides the graphics about the fitted consumption and relative risky portfolio process with the expected wealth and stock price path. Besides a larger sum of the squared relative distances for model $M_{a(t),b(t)}^{CRRA}$, especially the fitted risky investments $\hat{\pi}^{\star}(t ; v_{0})$ in Figure \[fig:Analysis:FittedConsumptionInvestment:Benchmark:RichInvestor:CRRA\] show that zero floors for consumption and wealth ($\bar{c}(t) \equiv 0$ and $F \equiv 0$) leads to an imprecise calibration and a large deviation from its given target curve due to a drop in model flexibility. Table \[tab:Analysis:ParameterEstimation:Error:Benchmark:RichInvestor:CRRA\] suggests that this drop in flexibility is attempted to be compensated by a higher risk aversion in terms of more negative estimated values for $\hat{b}$ and $b(t)$, see also Figure \[fig:Analysis:InputFunctions:Benchmark:RichInvestor:CRRA\].
\
Conclusion {#sec:Conclusion}
==========
This paper studies the optimal quantitative and dynamic consumption and investment strategies under age-dependent risk preferences (coefficient of risk aversion $b(t)$ and preference between consumption and terminal wealth $a(t)$). The findings demonstrate that strategies applied for life-cycle pension funds or pension insurance could significantly be improved by taking age-dependent risk preferences into account. For this reason, the paper combines the elements terminal wealth with a minimum level and consumption under time-varying risk preferences and minimum level into a dynamic life-cycle consumption-investment model. A sound economic understanding of the model parts is provided. In Section \[sec:SeparationTechnique\] the corresponding portfolio optimization problem is solved analytically with a separation approach which allows to solve the consumption and the terminal wealth part of the original consumption-investment problem separately. The formulas show that age-depending risk preferences in combination with terminal wealth considerations and minimum levels for consumption and wealth have a significant impact on the optimal controls.
Section \[sec:NCS\] investigates the optimal controls and provides a comparison with already existing and solved benchmark models. The analysis is divided into two parts. In the first part the risk preferences are calibrated towards given realistic curves for consumption and investment. The result emphasizes that only our proposed flexible model, in comparison with the other considered benchmark models, provides an adequate fit of the agent’s behavior. We draw the conclusion that time-varying preferences (risk aversion $b(t)$ and preference between consumption and terminal wealth $a(t)$) are necessary to provide a sufficient degree of flexibility to accurately fit the two control variables consumption and investment: Our proposed model turns out to be able to explain the given investor consumption and investment decisions, but the benchmark models fail. The very same result is obtained when time-dependent preference functions are considered, but the consumption and wealth floors are omitted. The second part focuses on the behavior analysis of the optimal consumption, investment and wealth under a positive and negative market environment.
Future research on this topic could deal with generalizations of the dynamic life-cycle model. For instance, investment constraints could be included to make the whole setup more applicable as budgets in practice are commonly exposed to constraints on allocation or risk. Furthermore, since unemployment risk and uncertain future income are essential for individuals, those risks and impacts on the optimal controls and wealth process could be further explored. Finally, including mortality and a life insurance product into the model could help people in determining their optimal individual life insurance investment embedded in a more realistic, flexible framework.
Acknowledgements {#acknowledgements .unnumbered}
================
Pavel V. Shevchenko acknowledges the support of Australian Research Council’s Discovery Projects funding scheme (project number DP160103489).
Proofs {#app:Proofs}
======
The consumption problem {#app:ProofsConsumptionProblem}
-----------------------
The Lagrangian of the Problem subject to is $$\begin{aligned}
\mathcal{L}(c,\lambda_{1}) = {} & {\mathbb{E}}\left[\int_{0}^{T} U_{1}(t,c(t)) dt\right] - \lambda_{1} \left({\mathbb{E}}\left[\int_{0}^{T} \tilde{Z}(t) \left(c(t) - y(t)\right) dt\right] - v_{1}\right) \\
= {} & {\mathbb{E}}\left[\int_{0}^{T} U_{1}(t,c(t)) - \lambda_{1} \left(\tilde{Z}(t) \left(c(t) - y(t)\right) - \frac{1}{T} v_{1}\right) dt\right].\end{aligned}$$ By the structure of the utility function, the optimal $c_{1}$ fulfills $c_{1}(t;v_{1}) > \bar{c}(t)$ and thus the first order conditions involve existence of a Lagrange multiplier $\lambda_{1} = \lambda_{1}(v_{1}) > 0$ such that the optimal $c_{1}$ maximizes $\mathcal{L}(c,\lambda_{1})$ and such that complementary slackness holds true. Hence it can be shown that the Karush-Kuhn-Tucker conditions besides the first derivative condition are satisfied.
Following [@Aase2017], let $\nabla_{h} \mathcal{L}(c,\lambda_{1};h)$ denote the directional derivative of $\mathcal{L}(c,\lambda_{1})$ in the feasible direction $h$. The directional derivative of a function $f$ in the direction $h$ is generally defined by $$\begin{aligned}
\nabla_{h} f(x) = \lim_{y \to 0} \frac{f(x + h y) - f(x)}{y}.\end{aligned}$$ If $f$ is differentiable at $x$ this results in $$\begin{aligned}
\nabla_{h} f(x) = f^{\prime}(x) h.\end{aligned}$$ In our case, for the inner function it holds $$\begin{aligned}
\nabla_{h} \left(U_{1}(t,c(t)) - \lambda_{1} \left(\tilde{Z}(t) \left(c(t) - y(t)\right) - \frac{1}{T} v_{1}\right)\right) = {} & \frac{\partial}{\partial c} \left(U_{1}(t,c(t)) - \lambda_{1} \left(\tilde{Z}(t) \left(c(t) - y(t)\right) - \frac{1}{T} v_{1}\right)\right) h(t) \\
= {} & \left(\frac{\partial}{\partial c} U_{1}(t,c(t)) - \lambda_{1} \tilde{Z}(t)\right) h(t).\end{aligned}$$ By the dominated convergence theorem, which allows interchanging expectation and differentiation, the first order condition gives $$\begin{aligned}
0 = {} & {\mathbb{E}}\left[\int_{0}^{T} \left(\frac{\partial}{\partial c} U_{1}(t,c(t)) - \lambda_{1} \tilde{Z}(t)\right) h(t) dt\right] \\
= {} & {\mathbb{E}}\left[\int_{0}^{T} \left(e^{- \beta t} a(t) \left(\frac{1}{1-b(t)} \left(c(t) - \bar{c}(t)\right)\right)^{b(t)-1} - \lambda_{1} \tilde{Z}(t)\right) h(t) dt\right]\end{aligned}$$ for all feasible $h$. In order to fulfill this condition for any $h$, the optimal consumption rate process must be $$\begin{aligned}
\label{eq:Solution:c1:ConsumptionOnly}
c_{1}(t;v_{1}) = (1-b(t)) \left(\lambda_{1} \frac{e^{\beta t}}{a(t)} \tilde{Z}(t)\right)^{\frac{1}{b(t)-1}} + \bar{c}(t),\ t \in [0,T].\end{aligned}$$ Since $U_{1}(t,c)$ strictly increases in $c$, the budget constraint for the optimal solution in turns to equality, i.e. $$\begin{aligned}
{\mathbb{E}}\left[\int_{0}^{T} \tilde{Z}(t) \left(c_{1}(t;v_{1}) - y(t)\right) dt\right] = v_{1}.\end{aligned}$$ When plugging in and by Fubini, the budget condition turns into $$\begin{aligned}
v_{1} = {} & {\mathbb{E}}\left[\int_{0}^{T} \tilde{Z}(t) \left((1-b(t)) \left(\lambda_{1} \frac{e^{\beta t}}{a(t)} \tilde{Z}(t)\right)^{\frac{1}{b(t)-1}} + \bar{c}(t) - y(t)\right) dt\right] \\
= {} & \int_{0}^{T} (1-b(t)) \left(\lambda_{1} \frac{e^{\beta t}}{a(t)}\right)^{\frac{1}{b(t)-1}} {\mathbb{E}}\left[\tilde{Z}(t)^{\frac{b(t)}{b(t)-1}}\right] dt + \int_{0}^{T} {\mathbb{E}}\left[\tilde{Z}(t)\right] \left(\bar{c}(t) - y(t)\right) dt \\
= {} & \int_{0}^{T} (1-b(t)) \left(\lambda_{1} \frac{e^{\beta t}}{a(t)}\right)^{\frac{1}{b(t)-1}} e^{- \frac{b(t)}{b(t)-1} \left(r + \frac{1}{2}\|\gamma\|^{2}\right) t + \frac{1}{2} \left(\frac{b(t)}{b(t)-1}\right)^{2} \|\gamma\|^{2} t} dt \\
& + \int_{0}^{T} e^{- \left(r + \frac{1}{2}\|\gamma\|^{2}\right) t + \frac{1}{2} \|\gamma\|^{2} t} \left(\bar{c}(t) - y(t)\right) dt \\
= {} & \int_{0}^{T} (1-b(t)) \left(\frac{e^{\left[\beta - b(t) \left(r - \frac{1}{2} \frac{1}{b(t)-1} \|\gamma\|^{2}\right)\right] t}}{a(t)}\right)^{\frac{1}{b(t)-1}} \lambda_{1}^{\frac{1}{b(t)-1}} dt + \int_{0}^{T} e^{- r t} \left(\bar{c}(t) - y(t)\right) dt \\
= {} & \int_{0}^{T} (1-b(t)) \left(\frac{e^{\left[\beta - b(t) \left(r - \frac{1}{2} \frac{1}{b(t)-1} \|\gamma\|^{2}\right)\right] t}}{a(t)}\right)^{\frac{1}{b(t)-1}} \lambda_{1}^{\frac{1}{b(t)-1}} dt + F_{1}(0).\end{aligned}$$ Here we used that $\tilde{Z}(t)$ is a log-normal random variable and so is $\tilde{Z}(t)^{\frac{b(t)}{b(t)-1}}$. For any $v_{1} > F_{1}(0) = \int_{0}^{T} e^{- r t} \left(\bar{c}(t) - y(t)\right) dt$, the above equality determines $\lambda_{1} > 0$ uniquely, since the integral in which $\lambda_{1}$ appears strictly decreases in $\lambda_{1}$ and has the limits $0$ and $\infty$ as $\lambda_{1}$ approaches $\infty$ and $0$. It follows immediately that the condition $v_{1} > \int_{0}^{T} e^{- r t} \left(\bar{c}(t) - y(t)\right) dt$ in is inevitable. The optimal wealth process $V_{1}(t ; v_{1})$ which arises by applying $c_{1}(t ; v_{1})$ is $$\begin{aligned}
V_{1}(t ; v_{1}) = {} & {\mathbb{E}}\left[\int_{t}^{T} \frac{\tilde{Z}(s)}{\tilde{Z}(t)} \left(c_{1}(s;v_{1}) - y(s)\right) ds \Big| \mathcal{F}_{t}\right] \\
= {} & \frac{1}{\tilde{Z}(t)} {\mathbb{E}}\left[\int_{t}^{T} \tilde{Z}(s) \left\{(1-b(s)) \left(\lambda_{1} \frac{e^{\beta s}}{a(s)} \tilde{Z}(s)\right)^{\frac{1}{b(s)-1}} + \bar{c}(s) - y(s)\right\} ds \Big| \mathcal{F}_{t}\right] \\
= {} & \frac{1}{\tilde{Z}(t)} \left\{{\mathbb{E}}\left[\int_{t}^{T} (1-b(s)) \left(\lambda_{1} \frac{e^{\beta s}}{a(s)}\right)^{\frac{1}{b(s)-1}} \tilde{Z}(s)^{\frac{b(s)}{b(s)-1}} ds \Big| \mathcal{F}_{t}\right] + {\mathbb{E}}\left[\int_{t}^{T} \tilde{Z}(s) \left(\bar{c}(s) - y(s)\right) ds \Big| \mathcal{F}_{t}\right]\right\} \displaybreak \\
= {} & \frac{1}{\tilde{Z}(t)} \left\{\int_{t}^{T} (1-b(s)) \left(\lambda_{1} \frac{e^{\beta s}}{a(s)}\right)^{\frac{1}{b(s)-1}} {\mathbb{E}}\left[\tilde{Z}(s)^{\frac{b(s)}{b(s)-1}} \Big| \mathcal{F}_{t}\right] ds + \int_{t}^{T} \left(\bar{c}(s) - y(s)\right) {\mathbb{E}}\left[\tilde{Z}(s) \Big| \mathcal{F}_{t}\right] ds\right\}.\end{aligned}$$ $\tilde{Z}(s)$ can be written as $\frac{\tilde{Z}(s)}{\tilde{Z}(t)} \tilde{Z}(t)$ where $\frac{\tilde{Z}(s)}{\tilde{Z}(t)}$ is independent of $\mathcal{F}_{t}$ and $\tilde{Z}(t)$ is $\mathcal{F}_{t}$-measurable. Therefore it follows $$\begin{aligned}
{\mathbb{E}}\left[\tilde{Z}(s) \Big| \mathcal{F}_{t}\right] = {} & \tilde{Z}(t) {\mathbb{E}}\left[\frac{\tilde{Z}(s)}{\tilde{Z}(t)}\right] = \tilde{Z}(t) e^{- \left(r + \frac{1}{2}\|\gamma\|^{2}\right) (s-t) + \frac{1}{2} \|\gamma\|^{2} (s-t)} = \tilde{Z}(t) e^{- r (s-t)}, \\
{\mathbb{E}}\left[\tilde{Z}(s)^{\eta} \Big| \mathcal{F}_{t}\right] = {} & \tilde{Z}(t)^{\eta} {\mathbb{E}}\left[\left(\frac{\tilde{Z}(s)}{\tilde{Z}(t)}\right)^{\eta}\right] = \tilde{Z}(t)^{\eta} e^{- \eta \left(r + \frac{1}{2}\|\gamma\|^{2}\right) (s-t) + \frac{1}{2} \eta^{2} \|\gamma\|^{2} (s-t)} = \tilde{Z}(t)^{\eta} e^{- \eta \left(r - \frac{1}{2} (\eta - 1) \|\gamma\|^{2}\right) (s-t)}\end{aligned}$$ for any $\eta \in {\mathbb{R}}$, where we used that $\frac{\tilde{Z}(s)}{\tilde{Z}(t)}$ and thus $\left(\frac{\tilde{Z}(s)}{\tilde{Z}(t)}\right)^{\eta}$ are log-normally distributed. Define the function $g$ by $$\begin{aligned}
g(s,t; v_{1}) = (1-b(s)) \left(\frac{e^{\beta s - b(s) \left(r - \frac{1}{2} \frac{1}{b(s)-1} \|\gamma\|^{2}\right) (s-t)}}{a(s)}\right)^{\frac{1}{b(s)-1}} \lambda_{1}^{\frac{1}{b(s)-1}},\end{aligned}$$ then the optimal wealth process is given by $$\begin{aligned}
\label{eq:V1:ConsumptionProblem}
V_{1}(t ; v_{1}) = \int_{t}^{T} g(s,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(s)-1}} ds + F_{1}(t)\end{aligned}$$ with $F_{1}(t)$ defined in . The dynamics can be calculated as $$\begin{aligned}
d V_{1}(t ; v_{1}) = {} & \left(- g(t,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(t)-1}} dt + \int_{t}^{T} d_{t} \left(g(s,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(s)-1}}\right) ds\right) \\
& + \left(- \left(\bar{c}(t) - y(t)\right) dt + \int_{t}^{T} d_{t} \left(e^{- r (s-t)} \left(\bar{c}(s) - y(s)\right)\right) ds\right) \\
= {} & - g(t,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(t)-1}} dt + \int_{t}^{T} d_{t} \left(g(s,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(s)-1}}\right) ds - \left(\bar{c}(t) - y(t)\right) dt \\
& + \left(\int_{t}^{T} r e^{- r (s-t)} \left(\bar{c}(s) - y(s)\right) ds\right) dt \\
= {} & \left(- g(t,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(t)-1}} - \left(\bar{c}(t) - y(t)\right) + \int_{t}^{T} r e^{- r (s-t)} \left(\bar{c}(s) - y(s)\right) ds\right) dt \\
& + \int_{t}^{T} d_{t} \left(g(s,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(s)-1}}\right) ds.\end{aligned}$$ Notice that by Itô’s formula, $$\begin{aligned}
d \left(\tilde{Z}(t)^{\frac{1}{b(s)-1}}\right) = {} & \frac{1}{b(s)-1} \tilde{Z}(t)^{\frac{1}{b(s)-1} - 1} d \tilde{Z}(t) + \frac{1}{2} \frac{1}{b(s)-1} \left(\frac{1}{b(s)-1} - 1\right) \tilde{Z}(t)^{\frac{1}{b(s)-1} - 2} \tilde{Z}(t)^{2} \|\gamma\|^{2} dt \\
= {} & \tilde{Z}(t)^{\frac{1}{b(s)-1}} \left\{\left[- \frac{1}{b(s)-1} r + \frac{1}{2} \frac{1}{b(s)-1} \left(\frac{1}{b(s)-1} - 1\right) \|\gamma\|^{2}\right] dt - \frac{1}{b(s)-1} \gamma' dW(t)\right\}.\end{aligned}$$ Moreover, it holds $$\begin{aligned}
d_{t} g(s,t; v_{1}) = {} & (1-b(s)) \left(\frac{e^{\beta s - b(s) \left(r - \frac{1}{2} \frac{1}{b(s)-1} \|\gamma\|^{2}\right) s}}{a(s)}\right)^{\frac{1}{b(s)-1}} \lambda_{1}^{\frac{1}{b(s)-1}} d_{t} \left(e^{\frac{b(s)}{b(s)-1} \left(r - \frac{1}{2} \frac{1}{b(s)-1} \|\gamma\|^{2}\right) t}\right) \\
= {} & (1-b(s)) \left(\frac{e^{\beta s - b(s) \left(r - \frac{1}{2} \frac{1}{b(s)-1} \|\gamma\|^{2}\right) s}}{a(s)}\right)^{\frac{1}{b(s)-1}} \lambda_{1}^{\frac{1}{b(s)-1}} \\
& \times \frac{b(s)}{b(s)-1} \left(r - \frac{1}{2} \frac{1}{b(s)-1} \|\gamma\|^{2}\right) e^{\frac{b(s)}{b(s)-1} \left(r - \frac{1}{2} \frac{1}{b(s)-1} \|\gamma\|^{2}\right) t} dt \\
= {} & \frac{b(s)}{b(s)-1} \left(r - \frac{1}{2} \frac{1}{b(s)-1} \|\gamma\|^{2}\right) (1-b(s)) \left(\frac{e^{\beta s - b(s) \left(r - \frac{1}{2} \frac{1}{b(s)-1} \|\gamma\|^{2}\right) (s-t)}}{a(s)}\right)^{\frac{1}{b(s)-1}} \lambda_{1}^{\frac{1}{b(s)-1}} dt \\
= {} & \frac{b(s)}{b(s)-1} \left(r - \frac{1}{2} \frac{1}{b(s)-1} \|\gamma\|^{2}\right) g(s,t; v_{1}) dt.\end{aligned}$$ With this we obtain $$\begin{aligned}
d_{t} \left(g(s,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(s)-1}}\right) = {} & g(s,t; v_{1}) d \left(\tilde{Z}(t)^{\frac{1}{b(s)-1}}\right) + \tilde{Z}(t)^{\frac{1}{b(s)-1}} d_{t} g(s,t; v_{1}) + 0 \\
= {} & g(s,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(s)-1}} \\
& \times \left\{\left[- \frac{1}{b(s)-1} r + \frac{1}{2} \frac{1}{b(s)-1} \left(\frac{1}{b(s)-1} - 1\right) \|\gamma\|^{2}\right] dt - \frac{1}{b(s)-1} \gamma' dW(t)\right\} \\
& + \tilde{Z}(t)^{\frac{1}{b(s)-1}} \frac{b(s)}{b(s)-1} \left(r - \frac{1}{2} \frac{1}{b(s)-1} \|\gamma\|^{2}\right) g(s,t; v_{1}) dt \\
= {} & g(s,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(s)-1}} \left\{\left(r - \frac{1}{b(s)-1} \|\gamma\|^{2}\right) dt - \frac{1}{b(s)-1} \gamma' dW(t)\right\}.\end{aligned}$$ Define $$\begin{aligned}
Y(t) = \int_{t}^{T} \frac{1}{b(s)-1} g(s,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(s)-1}} ds.\end{aligned}$$ In summary, the dynamics of the optimal wealth process is then given by $$\begin{aligned}
d V_{1}(t ; v_{1}) = {} & \left(- g(t,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(t)-1}} - \left(\bar{c}(t) - y(t)\right) + \int_{t}^{T} r e^{- r (s-t)} \left(\bar{c}(s) - y(s)\right) ds\right) dt \nonumber \\
& + \int_{t}^{T} g(s,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(s)-1}} \left\{\left(r - \frac{1}{b(s)-1} \|\gamma\|^{2}\right) dt - \frac{1}{b(s)-1} \gamma' dW(t)\right\} ds \nonumber \\
= {} & \Bigg(- g(t,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(t)-1}} - \left(\bar{c}(t) - y(t)\right) + \int_{t}^{T} r e^{- r (s-t)} \left(\bar{c}(s) - y(s)\right) ds \nonumber \\
& + \int_{t}^{T} \left(r - \frac{1}{b(s)-1} \|\gamma\|^{2}\right) g(s,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(s)-1}} ds\Bigg) dt \nonumber \\
& - \underbrace{\left(\int_{t}^{T} \frac{1}{b(s)-1} g(s,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(s)-1}} ds\right)}_{= Y(t)} \gamma' dW(t) \nonumber \\
= {} & \Bigg\{r \underbrace{\left(\int_{t}^{T} e^{- r (s-t)} \left(\bar{c}(s) - y(s)\right) ds + \int_{t}^{T} g(s,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(s)-1}} ds\right)}_{= V_{1}(t ; v_{1})} \nonumber \\
& - g(t,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(t)-1}} - \left(\bar{c}(t) - y(t)\right) - \|\gamma\|^{2} \underbrace{\int_{t}^{T} \frac{1}{b(s)-1} g(s,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(s)-1}} ds}_{= Y(t)}\Bigg\} dt \nonumber \\
& - Y(t) \gamma' dW(t) \nonumber \\
= {} & \left(r V_{1}(t ; v_{1}) - g(t,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(t)-1}} - \left(\bar{c}(t) - y(t)\right) - \|\gamma\|^{2} Y(t)\right) dt - Y(t) \gamma' dW(t) \nonumber \\
= {} & \mu_{V_{1}}(t) dt - Y(t) \gamma' dW(t) \label{eq:SDE:OptimalV:ConsumptionOnly}\end{aligned}$$ with drift $$\begin{aligned}
\mu_{V_{1}}(t) = {} & r V_{1}(t ; v_{1}) - g(t,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(t)-1}} - \bar{c}(t) + y(t) - \|\gamma\|^{2} Y(t).\end{aligned}$$ By it follows $$\begin{aligned}
c_{1}(t;v_{1}) = (1-b(t)) \left(\lambda_{1} \frac{e^{\beta t}}{a(t)} \tilde{Z}(t)\right)^{\frac{1}{b(t)-1}} + \bar{c}(t) = g(t,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(t)-1}} + \bar{c}(t).\end{aligned}$$ Hence $$\begin{aligned}
\mu_{V_{1}}(t) = {} & r V_{1}(t ; v_{1}) - c_{1}(t;v_{1}) + y(t) - \|\gamma\|^{2} Y(t).\end{aligned}$$ In order to determine the optimal investment strategy $\pi_{1}(t ; v_{1})$ to Problem we compare the optimal wealth dynamics in and : $$\begin{aligned}
d V_{1}(t ; v_{1}) = {} & V_{1}(t ; v_{1}) \left[\left(r + \hat{\pi}_{1}(t ; v_{1})' \left(\mu - r \mathbf{1}\right)\right) dt + \hat{\pi}_{1}(t ; v_{1})'\sigma dW(t)\right] - c_{1}(t;v_{1}) dt + y(t) dt, \\
d V_{1}(t ; v_{1}) = {} & \left(r V_{1}(t ; v_{1}) - c_{1}(t;v_{1}) + y(t) - \|\gamma\|^{2} Y(t)\right) dt - Y(t) \gamma' dW(t).\end{aligned}$$ Matching the diffusion terms yields the equality $$\begin{aligned}
\hat{\pi}_{1}(t ; v_{1}) = - \frac{Y(t)}{V_{1}(t ; v_{1})} \Sigma^{-1} (\mu - r \mathbf{1})\end{aligned}$$ which simultaneously matches the drift terms. By the first mean value theorem for integrals[^2] it furthermore follows that there exists $\tilde{t}_{1} \in (t,T)$ such that $$\begin{aligned}
Y(t) = {} & \int_{t}^{T} \frac{1}{b(s)-1} g(s,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(s)-1}} ds = \frac{1}{b(\tilde{t}_{1})-1} \int_{t}^{T} g(s,t; v_{1}) \tilde{Z}(t)^{\frac{1}{b(s)-1}} ds \\
\stackrel{\eqref{eq:V1:ConsumptionProblem}}{=} {} & \frac{1}{b(\tilde{t}_{1})-1} \left(V_{1}(t ; v_{1}) - F_{1}(t)\right).\end{aligned}$$ This determines the optimal investment strategy to be $$\begin{aligned}
\hat{\pi}_{1}(t ; v_{1}) = \frac{1}{1 - b(\tilde{t}_{1})} \Sigma^{-1} (\mu - r \mathbf{1}) \frac{V_{1}(t ; v_{1}) - F_{1}(t)}{V_{1}(t ; v_{1})}.\end{aligned}$$
Firstly, the value function of this problem is $$\begin{aligned}
\mathcal{V}_{1}(v_{1}) = {} & {\mathbb{E}}\left[\int_{0}^{T} U_{1}(t,c_{1}(t;v_{1})) dt\right] = {\mathbb{E}}\left[\int_{0}^{T} e^{- \beta t} \frac{1-b(t)}{b(t)} a(t) \left(\frac{1}{1-b(t)} \left(c_{1}(t;v_{1}) - \bar{c}(t)\right)\right)^{b(t)} dt\right] \\
= {} & {\mathbb{E}}\left[\int_{0}^{T} e^{- \beta t} \frac{1-b(t)}{b(t)} a(t) \left(\lambda_{1} \frac{e^{\beta t}}{a(t)} \tilde{Z}(t)\right)^{\frac{b(t)}{b(t)-1}} dt\right] \\
= {} & \int_{0}^{T} e^{- \beta t} \frac{1-b(t)}{b(t)} a(t) \left(\lambda_{1} \frac{e^{\beta t}}{a(t)}\right)^{\frac{b(t)}{b(t)-1}} {\mathbb{E}}\left[\tilde{Z}(t)^{\frac{b(t)}{b(t)-1}}\right] dt \\
= {} & \int_{0}^{T} \frac{1-b(t)}{b(t)} \left(\frac{e^{\beta t}}{a(t)}\right)^{\frac{1}{b(t)-1}} \lambda_{1}^{\frac{b(t)}{b(t)-1}} {\mathbb{E}}\left[\tilde{Z}(t)^{\frac{b(t)}{b(t)-1}}\right] dt \\
= {} & \int_{0}^{T} \frac{1-b(t)}{b(t)} \left(\frac{e^{\beta t}}{a(t)}\right)^{\frac{1}{b(t)-1}} \lambda_{1}^{\frac{b(t)}{b(t)-1}} e^{- \frac{b(t)}{b(t)-1} \left(r + \frac{1}{2}\|\gamma\|^{2}\right) t + \frac{1}{2} \left(\frac{b(t)}{b(t)-1}\right)^{2} \|\gamma\|^{2} t} dt \displaybreak \\
= {} & \int_{0}^{T} \frac{1-b(t)}{b(t)} \left(\frac{e^{\left[\beta - b(t) \left(r - \frac{1}{2} \frac{1}{b(t)-1} \|\gamma\|^{2}\right)\right] t}}{a(t)}\right)^{\frac{1}{b(t)-1}} \lambda_{1}^{\frac{b(t)}{b(t)-1}} dt,\end{aligned}$$ where $\lambda_{1}$ is subject to . From differentiating both sides of Equation with respect to $v_{1}$ we derive $$\begin{aligned}
1 = {} & \frac{\partial}{\partial v_{1}} \int_{0}^{T} (1-b(t)) \left(\frac{e^{\left[\beta - b(t) \left(r - \frac{1}{2} \frac{1}{b(t)-1} \|\gamma\|^{2}\right)\right] t}}{a(t)}\right)^{\frac{1}{b(t)-1}} \lambda_{1}^{\frac{1}{b(t)-1}} dt \nonumber \\
= {} & \int_{0}^{T} (1-b(t)) \left(\frac{e^{\left[\beta - b(t) \left(r - \frac{1}{2} \frac{1}{b(t)-1} \|\gamma\|^{2}\right)\right] t}}{a(t)}\right)^{\frac{1}{b(t)-1}} \frac{\partial}{\partial v_{1}} \left(\lambda_{1}^{\frac{1}{b(t)-1}}\right) dt. \label{eq:Lagrange:ConsumptionOnly:helpderivative}\end{aligned}$$ This helps to identify $\mathcal{V}_{1}^{\prime}(v_{1})$ to be $$\begin{aligned}
\mathcal{V}_{1}^{\prime}(v_{1}) = {} & \frac{\partial}{\partial v_{1}} \int_{0}^{T} \frac{1-b(t)}{b(t)} \left(\frac{e^{\left[\beta - b(t) \left(r - \frac{1}{2} \frac{1}{b(t)-1} \|\gamma\|^{2}\right)\right] t}}{a(t)}\right)^{\frac{1}{b(t)-1}} \lambda_{1}^{\frac{b(t)}{b(t)-1}} dt \\
= {} & \int_{0}^{T} \frac{1-b(t)}{b(t)} \left(\frac{e^{\left[\beta - b(t) \left(r - \frac{1}{2} \frac{1}{b(t)-1} \|\gamma\|^{2}\right)\right] t}}{a(t)}\right)^{\frac{1}{b(t)-1}} \frac{\partial}{\partial v_{1}} \left(\lambda_{1}^{\frac{b(t)}{b(t)-1}}\right) dt \\
= {} & \int_{0}^{T} \frac{1-b(t)}{b(t)} \left(\frac{e^{\left[\beta - b(t) \left(r - \frac{1}{2} \frac{1}{b(t)-1} \|\gamma\|^{2}\right)\right] t}}{a(t)}\right)^{\frac{1}{b(t)-1}} \frac{\partial}{\partial v_{1}} \left(\left(\lambda_{1}^{\frac{1}{b(t)-1}}\right)^{b(t)}\right) dt \\
= {} & \int_{0}^{T} \frac{1-b(t)}{b(t)} \left(\frac{e^{\left[\beta - b(t) \left(r - \frac{1}{2} \frac{1}{b(t)-1} \|\gamma\|^{2}\right)\right] t}}{a(t)}\right)^{\frac{1}{b(t)-1}} b(t) \left(\lambda_{1}^{\frac{1}{b(t)-1}}\right)^{b(t)-1} \frac{\partial}{\partial v_{1}} \left(\lambda_{1}^{\frac{1}{b(t)-1}}\right) dt \\
= {} & \lambda_{1} \int_{0}^{T} (1-b(t)) \left(\frac{e^{\left[\beta - b(t) \left(r - \frac{1}{2} \frac{1}{b(t)-1} \|\gamma\|^{2}\right)\right] t}}{a(t)}\right)^{\frac{1}{b(t)-1}} \frac{\partial}{\partial v_{1}} \left(\lambda_{1}^{\frac{1}{b(t)-1}}\right) dt \stackrel{\eqref{eq:Lagrange:ConsumptionOnly:helpderivative}}{=} \lambda_{1}.\end{aligned}$$ further implies concavity of $\mathcal{V}_{1}(v_{1})$ as $$\begin{aligned}
1 = {} & \int_{0}^{T} (1-b(t)) \left(\frac{e^{\left[\beta - b(t) \left(r - \frac{1}{2} \frac{1}{b(t)-1} \|\gamma\|^{2}\right)\right] t}}{a(t)}\right)^{\frac{1}{b(t)-1}} \frac{\partial}{\partial v_{1}} \left(\lambda_{1}^{\frac{1}{b(t)-1}}\right) dt \\
= {} & \int_{0}^{T} (1-b(t)) \left(\frac{e^{\left[\beta - b(t) \left(r - \frac{1}{2} \frac{1}{b(t)-1} \|\gamma\|^{2}\right)\right] t}}{a(t)}\right)^{\frac{1}{b(t)-1}} \frac{1}{b(t)-1} \lambda_{1}^{\frac{1}{b(t)-1} - 1} \lambda_{1}^{\prime} dt \\
= {} & - \lambda_{1}^{\prime} \int_{0}^{T} \left(\frac{e^{\left[\beta - b(t) \left(r - \frac{1}{2} \frac{1}{b(t)-1} \|\gamma\|^{2}\right)\right] t}}{a(t)}\right)^{\frac{1}{b(t)-1}} \lambda_{1}^{- \frac{b(t)-2}{b(t)-1}} dt\end{aligned}$$ and thus $$\begin{aligned}
\mathcal{V}_{1}^{\prime\prime}(v_{1}) = \lambda_{1}^{\prime} = - \left(\int_{0}^{T} \left(\frac{e^{\left[\beta - b(t) \left(r - \frac{1}{2} \frac{1}{b(t)-1} \|\gamma\|^{2}\right)\right] t}}{a(t)}\right)^{\frac{1}{b(t)-1}} \lambda_{1}^{- \frac{b(t)-2}{b(t)-1}} dt\right)^{-1} < 0.\end{aligned}$$
The terminal wealth problem {#app:ProofsTerminalWealthProblem}
---------------------------
The Lagrangian of the Problem subject to is $$\begin{aligned}
\mathcal{L}(V,\lambda_{2}) = {} & {\mathbb{E}}\left[U_{2}(V)\right] - \lambda_{2} \left({\mathbb{E}}\left[\tilde{Z}(T) V\right] - v_{2}\right) = {\mathbb{E}}\left[U_{2}(V) - \lambda_{2} \left(\tilde{Z}(T) V - v_{2}\right)\right].\end{aligned}$$ First of all, it is clear that ${c_{2}(t ; v_{2}) \equiv 0}$. By the structure of the utility function, the optimal $V_{2}$ fulfills $V_{2}(T;v_{2}) > F$ and thus the first order conditions involve existence of a Lagrange multiplier $\lambda_{2} = \lambda_{2}(v_{2}) > 0$ such that the optimal $V_{2}$ maximizes $\mathcal{L}(V,\lambda_{2})$ and such that complementary slackness holds true. Hence it can be shown that the Karush-Kuhn-Tucker conditions besides the first derivative condition are satisfied. By the dominated convergence theorem, the first order condition with respect to the directional derivative gives $$\begin{aligned}
0 = {} & {\mathbb{E}}\left[\left(\frac{\partial}{\partial V} U_{2}(V) - \lambda_{2} \tilde{Z}(T)\right) h\right] = {\mathbb{E}}\left[\left(e^{- \beta T} \hat{a} \left(\frac{1}{1-\hat{b}} (V-F)\right)^{\hat{b}-1} - \lambda_{2} \tilde{Z}(T)\right) h\right],\end{aligned}$$ which has to be satisfied for all suitable $h$; hence the optimal terminal wealth has to fulfill $$\begin{aligned}
\label{eq:Solution:V2T:TerminalWealthOnly}
V_{2}(T ; v_{2}) = (1-\hat{b}) \left(\lambda_{2} \frac{e^{\beta T}}{\hat{a}} \tilde{Z}(T)\right)^{\frac{1}{\hat{b}-1}} + F.\end{aligned}$$ Since $U_{2}(V)$ strictly increases in $V$, complementary slackness implies equality for the budget constraint $$\begin{aligned}
{\mathbb{E}}\left[\tilde{Z}(T) V_{2}(T ; v_{2})\right] = v_{2}.\end{aligned}$$ Using and Fubini this gives $$\begin{aligned}
v_{2} = {} & {\mathbb{E}}\left[\tilde{Z}(T) \left((1-\hat{b}) \left(\lambda_{2} \frac{e^{\beta T}}{\hat{a}} \tilde{Z}(T)\right)^{\frac{1}{\hat{b}-1}} + F\right)\right] = (1-\hat{b}) \left(\lambda_{2} \frac{e^{\beta T}}{\hat{a}}\right)^{\frac{1}{\hat{b}-1}} {\mathbb{E}}\left[\tilde{Z}(T)^{\frac{\hat{b}}{\hat{b}-1}}\right] + F {\mathbb{E}}\left[\tilde{Z}(T)\right] \\
= {} & (1-\hat{b}) \left(\lambda_{2} \frac{e^{\beta T}}{\hat{a}}\right)^{\frac{1}{\hat{b}-1}} e^{- \frac{\hat{b}}{\hat{b}-1} \left(r + \frac{1}{2}\|\gamma\|^{2}\right) T + \frac{1}{2} \left(\frac{\hat{b}}{\hat{b}-1}\right)^{2} \|\gamma\|^{2} T} + F e^{- \left(r + \frac{1}{2}\|\gamma\|^{2}\right) T + \frac{1}{2} \|\gamma\|^{2} T} \\
= {} & (1-\hat{b}) \left(\frac{e^{\left[\beta - \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)\right] T}}{\hat{a}}\right)^{\frac{1}{\hat{b}-1}} \lambda_{2}^{\frac{1}{\hat{b}-1}} + e^{- r T} F \displaybreak \\
= {} & (1-\hat{b}) \left(\frac{e^{\left[\beta - \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)\right] T}}{\hat{a}}\right)^{\frac{1}{\hat{b}-1}} \lambda_{2}^{\frac{1}{\hat{b}-1}} + F_{2}(0).\end{aligned}$$ Solving for $\lambda_{2}$ yields $$\begin{aligned}
\label{eq:Lagrange:TerminalWealthOnly}
\lambda_{2} = \left(\frac{v_{2} - F_{2}(0)}{(1-\hat{b}) \left(\frac{e^{\left[\beta - \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)\right] T}}{\hat{a}}\right)^{\frac{1}{\hat{b}-1}}}\right)^{\hat{b}-1} = e^{-\left[\beta - \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)\right] T} \left(1-\hat{b}\right)^{1-\hat{b}} \hat{a} \left(v_{2} - F_{2}(0)\right)^{\hat{b}-1}\end{aligned}$$ where ${v_{2} > F_{2}(0) = e^{- r T} F}$ in is required. Plugging this back into , the optimal terminal wealth is $$\begin{aligned}
V_{2}(T ; v_{2}) = {} & (1-\hat{b}) \left(\frac{e^{\beta T}}{\hat{a}} \tilde{Z}(T)\right)^{\frac{1}{\hat{b}-1}} \lambda_{2}^{\frac{1}{\hat{b}-1}} + F \nonumber \\
= {} & (1-\hat{b}) \left(\frac{e^{\beta T}}{\hat{a}} \tilde{Z}(T)\right)^{\frac{1}{\hat{b}-1}} \left(\frac{v_{2} - F_{2}(0)}{(1-\hat{b}) \left(\frac{e^{\left[\beta - \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)\right] T}}{\hat{a}}\right)^{\frac{1}{\hat{b}-1}}}\right) + F \nonumber \\
= {} & \left(v_{2} - F_{2}(0)\right) \left(e^{\hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) T} \tilde{Z}(T)\right)^{\frac{1}{\hat{b}-1}} + F \nonumber \\
= {} & \left(v_{2} - F_{2}(0)\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) T} \tilde{Z}(T)^{\frac{1}{\hat{b}-1}} + F. \label{eq:Solution:V2T:TerminalWealthOnly:lambdaInserted}\end{aligned}$$ The optimal wealth process replicates $V_{2}(T ; v_{2})$ and is uniquely given by $$\begin{aligned}
V_{2}(t ; v_{2}) = {} & {\mathbb{E}}\left[\frac{\tilde{Z}(T)}{\tilde{Z}(t)} V_{2}(T ; v_{2}) \Big| \mathcal{F}_{t}\right] = \frac{1}{\tilde{Z}(t)} {\mathbb{E}}\left[\tilde{Z}(T) \left\{\left(v_{2} - F_{2}(0)\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) T} \tilde{Z}(T)^{\frac{1}{\hat{b}-1}} + F\right\} \Big| \mathcal{F}_{t}\right] \\
= {} & \frac{1}{\tilde{Z}(t)} \left\{\left(v_{2} - F_{2}(0)\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) T} {\mathbb{E}}\left[\tilde{Z}(T)^{\frac{\hat{b}}{\hat{b}-1}} \Big| \mathcal{F}_{t}\right] + F {\mathbb{E}}\left[\tilde{Z}(T) \Big| \mathcal{F}_{t}\right]\right\} \\
= {} & \frac{1}{\tilde{Z}(t)} \left\{\left(v_{2} - F_{2}(0)\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) T} \tilde{Z}(t)^{\frac{\hat{b}}{\hat{b}-1}} e^{- \frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) (T-t)} + F \tilde{Z}(t) e^{- r (T-t)}\right\}.\end{aligned}$$ This finally gives $$\begin{aligned}
\label{eq:V2:TerminalWealthProblem}
V_{2}(t ; v_{2}) = \left(v_{2} - F_{2}(0)\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} + F_{2}(t)\end{aligned}$$ with $F_{2}(t)$ defined in . Recall that $$\begin{aligned}
d \left(\tilde{Z}(t)^{\frac{1}{\hat{b}-1}}\right) = \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} \left\{\left[- \frac{1}{\hat{b}-1} r + \frac{1}{2} \frac{1}{\hat{b}-1} \left(\frac{1}{\hat{b}-1} - 1\right) \|\gamma\|^{2}\right] dt - \frac{1}{\hat{b}-1} \gamma' dW(t)\right\}.\end{aligned}$$ It follows by Itô $$\begin{aligned}
d & \left(e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} \tilde{Z}(t)^{\frac{1}{\hat{b}-1}}\right) \\
& = e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} d \left(\tilde{Z}(t)^{\frac{1}{\hat{b}-1}}\right) + \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} d \left(e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t}\right) + 0 \\
& = e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} \left\{\left[- \frac{1}{\hat{b}-1} r + \frac{1}{2} \frac{1}{\hat{b}-1} \left(\frac{1}{\hat{b}-1} - 1\right) \|\gamma\|^{2}\right] dt - \frac{1}{\hat{b}-1} \gamma' dW(t)\right\} \\
& \quad + \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} \frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} dt \\
& = e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} \\
& \quad \times \left\{\left[- \frac{1}{\hat{b}-1} r + \frac{1}{2} \frac{1}{\hat{b}-1} \left(\frac{1}{\hat{b}-1} - 1\right) \|\gamma\|^{2}\right] dt - \frac{1}{\hat{b}-1} \gamma' dW(t) + \frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) dt\right\} \\
& = e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} \left\{\frac{1}{\hat{b}-1}\left[(\hat{b}-1) r + \frac{1}{2} \left(\frac{1}{\hat{b}-1} - 1 - \frac{\hat{b}}{\hat{b}-1}\right) \|\gamma\|^{2}\right] dt - \frac{1}{\hat{b}-1} \gamma' dW(t)\right\} \\
& = e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} \left\{\left(r - \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) dt - \frac{1}{\hat{b}-1} \gamma' dW(t)\right\}.\end{aligned}$$
Then the optimal wealth dynamics can be calculated as $$\begin{aligned}
d V_{2}(t ; v_{2}) = {} & \left(v_{2} - F_{2}(0)\right) d \left(e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} \tilde{Z}(t)^{\frac{1}{\hat{b}-1}}\right) + r F_{2}(t) dt \\
= {} & \underbrace{\left(v_{2} - F_{2}(0)\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} \tilde{Z}(t)^{\frac{1}{\hat{b}-1}}}_{\stackrel{\eqref{eq:V2:TerminalWealthProblem}}{=} V_{2}(t ; v_{2}) - F_{2}(t)} \left\{\left(r - \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) dt - \frac{1}{\hat{b}-1} \gamma' dW(t)\right\} + r F_{2}(t) dt \\
= {} & r V_{2}(t ; v_{2}) dt + \left(V_{2}(t ; v_{2}) - F_{2}(t)\right) \left\{- \frac{1}{\hat{b}-1} \|\gamma\|^{2} dt - \frac{1}{\hat{b}-1} \gamma' dW(t)\right\}.\end{aligned}$$ Comparing the diffusion term with the one from for $y(t) \equiv 0$ implies $$\begin{aligned}
\hat{\pi}_{2}(t ; v_{2}) = \frac{1}{1-\hat{b}} \Sigma^{-1} (\mu - r \mathbf{1}) \frac{V_{2}(t ; v_{2}) - F_{2}(t)}{V_{2}(t ; v_{2})}\end{aligned}$$ which automatically matches the drifts iff ${c_{2}(t ; v_{2}) \equiv 0}$.
The value function of this problem is given by $$\begin{aligned}
\mathcal{V}_{2}(v_{2}) = {} & {\mathbb{E}}\left[U_{2}(V_{2}(T ; v_{2}))\right] = {\mathbb{E}}\left[U_{2}\left(\left(v_{2} - F_{2}(0)\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) T} \tilde{Z}(T)^{\frac{1}{\hat{b}-1}} + F\right)\right] \\
= {} & e^{- \beta T} \frac{1-\hat{b}}{\hat{b}} \hat{a} \left(\frac{1}{1-\hat{b}}\right)^{\hat{b}} \left(v_{2} - F_{2}(0)\right)^{\hat{b}} e^{\frac{\hat{b}^{2}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) T} {\mathbb{E}}\left[\tilde{Z}(T)^{\frac{\hat{b}}{\hat{b}-1}}\right] \\
= {} & e^{- \beta T} \frac{\left(1-\hat{b}\right)^{1-\hat{b}}}{\hat{b}} \hat{a} \left(v_{2} - F_{2}(0)\right)^{\hat{b}} e^{\frac{\hat{b}^{2}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) T} e^{- \frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) T} \displaybreak \\
= {} & e^{\left[- \beta + \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)\right] T} \frac{\left(1-\hat{b}\right)^{1-\hat{b}}}{\hat{b}} \hat{a} \left(v_{2} - F_{2}(0)\right)^{\hat{b}}.\end{aligned}$$ This implies $$\begin{aligned}
\mathcal{V}_{2}^{\prime}(v_{2}) = {} & e^{\left[- \beta + \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)\right] T} \frac{\left(1-\hat{b}\right)^{1-\hat{b}}}{\hat{b}} \hat{a} \hat{b} \left(v_{2} - F_{2}(0)\right)^{\hat{b}-1} \\
= {} & e^{\left[- \beta + \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)\right] T} \left(1-\hat{b}\right)^{1-\hat{b}} \hat{a} \left(v_{2} - F_{2}(0)\right)^{\hat{b}-1} \\
\stackrel{\eqref{eq:Lagrange:TerminalWealthOnly}}{=} {} \lambda_{2} > 0.\end{aligned}$$ Due to the assumption ${v_{2} - F_{2}(0) > 0}$ in , it is straightforward that ${\mathcal{V}_{2}^{\prime\prime}(v_{2}) = \lambda_{2}^{\prime} < 0}$: $$\begin{aligned}
\mathcal{V}_{2}^{\prime\prime}(v_{2}) = - e^{\left[- \beta + \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)\right] T} \left(1-\hat{b}\right)^{2-\hat{b}} \hat{a} \left(v_{2} - F_{2}(0)\right)^{\hat{b}-2} < 0.\end{aligned}$$
Optimal merging of the individual solutions {#app:ProofsOptimalMerging}
-------------------------------------------
1. $\mathcal{V}(v_{0}) \ge \sup_{v_{1} \ge F_{1}(0),\ v_{2} \ge F_{2}(0),\ v_{1} + v_{2} = v_{0}} \left\{\mathcal{V}_{1}(v_{1}) + \mathcal{V}_{2}(v_{2})\right\}$:
Let $(\pi_{1}(t;v_{1}), c_{1}(t;v_{1}))$ and $(\pi_{2}(t;v_{2}), c_{2}(t;v_{2}))$ denote the optimal controls to Problems and with optimal wealth processes $V_{1}(t;v_{1})$ and $V_{2}(t;v_{2})$ to the initial wealths ${v_{1} \ge F_{1}(0)}$ and ${v_{2} \ge F_{2}(0)}$. Then, as the budget constraints for the optimal solutions to all three problems hold with equality, $$\begin{aligned}
\mathcal{V}_{1}(v_{1}) + \mathcal{V}_{2}(v_{2}) = {} & {\mathbb{E}}\left[\int_{0}^{T} U_{1}(t,c_{1}(t;v_{1})) dt + U_{2}(V_{2}(T;v_{2}))\right] \\
\le {} & \sup_{(\pi,c) \in \Lambda} J(\pi,c;v_{0}) = \mathcal{V}(v_{0})\end{aligned}$$ for all $v_{1},v_{2}$ with ${v_{1} + v_{2} = v_{0}}$. Thus $$\begin{aligned}
\mathcal{V}(v_{0}) \ge \sup_{v_{1} \ge F_{1}(0),\ v_{2} \ge F_{2}(0),\ v_{1} + v_{2} = v_{0}} \left\{\mathcal{V}_{1}(v_{1}) + \mathcal{V}_{2}(v_{2})\right\}.\end{aligned}$$
2. $\mathcal{V}(v_{0}) \le \sup_{v_{1} \ge F_{1}(0),\ v_{2} \ge F_{2}(0),\ v_{1} + v_{2} = v_{0}} \left\{\mathcal{V}_{1}(v_{1}) + \mathcal{V}_{2}(v_{2})\right\}$:
Let $(\pi^{\star}, c^{\star})$ denote the optimal controls which maximize $\mathcal{V}(v_{0})$ with optimal wealth process $V^{\star}$ to the initial wealth $v_{0} > 0$. Define $$\begin{aligned}
v_{1} = {\mathbb{E}}\left[\int_{0}^{T} \tilde{Z}(t) \left(c^{\star}(t) - y(t)\right) dt\right],\ v_{2} = {\mathbb{E}}\left[\tilde{Z}(T) V^{\star}(T)\right].\end{aligned}$$ Then, ${v_{1} + v_{2} = v_{0}}$ and $$\begin{aligned}
\mathcal{V}(v_{0}) = {} & {\mathbb{E}}\left[\int_{0}^{T} U_{1}(t,c^{\star}(t)) dt\right] + {\mathbb{E}}\left[U_{2}(V^{\star}(T))\right] \le \mathcal{V}_{1}(v_{1}) + \mathcal{V}_{2}(v_{2}).\end{aligned}$$ Hence $$\begin{aligned}
\mathcal{V}(v_{0}) \le \sup_{v_{1} \ge F_{1}(0),\ v_{2} \ge F_{2}(0),\ v_{1} + v_{2} = v_{0}} \left\{\mathcal{V}_{1}(v_{1}) + \mathcal{V}_{2}(v_{2})\right\}.\end{aligned}$$
In accordance with Theorem \[thm:ConnectionValueFunctions\] and by expressing $v_{2} = v_{0} - v_{1}$, the candidate for the optimal ${v_{1}^{\star}}$ is the one that satisfies the first order derivative condition on the budget $$\begin{aligned}
0 = \frac{\partial}{\partial v_{1}} \left(\mathcal{V}_{1}(v_{1}) + \mathcal{V}_{2}(v_{0} - v_{1})\right) = \mathcal{V}_{1}^{\prime}(v_{1}) - \mathcal{V}_{2}^{\prime}(v_{0} - v_{1})\end{aligned}$$ such that ${v_{1}^{\star} \ge F_{1}(0)}$, ${v_{2}^{\star} = v_{0} - v_{1}^{\star}}$ with ${v_{2}^{\star} \ge F_{2}(0)}$; thus ${F_{1}(0) \le v_{1}^{\star} \le v_{0} - F_{2}(0)}$. Theorems \[thm:Solution:ConsumptionOnly:ValueFunction:lambda\] and \[thm:Solution:TerminalWealthOnly:ValueFunction:lambda\] tell that $\mathcal{V}_{1}(v_{1})$ and $\mathcal{V}_{2}(v_{2})$ are strictly concave functions in $v_{1}$ respectively $v_{2}$. Therefore, it follows $$\begin{aligned}
0 = \frac{\partial^{2}}{\partial v_{1}^{2}} \left(\mathcal{V}_{1}(v_{1}) + \mathcal{V}_{2}(v_{0} - v_{1})\right) = \mathcal{V}_{1}^{\prime\prime}(v_{1}) + \mathcal{V}_{2}^{\prime\prime}(v_{0} - v_{1}) < 0.\end{aligned}$$ This implies that the candidates $v_{1}^{\star}$ and ${v_{2}^{\star} = v_{0} - v_{1}^{\star}}$ are the solution when the constraint $F_{1}(0) \le v_{1}^{\star} \le v_{0} - F_{2}(0)$ applies.
In accordance with Theorems \[thm:Solution:ConsumptionOnly:ValueFunction:lambda\] and \[thm:Solution:TerminalWealthOnly:ValueFunction:lambda\] we have $$\begin{aligned}
\mathcal{V}_{1}^{\prime}(v_{1}) = {} & \lambda_{1}, \\
\mathcal{V}_{2}^{\prime}(v_{2}) = {} & \lambda_{2} = e^{-\left[\beta - \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)\right] T} \left(1-\hat{b}\right)^{1-\hat{b}} \hat{a} \left(v_{2} - F_{2}(0)\right)^{\hat{b}-1}.\end{aligned}$$ By equating $\mathcal{V}_{1}^{\prime}(v_{1})$ and $\mathcal{V}_{2}^{\prime}(v_{0} - v_{1})$ we obtain $$\begin{aligned}
\eqref{eq:OptimalMerging:ValueFunction} \text{ in Lemma \ref{lemma:EqualityConditionValueFunctions}} \ \Leftrightarrow\ & \mathcal{V}_{1}^{\prime}(v_{1}) = \mathcal{V}_{2}^{\prime}(v_{0} - v_{1}) \\
\Leftrightarrow\ & \lambda_{1} = \lambda_{2} = e^{-\left[\beta - \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)\right] T} \left(1-\hat{b}\right)^{1-\hat{b}} \hat{a} \left(v_{0} - v_{1} - F_{2}(0)\right)^{\hat{b}-1}.\end{aligned}$$ Inserting $\lambda_{1}$ in Equation , the optimal $v_{1}^{\star}$ is the solution to $$\begin{aligned}
v_{1} - \int_{0}^{T} \chi(t) \left(v_{0} - v_{1} - F_{2}(0)\right)^{\frac{\hat{b}-1}{b(t)-1}} dt = F_{1}(0),\end{aligned}$$ where the continuous function $\chi(t)$ is defined by $$\begin{aligned}
\chi(t) = (1-b(t)) \left(1-\hat{b}\right)^{\frac{1-\hat{b}}{b(t)-1}} \left(\frac{\hat{a}}{a(t)}\right)^{\frac{1}{b(t)-1}} \left(\frac{e^{\left[\beta - b(t) \left(r - \frac{1}{2} \frac{1}{b(t)-1} \|\gamma\|^{2}\right)\right] t}}{e^{\left[\beta - \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)\right] T}}\right)^{\frac{1}{b(t)-1}} > 0.\end{aligned}$$ It remains to verify ${F_{1}(0) \le v_{1}^{\star} \le v_{0} - F_{2}(0)}$ and uniqueness of $v_{1}^{\star}$. For this sake, define the function $f$ by $$\begin{aligned}
f: (-\infty,v_{0} - F_{2}(0)],\ f(x) = x - \int_{0}^{T} \chi(t) \left(v_{0} - x - F_{2}(0)\right)^{\frac{\hat{b}-1}{b(t)-1}} dt - F_{1}(0).\end{aligned}$$ $v_{1}^{\star}$ is the root of the function $f$, i.e. ${f(v_{1}^{\star}) = 0}$, if it holds ${v_{1}^{\star} \ge F_{1}(0)}$. $f$ is continuous in $x$, the exponent ${\frac{\hat{b}-1}{b(t)-1}}$ within the first integral is positive. Furthermore, due to ${v_{0} > F(0)}$ claimed in and ${F(t) = F_{1}(t) + F_{2}(t)}$, we have for the limits $$\begin{aligned}
\lim_{x \searrow F_{1}(0)} f(x) = {} & - \int_{0}^{T} \chi(t) \left(v_{0} - F_{1}(0) - F_{2}(0)\right)^{\frac{\hat{b}-1}{b(t)-1}} dt = - \int_{0}^{T} \chi(t) \left(v_{0} - F(0)\right)^{\frac{\hat{b}-1}{b(t)-1}} dt < 0, \\
\lim_{x \nearrow v_{0} - F_{2}(0)} f(x) = {} & v_{0} - F_{2}(0) - \int_{0}^{T} \chi(t) \left(v_{0} - \left(v_{0} - F_{2}(0)\right) - F_{2}(0)\right)^{\frac{\hat{b}-1}{b(t)-1}} dt - F_{1}(0) \\
= {} & v_{0} - F_{2}(0) - F_{1}(0) = v_{0} - F(0) > 0.\end{aligned}$$ Note, ${F_{1}(0) \le v_{1} = v_{0} - v_{2} \le v_{0} - F_{2}(0)}$ for general $v_{1}$ and $v_{2}$. Additionally, $f$ is strictly monotone increasing in $x$ since $$\begin{aligned}
f^{\prime}(x) = 1 + \int_{0}^{T} \chi(t) \frac{\hat{b}-1}{b(t)-1} \left(v_{0} - x - F_{2}(0)\right)^{\frac{\hat{b}-b(t)}{b(t)-1}} dt > 0,\ \forall x \le v_{0} - F_{2}(0).\end{aligned}$$ We conclude that there exists a unique root ${x \in [F_{1}(0), v_{0} - F_{2}(0)]}$ such that ${f(x) = 0}$. Therefore, we conclude that the optimal $v_{1}^{*}$ and $v_{2}^{\star} = v_{0} - v_{1}^{*}$ exist and are unique. $v_{1}^{*}$ is the solution to Equation . The optimal Lagrange multiplier $\lambda_{1}^{\star} = \lambda_{1}(v_{1}^{\star})$ is then given by $$\begin{aligned}
\lambda_{1}^{\star} = \left(1-\hat{b}\right)^{1-\hat{b}} \hat{a} e^{-\left[\beta - \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)\right] T} \left(v_{0} - v_{1}^{\star} - F_{2}(0)\right)^{\hat{b}-1}.\end{aligned}$$
Starting with ${V^{\star}(t; v_{0}) = V_{1}(t;v_{1}^{\star}) + V_{2}(t;v_{2}^{\star})}$ we compare the dynamics of both sides of the equation: $$\begin{aligned}
\label{eq:Merging:dV:dV1+dV2}
d V^{\star}(t; v_{0}) = d V_{1}(t;v_{1}^{\star}) + d V_{2}(t;v_{2}^{\star}).\end{aligned}$$ Equation for $V^{\star}(t; v_{0})$, $V_{1}(t;v_{1}^{\star})$ and $V_{2}(t;v_{2}^{\star})$, with $y(t) \equiv 0$ for $V_{2}(t;v_{2}^{\star})$, provides $$\begin{aligned}
d V^{\star}(t; v_{0}) = {} & V^{\star}(t; v_{0}) \left[\left(r + \hat{\pi}^{\star}(t; v_{0})' \left(\mu - r \mathbf{1}\right)\right) dt + \hat{\pi}^{\star}(t; v_{0})'\sigma dW(t)\right] - c^{\star}(t; v_{0}) dt + y(t) dt, \\
d V_{1}(t;v_{1}^{\star}) = {} & V_{1}(t;v_{1}^{\star}) \left[\left(r + \hat{\pi}_{1}(t;v_{1}^{\star})' \left(\mu - r \mathbf{1}\right)\right) dt + \hat{\pi}_{1}(t;v_{1}^{\star})'\sigma dW(t)\right] - c_{1}(t;v_{1}^{\star}) dt + y(t) dt, \\
d V_{2}(t;v_{2}^{\star}) = {} & V_{2}(t;v_{2}^{\star}) \left[\left(r + \hat{\pi}_{2}(t;v_{2}^{\star})' \left(\mu - r \mathbf{1}\right)\right) dt + \hat{\pi}_{2}(t;v_{2}^{\star})'\sigma dW(t)\right].\end{aligned}$$ Comparing the diffusion terms in gives $$\begin{aligned}
\hat{\pi}^{\star}(t; v_{0}) = \frac{\hat{\pi}_{1}(t;v_{1}^{\star}) V_{1}(t;v_{1}^{\star}) + \hat{\pi}_{2}(t;v_{2}^{\star}) V_{2}(t;v_{2}^{\star})}{V^{\star}(t; v_{0})}.\end{aligned}$$ Inserting this back and comparing the drift terms finally leads to $$\begin{aligned}
c^{\star}(t; v_{0}) = c_{1}(t;v_{1}^{\star}).\end{aligned}$$ Notice that the pair ${(\hat{\pi}^{\star},c^{\star})}$ is admissible, i.e. ${(\hat{\pi}^{\star},c^{\star}) \in \Lambda}$ because ${(\hat{\pi}_{1},c_{1}) \in \Lambda_{1}}$ and ${(\hat{\pi}_{2},0) \in \Lambda_{2}}$ which implies $$\begin{aligned}
V^{\star}(t; v_{0}) = \underbrace{V_{1}(t;v_{1}^{\star})}_{\ge - \int_{t}^{T} e^{- r (s-t)} y(s) ds} + \underbrace{V_{2}(t;v_{2}^{\star})}_{\ge 0} \ge - \int_{t}^{T} e^{- r (s-t)} y(s) ds,\ {\mathbb{P}}-a.s.,\ \forall t \in [0,T].\end{aligned}$$ Using the solutions in Theorems \[thm:Solution:ConsumptionOnly\] and \[thm:Solution:TerminalWealthOnly:WithoutProbabilityConstraint\] we derive the following for the utility setup in : $$\begin{aligned}
\hat{\pi}^{\star}(t ; v_{0}) = {} & \Sigma^{-1} (\mu - r \mathbf{1}) \frac{\frac{1}{1 - b(\tilde{t}_{1}^{\star})} \left(V_{1}(t ; v_{1}^{\star}) - F_{1}(t)\right) + \frac{1}{1-\hat{b}} \left(V_{2}(t ; v_{2}^{\star}) - F_{2}(t)\right)}{V^{\star}(t ; v_{0})}, \\
c^{\star}(t;v_{0}) = {} & c_{1}(t;v_{1}^{\star}) = g(t,t; v_{1}^{\star}) \tilde{Z}(t)^{\frac{1}{b(t)-1}} + \bar{c}(t) = (1-b(t)) \left(\lambda_{1}^{\star} \frac{e^{\beta t}}{a(t)} \tilde{Z}(t)\right)^{\frac{1}{b(t)-1}} + \bar{c}(t), \\
V^{\star}(t ; v_{0}) = {} & V_{1}(t;v_{1}^{\star}) + V_{2}(t;v_{2}^{\star}) \\
= {} & \int_{t}^{T} g(s,t; v_{1}^{\star}) \tilde{Z}(t)^{\frac{1}{b(s)-1}} ds + F_{1}(t) + \left(v_{2}^{\star} - F_{2}(0)\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} + F_{2}(t) \\
= {} & \int_{t}^{T} g(s,t; v_{1}^{\star}) \tilde{Z}(t)^{\frac{1}{b(s)-1}} ds + \left(v_{2}^{\star} - F_{2}(0)\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} + F(t), \\
V^{\star}(T ; v_{0}) = {} & \left(v_{2}^{\star} - F_{2}(0)\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) T} \tilde{Z}(T)^{\frac{1}{\hat{b}-1}} + F, \\
V_{1}(t ; v_{1}^{\star}) = {} & \int_{t}^{T} g(s,t; v_{1}^{\star}) \tilde{Z}(t)^{\frac{1}{b(s)-1}} ds + F_{1}(t), \\
V_{2}(t ; v_{2}^{\star}) = {} & \left(v_{2}^{\star} - F_{2}(0)\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} + F_{2}(t),\end{aligned}$$ for all $t \in [0,T]$, with $$\begin{aligned}
g(s,t; v_{1}^{\star}) = {} & (1-b(s)) \left(\frac{e^{\beta s - b(s) \left(r - \frac{1}{2} \frac{1}{b(s)-1} \|\gamma\|^{2}\right) (s-t)}}{a(s)}\right)^{\frac{1}{b(s)-1}} \left(\lambda_{1}^{\star}\right)^{\frac{1}{b(s)-1}} \\
= {} & (1-b(s)) \left(1-\hat{b}\right)^{\frac{1-\hat{b}}{b(s)-1}} \left(\frac{\hat{a}}{a(s)}\right)^{\frac{1}{b(s)-1}} \left(\frac{e^{\beta s - b(s) \left(r - \frac{1}{2} \frac{1}{b(s)-1} \|\gamma\|^{2}\right) (s-t)}}{e^{\left[\beta - \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)\right] T}}\right)^{\frac{1}{b(s)-1}} \left(v_{0} - v_{1}^{\star} - F_{2}(0)\right)^{\frac{\hat{b}-1}{b(s)-1}} \\
\stackrel{\eqref{eq:definition:chi}}{=} {} & \chi(s) e^{\frac{b(s)}{b(s)-1} \left(r - \frac{1}{2} \frac{1}{b(s)-1} \|\gamma\|^{2}\right) t} \left(v_{0} - v_{1}^{\star} - F_{2}(0)\right)^{\frac{\hat{b}-1}{b(s)-1}}.\end{aligned}$$ Furthermore, $\tilde{t}_{1}^{\star} = \tilde{t}_{1}(v_{1}^{\star}) \in (t,T)$ solves : $$\begin{aligned}
b(\tilde{t}_{1}^{\star}) = {} & 1 + \frac{\int_{t}^{T} g(s,t; v_{1}^{\star}) \tilde{Z}(t)^{\frac{1}{b(s)-1}} ds}{\int_{t}^{T} \frac{1}{b(s)-1} g(s,t; v_{1}^{\star}) \tilde{Z}(t)^{\frac{1}{b(s)-1}} ds}.\end{aligned}$$
The formula for the optimal investment strategy is straightforward from Theorem \[thm:Solution:Merging:OriginalProblem\] as ${b(\tilde{t}_{1}^{\star}) \equiv \hat{b}}$ and ${V^{\star}(t ; v_{0}) = V_{1}(t ; v_{1}^{\star}) + V_{2}(t ; v_{2}^{\star})}$ for any $t \in [0,T]$. The optimal $v_{1}^{\star}$ can be determined by Lemma \[lemma:EqualityConditionValueFunctions:Solution\] as the solution to Equation : $$\begin{aligned}
v_{1}^{\star} - \left(v_{0} - v_{1}^{\star} - F_{2}(0)\right) \int_{0}^{T} \chi(t) dt = F_{1}(0),\end{aligned}$$ where $$\begin{aligned}
\chi(t) = \left(\frac{\hat{a}}{a(t)}\right)^{\frac{1}{\hat{b}-1}} e^{- \frac{1}{\hat{b}-1} \left[\beta - \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)\right] (T-t)}.\end{aligned}$$ Therefore, $$\begin{aligned}
v_{1}^{\star} = \frac{\left(v_{0} - F_{2}(0)\right) \int_{0}^{T} \chi(t) dt + F_{1}(0)}{\int_{0}^{T} \chi(t) dt + 1}.\end{aligned}$$ is the optimal budget to the consumption problem, ${v_{2}^{\star} = v_{0} - v_{1}^{\star}}$ is the optimal budget to the terminal wealth problem. Furthermore, by Lemma \[lemma:EqualityConditionValueFunctions:Solution\] one knows $$\begin{aligned}
\left(\lambda_{1}^{\star}\right)^{\frac{1}{\hat{b}-1}} = \frac{1}{1-\hat{b}} \left(\frac{\hat{a}}{e^{\left[\beta - \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)\right] T}}\right)^{\frac{1}{\hat{b}-1}} \left(v_{0} - v_{1}^{\star} - F_{2}(0)\right).\end{aligned}$$ This enables us to calculate $g(s,t; v_{1}^{\star})$ to be $$\begin{aligned}
g(s,t; v_{1}^{\star}) = {} & (1-\hat{b}) \left(\frac{e^{\beta s - \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) (s-t)}}{a(s)}\right)^{\frac{1}{\hat{b}-1}} \left(\lambda_{1}^{\star}\right)^{\frac{1}{\hat{b}-1}} \\
= {} & \left(\frac{\hat{a}}{a(s)}\right)^{\frac{1}{\hat{b}-1}} e^{- \frac{1}{\hat{b}-1} \left[\beta (T-s) + \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) (s-t-T)\right]} \left(v_{0} - v_{1}^{\star} - F_{2}(0)\right) \\
= {} & \left(\frac{\hat{a}}{a(s)}\right)^{\frac{1}{\hat{b}-1}} e^{- \frac{1}{\hat{b}-1} \left[\beta (T-s) + \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) (s-t-T)\right]} \left(v_{0} - \frac{\left(v_{0} - F_{2}(0)\right) \int_{0}^{T} \chi(t) dt + F_{1}(0)}{\int_{0}^{T} \chi(t) dt + 1} - F_{2}(0)\right) \\
= {} & \left(\frac{\hat{a}}{a(s)}\right)^{\frac{1}{\hat{b}-1}} e^{- \frac{1}{\hat{b}-1} \left[\beta (T-s) + \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) (s-t-T)\right]} \left(\frac{v_{0} - F_{2}(0) - F_{1}(0)}{\int_{0}^{T} \chi(t) dt + 1}\right) \\
= {} & \left(\frac{\hat{a}}{a(s)}\right)^{\frac{1}{\hat{b}-1}} e^{- \frac{1}{\hat{b}-1} \left[\beta (T-s) + \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) (s-t-T)\right]} \left(\frac{v_{0} - F(0)}{\int_{0}^{T} \chi(t) dt + 1}\right)\end{aligned}$$ with $F(0) = \int_{0}^{T} e^{- r s} \left(\bar{c}(s) - y(s)\right) ds + e^{- r T} F$ defined in , and thus using Theorem \[thm:Solution:Merging:OriginalProblem\]: $$\begin{aligned}
V_{1}(t ; v_{1}^{\star}) = {} & \int_{t}^{T} g(s,t; v_{1}^{\star}) \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} ds + F_{1}(t) \\
= {} & \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} \left(\frac{v_{0} - F(0)}{\int_{0}^{T} \chi(t) dt + 1}\right) \int_{t}^{T} \left(\frac{\hat{a}}{a(s)}\right)^{\frac{1}{\hat{b}-1}} e^{- \frac{1}{\hat{b}-1} \left[\beta (T-s) + \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) (s-t-T)\right]} ds + F_{1}(t) \\
= {} & \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} \left(v_{0} - F(0)\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} \frac{\int_{t}^{T} \chi(s) ds}{\int_{0}^{T} \chi(t) dt + 1} + F_{1}(t).\end{aligned}$$ With, again from Theorem \[thm:Solution:Merging:OriginalProblem\], $$\begin{aligned}
V_{2}(t ; v_{2}^{\star}) = {} & \left(v_{2}^{\star} - F_{2}(0)\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} + F_{2}(t) \\
= {} & \left(v_{0} - v_{1}^{\star} - F_{2}(0)\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} + F_{2}(t) \\
= {} & \left(\frac{v_{0} - F_{2}(0) - F_{1}(0)}{\int_{0}^{T} \chi(t) dt + 1}\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} + F_{2}(t) \\
= {} & \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} \left(v_{0} - F(0)\right) \frac{e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t}}{\int_{0}^{T} \chi(t) dt + 1} + F_{2}(t)\end{aligned}$$ because ${F(t) = F_{1}(t) + F_{2}(t)}$ $\forall t \in [0,T]$, it follows $$\begin{aligned}
V^{\star}(t ; v_{0}) = {} & V_{1}(t ; v_{1}^{\star}) + V_{2}(t ; v_{2}^{\star}) \\
= {} & \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} \left(v_{0} - F(0)\right) \frac{1}{\int_{0}^{T} \chi(t) dt + 1} \left\{e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} \int_{t}^{T} \chi(s) ds + e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t}\right\} \\
& + F_{1}(t) + F_{2}(t) \displaybreak \\
= {} & \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} \left(v_{0} - F(0)\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} \frac{\int_{t}^{T} \chi(s) ds + 1}{\int_{0}^{T} \chi(t) dt + 1} + F(t).\end{aligned}$$ Finally, the optimal consumption rate process can then be determined from Theorem \[thm:Solution:Merging:OriginalProblem\] as $$\begin{aligned}
c^{\star}(t;v_{0}) = {} & (1-\hat{b}) \left(\lambda_{1}^{\star} \frac{e^{\beta t}}{a(t)} \tilde{Z}(t)\right)^{\frac{1}{\hat{b}-1}} + \bar{c}(t) \\
= {} & \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} \left(\frac{e^{\beta t}}{a(t)}\right)^{\frac{1}{\hat{b}-1}} \left(\frac{\hat{a}}{e^{\left[\beta - \hat{b} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)\right] T}}\right)^{\frac{1}{\hat{b}-1}} \left(\frac{v_{0} - F(0)}{\int_{0}^{T} \chi(t) dt + 1}\right) + \bar{c}(t) \\
= {} & \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} \left(v_{0} - F(0)\right) \left(\frac{\hat{a}}{a(t)}\right)^{\frac{1}{\hat{b}-1}} e^{- \frac{1}{\hat{b}-1} \beta (T-t) + \frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) T} \left(\frac{1}{\int_{0}^{T} \chi(t) dt + 1}\right) + \bar{c}(t) \\
= {} & \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} \left(v_{0} - F(0)\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)
t} \frac{\chi(t)}{\int_{0}^{T} \chi(t) dt + 1} + \bar{c}(t) \\
= {} & \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} \left(v_{0} - F(0)\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right)
t} \frac{\int_{t}^{T} \chi(s) ds + 1}{\int_{0}^{T} \chi(t) dt + 1} \frac{\chi(t)}{\int_{t}^{T} \chi(s) ds + 1} + \bar{c}(t) \\
= {} & \frac{\chi(t)}{\int_{t}^{T} \chi(s) ds + 1} \left(V^{\star}(t ; v_{0}) - F(t)\right) + \bar{c}(t).\end{aligned}$$ By defining $$\begin{aligned}
\zeta(t) = \frac{\chi(t)}{\int_{t}^{T} \chi(s) ds + 1} > 0\end{aligned}$$ we obtain $$\begin{aligned}
c^{\star}(t;v_{0}) = \zeta(t) \left(V^{\star}(t ; v_{0}) - F(t)\right) + \bar{c}(t).\end{aligned}$$ With the definition of $\zeta(t)$, the optimal wealth process finally can be written as $$\begin{aligned}
V^{\star}(t ; v_{0}) = {} & \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} \left(v_{0} - F(0)\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} \frac{\int_{t}^{T} \chi(s) ds + 1}{\int_{0}^{T} \chi(t) dt + 1} + F(t) \\
= {} & \frac{1}{\zeta(t)} \tilde{Z}(t)^{\frac{1}{\hat{b}-1}} \left(v_{0} - F(0)\right) e^{\frac{\hat{b}}{\hat{b}-1} \left(r - \frac{1}{2} \frac{1}{\hat{b}-1} \|\gamma\|^{2}\right) t} \frac{\chi(t)}{\int_{0}^{T} \chi(t) dt + 1} + F(t).\end{aligned}$$
[^1]: *Preprint*
[^2]: For two integrable functions $f(x)$ and $g(x)$ on the interval $(a,b)$, where $f(x)$ is continuous and $g(x)$ does not change sign on $(a,b)$, there exists $d \in (a,b)$ such that $$\begin{aligned}
\int_{a}^{b} f(x) g(x) dx = f(d) \int_{a}^{b} g(x) dx.\end{aligned}$$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we study the effects of noncommutative phase space deformations on the Schwarzschild black hole. This idea has been previously studied in [Friedmann-Robertson-Walker (FRW)]{} cosmology, where this “noncommutativity” provides a simple mechanism that can explain the origin of the cosmological constant. In this paper we obtain the same relationship between the cosmological constant and the deformation parameter that appears in deformed phase space cosmology, but in the context of the deformed phase space black holes. This was achieved by comparing the entropy of the deformed Schwarzschild black hole with the entropy of the Schwarzschild-de Sitter black hole.'
author:
- 'A. Crespo-Hernández'
- 'E. A. Mena-Barboza'
- 'M. Sabido'
title: On the Entropy of Deformed Phase Space Black Hole and the Cosmological Constant
---
Introduction
============
The cosmological constant problem has been addressed for several years and remains as one of the central issues in physics [@polchinski]. The discovery of the acceleration of the universe is usually attributed to a small, non-vanishing cosmological constant $\Lambda$. The different contributions to the vacuum energy density, from ordinary particle physics gives a value for $\langle\rho\rangle$ of order $M_p^4$ and should be canceled by the bare value of $\Lambda$. This cancellation has to be better than $10^{-120}$, if we compare the zero-point energy of a scalar field with the experimental value of $\langle\rho_{obs}\rangle$. This incredible degree of fine tuning suggests that we are missing important physics [@cliff]. It is likely that the correct way to interpret the tiny value of the cosmological constant by conventional quantum field theory is not the whole story. This has lead many authors to suggest that the solution will come from an unconventional approach in fundamental physics. Recently there has been mounting evidence of a late time accelerating universe, when one considers deformed phase space models [@vakili; @obregon1; @tachy1; @maleko]. One is tempted to seriously consider the idea of noncommutative space-time to study the cosmological constant [@eri].
The idea of a noncommutative space-time is not new, it has been studied from the physics perspective by Snyder [@wigner; @snyder] and formal mathematical perspective was given by Connes [@connes], during the last decades a lot of work has been done in noncommutative physics [@nekra; @connes_2; @douglas; @nc_grav; @nc_grav2; @nc_grav3; @nc_grav4; @nc_grav5; @nc_grav6; @connes2]. The noncommutative relationship between the space-time coordinates $[\hat{x}^i,\hat{x}^j]=i\Theta^{ij}$, gives an uncertainty principle between the coordinates that can be interpreted as length limit. Furthermore, noncommutativity is expected to be relevant at the Planck scale where it is known that usual semi classical considerations breakdown.
From these ideas noncommutative theories of gravity were proposed [@nc_grav; @nc_grav2; @nc_grav3; @nc_grav4; @nc_grav5; @nc_grav6]. All of these formulations showed that the end result of a noncommutative theory of gravity, is a highly nonlinear theory. To avoid the difficulties of working with these highly nonlinear theories, and to study the effects of noncommutativity on different aspects of the universe, noncommutative cosmology was proposed [@ncqc]. The authors noticed that the noncommutative deformations modify the noncommutative fields and conjectured that the effects of the full noncommutative theory of gravity should be reflected in the minisuperspace variables. This was achieved by introducing the Moyal product of functions in the Wheeler–DeWitt (WDW) equation of the Kantowski–Sachs (KS) cosmological model, in the same manner as is done in noncommutative quantum mechanics. This approach to noncommutativity has been used in connection to the Hierarchy problem and to the cosmological constant problem [@darabi; @eri]. In the last five years, more general minisuperspace deformations have been used in cosmology. In several models, the authors find an explicit relationship between the cosmological constant $\Lambda$ and the deformation parameter $\eta$ and conjecture that the deformation parameter plays the role of a cosmological constant [@vakili; @tachy1; @maleko]. They arrive to this conclusion by comparing the scale factor of the deformed (“noncommutative”) model with the scale factor of a commutative model that includes $\Lambda$ and show that the deformation parameter plays the role of the cosmological constant of the commutative model.
To further explore this relationship we consider a different gravitational system, the black hole. Black holes are the perfect systems to study fundamental problems in gravity and therefore natural candidates to further explore the relationship between the cosmological constant and noncommutivity. In our approach we intend to study the origin of the cosmological constant and show that minisuperspace deformations are a feasible mechanism that can be responsible for its origin.
The main goal of this paper is to see if a relationship between the deformation parameter $\eta$ and the cosmological constant $\Lambda$ arises in the context of deformed phase space black holes. We will introduce noncommutativity in the phase space constructed from the minisuperspace variables, by modifying the symplectic structure (Poisson’s algebra of the minisuperspace) [@darabi; @ncqc; @vakili; @tachy1; @maleko]. To calculate the entropy of the deformed black hole, we follow the procedure in [@ncbh; @bertolami], where the authors study noncommutativity in black holes. We start with the phase space deformation [@darabi; @vakili; @tachy1; @maleko], in the KS cosmological model [@ncbh; @piratas]. Interchanging the coordinates $t \leftrightarrow r$, we go from a cosmological solution to a black hole solution [@ncbh; @bertolami] and find the WDW equation for the noncommutative Schwarzschild black hole. As we are interested in obtaining the modification to the entropy of the black hole due to the phase space deformation . We use the Feynman–Hibbs path integrals formalism [@hibbs] to calculate the thermodynamical properties of the black hole [@ncbh; @tkach]. In the spirit of [@vakili; @tachy1; @maleko], we compare the entropy of the deformed model with the entropy of a commutative model (Schwarzschild-de Sitter black hole) to find a relationship between $\Lambda$ and the deformation parameters.
The paper is organized as follows, in Section \[s2\], we present the model and introduce the noncommutative deformation. In Section \[s3\], we will show that, in our approach the origin of the cosmological constant and the deformation parameters are related. Section \[s4\], is devoted for final remarks.
Deformed Black Hole Model {#s2}
=========================
In this section we present the model for the deformed phase space black hole. It is based on the relationship between the cosmological KS metric and the Schwarzschild metric [@ro]. The Schwarzschild solution can be written as $$ds^{2}=-\left(1-\frac{2m}{r}\right)dt^{2}+\left( 1-\frac{2m}{r}\right)^{-1}dr^{2}
+r^{2}\left(d\theta^{2}+\sin^{2}\theta d\varphi^{2}\right) .$$ For the case $r<2m$, the $g_{tt}$ and $g_{rr}$ components of the metric change their sign and $\partial_{t}$ becomes a space-like vector. If we make the coordinate transformation $t\leftrightarrow r$, we find$$ds^{2}=-\left(\frac{2m}{t}-1\right)^{-1} dt^{2}+\left( \frac{2m}{t}-1\right)dr^{2}
+t^{2}\left(d\theta^{2}+\sin^{2}\theta d\varphi^{2}\right),$$ when compared with the parametrization by Misner of the KS metric$$ds^{2}=-N^{2}dt^{2}+e^{\left( 2\sqrt{3}\gamma\right)} dr^{2}+
e^{\left(-2\sqrt{3}\gamma\right)}
e^{\left( -2\sqrt{3}\Omega\right)} \left(
d\theta^{2}+\sin^{2}\theta d\varphi^{2}\right),
\label{KSmetric}$$ we identify $$e^{ -2\sqrt{3}\gamma}e^{
-2\sqrt{3}\Omega}=t^{2},\quad e^{2\sqrt{3}\gamma}
=\frac{2m}{t}-1,\quad N^{2}=\left(\frac{2m}{t}-1\right)^{-1}.
\label{dif}$$ This metric with the identification of the $N, \gamma$ and $\Omega$ functions is also a classical solution to Einstein’s equations. The WDW equation for the KS metric, with some particular factor ordering, is$$\left[ -\frac{\partial^{2}}{\partial\Omega^{2}}+\frac{\partial^{2}}{\partial\gamma^{2}}+48e^{ -2\sqrt{3}\Omega } \right]
\psi(\Omega,\gamma)=0.\label{ks}$$ In [@ro1], the authors use the WDW equation as a quantum equation for the Schwarzschild black hole and used as the starting point to calculate the entropy of black holes [@tkach].
Deformed phase space black holes, satisfy a deformed Poisson algebra. To construct the deformed algebra we will follow the approach in [@tachy1]. We start with a transformation on the classical phase space variables ${\{\Omega, \gamma, P_\Omega, P_\gamma\}}$, that satisfies the usual Poisson algebra $$\begin{aligned}
\hat{\Omega}\rightarrow\Omega-\frac{\theta}{2}{P}_{\gamma},\qquad && \hat{\gamma}\rightarrow{\gamma} +\frac{\theta}{2}{P}_{\Omega}, \nonumber\\
\hat{P}_\Omega\rightarrow{P}_{\Omega} +\frac{\eta}{2}{\gamma}, \qquad && \hat{P}_\gamma\rightarrow{P}_{\gamma}-\frac{\eta}{2}{\Omega}. \label{12}\end{aligned}$$ $\hat{\Omega}, \hat{\gamma}, \hat{P}_{\Omega}, \hat{P}_{\gamma}$ are the deformed phase space variables and $\theta,\eta$ are the deformation parameters. It is easy to show that the deformed minisuperspace variables satisfy the new algebra $$[\hat{\Omega},\hat{\gamma}]=i\theta, [\hat{\Omega},\quad\hat{P}_{\Omega}]=[\hat{\gamma},\quad\hat{P}_{\gamma}]=i+\sigma,\quad [\hat{P}_{\Omega},\hat{P}_{\gamma}]=i\eta, \label{11}$$ where $\sigma=\theta\eta/4$. Now that we have introduced the deformed phase space, we use the hamiltonian constraint for the KS model, but constructed with the variables that satisfy the modified algebra. When working with phase space deformations, two physical description arise. One is based on the original variables $\Omega$, $\gamma$ and another based on the new variables $\hat{\Omega}$, $\hat{\gamma}$. The first description has the interpretation of a commutative theory where the effects of the deformation are encoded in a modified interaction, this description is referred as the “[*C-frame*]{}” interpretation. The second theory privilege the deformed variables, this is a theory with noncommutative variables but with the original interaction and is usually referred as the “[*[NC]{}-frame*]{}” formulation. In this paper we will work in the “[*C-frame*]{}”, therefore we can use a commutative space where the noncommutative effects are encoded in a modified potential [@barbosa_1; @tachy2]. This allows us to use the Feynman-Hibbs method to calculate the entropy of the deformed phase space black hole.
With the transformations (\[12\]), the WDW equation is deformed and exhibits an explicit dependence on the noncommutative parameters $$\left[-\left(-i\frac{\partial}{\partial \Omega}+\frac{\eta}{2}\gamma\right)^2+\left(-i\frac{\partial}{\partial \gamma}-\frac{\eta}{2}\Omega\right)^2
-48e^{-2\sqrt{3}\left[\Omega+\frac{i\theta}{2}\frac{\partial}{\partial \gamma}\right]}\right]\psi(\Omega,\gamma)=0.\label{14}$$ We use the anzats $\psi_\nu(\Omega,\gamma)=\Xi(\Omega)e^{\left[i\left(\nu-\frac{\eta}{2}\Omega\right)\right]\gamma}$ in Equation (\[14\]) to obtain $$-\Xi''(\Omega)-(\eta\Omega-\nu)^2\Xi(\Omega)
+48\text{exp}\left[-2\sqrt{3}(1+\sigma)\Omega+\sqrt{3}\theta \nu\right]\Xi(\Omega)=0.\label{20}$$ This second order ordinary differential equation can be solved numerically and the solutions depend on the deformation parameters $\theta$ and $\eta$. For our purposes we don’t need analytical solutions. To compute the partition function of the deformed black hole, we use the Feynman–Hibbs procedure [@hibbs]. This procedure is based in exploiting the similarities of the expression of the density matrix and the kernel of Feynman’s path integral approach to quantum mechanics. By doing a Wick rotation $t\to i\beta$ we get the Boltzmann factor and the kernel is transformed to the density matrix. [ The kernel is calculated along the paths that go from $x_1$ to $x_2$, if we consider small $\Delta t$ (small $\beta$). Then when calculating the partition function, only the paths that stay near $x_1$ have a non negligible contribution (the exponential in the expression for the density matrix gives a negligible contribution to the sum from the other paths). Therefore, the potential to a first order approximation can be written as $V(x)\approx V(x_1)$, for all the contributing paths. In this approximation, ]{}we can formally establish a map from the path integral formulation of quantum mechanics to the classical canonical partition function. To introduce quantum mechanical effects, we [ must incorporate the changes to the potential along the path]{}, in particular we are interested in the first order effects. For this we start by doing a Taylor expansion around the mean position $\tilde x$ along any path. Calculating the kernel with $\tilde x$ and doing the Wick rotation, we get the modified partition function. [ This partition function is calculated in a classical manner but with the corrected potential, and the quantum effects are encoded in the corrected potential (the potential calculated along the mean value of the path). The effective potential is a mean value of the potential $V(x)$ averaged over points near $\tilde x$ with a gaussian distribution. Therefore to incorporate the quantum effects we only need to calculate the corrected potential]{}, and simply calculate the partition function using the corrected potential [@ncbh; @tkach]. Because we are using the “[*C-frame*]{}” interpretation for the deformed phase space model, we can assume that we are working with commutative variables and the noncommutative effects are encoded in the potential $$V(\Omega)=48e^{\left[-2\sqrt{3}(1+\sigma)\Omega + \sqrt{3}\theta \nu\right]}-(\eta\Omega-\nu)^2.$$ Furthermore, to calculate the canonical partition function we can use the Feynman–Hibbs approach to the noncommutative potential [@tkach; @ncbh; @bertolami]. We start by expanding the potential to second order in $\Omega$ and making the change of variable $\Omega\rightarrow \frac{1}{\sqrt{6}}\chi+b$ where $b$ is a constant $$b=\frac{a\eta-48\sqrt{3}(1+\sigma)e^{\sqrt{3}\theta a}}{\eta^2-288(1+\sigma)^2e^{\sqrt{3}\theta a}}.$$ Now multiplying by Planck[’]{}s energy ${E_p}$ and setting $\chi=\frac{x}{l_p}$ Equation (\[20\]) takes the form $$\begin{aligned}
&&-\frac{l_p^2E_p}{2}\Xi''(x)+\left(4(1+\sigma)^2e^{\sqrt{3}\theta \nu}-\frac{\eta^2}{72}\right)\frac{E_p}{l_p^2}x^2\Xi(x)\\
&&=\frac{E_p}{12}\left[\nu^2+\frac{\left(\nu\eta -48\sqrt{3}(1+\sigma)e^{\sqrt{3}\theta \nu}\right)^2}{288(1+\sigma)^2e^{\sqrt{3}\theta \nu}-\eta^2}-48e^{\sqrt{3}\theta \nu}\right]\Xi(x). \nonumber\label{27}\end{aligned}$$ The Feynman-Hibbs procedure allows to incorporate the quantum corrections to the partition function through the corrected potential, which results in $$U(x)=\frac{3}{4\pi}\left((1+\sigma)^2e^{\sqrt{3}\theta \nu}-\frac{\eta^2}{288}\right)\frac{E_p}{l_p^2}\left[x^2+\frac{\beta l_p^2E_p}{12}\right],$$ the corrected partition function is $$Z=\sqrt{\frac{2\pi}{3}}\left((1+\sigma)^2e^{\sqrt{3}\theta \nu}-\frac{\eta^2}{288}\right)^{-\frac{1}{2}}
\frac{e^{-\frac{1}{16\pi}\left((1+\sigma)^2e^{\sqrt{3}\theta \nu}-\frac{\eta^2}{288}\right)\beta^2E_p^2}}{\beta E_p}, \label{38}$$ from which we can proceed to calculate thermodynamic properties.
The internal energy is $\langle E\rangle=-\frac{\partial}{\partial\beta}\ln Z$: $$\langle E\rangle=\frac{1}{\beta}+\frac{1}{8\pi}\left((1+\sigma)^2e^{\sqrt{3}\theta \nu}-\frac{\eta^2}{288}\right)\beta E_p^2. \label{39}$$ For the black hole $\langle E\rangle=mc^2$ and solving the quadratic equation for $\beta$ we get the temperature of the deformed phase space black hole. The temperature of the black hole can be written in terms of the deformation parameters and mass of the black hole and defining the “noncommutative” temperature $\hat{\beta}_H$ $$\hat{\beta}_H=\frac{8\pi mc^2}{\left((1+\sigma)^2e^{\sqrt{3}\theta \nu}-\frac{\eta^2}{288}\right)E_p^2},\label{temp}$$ then we can write a simple expression for the deformed phase space black hole temperature. $$\beta=\hat{\beta}_H\left(1-\frac{1}{mc^2\hat{\beta}_H}\right).$$ It is easy to verify that for the appropriate values of the deformation parameter reproduce results in [@tkach; @ncbh; @kastrup]. [ Usually the temperature is defined by the elimination of conical singularities at the horizon of the Euclidean black hole geometry. In order to determine the temperature of the horizon of the deformed black hole, we need to take the full noncomutative gravity action (which is a highly no linear [@nc_grav; @nc_grav2; @nc_grav3; @nc_grav4; @nc_grav5; @nc_grav6]), find the deformed black hole metric and calculate the temperature by resolving the conical singularity. Due to the complexity of noncommutative theories of gravity, we choose to follow the approach of noncommutative cosmology: introduce noncommutativity in the minisuperspace. Because we are using a well defined formalism to calculate the thermodynamics of the black hole that in the commutative case gives the correct temperature [@tkach; @ncbh; @kastrup]. Therefore, we can be confident that the calculated temperature for the deformed phase space black hole, is well defined in the context of the Feynman–Hibbs approach, but we can not prove that it is equivalent to the removal of a conical singularity.]{}
The entropy is calculated from $S=k\beta\langle E\rangle+k\ln Z,$ and find the deformed Hawking–Bekenstein entropy as $$\begin{aligned}
\frac{S}{k}=\frac{\hat{S}_{BH}}{k}-\frac{1}{2}\ln\left[\frac{\hat{S}_{BH}}{k}\right]+\mathcal{O}\left[\left(\frac{\hat{S}_{BH}}{k}\right)^{-1}\right], \label{endagts1}\end{aligned}$$ where $$\frac{\hat{S}_{BH}}{k}=\frac{4\pi m^2c^4}{\left((1+\sigma)^2e^{\sqrt{3}\theta \nu}-\frac{\eta^2}{288}\right)E_p^2},\nonumber$$ as with the temperature, for $\theta=0$ and $\eta=0$, we recover the commutative result. For $\eta=0$ and $\theta \ne 0$, we obtain the results in [@ncbh]. Finally, when the deformation is only present on the canonical momentum ($\theta=0$ and $\eta\ne 0$), with $E_p=c=k=1$, we get $$S_{BH}^{\eta}=4\pi m^2\left(1+\frac{\eta^2}{288}+\dots\right),\label{nc}$$ to first order in the $(\eta)^2$.
As expected the phase space deformation gives a correction to the Schwarzschild black hole entropy.
Deformed Phase Space Parameters and $\Lambda$ {#s3}
=============================================
During the last few years mounting evidence on the relationship between $\Lambda$ and the deformation parameter $\beta$ has appeared [@vakili; @tachy1; @maleko]. To find an expression for $\Lambda_{eff}$, the authors compare the de Sitter cosmology scale factor with the scale factor of the deformed phase space model in the limit $t\to \infty$. In this limit, the scale factor of the deformed model behaves as the de Sitter scale factor and an effective cosmological constant $\Lambda_{eff}$ can be defined. This effective cosmological constant is a function of the deformation parameters. In order to obtain the relationship between the deformation parameters and the cosmological constant, we will compare a commutative model that includes $\Lambda$ with a deformed phase space model that does not include $\Lambda$. To establish the connection between $\Lambda$ and the parameter $\eta$, we choose as the commutative gravitational system with a cosmological constant, the Schwarzschild-de Sitter black hole.
The metric for the Schwarzschild-de Sitter black hole model is $$ds^2=-\left(1-\frac{2m}{r}+\frac{\Lambda r^2}{3}\right)dt^2
+\frac{dr^2}{ 1-\frac{2m}{r}+\frac{\Lambda r^2}{3}}
+r^2\left(\text{d}\theta^2+\sin^2\theta\text{d}\phi^2\right),$$ which it depends on the mass and the cosmological constant. There is a singularity at $r=0$ and two horizons. The smaller one $r_b$ is called the black hole horizon and the larger one $r_c$ the cosmological horizon. As explained in [@teitelboim], we have two options, either use $r_c$ or $r_b$ as boundaries.
When calculating thermodynamic properties, this can be physically interpreted as two horizons that are not in thermal equilibrium [@t2]. For our purposes we will restrict to the black hole horizon.
The black hole horizon radius can be written as $$r_b=\frac{2}{\sqrt{\Lambda}}\cos\left(\frac{\pi+\chi}{3}\right),\label{rh}$$ where $\chi=\cos^{-1}(3\sqrt{\Lambda m^2})$.
The mass can be defined in terms of the black hole horizon [@bousso; @julio] $$m=\frac{r_b}{2}\left(1-\frac{\Lambda r^2}{3}\right),$$ $m$ is only defined when $r$ takes values between the two horizons, for vanishing $\Lambda$ we get the value for Schwarzschild black hole.
To calculate the entropy we follow as usual, thus the entropy is proportional to the area $S\sim \pi r_b^2$. Expanding the black hole radius $r_b$ to third order in $\sqrt{\Lambda}m$, we get for the entropy
$$S=
\frac{\pi r_b^2}{4}\approx\pi m^2\left(1+\frac{8\Lambda m^2}{3}\right)+\cdots.\label{21}$$
As already stated, where interesting of comparing the deformed phase space Schwarzschild black hole with the commutative Schwarzschild-de Sitter black hole. Comparing (\[21\]) with the entropy of the deformed model (\[nc\]) we find that $\Lambda\sim\frac{\eta^2}{m^2}$.
Conclusions {#s4}
===========
In this paper we have explored the possible relationship between the cosmological constant $\Lambda$ and the noncommutative deformation parameter of the canonical momentum. This relationship has been suggested when studying the late time behaviour of several cosmological models and comparing with the commutative de Sitter cosmology [@tachy1]. The results point out a quadratic dependence of the cosmological constant with the noncommutative parameter suggesting that the origin of the cosmological constant could be related to phase space deformations.
The approach we followed is similar, we compare the deformed phase space Schwarzschild black hole with the commutative [ Schwarzschild-]{}de Sitter black hole. We take advantage of the relationship between the Schwarzschild metric and the KS cosmological model, so that we can use the temperature and entropy of the deformed phase space black hole, calculated from the deformed KS-WDW equation [@bertolami]. Comparing with the entropy of the Schwarzschild-de Sitter black hole in the limit $\eta<<1$ and $\theta=0$, we find that $\Lambda\sim\frac{\eta^2}{m^2}$. This is consistent with the results obtained in the cosmological scenario [@tachy1], where a quadratic relationship between the cosmological constant and the deformation parameter has been obtained. It is encouraging that we have obtained the same result in a different gravitational system, but we can not say that the deformation parameter can always replace $\Lambda$. We need to establish the relationship between $\Lambda$ and $\eta$ under less restrictive conditions. This is under research and will be reported elsewhere.
\
[**Author contributions:**]{} All of the authors conceived of, designed and performed the calculations together. In addition, they analyzed the results and wrote the paper together. All authors have read and approved the final manuscript.
The authors contributed to this work equally.\
[**Conflict of interests:**]{} The authors declare no conflict of interest.
[999]{} J. Polchinski, The cosmological constant and the string theory landscape, arXiv hep-th:0603024. C. P. Burgess, The Cosmological Constant Problem: Why it’s hard to get Dark Energy from Micro-physics, doi:10.1093/acprof:oso/9780198728856.003.0004 arXiv:1309.4133 \[hep-th\]. O. Obregon, I. Quiros, Can noncommutative effects account for the present speed up of the cosmic expansion? Phys. Rev. [*D84*]{}, 044005 (2011). B. Vakili, P. Pedram, S. Jalalzadeh, Late time acceleration in a deformed phase space model of dilaton cosmology. Phys. Lett. [*B687*]{}, 119-123 (2010). S. Pérez-Payan, M. Sabido and C. Yee-Romero, Effects of deformed phase space on scalar field cosmology. Phys. Rev. D [*88*]{}, no. 2, 027503 (2013). B. Malekolkalami, K. Atazadeh and B. Vakili, Late time acceleration in a non-commutative model of modified cosmology. Phys. Lett. B [*739*]{}, 400 (2014). O. Obregon, M. Sabido and E. Mena, On noncommutative minisuperspace, cosmology and Lambda. Mod. Phys. Lett. A [*24*]{}, 1907 (2009). E. Wigner, Phys. Rev. 40 (1932) 749. H. S. Snyder, Phys. Rev. 71 (A. Connes, Noncommutative geometry: Year 2000. math/0011193 \[math-qa\].1947) 38. A. Connes, Noncommutative geometry. ISBN-9780121858605; A. Connes A Short survey of noncommutative geometry. J. Math. Phys. [*41*]{}, 3832 (2000). N. Seiberg and E. Witten, String theory and noncommutative geometry. JHEP [*9909*]{}, 032 (1999). A. Connes, M. R. Douglas and A. S. Schwarz, Noncommutative geometry and matrix theory: Compactification on tori. JHEP [*9802*]{}, 003 (1998). M. R. Douglas and N. A. Nekrasov, Noncommutative field theory. Rev. Mod. Phys. [*73*]{}, 977 (2001). H. Garcia-Compean, O. Obregon, C. Ramirez and M. Sabido, Noncommutative selfdual gravity. Phys. Rev. D [*68*]{}, 044015 (2003). H. Garcia-Compean, O. Obregon, C. Ramirez and M. Sabido, Noncommutative topological theories of gravity. Phys. Rev. D [*68*]{}, 045010 (2003). P. Aschieri, M. Dimitrijevic, F. Meyer and J. Wess, Noncommutative geometry and gravity. Class. Quant. Grav. [*23*]{}, 1883 (2006). X. Calmet and A. Kobakhidze, Noncommutative general relativity. Phys. Rev. D [*72*]{}, 045010 (2005). L. Alvarez-Gaume, F. Meyer and M. A. Vazquez-Mozo, Comments on noncommutative gravity. Nucl. Phys. B [*753*]{}, 92 (2006). S. Estrada-Jimenez, H. Garcia-Compean, O. Obregon and C. Ramirez, Twisted Covariant Noncommutative Self-dual Gravity. Phys. Rev. D [*78*]{}, 124008 (2008). A. H. Chamseddine and A. Connes, Noncommutative Geometry as a Framework for Unification of all Fundamental Interactions including Gravity. Part I. Fortsch. Phys. [*58*]{}, 553 (2010). H. Garcia-Compean, O. Obregon and C. Ramirez, Noncommutative quantum cosmology. Phys. Rev. Lett. [*88*]{}, 161301 (2002). A. Rezaei-Aghdam, F. Darabi, A. R. Rastkar, Noncommutativity in quantum cosmology and the cosmological constant problem. Phys. Lett. [*B615*]{}. J. C. Lopez-Dominguez, O. Obregon, M. Sabido and C. Ramirez, Towards Noncommutative Quantum Black Holes. Phys. Rev. D [*74*]{}, 084024 (2006). C. Bastos, O. Bertolami, N. C. Dias and J. N. Prata, Black Holes and Phase Space Noncommutativity. Phys. Rev. D [*80*]{}, 124038 (2009). C. Bastos, O. Bertolami, N. C. Dias and J. N. Prata, Phase-Space Noncommutative Quantum Cosmology. Phys. Rev. D [*78*]{}, 023516 (2008). R. P. Feynman and A. R. Hibbs, Quantum Mechanics and Path Integrals (McGraw-Hill, New York, 1965). O. Obregon, M. Sabido and V. I. Tkach, Entropy using path integrals for quantum black hole models. Gen. Rel. Grav. [*33*]{}, 913 (2001). R. Kantowski and R. K. Sachs, Some Spatially Homogeneous Anisotropic Relativistic Cosmological Models. J. Math. Phys. [*7*]{}, 443 (1966). O. Obregón and M. P. Ryan. Mod. Phys. Let A, [*40*]{}, 3251(1998). G. D. Barbosa, Noncommutative conformally coupled scalar field cosmology and its commutative counterpart. Phys. Rev. D [*71*]{}, 063511 (2005). S. Perez-Payan, M. Sabido, E. Mena and C. Yee-Romero, Analysis of Scalar Field Cosmology with Phase Space Deformations. Adv. High Energy Phys. [*2014*]{}, 958137 (2014). H. A. Kastrup, Canonical quantum statistics of an isolated Schwarzschild black hole with a spectrum E(n) = sigma n\*\*(1/2) E(p). Phys. Lett. B [*413*]{}, 267 (1997). C. Teitelboim, Gravitational thermodynamics of Schwarzschild-de Sitter space. hep-th/0203258. A. Gomberoff and C. Teitelboim, de Sitter black holes with either of the two horizons as a boundary. Phys. Rev. D [*67*]{}, 104024 (2003) R. Bousso, Bekenstein bounds in de Sitter and flat space. JHEP [*0104*]{}, 035 (2001). J. C. Lopez-Dominguez, O. Obregon and S. Zacarias, Towards a supersymmetric generalization of the Schwarzschild- (anti) de Sitter space-times. Phys. Rev. D [*84*]{} (2011) 024015.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
A theory of Ruelle-Pollicott (RP) resonances for stochastic systems is introduced. These resonances are defined as the eigenvalues of the generator (Kolmogorov operator) of a given stochastic system. By relying on the theory of Markov semigroups, decomposition formulas of correlation functions and power spectral densities (PSDs) in terms of RP resonances are then derived. These formulas describe, for a broad class of stochastic differential equations (SDEs), how the RP resonances characterize the decay of correlations as well as the signal’s oscillatory components manifested by peaks in the PSD.
It is then shown that a notion reduced RP resonances can be rigorously defined, as soon as the dynamics is partially observed within a reduced state space $V$. These reduced resonances are obtained from the spectral elements of reduced Markov operators acting on functions of the state space $V$, and can be estimated from series. When the sampling rate (in time) at which the observations are collected is either sufficiently small or large, it is shown that the reduced RP resonances approximate the RP resonances of the generator of the conditional expectation in $V$, i.e. the optimal reduced system in $V$ obtained by averaging out the contribution of the unobserved variables. The approach is illustrated on a stochastic slow-fast system for which it is shown that the reduced RP resonances allow for a good reconstruction of the correlation functions and PSDs, even when the time-scale separation is weak.
The companions articles, Part II[@PartII] and Part III [@PartIII], deal with further practical aspects of the theory presented in this contribution. One important byproduct consists of the diagnosis usefulness of stochastic dynamics that RP resonances offer. This is illustrated in the case of a stochastic Hopf bifurcation in Part II. There, it is shown that such a bifurcation has a clear signature in terms of the geometric organization of the RP resonances in the left half plane. This analysis provides thus an unambiguous signature of nonlinear oscillations contained in a noisy signal and that can be extracted from time series. By relying then on the theory of reduced RP resonances presented in this contribution, Part III addresses then the question of detection and characterization of such oscillations in a high-dimensional stochastic system, namely the Cane-Zebiak model of El Niño-Southern Oscillation subject to noise modeling fast atmospheric fluctuations.
author:
- 'Mickaël D. Chekroun'
- Alexis Tantet
- 'Henk A. Dijkstra'
- 'J. David Neelin'
bibliography:
- 'atantet.bib'
title: 'Ruelle-Pollicott Resonances of Stochastic Systems in Reduced State Space. Part I: Theory'
---
Introduction
============
The determination of reduced equations that aims at mimicking in a reduced state space the dynamics issued from a system of [*stochastic differential equations (SDEs)*]{} has become a central question in several disciplines ranging from molecular dynamics [@schutte1999direct; @schutte1999conformational; @schutte2001transfer] to epidemic models [@FS09; @forgoston2009accurate] and climate dynamics over the past decades; see e.g. [@MajdaEtAl_CPAM01; @CLW15_vol2; @gottwald_crommelin_franzke_2017] and references therein. Various approaches have been proposed, and in most of the cases, different metrics associated with moment estimates, probability density functions or time marginals are used to compare the reduced dynamics obtained from the surrogate system to that of the original system of stochastic differential equations (SDEs).
Here, we adopt a completely different approach which consists of using correlation functions and power spectral densities as “metrics” of comparison. In that respect, we frame hereafter a theory of Ruelle-Pollicott (RP) resonances for stochastic systems. Initially introduced for discrete and continuous chaotic deterministic systems [@ruelle1986locating; @pollicott1986meromorphic], these resonances are known to provide rigorous decomposition formulas of correlation functions; see e.g. [@Baladi2000; @Butterley_Liverani]. However, the derivation of such formulas in the deterministic context is made difficult by the singular character of the underlying invariant measures. In presence of noise, smoothing effects are known to occur for a broad class of SDEs (see Appendix \[Sec\_Feller\]), and the derivation of such formulas is thus facilitated. Section \[sec:ErgodicSpectrumSDE\] presents an extension of such decomposition formulas to the stochastic context.
The RP resonances are then defined naturally as the eigenvalues of the generator (Kolmogorov operator) of a given stochastic system. As shown in Sec. \[sec:ErgodicSpectrumSDE\], these resonances allow for decomposition formulas of correlation functions and PSDs in terms of these resonances; see Corollary \[Cor\_decomp\] and formula below. The resulting formulas describe how the RP resonances characterize the decay of correlations as well as the signal’s oscillatory components manifested by peaks in the PSD. The obtention of these formulas relies on tools and concepts from stochastic analysis, based on the theory of Markov semigroups on one hand, and the spectral theory of semigroups, on the other. These tools are briefly surveyed in Appendix \[Sec\_Appendix\], having in mind a wider audience in the geosciences and macroscopic physics.
From a practical viewpoint, our decomposition formulas benefit furthermore from a natural dynamical interpretation as they relate to the spectral elements of the Kolmogorov operator and thus to the SDE’s coefficients themselves. Such dynamical interpretations (and thus decompositions) are potentially useful for identifying physical processes responsible for power excess bumps or other broad band peaks in the PSD of noisy observations, a topic of active research in various areas of physics such as asteroseismology [@kallinger2010asteroseismology; @gruberbauer2009detection], supersonic flows [@tam1995supersonic], or climate dynamics [@ghil40advanced].
In Section \[Sec\_Theory\_reduced\_spec\], we present the main contribution of this article, by inquiring whether one can extract useful resonances from partial observations in a reduced state space, that still relate to the correlation functions and PSDs. First, we show that a notion reduced RP resonances can be indeed rigorously framed, as soon as the (stochastic) dynamics is partially observed within a reduced state space $V$. These reduced resonances are obtained from the spectral elements of reduced Markov operators acting on functions of the state space $V$, and can be estimated from series; see Sec. \[Sec\_reducedRP\]. The reduced Markov operators are ensured by Theorem \[Theorem\_A\_stoch\] that extends [@Chek_al14_RP Theorem A] within the stochastic context of this article. [*Mutatis mutandis*]{}, Theorem \[Theorem\_A\_stoch\] as [@Chek_al14_RP Theorem A], shows that — given a reduced state space $\V$, a mapping $h$ from the full state $\cX$ space onto $\V$, and a Markov semigroup $P_t$ possessing an invariant measure on $\cX$ — a family of reduced Markov operators (acting on functions of $\V$) can be naturally associated with $P_t$, $\V$ and $h$. This family characterizes a coarse-graining in the reduced state space $\V$ — such as induced by the map $h$ — of the actual transition probabilities associated with $P_t$ and taking place within the full state space $\cX$. As a result, Theorem \[Theorem\_A\_stoch\] provides, for instance, an extension to the stochastic context of Markov representations used in [@schutte1999direct] for the description of coarse-grained dynamical features of Hamiltonian systems arising in molecular dynamics.
In a second step, we show, under ergodicity (or mixing) assumptions satisfied for a broad class of SDEs, that the reduced RP resonances approximate the RP resonances of the generator of the conditional expectation in $V$, i.e. the optimal reduced system in $V$ obtained by averaging out the contribution of the unobserved variables; see Theorem \[Thm\_assymp\_gene\]. These resonances can be estimated from time series; see Sec. \[Sec\_reducedRP\]. Theorem \[Thm\_assymp\_gene\] shows then that the relationship between the reduced RP resonances and the conditional expectation is ensured when the sampling rate (in time) at which the observations are collected, is either sufficiently small or large.[^1] The fact that the reduced RP resonances relate to the conditional expectation offers useful insights for the investigation of whether non-Markovian effects [*à la*]{} Mori-Zwanzig should be included or not for the derivation of efficient reduced systems [@wouters2012; @wouters2013].
In that perspective, we analyze in Sec. \[Sec\_Langevin\] correlation functions and PSDs as reconstructed from reduced RP resonances of a stochastic slow-fast system, by using the formulas of Sec. \[sec:ErgodicSpectrumSDE\] in which the reduced RP resonances replace the genuine eponymous resonances. In the case of a strong time-scale separation, without any surprise these reduced RP resonances, as associated with the conditional expectation — well-approximated by a slow manifold reduction for the example of Sec. \[Sec\_Langevin\] —, provide excellent reproduction of correlation functions and PSDs obtained by standard sample estimates. The surprise arises when the system is placed in a regime without time-scale separation. In this case, the reduced RP resonances allow still to reconstruct to a very good accuracy level the correlation functions and PSDs providing thus an unambiguous diagnosis that the conditional expectation is sufficient to close the system and that the inclusion of non-Markovian effects is here useless. Such a diagnosis regarding the conditional expectation drawn from RP resonance analysis offers thus promising perspectives in terms of reduced order modeling, and may add insights to reduction approaches exploiting conditional expectations such as in e.g. [@weinan2004metastability; @legoll2010effective; @lu2014exact].
The usefulness of reduced RP resonance analysis is not limited to providing useful diagnosis for model reduction. The companions articles, Part II[@PartII] and Part III [@PartIII], deal with further practical aspects of the theory presented in this article. One important byproduct consists of the diagnosis usefulness of stochastic dynamics that RP resonances themselves, offer. This is illustrated in the case of a stochastic Hopf bifurcation in Part II. There, it is shown that such a bifurcation has a clear signature in terms of the geometric organization of the RP resonances in the left half plane. This analysis provides thus an unambiguous signature of nonlinear oscillations contained in a noisy signal and that can be extracted from time series. By relying then on the theory of reduced RP resonances presented in this contribution, Part III addresses then the question of detection and characterization of such oscillations in a high-dimensional stochastic system, namely a system obtained from a semispectral approximation of the Cane-Zebiak model of El Niño-Southern Oscillation, with the addition of noise modeling fast atmospheric fluctuations.
Ruelle-Pollicott resonances and decay of correlations from stochastic differential equations {#sec:ErgodicSpectrumSDE}
============================================================================================
The authors in [@Chek_al14_RP] have introduced a new mathematical framework to (i) understand and diagnose — through partial observations — the variability of chaotic flows, and (ii) to analyze parameter sensitivity that may occur in the modeling of such observations. The framework relied on the theory of Ruelle-Pollicott (RP) resonances introduced in the mid-80’s [@ruelle1986locating; @pollicott1986meromorphic] and that, at the time of the publication of [@Chek_al14_RP], was known only by a little group of experts working in the field dynamical system theory and the mathematical study of scattering resonances [@Zworski2017]. Initially introduced for discrete and continuous chaotic deterministic systems (see also [@Butterley_Liverani] for the case of Anosov flows), the RP resonances extend to stochastic systems. This section presents such an extension by relying on tools from stochastic analysis, based on the theory of Markov semigroups on one hand, and the spectral theory of semigroups, on the other; see also [@gaspard2002trace; @dyatlov2015stochastic] for complementary approaches.
As we will see, the RP resonances characterize important features of the solution’s variability (e.g. its oscillatory behavior), such as typically reflected in power spectra or correlation functions, in terms of the spectrum of the underlying Liouville operator for deterministic systems or the Fokker-Planck operator for stochastic systems, but are in general difficult to estimate especially if the dimension of the state space is large. Section \[Sec\_Theory\_reduced\_spec\] below addresses the implication of estimating resonances from partial observations in a reduced state space, and how these reduced resonances relate to the full RP resonances.
Generalities
------------
We consider Stochastic Differential Equations (SDEs) of the form: $$\label{Eq_SDE}
\d X= F(X) \d t + D(X) \d W_t, \qquad X \in \mathbb{R}^p.$$ Here $W_t=(W_t^1,\cdots,W_t^q)$ denotes an $\mathbb{R}^q$-valued Wiener process ($q$ not necessarly equal to $p$) whose components are mutually independent Brownian motions.
In Eq. , the drift part is provided by a (possibly nonlinear) vector field $F$ of $\mathbb{R}^p$, and the (also possibly nonlinear) stochastic diffusion in its Itô version, given by $D(X) \d W_t $, has its $i^{th}$-component ($1\leq i\leq p$) given by
$$\bigg[D(X) \d W_t \bigg ]_i=\sum_{j=1}^q D_{ij} (X) \d W_t^j, \; \; q\geq 1.$$
Throughout this article, we assume that the vector field $F$ and the matrix-valued function $$D:\mathbb{R}^p \rightarrow \mbox{Mat}_{\mathbb{R}}(p\times q),$$ satisfy regularity conditions that guarantee the existence and the uniqueness of mild solutions, as well as the continuity of the trajectories; see, e.g. [@cerrai2001second; @flandoli2010flow] for such conditions in the case of locally Lipschitz coefficients. Sometimes the resulting stochastic process solving Eq. emanating from $x$ at $t=0$, will be denoted by $X_t^x$, when dependence on the initial datum would be stressed, otherwise simply $X_t$.
It is well-known that the evolution of the probability density of the stochastic process, $X_t$, solving Eq. , is governed by the [*Fokker-Planck equation*]{} $$\label{Eq_FKP}
\partial_t \rho(X,t) =\mathcal{A} \rho(X,t)= -\mbox{div}(\rho(X,t)F(X)) +\frac{1}{2} \mbox{div} \nabla (\mathbf{\Sigma}(X) \rho(X,t)), \;\; X\in \mathbb{R}^p,$$ with $\mathbf{\Sigma}(X)=D(X) D(X)^T$ denoting the diffusion tensor. In practice, one is interested in stationary solutions $\mu$ to Eq. that provide a statistical description of the asymptotic behavior of the solutions to Eq. , and thus of probability density functions (PDFs), typically obtained as marginal distributions of $\mu$.
What is less-known however is that the spectral properties of the 2nd-order differential operator, $\mathcal{A}$, informs about fundamental objects such as the power spectra or correlation functions computed typically along a stochastic path of Eq. . To understand these relationships, we derive in the next subsection, decomposition formulas of correlations and power spectra in terms of the spectral elements of an operator closely related to $\mathcal{A}$, namely the generator $K$ of the associated Markov semigroup.
Ruelle-Pollicott (RP) resonances and the spectral decomposition of correlation functions {#Sec_spec_Pt}
----------------------------------------------------------------------------------------
As recalled in Appendix \[Sec\_ergodic+mixing\] (Theorem \[Thm\_cond\_mixing\]), the existence of an invariant measure $\mu$ ensures that the Markov semigroup $P_t$ (see Appendix \[Sec\_Marov\_smg\]) associated with Eq. , is a strongly continuous semigroup in $L^p_\mu(\cX)$, for any $p\geq 1$. This property alone allows us to get useful decomposition formulas of correlation functions $C_{f,g}(t)$ given by \[Cor\_func\_def\] C\_[f,g]{}(t)=fP\_t g - f g .
This is the content of Corollary \[Cor\_decomp\] below. It makes use of elements of the spectral theory of strongly continuous semigroups that we recall below and apply to Markov semigroups.
In that respect, recall that the [*essential growth bound*]{} of a semigroup $\mathcal{T}=(T(t))_{t\geq 0}$ is defined as \[Def\_ess-bound\] \_[ess]{}()= T(t)\_[ess]{}, where \[Def\_essnorm\] T(t)\_[ess]{} ={ T(t)- \_[()]{} : =L\^p\_()}. In other words, $\| T(t)\|_{ess}$ measures the distance of $T(t)$ to the set of linear and compact operators of $L^p_{\mu}(\cX)$. In the case this distance approaches zero as $t \rightarrow \infty$, the strongly continuous semigroup $\mathcal{T}$ is called [*quasi-compact.*]{} The theory of quasi-compact strongly continuous semigroups shows that the latter property is equivalent to $-\infty<\omega_{ess}(\mathcal{T})<0$; see [@Engel_Nagel Prop. V.3.5].
A semigroup is [*eventually compact*]{} if there exists $t_0>0$ such that $T(t_0)$ is compact. A semigroup that is eventually compact satisfies thus $\omega_{ess}(\mathcal{T})=-\infty$. At the same time, the definition of allows for semigroups that are not eventually compact while their essential growth bound is negative infinity[^2].
In all cases, we have the following decomposition theorem of strongly continuous semigroups [@Engel_Nagel Thm. V.3.1] that we apply to the case of Markov semigroups possessing an invariant measure.
\[Thm\_decomp\] Let $\mu$ be an invariant measure of a Markov semigroup $\mathcal{P}=(P_t)_{t\geq 0}$ and let $K$ be the corresponding generator in $L^2_\mu$. Let $\lambda_1,\cdots,\lambda_N$ be in $\sigma(K)$ satisfying $\Re \lambda_1,\cdots,\Re \lambda_N > \omega_{ess}(\mathcal{P})$. Then the following properties hold:
The $\lambda_j$’s are isolated spectral values of $K$ with finite algebraic multiplicity.
If $\Pi_1,\cdots,\Pi_N$ denote the corresponding spectral projections and $m_1,\cdots,m_N$ the corresponding orders of poles of the resolvent of $K$, then \[P-decomp1\] P\_t =\_[j=1]{}\^N T\_j(t) +R\_N(t) where \[P-decomp2\] T\_j(t)=e\^[\_j t]{}\_j, and where for every $\epsilon >0$, there exists $M >0$ such that \[P-decomp3\] R\_N(t)\_[(L\^2\_())]{} M e\^[(\^\_N+) t]{}, t0, with \[Est\_omegaN\] \^\_N ={ \_[ess]{}()} { : (K)\\{\_1,,\_N}}.
In what follows we denote by $\langle \cdot ,\cdot \rangle_{\mu}$ the inner product in $L^2_\mu$. We have then the following corollary regarding the decomposition of correlation functions. \[Cor\_decomp\] Let $\mu$ be an invariant measure of a Markov semigroup $\mathcal{P}=(P_t)_{t\geq 0}$ and let $K$ be the corresponding generator in $L^2_\mu$. Let $f$ and $g$ in $L^2_{\mu}$ such that $\langle f\rangle_{\mu}=\langle g\rangle_{\mu}=0$, then given the $\lambda_js$ and the $\Pi_js$ as defined in Theorem \[Thm\_decomp\], the correlation function $C_{f,g}(t)$ associated with the observables $f$ and $g$ has the following expansion \[Eq\_decomp\_corr1\] with |\_N(t)|M f\_[L\_\^2]{}g\_[L\_\^2]{} e\^[(\_N\^+) t]{}, t0. In the particular case of a discrete spectrum, $\sigma(K) =\{\lambda_1,\lambda_2,\cdots\}$, with $\Re \lambda_{j+1} \leq \Re \lambda_{j} $ for all $j\in\mathbb{N}^\ast$, and $\Re \lambda_j \rightarrow -\infty$, if $m_j=1$ for all $j\geq 1$, then the expansion takes the form: \[Eq\_decomp\_corr2\] where $\psi_j$ denotes the $L^2_{\mu}$-eigenfunction associated with $\lambda_j$ and $\psi_j^\ast$, the associated $L^2_{\mu}$-eigenfunction of the adjoint operator $K^\ast$ of $K$.
The decomposition is a direct consequence of -. The proof of consists of noting for all $j\geq 1$ \[Eq\_coeff\] f (x) (\_j g) (x) (x)&= f(x) \_j\^, g \_\_j(x) (x),\
&=f, \_j \_\_j\^, g \_, and that $\mathcal{Q}_N(t)\underset{N\rightarrow \infty}\longrightarrow 0$ in virtue of our assumptions about $\sigma(K)$ and of and .
The $\lambda_j$’s such that $\Re \lambda_j > \omega_{ess}(\mathcal{P})$, are called the [**RP resonances**]{}. In other words they correspond to the point spectrum of $K$. Note that there exist stochastic processes for which $R_N=0$ (i.e. no essential spectrum), for instance Ornstein-Uhlenbeck processes [@metafune2002spectrum]. Remark \[Rmk\_discrete\_spec\]-(ii) points out other (nonlinear) stochastic processes that do not have an essential spectrum. Panel (a) of Fig. \[RP\_schema\] shows a schematic of a general case, where the RP resonances are contained within a strip of the complex plane, away from the imaginary axis and the essential spectrum of $K$. The rate of decay of correlations is controlled by the spectral gap, $\tau=\mbox{gap}(K)$ defined in . Appendix \[Sec\_spec\_gap\] reports on a broad class of Markov semigroups (and thus SDEs) whose generator possesses a spectral gap; see also [@froyland1997computer; @Baladi2000; @Butterley_Liverani; @Chek_al14_RP] for results in the deterministic context.
\[Rmk\_Dynkin\]
[**Decay of correlations**]{}. Under the conditions of Corollary \[Cor\_decomp\], one obtains that $C_{f, g}(t) \underset{t\rightarrow \infty}\longrightarrow 0$, for any $f,g$ in $L^2_{\mu}$ such that $\int f \d \mu =\int g\d \mu =0$, and without assuming $P_t$ to be strong Feller and irreducible.
[**Transfer operator**]{}. By working in the Hilbert space $L^2_{\mu}$ allows us to define the transfer operator $\mathfrak{L}_t $ acting on functions in $L^2_{\mu}$, by the duality formula: \[Eq\_exchange2\] fP\_t g = \_t f g , f, gL\^2\_, t0. In a similar fashion that $P_t$ is related to the Kolmogorov operator $\mathcal{K}$ via its generator (see (iii) below), the transfer operator $\mathfrak{L}_t$ is related to the Fokker-Planck operator $\mathcal{A}$ defined in .
[**RP resonances and Kolmogorov operator**]{}. There are relationships between the abstract operator $K$ and the concrete Kolmogorov operator $\mathcal{K}$ defined in . In that respect, a theorem of Dynkin ensures (e.g. [@rogers2000diffusions p. 258]) that if $(P_t)_{t\geq0}$ is a [*Feller semigroup*]{} of a Markov process with continuous paths such that $C_0^{\infty}(\mathbb{R}^N)$ is contained in the domain $D(K)$ of the generator, then there exist functions $a_{ij}$, $b_i$, and $c$ in $C(\mathbb{R}^N)$ $(i, j = 1, . . . , N)$ such that for any $x$, the matrix $\big(a_{ij}(x)\big)_{1\leq i,j\leq N}$ is non-negative definite, $c(x) \leq 0$,[^3] and \[Eq\_K\] Kf (x) =\_[i,j=1]{}\^N a\_[ij]{}(x) \_[ij]{}\^2 f + \_[i=1]{}\^N b\_i (x)\_i f + c(x) f, fC\_0\^(\^N). Furthermore, \[Expression\_aij\] b\_i(x)=F\_i(x), a\_[ij]{}(x)=(D(x) D(x)\^T)\_[ij]{}. For such reasons we will sometimes refer to the RP resonances as the RP resonances of the Kolmogorov operator $\mathcal{K}$, or the Fokker-Planck operator $\mathcal{A}$ (by duality). This language is often adopted in Part III, more oriented towards applications than the present Part I.
However, the spectral analysis of general 2nd-order operators, typically non-self adjoint such as given in , is not an easy task in practice, especially when $N$ is large; see [@davies1996spectral; @davies2007linear]. We refer nevertheless to [@eckmann2003spectral; @herau2004isotropic] for a detailed (and enlightening) study of the “shape” of the spectrum (in $L^2_\mu$ for [@herau2004isotropic] and Sobolev spaces for [@eckmann2003spectral]) of a broad class of hypoelliptic operators such as arising in the theory of Langevin dynamics; see also [@ottobre2012exponential].
\[Rmk\_discrete\_spec\]
Note that the sum in starts actually at $j=2$ since $\psi_1=\mathbbm{1}_{\cX}$ (since $1\in \sigma(P_t)$) and $\langle f\rangle_{\mu}=0$, by assumption.
When $\sigma(K) =\{\lambda_1,\lambda_2,\cdots\}$ with $\Re \lambda_{j+1} \leq \Re \lambda_{j} $ for all $j\in\mathbb{N}^\ast$, and $\Re \lambda_j \rightarrow -\infty$, is only assumed (without requiring $m_j=1$) then holds with $N=\infty$ and $\mathcal{Q}_N=0$. Such an hypothesis on the spectrum of $K$ is systematically satisfied for the Markov semigroups that are eventually compact and possess an invariant measure $\mu$; see [@Engel_Nagel Cor. V.3.2]. A large subclass of such Markov semigroups are the compact ones. It includes a broad class of SDEs such as the following gradient dynamics on $\mathbb{R}^d$ subject to noise: \[SDE\_gradient\] x=-V(x) t + W\_t, >0, where $W_t$ is a $d$-dimensional Wiener process. The diffusion is in that case elliptic, so it is straightforward that the assumptions of Hörmander’s theorem are satisfied, and the corresponding transition probabilities are smooth, provided that $V$ is smooth. In fact for a broad class of smooth potentials $V$ that satisfy appropriate growth conditions, the compactness of $P_t$ is ensured; e.g. [@lorenzi2006analytical Thm. 8.5.3] and [@ganidis1999convergence Lem. 1.2]. See also [@metafune2002compact] for conditions ensuring the compactness of Markov semigroups that are not necessarily associated with gradient dynamics.
Note that if $P_t$ is eventually compact and possesses a unique ergodic invariant measure $\mu$, then there exists $\alpha >0$ and $M\geq 1$ and \[Eq\_expo\_conv2\] P\_t -\_\_[L\^2\_]{}M e\^[-t]{} -\_\_[L\^2\_]{}. Note that if $P_t$ has a regularizing effect sending $L_{\mu}^2$ into the Sobolev space $W_{\mu}^{1,2}$ in finite time and $W_{\mu}^{1,2}$ is compactly embedded into $L_{\mu}^2$ then the Poincaré inequality holds [@lorenzi2006analytical Prop. 8.6.1] and $P_t$ is eventually compact. The proof of is then a consequence of the theory of eventually compact semigroups [@Engel_Nagel Cor. V.3.3] and the existence of a unique ergodic measure which ensures that $\mathbbm{1}_{\cX}$ is the unique $L^2_{\mu}$-eigenfunction of $P_t$ (up to a multiplication factor) associated with the eigenvalue $1$.
[0.48]{} ![[ The Ruelle-Pollicott (RP) resonances are the isolated eigenvalues of the Markov semigroup generator, $K$; they are represented by red dots in Panel (b) and by black dots here. The rightmost vertical line represents the imaginary axis above which the power spectrum lies; see Panel (a) for another perspective. The rate of decay of correlations is controlled by the spectral gap $\tau$; see Appendix \[Sec\_spec\_gap\]. [**Panel (b):**]{} \[Courtesy of Maciej Zworski\] The imaginary part of the RP resonances corresponds to the location of a peak in the PSD (black curve lying above the imaginary axis) and the real part to its width. In blue is represented a reconstruction of the PSD based on RPs; a discrepancy is shown here to emphasize that in practice the RPs are very often only estimated/approximated; see [@Chek_al14_RP]. ]{}[]{data-label="RP_schema"}](RP_spectrum_fig.pdf "fig:"){height=".8\textwidth" width=".85\textwidth"}
[0.48]{} ![[ The Ruelle-Pollicott (RP) resonances are the isolated eigenvalues of the Markov semigroup generator, $K$; they are represented by red dots in Panel (b) and by black dots here. The rightmost vertical line represents the imaginary axis above which the power spectrum lies; see Panel (a) for another perspective. The rate of decay of correlations is controlled by the spectral gap $\tau$; see Appendix \[Sec\_spec\_gap\]. [**Panel (b):**]{} \[Courtesy of Maciej Zworski\] The imaginary part of the RP resonances corresponds to the location of a peak in the PSD (black curve lying above the imaginary axis) and the real part to its width. In blue is represented a reconstruction of the PSD based on RPs; a discrepancy is shown here to emphasize that in practice the RPs are very often only estimated/approximated; see [@Chek_al14_RP]. ]{}[]{data-label="RP_schema"}](PSD_schema.pdf "fig:"){width=".8\textwidth" height=".85\textwidth"}
Decomposition of the power spectrum {#Sec_PSD}
-----------------------------------
Given an observable $h:\mathbb{R}^p\rightarrow \mathbb{R}$ for the system $\eqref{Eq_SDE}$, we recall that the [*correlation spectrum*]{} $S_h(f)$ is obtained by taking the Fourier transform of the correlation function $C_h (t)$, namely
$$\label{Eq_corr}
S_h(f) =\widehat{C_h}(f), $$
where $C_h$ given by (for a given invariant measure $\mu$) with $f=g=h$, therein. For a broad class of SDEs that possess an ergodic probability distribution $\mu$, the spectrum in $L^2_\mu$, $\sigma(K)$, of the Markov semigroup generator, is typically contained in the left-half complex plane, $\{z\in \mathbb{C}\;:\: \Re(z)\leq 0\}$ and its resolvent $R(z)=(z \mbox{Id}-K)^{-1}$, is a well-defined linear operator that satisfies
$$\label{Eq_PSD_RPs}
S_h(f) =\int_{\mathbb{R}^p} h(X) \big[R(i f) h \big](X)\d \mu.$$
In , the frequency $f$ lies in the complex plane $\mathbb{C}$, and the poles of the resolvent $R(i f)$ which correspond to the RP resonances introduce singularities into $S_h(f)$. Once the power spectral density (PSD) is calculated, i.e. once $|S_h(f)|$ is computed with $f$ taken to be real, these poles manifest themselves as peaks that stand out over a continuous background at the frequency $f$ if the corresponding RP resonances with imaginary part $f$ (or nearby) are close enough to the imaginary axis. The continuous background may have different origins. In the case of a pure point spectrum, it is due to RP resonances located far from the imaginary axis. In the presence of a continuous spectrum and RP resonances (such as shown in Panel (a) of Fig. \[RP\_schema\]), if the latter are close to the imaginary axis, then the continuous background of the PSD is mainly due to the contribution of the continuous part of $\sigma(K)$ lying typically in a sector $\{z\in \mathbb{C}\;:\: \Re(z)\leq -\gamma\}$, for some $\gamma >0$.[^4]
Formula establishes furthermore a useful correspondence between PSD and RP resonances. Indeed, from we see that the imaginary part of the RP resonances corresponds to the location of a peak in the PSD and the absolute value of the real part to its width; see Panel (b) of Fig. \[RP\_schema\].
Let us take $f=g=h$ in . By denoting by $a_j(h)$’s the coefficients arising in , the latter decomposition formula writes $$\label{Eq_decomp_corr1_b}
C_{h}(t)= \sum_{j=1}^{N_p} \Big[\sum_{k=0}^{m_j-1} \frac{t^k}{k!} (K-\lambda_j \textrm{Id})^{k}\Big] a_j(h) e^{\lambda_j t} + \mathcal{Q}_N(t),$$ where $\mathcal{Q}_N(t)$ exhibits typically a decay property associated with properties of the essential spectrum of $\mathcal{A}$. Note that the $\lambda_j$’s do not depend on the observable $h$, but that the $a_j(h)$’s do.
If we assume that $\Re(\lambda_j) < 0$ for $j > 0$, each $\lambda_j$ is simple and the absence of an essential spectrum for $K$, then the correlation $C_{h}(t)$ in takes the simpler form of a weighted sum of complex exponentials (i.e. ), and the corresponding correlation spectrum $S_{h}(f)$ possesses itself a similar decomposition in terms of Lorentzian functions, namely: S\_[h]{}(f) = - \_[j = 1]{}\^a\_j(h) , f. \[eq:spectralPower\] It is noteworthy that such Lorentzian decompositions of the PSD are frequently encountered in applications such as for instance in spectroscopy due to the presence of resonances; e.g. [@gruberbauer2009detection; @kallinger2010asteroseismology].
In summary, the decompositions and inform us about the following features:
- Each RP resonance is associated with an exponential contribution to the decay of correlation.
- The closer an eigenvalue to imaginary axis, the slower the decay.
- In the limit of purely imaginary eigenvalues, the associated contributions to the correlation functions are purely oscillatory and prevent the decay of correlations.
- The angular frequency at which each contribution oscillates is given by the imaginary part of the associated eigenvalue.
- Eigenvalues close to the imaginary axis are associated with resonances (i.e. peaks) in the power spectrum. The spectral peak is located at the frequency given by the imaginary part of the eigenvalue and its width is proportional to the absolute value of the real part.
- The contribution of each eigenvalue to a correlation function or a power spectrum is weighted as in , corresponding to the projection of the observables $h$ onto the eigenfunctions of $K$ and its adjoint.
Reduced Ruelle-Pollicott resonances {#Sec_Theory_reduced_spec}
===================================
We present in this section the main results that serve us as a foundational basis for the applications discussed hereafter and in Part III [@PartIII]. The goal is to characterize the resonances that can be extracted from low-dimensional observations in a reduced state space and how these resonances relate to the RP resonances introduced in Sec. \[Sec\_spec\_Pt\] above.
The results of this section rely substantially on the general disintegration theorem of probability measures; see e.g. [@dellacherie1978probabilities p. 78]. The disintegration theorem states that given a probability measure $\mu$ on $\mathbb{R}^N$, a finite-dimensional Euclidean space $V$ for which $1\leq \,$dim$(\V)<N$, and a map $h: \mathbb{R}^N \rightarrow V$ (Borel-measurable), then there exists a uniquely determined family of probability measures $\{\mu_v\}_{v\in V}$ such that, for $\m$-almost all [^5] $v$ in $V$, $\mu_v$ is concentrated on the pre-image $h^{-1} (\{v\})$ of $v$, i.e. $\mu_v \left(\mathbb{R}^N\setminus h^{-1} (\{v\}) \right) = 0,$ and such that for every Borel-measurable function $\phi: \mathbb{R}^N \rightarrow \mathbb{R}$, \[Eq\_desint\] \_[\^N]{} (x) (x) = \_[V]{} (\_[xh\^[-1]{} ({v})]{} (x) \_[v]{} (x) )(v). Here $\m$ denotes the push-forward in $V$ of the measure $\mu$ in $\mathbb{R}^N$, by the map $h$.
Hereafter, we apply this result for the proof of Theorem \[Thm\_assymp\_gene\] when the reduced state space, $V$, is a subspace of $\mathbb{R}^N$, and the mapping $h$ is the projector $\Pi_V$ onto $V$. In this case, a decomposition analogous to holds for the measure $\mu$ itself, namely \[Eq\_desint2\] (BF)=\_[F]{} \_[v]{} (F) (), BF (V)(W), with $W$ being the subspace such that $V\oplus W=\mathbb{R}^N$. For later notation convenience, we denote by $\cX_v$ the pre-image $h^{-1} (\{v\})$.
Reduced Markov operators from partial observations of stochastic systems
------------------------------------------------------------------------
We start with Theorem \[Theorem\_A\_stoch\] below that extends [@Chek_al14_RP Theorem A] within the stochastic context of this article. [*Mutatis mutandis*]{}, Theorem \[Theorem\_A\_stoch\] as [@Chek_al14_RP Theorem A], shows that — given a reduced state space $\V$, a continuous map $h$ from $\mathbb{R}^N$ to $\V$, and a Markov semigroup $(P_t)_{t\geq 0}$ possessing an invariant measure on $\mathbb{R}^N$ — a family of reduced Markov operators (acting on functions of $\V$) can be naturally associated with $(P_t)_{t\geq 0}$, $\V$ and $h$. This family characterizes a coarse-graining in the reduced state space $\V$ — such as induced by the map $h$ — of the actual transition probabilities associated with $(P_t)_{t\geq 0}$ and taking place within the full state space $\mathbb{R}^N$. As a result, Theorem \[Theorem\_A\_stoch\] provides, for instance, an extension to the stochastic context of Markov representations used in [@schutte1999direct] for the description of coarse-grained dynamical features of Hamiltonian systems arising in molecular dynamics. Furthermore, [empirical estimates of]{} the transition probabilities appearing in of Theorem \[Theorem\_A\_stoch\] are important for applications, as briefly discussed in Section \[Sec\_reducedRP\] below and in more details in Part III [@PartIII].
\[Theorem\_A\_stoch\] Let $(P_t)_{t\geq 0}$ be a Markov semigroup that possesses an invariant measure $\mu$, and let $V$ be a reduced state space for which $1\leq \mbox{dim}(V) < \mbox{dim}(\mathbb{R}^N)$. Let $h:\mathbb{R}^N \rightarrow V$ be a continuous surjective function and let us denote by $\m=h_{\ast} \mu$, the push-forward of the measure $\mu$ by $h$.
Then there exists a time-dependent family of Markov operators $T_t$ acting on $L^1_{\m}(V)$ such that, for any Borel sets $B$ and $C$ of $V$, and any $t>0$, \[Eq\_main\] =(S(t, ) xh\^[-1]{}(C) | x h\^[-1]{}(B)), where $S(t,\omega)$ denotes the stochastic flow associated with $P_t$.
Furthermore, if $\mu$ is ergodic then for any $f$ in $L^2_{\m}(V)$, it holds \[weak\_ergodic\] \_0\^t \_[V]{} T\_s f (v) (v) s \_[V]{} f(v) (v).
Let us introduce for any $t>0$ the operator $T_t$ which maps $f$ in $L_{\m}^1(V)$ to $T_t f$ given by \[Def\_Tt\] T\_t f(v)=\_[x\_v]{} P\_t \[f h\](x) \_v(x),where $\cX_v$ denotes the pre-image (in $\mathbb{R}^N$) of $v$ by $h$, and $\mu_v$ denotes the disintegrated probability measure associated with $h$, such as given by the disintegration theorem recalled above. The bracket-notation, $[\cdot]$, is used here in to emphasize that the formula should be read as follows: $P_t$ is first applied on $f \circ h$, then the integration with respect to $\mu_v$, is undertaken. Note that $T_t$ acts on observables of $V$, i.e. on functions $\phi(v)$, with $v$ lying within the reduced state space.
By integrating with respect to the measure $\m$, and applying with $\phi=P_t [f\circ h]$, one notes that \_[V]{}T\_t f(v) (v)=\_[\^N]{} P\_t \[fh\] (x) (x).
This last identity allows us to get the following estimates |\_[V]{}T\_t f(v) (v)|& \_[\^N]{} | P\_t \[f h\](x) |(x)\
&\_[\^N]{} |f h (x)| (x), the last inequality resulting from applied with $p=1$ and the invariance of $\mu$. We have thus the estimate T\_t f \_[L\^1\_(V)]{} \_[V]{} |f(v)| \_[\_v]{} \_[V]{} \_v (v)= f\_[L\^1\_(V)]{}, which shows that $T_t$ maps $L^1_{\m}(V)$ into itself, i.e. $T_t$ is a Markov operator.
The rest of the proof is also straightforward, and consists of noting that for any Borel sets $B$ and $C$ in $V$, one has T\_t \_B, \_C\_[L\^1\_,L\^\_]{} =&\_[V]{} \_C(v) (v)\
&= \_[\^N]{} \_[h\^[-1]{}(C)]{}(x)P\_t\[\_B h\](x) (x)\
& = \_[\^N]{} \_t \_[h\^[-1]{}(C)]{} \_B h\
& = \_[\^N]{} \_t \_[h\^[-1]{}(C)]{} \_[h\^[-1]{}(B)]{}\
Since $\m=h_{\ast }\mu$, we deduce that $\m(B)=\mu(h^{-1}(B))$, and thus by dividing by $\m(B)$, one obtains that = (S(t, ) xh\^[-1]{}(C) | x h\^[-1]{}(B)). B, C (V).
Finally the proof of consists of applying again the characterization (iii) of the ergodicity recalled in Definition \[Def\_ergodic\].
\[Rmk\_spectral-issue\] For any $f$ and $g$ in $L^2_{\m}(V)$, one can define correlation functions (in the reduced state space $V$) associated with the family of Markov operators $T_t$ by \[Cor\_func2\] \_[f,g]{}(t)=\_V fT\_t g - \_V f \_V g . These correlations inherit a nice property resulting from the disintegration formula , namely \[Cor\_func\] \_[f,g]{}(t)=C\_[fh,gh]{}(t), where $C_{f\circ h,g\circ h}(t)$ is given by in which $f$ (resp. $g$) is replaced by $f\circ h$ (resp. $g\circ h$). In particular the decomposition formulas of Corollary \[Cor\_decomp\] apply to the observables $f\circ h$ and $g\circ h$, and provide in turn a spectral decomposition of correlation functions in the reduced state space $V$, in terms of the spectral elements of the generator $K$ of the Markov semigroup $(P_t)_{t \geq 0}$.
The issue, however, is that such a decomposition involves a spectral analysis of $K$ that is typically out of reach numerically when the dimension of state space is large, and is difficult analytically in the general case. One would like instead to have at our disposal a decomposition in terms of the spectral elements associated with the reduced Markov operators $(T_t)_{t\geq0}$ rather than with full Markov semigroup $(P_t)_{t\geq0}$. The stumbling block is that $(T_t)_{t\geq0}$ does not form a semigroup in general i.e. \[non-Markov\] T\_[t+s]{}T\_t T\_s, (see [@tantet2015early]) and one cannot rely on the spectral theory of semigroups such as used in Section \[Sec\_spec\_Pt\], to reach this goal. The next sections analyzes what type of useful information that may still be extracted from $T_t$.
Pseudo generators of the family of reduced Markov operators and asymptotic behavior {#Sec_cond_exp}
-----------------------------------------------------------------------------------
We begin with the following Lemma that characterizes the pseudo-generators $G_t$ that can be associated with the family of reduced Markov operators $(T_t)_{t\geq 0}$, although the latter family does not form a semigroup in general. For related material, we refer to the recent works [@bittracher2015pseudogenerators; @bittracher2015pseudo] concerned with the metastability analysis of Langevin dynamics. In a certain sense, Theorem \[Thm\_assymp\_gene\] below along with the notion of reduced RP resonances proposed in Sec. \[Sec\_spec\_Pt\], provide a complementary approach for the analysis of reduced Markov operators for situations not necessarily limited to Langevin dynamics. Theorem \[Thm\_assymp\_gene\] shows in particular that the asymptotic behavior of $G_t$ as $t\rightarrow \infty$ (in a weak sense) is governed by the generator of a reduced SDE in which the dependence on the unobserved variables (lying outside of the reduced state space $\V$) has been averaged out; see below. In other words, Theorem \[Thm\_assymp\_gene\] provides a useful relationship between the reduced non-Markovian process associated with $(T_t)_{t\geq 0}$ and the Markov process associated with the conditional expectation induced by the observable $h$.
\[Lem\_pseudo\] Let $K$ denote the generator of $P_t$ in $L^2_\mu(\mathbb{R}^N)$. Then, the Markov operator $T_t$ of Theorem \[Theorem\_A\_stoch\] possesses for each time $t$, a pseudo-generator, i.e. for any $f$ in $L^2_{\m}(V)$ for which $f\circ h$ lies in $D(K)$, $\underset{s\rightarrow 0}\lim (T_{t+s}f-T_tf)/s$ exists and is given by \[Eq\_Gt\] G\_t f (v)= \_[\_v]{} P\_t K\[fh\] (x) \_v(x), whose domain is thus \[dom\_X\] D(G\_t)={f L\^2\_(V) : fh D(K)}.
The proof is elementary and consists of noting that for any $f$ in $L^2_{\m}(V)$ for which $f\circ h$ lies in $D(K)$, we have =\_[\_v]{} \[fh\] (x) \_v(x) \_[\_v]{} P\_t K\[fh\] (x) \_v(x), from the basic properties of a strongly continuous semigroup such as applied to $P_t$ (e.g. [@Engel_Nagel Lem. II.1.3]) and the dominated convergence theorem.
\[Thm\_assymp\_gene\] Assume that the assumptions of Theorem \[Theorem\_A\_stoch\] hold. If $\mu$ is ergodic, then there exists a closed linear operator $\mathcal{G}$ acting on $L^2_{\m} (V)$ such that for any $f$ in $L^2_{\m}(V)$ \[Eq\_assymp\_Gt\] \_0\^t\_[V]{} G\_t f (v) (v) \_[V]{} f (v) (v), and for which $\mathcal{G}$ is dissipative in the sense that \[Eq\_dissip\_G\] f,f \_[L\^2\_(V)]{} 0, f D().
Furthermore, if $\mu$ is strongly mixing then \[Eq\_assymp\_Gt\_strong-mixing\] \_[t ]{} \_[V]{} G\_t f (v) (v)=\_[V]{} f (v) (v).
Recall that the Markov semigroup $P_t$ from which $G_t$ is defined, is associated to Eq. . If $h$ is a projector, the operator $\mathcal{G}$ is densely defined, and if $P_t$ is a Feller semigroup for which $C_0^{\infty}(\mathbb{R}^N) \subset D(K)$, then $\mathcal{G}$ possesses the following differential expression \[Eq\_G\] f (v) =\_[i,j=1]{}\^[[dim]{}()]{} \_[ij]{}(v) \_[ij]{}\^2 f + \_[i=1]{}\^[[dim]{}()]{} \_i (v)\_i f , vV, fC\_0\^[2]{}(). where \[coeff\_G\] \_[ij]{}(v)=\_[\_v]{} \_[ij]{}(x) \_v (x), \_[i]{}(v)=\_[\_v]{} F\_[i]{}(x) \_v (x), x\^N, v , where $F$ and $\mathbf{\Sigma}$ denote respectively the drift part and the diffusion tensor associated with Eq. .
In other words, $\G$ is the generator of the Markov process associated with the [**conditional expectation**]{} induced by $h$, namely with the following reduced SDE in $\V$: \[Eq\_G\_Markov\] with $W_t^{\V}$ denoting a Brownian motion in $\V$ and for $1\leq i,j\leq \mbox{dim}(\V)$, \[Eq\_diffusion\_averaged\] \_[ij]{}(v)=((v)(v)\^T)\_[ij]{}=\_[k=1]{}\^M , M1, where the $D_{\ell k}$’s are the diffusion coefficients of the SDE .
[**Step 1: Proof of and .**]{} Let $f$ be in $\mathcal{X}$ given in . Then \_[V]{} G\_t f (v) (v) &= \_[V]{}\_[\_v]{} P\_t K\[fh\] (x) \_v(x) (v),\
&=\_[\^N]{} P\_t K\[fh\] (x) (x), by application of the disintegration formula .
Assuming the measure $\mu$ to be ergodic, the characterization (iii) of Definition \[Def\_ergodic\] allows us to infer with $\mathcal{G}$ given by \[Eq\_defG\] f (v) =\_[\_v]{} K\[fh\] (x) \_v (x), by application once more of the disintegration formula.
The same formula ensures furthermore that \_[V]{} f f = \_[\^N]{} K \[fh\] and by taking the real part, follows from .
[**Step 2: $\mathcal{G}$ is closed.**]{} Let $\{f_n\}$ be a sequence in $D(\mathcal{G})$ converging to $f$ in $L_{\m}^2(V)$ such that $\G f_n \rightarrow g $ in $L_{\m}^2(V)$ as $n\rightarrow \infty$.
Then from , one has \[Eq\_step\_closed1\] \_ |f\_n -g|\^2 = -2\_[\^N]{} & K \[f\_nh\] (x) (g h (x)) +\
&\_[\^N]{} |K \[f\_nh\] (x) |\^2+ \_[\^N]{} |gh (x)|\^2 (x).\
Now since $K$ is closed in $L^2_{\mu}(\mathbb{R}^N)$, $K [f_n\circ h] \underset{n\rightarrow\infty}\longrightarrow K [f\circ h]$ and $f \circ h \in D(K),$ namely $f \in D(\G)$. This shows furthermore that one can pass to the limit in , which gives \[Eq\_step\_closed2\] 2\_[\^N]{} K \[fh\] (x) (g h (x)) =\_[\^N]{} |K \[fh\] (x) |\^2+\_[\^N]{} |gh (x)|\^2 (x), The identity says nothing else than K \[fh\] -gh \^2\_[L\^2\_]{}=0, i.e. $K [f\circ h] =g\circ h$, $\mu$-almost everywhere, which gives \_[\_v]{} K \[fh\] (x) \_v(x) =\_[\_v]{} gh (x) \_v(x)=g(v), v . We have thus proved $\G f =g$, with $f$ in $D(K)$, i.e. that $\G$ is closed.
[**Step 3: $\mathcal{G}$ is densely defined if $h$ is a projector.**]{} The purpose is to prove that $D(\G)$ is dense in $L^2_{\m}(\V)$. Assume by contradiction that there exist $g$ in $L^2_{\m}(\V)$ for which any sequence in $D(\G)$ would stay at a certain distance (in $L^2_{\m}(\V)$) from $g$ or would have $g$ as a point of accumulation, but not unique.
On the other hand, since $D(K)$ is dense in $L^2_{\mu}(\mathbb{R}^N)$ here exists a sequence $\{\varphi_n\}$ in $D(K)$ that converges to $g\circ h$ in $L^2_{\mu}(\mathbb{R}^N).$ Let us define a sequence $\{f_n\}$ (in $D(\mathcal{G})$) from the semiconjugacy relation f\_nh=h\_n, i.e. such that the following diagram is commutative
\^N @>\_n D(K)>> \^N\
@VVhV @VVhV\
@>f\_n D()>>
Now let us note that $ h \circ \varphi_n \underset{n\rightarrow \infty}\longrightarrow h \circ g \circ h = g \circ h$ if $h$ is a projector, and therefore \_[\^N]{} |(f\_nh) (x)-(gh) (x)|\^2 (x)= \_ |f\_n(v) -g(v)|\^2 (v) 0, by application of the disintegration formula, leading thus to a contradiction.
[**Step 4: $\mathcal{G}$ provides the conditional expectation.**]{} This step is a straightforward consequence of the Dynkin theorem (see Remark \[Rmk\_Dynkin\]-(iii)) and the representation formula which leads to the expression of $\mathcal{G}$, by integration with respect to $\mu_v$.
\[Rmk\_discrete\_spec\]
[**$\mathcal{G}$ describes long- and short-time behaviors in $V$.**]{} Due to , the stochastic process associated with $T_t$ in the reduced state space $V$ is in general non-Markovian. Theorem \[Thm\_assymp\_gene\] shows however that $\mathcal{G}$ plays an important role in the description of the asymptotic behavior of this stochastic process. Noteworthy is that $\mathcal{G}$ as defined in is exactly the pseudo-generator $G_t$ of $T_t$, as $t\rightarrow 0$.
The operator $\mathcal{G}$ relates thus the long time assymptotics ($t\rightarrow \infty$) of the reduced stochastic process to its infinitesimal characteristics which describe the short-time behavior ($t\rightarrow 0$), as for time homogeneous Markov processes.
[**Reduced Kolmogorov operator.**]{} Given a projector $h=\Pi_V$ onto $V$, another useful object related to the conditional expectation is the following reduced Kolmogorov operator: \[Eq\_reduced\_Kolmo\] f(v)=\_[\_v]{} \[fh\] (x) \_v(x),vV, fC\_0\^[2]{}(), where $\mathcal{K}$ is the Kolmogorov operator defined in Appendix \[Sec\_ergodic+mixing\] for the full SDE.
This operator has an interesting interpretation: although it is not the “generator” of the non-Markovian stochastic process $v_t$ associated with $T_t$, it provides the generator of the Markov stochastic process describing the average behavior of $v_t$, when the averaging is taken over the unobserved variables (i.e. those lying outside of $V$).
For the sake of clarity, Theorem \[Thm\_assymp\_gene\] has been articulated (in part) for the case where $h$ is a projection. Of course Theorem \[Thm\_assymp\_gene\] can be proved for more general mappings $h$, and for $V$ that is not necessarily a subspace of $\mathbb{R}^N$. In such cases, the corresponding reduced Kolmogorov operator, $\overline{\mathcal{K}}$, becomes typically a non-local operator. Such considerations about the choice of reduced state space has its importance in practice. Indeed non-local features of $\overline{\mathcal{K}}$ can intuitively help keep a “trace” in the reduced state space of certain interactions between the observed and unobserved variables that would be otherwise averaged out by the conditional expectation by using standard projections.
Reduced RP resonances and conditional expectation {#Sec_reducedRP}
-------------------------------------------------
As pointed out earlier, RP resonances although useful to describe power spectra and correlation functions, are difficult to estimate in practice when the dimension of the state space gets large. In practice, only partial observations of the solutions to Eq. are available, e.g. few solution’s components. Theorem \[Theorem\_A\_stoch\] shows that from partial observations of a complex system that lie within a reduced state space $V$ and are collected at a sampling rate $\tau$, a (reduced) Markov operator $T_\tau$ with state space $V$ can be inferred from these observations. Theorem \[Theorem\_A\_stoch\] shows then that this operator $T_\tau$ characterizes the coarse-graining in $V$ of the transition probabilities in the full state space. At an intuitive level if a dominant recurrent behavior occurs within an irregular background, then $T_\tau$ must still “feel” this recurrent behavior within $V$, in case this dominant behavior is reflected in $V$. As pointed out already in [@Chek_al14_RP] such a recurrent behavior is manifested by eigenvalues of $T_\tau$ distributed evenly along an inner circle typically close to the unit circle, or by forming a parabola in the complex plane depending on the representation adopted; see Part II [@PartII] and Part III [@PartIII] .
An issue though, is that in general $T_\tau$ does not come from a Markov semigroup (acting on functions of the reduced state space $V$), and thus one cannot [*per se*]{} rely on the theory of RP resonances presented in Sec. \[Sec\_spec\_Pt\] to decompose e.g. correlation functions of $V$; see Remark \[Rmk\_spectral-issue\]. Theorem \[Thm\_assymp\_gene\] on the other hand, shows that $T_\tau$ has a pseudo-generator $G_\tau$ that is close to the generator of the conditional expectation , when either $\tau$ is sufficiently small or large; see Remark \[Rmk\_discrete\_spec\]-(ii). Thus, in such cases, the resonances that can be estimated from data collected at the sampling rate $\tau$ in the reduced state space have an interesting interpretation. They approximate the RP resonances associated with the reduced system , and are called the [**reduced RP resonances**]{}. As such, these resonances describe the solution’s variability captured by the conditional expectation .
In practice, the dimension of $V$ is kept low so that $T_\tau$ can be efficiently estimated via a maximum likelihood estimator (MLE). Note that the reduced state space $V$ should be also chosen such that the observed dynamics in $V$ carry relevant information on e.g. the variability of interest contained in the original system’s solutions.
We detail below our estimation procedure of reduced RP resonances. First a bounded domain $\mathcal{D}$ of $V$ should be chosen large enough so that “most realizations” of the stochastic process $X_t$ solving Eq. fall inside $\mathcal{D}$ after application of the observable $h: \mathbb{R}^p \rightarrow V$, i.e. $\mathcal{D}$ must be chosen so that $h(X_t)$ belongs to $\mathcal{D}$ for many realizations of the noise in Eq. . This domain is then discretized as the union of $M$ disjoint boxes $B_j$, forming thus a partition.
We assume that our observations are made out of the stochastic process $X_t$ (solving Eq. ) at discrete time instants $t=t_n$, given as multiple of a sampling time $\tau$, i.e. $t_n=n\tau$ with $1\leq n\leq N$, with $N$ assumed to be large. These observations made in the observation space $V$ are denoted by $Y_n=h(X_{t_n})$. By adapting the material contained in the Supporting Information of [@Chek_al14_RP] (see also [@tantet2015early]), the Markov operator $T_\tau$ is approximated by the $M\times M$ transition matrix $\Gamma_\tau$ whose entries are given $$\label{P_estimator}
(\Gamma_{\tau})_{ij}=\frac{\#\bigg\{\Big( Y_{n}\in B_j \Big)\wedge \Big(Y_{n+1} \in B_i \Big)\bigg\}}{\#\Big\{Y_{n} \in B_j \Big\}},$$ where the $B_j$’s form a partition (composed of $M$ disjoint boxes) of the aforementioned domain $\mathcal{D}$ in $V$; see e.g. [@schutte1999direct; @crommelin2006fitting; @Chek_al14_RP; @tantet2015early]. In , the notation $\#\{(Y_{n}\in B_k)\}$ gives the number of observations $Y_{n}$ visiting the box $B_k$, and the logical symbol “$\wedge$” means “and." The leading eigenvalues of the transition matrix $\Gamma_\tau$ can then be computed with an iterative algorithm such as ARPACK [@Lehoucq1997].
The reduced RP resonances are then obtained as the eigenvalues $\lambda_k(\tau)$ obtained from the eigenvalues $\zeta_k(\tau)$ of the Markov matrix $T_\tau$, according to $$\label{Eq_lambda}
\lambda_k(\tau) = \frac{\log \big(| \zeta_k(\tau)|\big)}{\tau}
+ i \frac{\arg \big(\zeta_k(\tau)\big)}{\tau}, \qquad 1\leq k\leq M,$$ where $\arg(z)$ (resp. $\log(z)$) denotes the principal value of the argument (that we adopt to lie in $[-\pi, \pi)$ in this article) (resp. logarithm) of the complex number $z$. At a basic level, the motivation behind is that the eigenvalues of $\Gamma_\tau$ as the eigenvalues of a Markov matrix, lie within the unit circle (representation that was adopted in [@Chek_al14_RP]) whereas we want here to relate these eigenvalues with the RP resonances associated with the generator $K$ of the original Eq. . This way, the $\lambda_k(\tau)$’s given by lie naturally within the left-half complex plane.
For $\tau$ sufficiently small or large, i.e. when the generator of the conditional expectation is well approximated by pseudo-generator of $T_\tau$, one can thus proceed as follows to measure the amount of solution’s variability captured by the conditional expectation:
Given an observable and a reduced state space $V$, one forms the right-hand side (RHS) of in which the $\lambda_j$’s are replaced by the $\lambda_j(\tau)$’s given by , and the eigenfunctions by those of the Markov operator $T_\tau$.
One compares the resulting “reduced" correlation function obtained in step 1, to the correlation function as estimated from standard techniques. If the correlations are close, one can conclude that the conditional expectation provides actually a good reduced system and that the non-Markovian effects are negligible to obtain an efficient closure in $V$, for this observable. The next section illustrates such situations. Note that by replacing the RP resonances by the reduced RP resonances in , power spectral densities can also be approximated in the same fashion.
When $\tau$ is neither small or large, the precise relationships between the $\lambda_k(\tau)$’s and the actual RP resonances are non-trivial to characterize in general. Nevertheless, in certain cases, as shown in Part III, the reduced RP resonances are very useful to diagnose and characterize important dynamical features such as nonlinear oscillations embedded within a stochastic background.
Applications to a stochastic slow-fast system {#Sec_Langevin}
=============================================
The model
---------
We consider the following stochastic system
[\[Eq\_3DHopf\]]{} x =(x- f y -xz) t + W\^1\_t\
y =(f x+y -yz) t+ W\^2\_t\
z =-(z-x\^2-y\^2)t+ W\^3\_t \[Eq\_3DHopf\_z\].
The stochastic processes $W_t^1, W_t^2, W_t^3$ are independent Brownian motions. The parameters $\lambda, f, \gamma$ and $\epsilon$ are specified below. In the case $\sigma = 0$, system arises in fluid dynamics and is investigated in [@noack_al_2003] as low-dimensional reduced model for a flow past a circular cylinder. Hereafter, we deal exclusively with the stochastic case, i.e. when $\sigma>0$. The reduction problem of this system is analyzed rigorously in [@chekroun2019grisanov] using a different approach inspired from [@CLW15_vol2]. In particular it is proved that this system generates a Markov semigroup which is strong Feller and irreducible that possesses a unique ergodic invariant measure $\mu$ and thus the theory of RP resonances presented in this article applies.
System is a slow-fast system driven by additive noise. The theory of slow-fast systems in the deterministic case, i.e. when $\sigma = 0$ in system , is well established when the time-scale separation is strong; see for instance [@Jones1995; @nipp2013invariant] or the recent monograph [@Kuehn:2015tol] and references therein. Indeed, a typical behavior of such systems is characterized by a separation of time scales between the so-called “slow” and “fast” variables,as controlled by the parameter $\epsilon$. For our present system, as $\epsilon$ gets small, the $z$-variable exhibits fast fluctuations on timescales over which the $x$- and $y$-variables vary more and more slowly.
Central to the study of such systems in the deterministic context, is the existence of a *slow manifold* which expresses often a (possibly approximate) slaving relationships between the slow- and fast-variables; the latter lying typically in an “$\epsilon$-neighborhood” (for $\epsilon$ sufficiently small) to the graph of a function $h$ of the slow variables. In the case of system the slow manifold is explicitly given, for $\sigma=0$, by \[Eq\_slow\_mnf\_int\] : & ,\
&(x,y) x\^2 + y\^2. More precisely, for $\sigma=0$, any solution $(x(t),y(t),z(t))$ to is attracted exponentially fast to the manifold given by $\mathcal{M}_\epsilon=\mbox{graph} (\Phi) + \mathcal{O}(\epsilon)$, and in particular the near to slaving relationship $z(t)=h(x^2(t)+y^2(t)) + \mathcal{O}(\epsilon)$ holds for $t$ sufficiently large. Foundations of such a dynamical behavior are found in [@FENICHEL197953; @10.2307/24890380] for much more general systems than , and we refer to [@nipp2013invariant Chap. 10] for an introduction to the theory of invariant manifolds (IMs) for singularly perturbed ordinary differential equations (ODEs); see also [@Kuehn:2015tol].
In what follows we determine the reduced RP resonances as described in Sec. \[Sec\_reducedRP\] not only for the system , but also for the reduced system based on the slow manifold $\Phi$, namely by integrating \[Eq\_2DHopf\] u&=(u- f v -u ( u\^2 + v\^2) )t + W\^1\_t\
v &=(f v +v -v (u\^2 + v\^2) )t+ W\^2\_t. The RP resonances of this system, a Hopf normal form subject to an additive white noise, are analyzed in details in Part II [@PartII]. This system is a [*paragon*]{} of nonlinear oscillations in presence of noise.
The reduced state space $V$ is taken to be the $(x,y)$-plane, being thus only a reduced state space for the original system. As shown below, the estimation of (reduced) RP resonances allow also for comparing the original system with its slow manifold reduction .
Numerical results
-----------------
Both systems are numerically integrated via an Euler-Maruyama scheme with a time step of size $\delta t =10^{-5}$. In each experiment, the systems are simulated up to $T=8\times10^4$ after removal of a transient dynamics of length $T_s=10^3$. For a given sampling rate $\tau$ (specified below) and $t_n=n \tau$, the variables collected are $Y_n=(x_n,y_n)$ where $x_n=x(t_n)$ and $y_n=y(t_n)$ for system , on one hand, and $Y_n=(u_n,v_n)$ where $u_n=u(t_n)$ and $v_n=v(t_n)$ for system , on the other. The domain $\mathcal{D}$ used for estimating the Markov matrix $\Gamma_\tau$ in , is taken to be $\mathcal{D}=[-6, 6] \times [-6, 6]$, decomposed into a uniform grid constituted of $300 \times 300$ cells.
Three parameter regimes are considered hereafter: two regimes with a strong time-scale separation (Cases I and II), and one with no apparent time-scale separation (Case III); see Tables \[Table\_CaseI\] and \[Table\_CaseIII\]. The sampling rate $\tau$ is chosen depending on these regimes as follows: $\tau=10^{-3}$ for Case I, and $\tau=10^{-2}$ for Cases II and III.
Within this experimental protocol, we first estimate the reduced RP resonances. To do so, we first estimate the Markov matrix entries of $\Gamma_\tau$ according to , and then determine the reduced RP resonances according to . For the three parameter regimes considered here, these resonances are shown by blue ’+’ signs in each Panel (a) of Figures \[Viz\_1\], \[Viz\_2\], and \[Viz\_3\]. The resonances for the slow manifold system are estimated according the same procedure. They are no longer reduced but genuine RP resonances since the system is two-dimensional. Thus, the RP resonances associated with system are shown by orange ’x’ signs in each Panel (a) of Figures \[Viz\_1\], \[Viz\_2\], and \[Viz\_3\]. In each of these Panels (a), the rightmost narrow panel shows the spectral reconstructions of the PSD of $u(t)$ (resp. $x(t)$) based on RP resonances (resp. reduced RP resonances) according to procedure ($\mathbf{C_1}$) described in Sec. \[Sec\_reducedRP\], by a dashed orange (resp. blue) curve, while the black curve shows its standard PSD sample estimate. In each Panel (b) of Figures \[Viz\_1\], \[Viz\_2\], and \[Viz\_3\] are shown the autocorrelation function (ACF) of $u(t)$ (resp. $x(t)$) based on RP resonances (resp. reduced RP resonances) based on the procedure ($\mathbf{C_1}$), and here again the black curve its standard ACF sample estimate.
The conclusions of these numerical experiments are without ambiguity. First, in the case of a strong time-scale separation between the observed variables ($x$,$y$) and the unobserved variable ($z$), i.e. in Cases I and II, the reduced RP resonances allow for an almost exact reconstruction of the ACF and PSD of $x(t)$ (as for $y(t)$, not shown). This is explained from the theoretical understanding provided by Secns. \[Sec\_cond\_exp\] and \[Sec\_reducedRP\]. Indeed, as explained therein, since the sampling rate $\tau$ is small here, it is expected that the reduced RP resonances provide a good approximation of the resonances associated with the generator of the conditional expectation . On the other hand, the conditional expectation is known to provide a (very) good approximation of the dynamics of the slow variables when $\epsilon$ is small, and that it coincides with the slow manifold reduced system as $\epsilon \rightarrow 0$; see [@chekroun2019grisanov]. Thus without any surprise the RP resonances of system coincide with the reduced RP resonances of for a strong time-scale separation. The fact that these resonance are organized are organized along parabolas in the left half plane, is thoroughly analyzed in Part II [@PartII].
The conclusion for the case of no time-scale separation (Case III) is more subtle, but reveals an interesting usefulness of the RP resonance analysis. For this case, the RP resonances of the slow manifold reduced system differ from the reduced RP resonances (see Panel (a) in Fig. \[Viz\_3\]) which comes here with a noticeable mismatch at the level of ACFs although not revealed at the level of PSDs. On the contrary, the reduced RP resonances allow for an almost perfect reconstruction of the ACF. As a consequence, one can conclude that the slow manifold is no longer a valid parameterization of the $z$-variable in terms of the $x$- and $y$-variables (also observed in [@chekroun2019grisanov] for this regime) but that still, the conditional expectation provides a valid approximation for the reduced dynamics.
Such a diagnosis regarding the conditional expectation drawn from RP resonance analysis offers promising perspectives in terms of reduced order modeling. As illustrated here, the latter resonances could indeed serve to determine whether efforts on non-Markovian effects must be invested or not. In that respect, we mention the recent variational approach relying on optimal manifolds minimizing a parameterization defect [@CLM19_closure] which offers new perspectives to approximate analytically the conditional expectation; see also [@CLM16_Lorenz9D]. Applied to Case III, this approach provides an analytical substitute to the slow manifold which allows in turn for the derivation of an efficient 2D reduced system of SDEs (without non-Markovian terms), confirming the conclusions drawn from RP resonance analysis.
[llllll]{} & $\lambda$ & $f$ & $\gamma$ & $\epsilon$ & $\sigma$\
Case I & $10^{-3}$ & $10^2$ & $5.6 \times 10^{-2}$ & $10^{-2}$ & $0.55$\
Case II & $10^{-3}$ & $10$ & $1$ & $10^{-2}$ & $0.2$\
[llllll]{} & $\lambda$ & $f$ & $\gamma$ & $\epsilon$ & $\sigma$\
Case IV & $10^{-3}$ & $10$ & $1$ & $10$ & $0.3$\
[0.48]{} ![[ RP resonances of (orange ’x’ sign) and reduced RP resonances of (blue ’+’ sign) in the $(x,y)$-plane. [**Rightmost frame in Panel (a):**]{} Power spectral densities (PSDs). [**Panel (b):**]{} Autocorrelation function (ACF) of $x(t)$ (resp. $u(t)$). The spectral reconstructions of the PSD and ACF of $u(t)$ (resp. $x(t)$) based on RP resonances (resp. reduced RP resonances) according to procedure ($\mathbf{C_1}$) described in Sec. \[Sec\_reducedRP\], are shown by a dashed orange (resp. blue) curve, while the black curves show their standard sample estimates.]{}[]{data-label="Viz_1"}](caseI_PSD-eps-converted-to.pdf "fig:"){width="100.00000%" height="70.00000%"}
[0.48]{} ![[ RP resonances of (orange ’x’ sign) and reduced RP resonances of (blue ’+’ sign) in the $(x,y)$-plane. [**Rightmost frame in Panel (a):**]{} Power spectral densities (PSDs). [**Panel (b):**]{} Autocorrelation function (ACF) of $x(t)$ (resp. $u(t)$). The spectral reconstructions of the PSD and ACF of $u(t)$ (resp. $x(t)$) based on RP resonances (resp. reduced RP resonances) according to procedure ($\mathbf{C_1}$) described in Sec. \[Sec\_reducedRP\], are shown by a dashed orange (resp. blue) curve, while the black curves show their standard sample estimates.]{}[]{data-label="Viz_1"}](caseI_Cor-eps-converted-to.pdf "fig:"){width="100.00000%" height="70.00000%"}
[0.48]{} ![Same as in Fig. \[Viz\_1\] but for Case II.[]{data-label="Viz_2"}](caseII_PSD-eps-converted-to.pdf "fig:"){width="100.00000%" height="70.00000%"}
[0.48]{} ![Same as in Fig. \[Viz\_1\] but for Case II.[]{data-label="Viz_2"}](caseII_Cor-eps-converted-to.pdf "fig:"){width="100.00000%" height="70.00000%"}
[0.48]{} ![Same as in Fig. \[Viz\_1\] but for Case III.[]{data-label="Viz_3"}](caseIV_PSD-eps-converted-to.pdf "fig:"){width="100.00000%" height="70.00000%"}
[0.48]{} ![Same as in Fig. \[Viz\_1\] but for Case III.[]{data-label="Viz_3"}](caseIV_Cor-eps-converted-to.pdf "fig:"){width="100.00000%" height="70.00000%"}
This work has been partially supported by the Office of Naval Research (ONR) Multidisciplinary University Research Initiative (MURI) grant N00014-12-1-0911 and N00014-16-1-2073 (MDC), by the National Science Foundation grants DMS-1616981(MDC) and AGS-1540518 (JDN), by the LINC project (No. 289447) funded by EC’s Marie-Curie ITN (FP7-PEOPLE-2011-ITN) program (AT and HD) and by the Utrecht University Center for Water, Climate and Ecosystems (AT).
Elements of stochastic analysis {#Sec_Appendix}
===============================
In this appendix we present a short survey of elements of stochastic analysis used in the Main Text. The main objective is to introduce the key concepts and tools of stochastic analysis for stochastic differential equations (SDEs), to a wider audience in the geosciences and macroscopic physics.
Markov semigroups {#Sec_Marov_smg}
-----------------
Two approaches dominate the analysis of stochastic dynamics. We are here concerned with the approach rooted in Stochastic Analysis which, contrary to the random dynamical system (RDS) approach [@Arnold98; @CSG11; @CLW15_vol1], does not substitute a deterministic (nonlinear) flow $S(t)$ by a stochastic flow $S(t,\omega)$ acting[^6] on the state space $\cX$ but rather by a family of linear operators $P_t$, acting on a space of observables of the state space, i.e. on functions of $\cX$. A typical choice of observables is given by $\mathcal{C}_b(\cX)$, the space of bounded and continuous functions on $\cX$.
More precisely, this family $P_t$ reflects the (averaged) action of the stochastic flow at the level of functions and is given as the mapping which to each function $\phi$ in $\mathcal{C}_b(\cX)$ associates the function: \[Pt\_def\] P\_t (x) =( (S(t,)x))=\_ (S(t,) x) (), t0, xX. In , the function $\phi$ is the aforementioned [*observable*]{}. Its physical meaning could be, for instance, the potential vorticity or the temperature of a fluid at a given location or averaged over a volume. The RHS of involves averaging over the realizations $\omega$, i.e. expectation. For deterministic flow it reduces to $P_t\phi(x)=\phi(S(t)x)$ and is known as the Koopman operator. Note that $P_t$ such as defined in is not limited to stochastic flow, more generally $P_t \phi (x) =\mathbb{E}( \phi(X_t^x))$ where $X_t^x$ denotes a stochastic process that solves Eq. (as associated with $P_t$) and emanates from $x$ in $\cX$.
Under general assumptions on $F$ and $D$, the stochastic process $X_t$ solving Eq. is Markovian (i.e. the future is determined only by the present value of the process) which translates at the level of $P_t$ into the following semigroup property \[Eq\_smg\_pty\] P\_0=, P\_t P\_s=P\_[t+s]{}, t,s0. A breakdown of indicates thus that the underlying stochastic process is non-Markovian.
It is noteworthy to mention that even when $P_t$ satisfies , it does not ensure that $P_t$ is a [*strongly continuous*]{} semigroup [@Pazy1983] on $\mathcal{C}_b(\cX)$. Nevertheless, $(P_t)_{t\geq 0}$ is extendable to a strongly continuous semigroup in $L^2_{\mu}$ as soon as $\mu$ is an invariant measure of the Markov semigroup; see Theorem \[Thm\_cond\_mixing\] below. The spectral theory of such semigroups [@Engel_Nagel] [is at the core of the]{} description of mixing properties in $L^2_{\mu}$, such as presented in Sec. \[Sec\_spec\_Pt\] in the Main Text.
Ergodic invariant measures and the strong Feller-Irreducibility approach {#Sec_Feller}
------------------------------------------------------------------------
The Fokker-Planck equation may support several weak stationary solutions. An important question, is thus the identification of stationary measures that describe the asymptotic statistical behavior of the solutions of Eq. , in a typical fashion. The notion of ergodic invariant measures plays a central role in that respect, and relies on the following important characterization of ergodic measures for (stochastically continuous) Markov semigroups [@DaPrato1996 Theorem 3.2.4]. \[Def\_ergodic\] An invariant measure is ergodic if one of the following three equivalent statements holds:
For any $f \in L^2_\mu(\cX)$, if $P_t f =f$, almost surely w.r.t $\mu$ ($\mu$-a.s.) for all $t\geq 0$, then $f$ is constant $\mu$-a.s.
For any Borel set $\Gamma$ of $\cX$, if $P_t \mathbbm{1}_{\Gamma} =\mathbbm{1}_{\Gamma}$ $\mu$-a.s. for all $t\geq 0$, then $\mu(\Gamma)=0$ or $1$.
For any $f \in L^2_\mu (\cX)$, $\frac{1}{T} \int_0^T P_s f \d s \underset{T\rightarrow \infty}\longrightarrow \int f \d \mu$ in $L^2_\mu(\cX).$
In practice, an efficient approach to show the existence of an ergodic measure consists of showing the existence of a [*unique*]{} invariant measure, since in this case such an invariant measure is necessarily ergodic [@DaPrato1996 Theorem 3.2.6]. Various powerful approaches exist to deal with the existence of a unique invariant measure. The next section discusses the classical approach based on the theory of [*strong Feller Markov semigroups*]{} and [*irreducibility*]{}.
The main interest of the strong Feller-Irreducibility approach lies in its usefulness for checking the conditions of the Doob-Khasminskii Theorem [@doob1948asymptotic; @khas1960ergodic; @DaPrato1996], the latter ensuring the existence of [*at most*]{} one ergodic invariant measure. This strategy requires the proof of certain smoothing properties of the associated Markov semigroup, and to show that any point can be (in probability) reached at any time instant by the process regardless of initial data. This property is known as [*irreducibility*]{}. It means that $P_t \mathbbm{1}_{U} (x)> 0$ for all $x$ in $\cX$, every $t>0$, and all non-empty open sets $U$ of $\cX$, which is equivalent to say that (S(t,)x-z < ) >0, for any $z$ in $\cX$, $\epsilon >0$ and $t>0$; see [@cerrai2001second p. 67]. In other words the irreducibility condition expresses the idea that any neighborhood of any point $z$ in $\cX$, is reachable at each time, with a positive probability.
Remarkably, the irreducibility is usually inferred from the controllability of the associated control system $\dot{x}=F(x) + D(X) u(t)$; see [@chekroun2019grisanov] for a simple illustration. This approach is well-known and based on the support theorem of Stroock and Varadhan [@Stroock71] (see also [@ikeda2014stochastic Theorem 8.1]) that shows that several properties of the SDEs can be studied and expressed in terms of the control theory of ordinary differential equations (ODEs); see [@DaPrato1996 Secns. 7.3 and 7.4] for the case of additive (non-degenerate) noise and [@arnold1987unique; @kliemann1987recurrence] for the more general case of nonlinear degenerate noise, i.e. in the case where the noise acts only on part of the system’s equations, corresponding to ker$(Q)\neq \{0\}$.
The strong Feller property means that the semigroup maps bounded measurable functions into bounded continuous functions. This property, related to a regularizing effect of the Markov semigroup $(P_t)_{t\geq 0}$, is a consequence of the [*hypoellipticity*]{} of the [*Kolmogrorov operator*]{} $\mathcal{K}$ defined on smooth functions $\psi$ (of class $C^2$): \[Def\_K0-additive\] (x)= ( \^2 (x)) +F(x), , where ( \^2 ())=\_[i,j=1]{}\^N \_[ij]{}\_[ij]{}\^2 . Note that hypoelliptic operators include those that are uniformly elliptic for which the Weyl’s smoothing lemma applies; e.g. [@dacorogna2004introduction Theorem 4.7]. Hypypoellipticity allows nevertheless for dealing with the case of degenerate noise, which is important in applications. A very efficient criteria for hypoellipticity is given by Hörmander’s theorem [@hormander1967hypoelliptic; @norris1986simplified]; see also [@CSG11 Appendix C1] for a discussion on the related Hörmander’s bracket condition and its implications to the existence of other types of meaningful measures for SDEs, namely the Sinaï-Ruelle-Bowen (SRB) random measures. We refer also to Part II [@PartII], for an instructive verification of the Hörmander’s condition in the case of the Hopf normal form subject to additive noise.
From a geophysical perspective, it is noteworthy to mention that the strong Feller-Irreducibility approach allows for dealing with a broad class of truncations of fluid dynamics models that would be perturbed by noise, possibly degenerate. For instance, in the case of truncations of 2D or 3D Navier-Stokes equations, the strong Feller-Irreducibility approach has been shown to be applicable even for an additive noise that forces only very few modes [@agrachev2005navier; @romito2004ergodicity]. The delicate point of the analysis is the verification of the controllability (and thus irreducibility) of the associated control system, by techniques typically adapted from [@jurdjevic1985polynomial] or rooted in chronological calculus as in [@agrachev2005navier]. Whatever the approach, the analysis requires the appropriate translation into geometrical terms of the cascade of energy in which the nonlinear terms transmit the forcing from the few modes to all the others [@romito2005geometric].
Markov semigroups and mixing {#Sec_ergodic+mixing}
----------------------------
We recall here standard results about Markov semigroups. It states that any Markov semigroup that is strong Feller and irreducible and for which an invariant measure exists (which is thus unique) is not only ergodic but also strongly mixing for the total variation norm of measures. Given two probability measures $\mu_1$ and $\mu_2$ on $\cX$, we recall that the latter is defined as [@hairer2009introduction Eq. (3.1)] \[mix\_1b\] [TV]{}(\_1,\_2)=\_ | g \_1-\_ g \_2|, where $\mathcal{B}_b(\cX)$ denotes the set of Borel measurable and bounded functions on $\cX.$
\[Thm\_cond\_mixing\] Let $\mu$ be an invariant measure of a Markov semigroup $(P_t)_{t\geq 0}$. For any $p\geq 1$ and $t\geq 0$, $P_t$ is extendable to a linear bounded operator on $L^p_\mu(\cX)$ still denoted by $P_t$. Moreover
$\|P_t\|_{\mathcal{L}(L^p_\mu(\cX))} \leq 1$
$P_t$ is strongly continuous semigroup in $L^p_\mu(\cX)$.
If furthermore $(P_t)_{t\geq 0}$ is strong Feller and irreducible, then $\mu$ is ergodic and for any $x$ in $\cX$ and $g$ in $L^1_{\mu}$ \[strong\_ergo\] \_0\^T g(X\_\^x) = \_ g(x), , where $X_t^x$ denotes the stochastic process solving the SDE associated with $P_t$.
[In this case, the invariant measure]{} $\mu$ is also [**strongly mixing**]{} in the sense that for any measure $\nu$ on $\cX$, we have: \[mix\_1\] [TV]{}(\_t ,)0.
For the definition of a strongly continuous semigroup also known as $C_0$-semigroup we refer to [@Engel_Nagel p. 36]; see also [@CK17]. For an introduction to semigroup theory we refer to [@van2012asymptotic; @engel2006short].
We prove first $(i)$. [The proof is standard and can be found e.g. in [@guionnet2003lectures Prop. 1.14] but is reproduced here for the reader’s convenience]{}. Let $g$ be in $\mathcal{C}_b(\cX)$. By the Hölder inequality, we have \[Ineq\_base\] |P\_t g (x) |\^p P\_t (|g|\^p) (x).
If we now integrate both sides of this inequality with respect to $\mu$, we obtain \[Eq\_Holder\] \_ |P\_t g (x) |\^p (x) \_ P\_t (|g |\^p) (x) (x)= \_ | g|\^p (x) (x), the latter equality resulting from the invariance of $\mu$. Since $\mathcal{C}_b(\cX)$ is dense in $L^p_\mu(\cX)$, the inequality can be extended to any function in $L^p_\mu(\cX)$, and thus $(P_t)_{t\geq 0}$ can be uniquely extended to a contraction semigroup in $L^p_\mu(\cX)$, and property $(i)$ is proved.
Let us show now that $(P_t)_{t\geq 0}$ is strongly continuous in $L^p_\mu(\cX)$. Since $(P_t)_{t\geq 0}$ is a Markov semigroup, for any $g$ in $\mathcal{C}_b(\cX)$ and $x$ in $\cX$, we have that the mapping $t \mapsto P_t g (x)$ is continuous. Therefore by the dominated convergence theorem \_[t0]{} P\_t g =g L\^p\_(). The density of $\mathcal{C}_b(\cX)$ in $L^p_\mu(\cX)$ allows us to conclude that this convergence holds when $g$ is in $L^p_\mu(\cX)$.
The ergodicity of $\mu$ results from the aforementioned Doob’s theorem. The time-average property and the mixing property can be obtained as a consequence of e.g. [@seidler1997ergodic Cor. 2.3]; see also [@stettner1994 Cor. 1].
Generator of a Markov semigroup {#Sec_gen_Markov}
-------------------------------
Recall that the generator $A$ of any strongly continuous semigroup $(T(t))_{t\geq 0}$ on a Hilbert space $\mathcal{H}$ is defined as the operator $A:D(A)\subset \mathcal{H} \rightarrow \mathcal{H}$, such that A = \_[t0\^+]{} (T(t) -), defined for every $\varphi$ in the domain D(A)={ | \_[t0\^+]{} (T(t) -) }.
As any generator of a contraction semigroup, given an invariant measure $\mu$, the generator $K$ of the contraction semigroup $(P_t)_{t\geq 0}$ in $L^2_{\mu}$ (Theorem \[Thm\_cond\_mixing\]-(i)) is dissipative, which is equivalent to say, since $L^2_{\mu}$ is a Hilbert space, that \[Eq\_dissip\_K\] K f,f \_[L\^2\_]{} 0, f D(K), where $D(K)$ denotes the domain of $K$; see e.g. [@Engel_Nagel Prop. II.3.23]. The domain $D(K)$ is furthermore dense in $L^2_{\mu}$ and $K$ is a closed operator; see [@Pazy1983 Cor. 2.5 p. 5]. The isolated part of the spectrum of $K$ provides the Ruelle-Pollicott resonances; see Sec. \[Sec\_spec\_Pt\].
Return to equilibrium and spectral gap {#Sec_spec_gap}
--------------------------------------
We present here some useful results concerning (i) the exponential return to equilibrium for strong Feller and irreducible Markov semigroups, and (ii) spectral gap in the spectrum of the Markov semigroup generator $K$; see Theorems \[Thm\_cve1\] and \[Thm\_cve2\] below. Theorem \[Thm\_cve1\] deals with semigroups that become [*quasi-compact*]{} after a finite time, and Theorem \[Thm\_cve2\] addresses the exponential $L^2$-convergence and lower bound of the spectral gap. For Theorem \[Thm\_cve1\], the approach is based on Lyapunov functions such as formulated in [@bellet2006ergodic]. We propose a slightly different presentation for which we provide the main elements of the proof. We refer to [@douc2009subgeometric] for an efficient (and beautiful) generalization of such Lyapunov-type criteria allowing for [*sub-exponential convergence*]{} towards the equilibrium.
Recall that the [*essential spectral radius*]{} $\mathbf{r}_{ess}(T)$ of a linear bounded operator $T$ on a Banach space $\mathcal{X}$ satisfies [@Engel_Nagel p. 249] the Hadamard formula \[Spec\_rad\] \_[ess]{}(T)= T\^n\_[ess]{}\^[1/n]{}, where T\_[ess]{}={ T- \_[()]{} : }. We have then the following convergence result.
\[Thm\_cve1\] Let $\mathcal{P}=(P_t)_{t\geq 0}$ be a strong Feller and irreducible Markov semigroup generated by an SDE given by Eq. for which $F$ and $G$ are locally Lipschitz. Assume that there exists a Lyapunov function[^7] $U$ and a compact set $\mathfrak{A}$ for which there exist $a >0$, $0<\kappa<1$ and $b<\infty$, such that
\[Eq\_Lyap\_cond\] $$\begin{aligned}
\mathcal{K} U & \leq a U,\label{Eq_Lyap_cond-a}\\
P_{t_0} U &\leq \kappa U + b \mathbbm{1}_{\mathfrak{A}}, \; \mbox{ for some } t_0>0,\label{Eq_Lyap_cond-b}\end{aligned}$$
where $\mathcal{K}$ is the Kolmogorov differential operator generating the Markov process associated with $\mathcal{P}$. Then for all $t>t_0$, $P_t$ becomes quasi-compact, i.e. \[Eq\_quasi\] \_[ess]{}(P\_t), where the essential spectral radius is taken for $P_t$ as acting on $\mathcal{X}=\mathcal{F}_{U}$ given by \[Eq\_FU\] \_[U]{}={f:\^N | f\_[U]{} <}, and endowed with the norm f\_U=.
Furthermore $(P_t)_{t\geq 0}$ has a unique invariant measure $\mu$, and the inequality ensures that there exist $C>0$ and $\lambda >0$ such that for all $f$ in $\mathcal{F}_{U}$, \[CVE\_pre\_totvar\] | P\_t f(x) -f | C e\^[-t]{} U(x), t>t\_0, x\^N.
[The proof of this result is found in Appendix \[Sect\_proof\]. ]{} The assumption verified from moment estimates in practice. For instance if there exist $k_0>0$ and $k_1>0$ such that \[Eq\_toto0\] |X\_t\^x| k\_0 e\^[-k\_1 t]{}|x| +c, t0, then for any $t\geq -\frac{1}{k_1} \log(\frac{1}{4 k_0})$, we have $ \mathbb{E} (|X_t^x| +1)\leq \frac{1}{2} (|x|+1)-\frac{1}{4} |x|+c+\frac{1}{2}$, which leads to (|X\_t\^x| +1) (|x|+1) +(c+ )\_[B\_r]{}, for all $r>4(c+\frac{1}{2}),$ and thus holds with $U(x)=|x|+1.$
More generally, if \[Meta\_Lyap\_cond\] U -U +, >0, 00, and similarly holds. In addition, implies . [Note that and are quite standard; see e.g. [@dragoni2012ergodicity Lemma 2.11].]{}
[Finally, note also that finding a Lyapunov function may be easier than proving inequalities of the form . For instance, if there is a Lyapunov function which grows polynomially like $\|p\|^q$, then one knows that the process has moments of order $q$; see [@meyn1993stability2; @meyn1993stability3]. ]{}
Finally, [lower bounds of the spectral gap in $L^2_{\mu}$ may be derived for a broad class of SDEs]{}. Recall that the generator $K$ has a spectral gap in $L^2_{\mu}$ if there exists $\delta>0$ such that \[Cond\_gap\] (K){: ()>-}={0}. The largest $\delta>0$ with this property is denoted by $\mathrm{gap}(K)$, namely \[Formula\_gap\] (K)={ >0 }.
The following result is a consequence in finite dimension of more general convergence results [@goldys2005 Theorems 2.5 and 2.6]. Since $(P_t)_{t\geq 0}$ is a C$_0$-semigroup in $L^2_{\mu}$, the theory of asymptotic behavior of a semigroup with a strictly dominant, algebraically simple eigenvalue (e.g. [@van2012asymptotic Thm. 3.6.2]) implies the spectral gap property stated in the following.
\[Thm\_cve2\] Assume that $(P_t)_{t\geq 0}$ is strong Feller and irreducible. Assume furthermore that the following ultimate bound holds for the associated stochastic process $X_t^x$, i.e. there exist $c,k,\alpha>0$ such that |X\_t\^x|\^2< k |x|\^2e\^[- t]{} +c, t0, x\^N. Then there exists a unique invariant measure $\mu$ for which the $U$-uniform ergodicity holds with $U(x)=1+|x|^2$, as well as the following exponential $L^2$-convergence \[CVE\_inL2\] P\_t - \_[L\^2\_]{} C e\^[-t]{} \_[L\^2\_]{}, t0, L\^2\_, with $C$ and $\lambda$ positive constants independent of $\varphi$; the latter rate of convergence being the same as that of . Furthermore, one has the following lower bound for the $L^2_{\mu}$-spectrum of the generator $K$: 0<(K).
We will see in Part II [@PartII] of this three-part article that Theorem \[Thm\_cve2\] has important practical consequences. In particular it shows for a broad class of controllable ODEs, perturbed by a white noise process for which the Kolmogorov operator is hypoelliptic, that an $L_{\mu}^2$-spectral gap is naturally induced by the noise whereas in absence of the latter the gap may be zero, leading thus to a form of mixing enhancement by the noise. We finally mention [@hairer2014spectral] for other conditions, ensuring an $L^2_\mu$-gap based on spectral gaps in Wasserstein distances, verifiable in practice by following the approach of [@hairer2011asymptotic]. We show in the next section that Wasserstein distances are also suitable to measure the decay of correlations.
Proof of Theorem \[Thm\_cve1\] {#Sect_proof}
------------------------------
It is standard from the theory of Lyapunov functions that the existence of a unique invariant measure $\mu$ is ensured by the condition together with the irreducibility and strong Feller properties. The rest of the proof is thus concerned with and the exponential convergence .
[**Step 1.**]{} First, note that the Itô formula gives U = U t + , which leads (since $\mathcal{K} U \leq a U$) to (U(x(t;x))) =P\_t U(x)e\^[a t]{}U(x), and therefore $P_t$ [is extendable]{} to a linear operator on $\mathcal{F}_U$ (defined in ) with norm $\|P_t\|\leq e^{a t}$.
The second inequality in ensures that for any $t>t_0$, \[Eq\_est0\] P\_t U (x) U(x) +b, x\^N. By definition, a Markov semigroup is monotone, thus one may iterate to obtain (by using $P_t \mathbbm{1}_{\mathbb{R}^N}=\mathbbm{1}_{\mathbb{R}^N}$), \[Eq\_est1\] P\_[nt]{} U(x) \^n U(x) +, n1. Consider now an arbitrary compact set $\mathfrak{B}$ in $\mathbb{R}^N$ and $f$ in $\mathcal{F}_{U}$, we have the bound | P\_[nt]{}f(x) -\_(x) P\_[nt]{} \_ f(x) | & U(x) ,\
& U(x) f\_[U]{} , where we have used the basic inequality (with $p=1$). This last inequality with leads to | P\_[nt]{}f(x) -\_(x) P\_[nt]{} \_ f(x) | U(x) f\_[U]{} (\^n + ). Since $\lim_{|x|\rightarrow \infty} U(x)=\infty$, given $\epsilon>0$ and $n>1$ one may thus choose a compact set $\mathfrak{B}_n$ such that \[Eq\_compact\_bounds\] P\_[nt]{}f -\_[\_n]{} P\_[nt]{} \_[\_n]{}f \_U f\_[U]{} (+)\^n, which leads to P\_[nt]{}-\_[\_n]{} P\_[nt]{}\_[\_n]{} \_[(\_U)]{} (+)\^n.
[**Step 2.**]{} Let us show that the linear operator =\_ P\_[t]{} \_: \_U \_U, is compact for any compact set $\mathfrak{B}$ of $\mathbb{R}^N$. This is equivalent to showing that for any sequence $g_k$ in $\mathcal{F}_U$ such that $\|g_k\|_U\leq 1$, one can extract a subsequence such that $\Lambda g_k$ is convergent in $\mathcal{F}_U$. Since $P_t$ is strongly Feller and $\mathbbm{1}_{\mathfrak{B}} g_k$ is bounded for each $k$, then $P_{t} \mathbbm{1}_{\mathfrak{B}} g_k$ belongs to $\mathcal{C}_b(\mathfrak{B})$, by definition. Thus the sequence $(\Lambda g_k)$ lives in $\mathcal{C}(\mathfrak{B}).$
We have |g\_k (x)|g\_k\_U P\_t U(x) U(x) + b U(y) +b, x, which shows that $\{\Lambda g_k\}$ is equibounded.
Furthermore, since $P_t$ is strong Feller, it has a smooth kernel[^8] and we have for all $x, x' \in \mathfrak{B}$ |g\_k (x)-g\_k (x’)|&\_[y]{}|\_t(x,y)- \_t(x’,y)| |f(y)|y,\
&|x-x’| |\_u \_t(u,v)| g\_k\_U \_ U(y) y, which shows that $\{\Lambda g_k\}$ is equicontinuous.
[Thus, the Ascoli-Arzelà theorem [@yosida1995functional p. 85] applies and]{} guarantees that [a subsequence from]{} $\Lambda g_k$ converges in $\mathcal{C}(\mathfrak{B})$ to $g$. Now since $U\geq 1$, [the same extraction from]{} $\Lambda g_k$ converges to $g\mathbbm{1}_{\mathfrak{B}}$ in $\mathcal{F}_U$. We conclude that $\mathbbm{1}_{\mathfrak{B}} \,P_{t} \mathbbm{1}_{\mathfrak{B}}$ is a compact mapping for any compact set $\mathfrak{B}$ of $\mathbb{R}^N$.
[**Step 3.**]{} Let $\mathfrak{B}_n$ be a sequence of compact sets satisfying , and let us consider the compact operators (from Step 2) $\mathfrak{C}_n$ defined by $\mathbbm{1}_{\mathfrak{B}_n} \,P_{n t} \mathbbm{1}_{\mathfrak{B}_n}$. We have then \[Def\_essnorm2\] P\_[nt]{}\_[ess]{}={ P\_[nt]{}- \_[(\_U)]{} : \_[U]{}}& P\_[n t]{}- \_n\
&(+)\^n. By applying to $P_t$ the Hadamard formula recalled in , we have thus for $t>t_0$ \_[ess]{}(P\_t)=\_[n]{} P\_[nt]{}\_[ess]{}\^[1/n]{}+, for all $\epsilon>0,$ and we deduce .
The exponential convergence is then ensured by showing that there is no other eigenvalue than 1 on the unit disk (or outside the unit disk) and that 1 is a simple eigenvalue; see [@bellet2006ergodic].
[^1]: In practice however it is often observed that the reduced RP resonances still provide useful information for “intermediate" sampling rates; see Part III[@PartIII].
[^2]: For instance any semigroup $\mathcal{T}$ such that $\| T(t)\|_{ess}\leq M \exp{(-\epsilon t^{\alpha})}$, with $\epsilon >0$ and $0<\alpha, M <1$.
[^3]: Furthermore if the process is non-explosive then $c\equiv 0.$ This excludes the cases for which the underlying Markov process leaving at time 0 from $x$ in $\mathbb{R}^N$ escapes to infinity at some finite time $t > 0$. This article is not concerned with explosive stochastic processes.
[^4]: While we recall that in such a case, the RP resonances are the isolated eigenvalues of finite multiplicity, lying within a strip $-\gamma < \Re(z) \leq 0$; see Panel (a) of Fig. \[RP\_schema\].
[^5]: i.e. up to an exceptional set of null measure with respect to $\m$.
[^6]: $\omega$ labelling the noise realization.
[^7]: Recall that a $C^2$ function $U$ is called a Lyapunov function $U(x)\geq1$ and $\lim_{|x|\rightarrow \infty} U(x)=\infty$, ensuring thus that the level sets $\{U\leq \alpha\}$ are compact.
[^8]: A probability kernel $ \mathfrak{T}_t$ allows for representing the Markov semigroup $P_t$ as $P_t f(x)=\int \mathfrak{T}_t(x,\d y) f(y)$; e.g. [@bakry2013analysis Prop. 1.2.3]. Having a smooth kernel means that $\mathfrak{T}_t(x,\d y)=\mathfrak{p}_t(x,y)\d y$ with $\mathfrak{p}_t$ infinitely differentiable, i.e. smooth.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
The pseudo-Lindley distribution which was introduced in Zeghdoudi and Nedjar (2016) is studied with regards to its upper tail. In that regard, and when the underlying distribution function follows the Pseudo-Lindley law, we investigate the behavior of its values, the asymptotic normality of the Hill estimator and the double-indexed generalized Hill statistic process (Ngom and Lo), the asymptotic normality of the records values and the moment problem.\
$^{\dag}$ Gane Samb LO.\
LERSTAD, Gaston Berger University, Saint-Louis, Sénégal (main affiliation).LSTA, Pierre and Marie Curie University, Paris VI, France.AUST - African University of Sciences and Technology, Abuja, Nigeria\
[email protected], [email protected], [email protected]\
Permanent address : 1178 Evanston Dr NW T3P 0J9,Calgary, Alberta, Canada.\
$^{\dag}$ Dr. Modou NGOM\
Work Affiliation : Ministery of High School (SENEGAL)\
LERSTAD, Gaston Berger University, Saint-Louis, Séné[email protected]\
$^{\dag}$ Dr. Moumouni Diallo\
Université des Sciences Sociale et de Gestion de Bamako ( USSGB)\
Faculté des Sciences Économiques et de Gestion (FSEG)\
Email: [email protected].\
**Keywords**. Lindley’s distribution; Pseudo-Lindley distribution; Extreme value theory; record values; Hill’s estimator; asymptotic laws **AMS 2010 Mathematics Subject Classification:** 6oG70; 62G20;62H10;62H15\
title: 'Extremes, extremal index estimation, records, moment problem for the Pseudo-Lindley distribution and applications'
---
Introduction {#sec1}
============
**1. General facts**.\
The following probability distribution function (*pdf*), named as the Pseudo-Lindley *pdf*,
$$f(x)=f(x,\theta,\beta)=\frac{\theta(\beta-1+\theta x) e^{-\theta x}}{\beta} 1_{(x\geq 0)} \label{glind}$$
with parameters $\theta>0$ and $\beta>1$, has been introduced by [@Zeghoudi3] as a generalization of the Lindley *pdf* :
$$\ell(x)=\frac{\theta^2(1+x) e^{-\theta x}}{1+\theta} 1_{(x\geq 0)}. \label{olind}$$
in the sense that for $\beta=1+\theta$, $f(\circ)$ is identical to $\ell(\circ)$.\
Actually, $f$ derives from $\ell$ by a mixture of a Lindley distributed random variable and an independent $\Gamma(2,\theta)$ random variables with mixture coefficients $r_1=(\beta-1)/\beta$ and $r_2=1/\beta$, where $1<r_1, \ r_2<1$ and $r_1+r_2=1$.\
The cumulative distribution *cdf* function is given by
$$1-F(x)=\left(\beta^{-1}(\beta+ \theta x) e^{-\theta x}\right) 1_{(x\geq 0)}.$$
The Lindley original distribution is an important law that been used and still being used in Reliability, in Survival analysis and other important disciplines. Because of its original remarkable qualities, it kicked off a considerable number generalizations as pointed out by [@Zeghoudi3]. The current generalization has been tested on real data and simulated and shows real interest in survival analysis, on the Guinean Ebola data for example ([@Zeghoudi3]). The paper of focused on asymptotic tests of that law based on moments estimators. The interest that distribution demonstrated in real data modeling motivated us to give some asymptotic theories on it, in view of statistical tests. In this paper, we deal with the properties of the upper tail, the extreme value distribution and the record values. etc., which of them providing statistical tests.\
Throughout the paper, $X$, $X_1$, $X_2$, $\cdots$ is a sequence independent real-valued random variables (*rv*), defined on the same probability space $(\Omega, \mathcal{A},\mathbb{P})$, with common cumulative distribution function $F$, with the first asymptotic moment function and the generalized inverse function defined by
$$R(x,F)=\frac{1}{1-F(x)}\int_{x}^{+\infty}(1-F(y)) \ dy, \ x \in ]0,+\infty[$$
and
$$F^{-1}(u)=\inf \{x \in \mathbb{R}, \ F(x)\geq u\} \ for \ u \in ]0,1[ \ and \ F^{-1}(0)=F^{-1}(0+).$$
For each $n\geq 1$, we denote the ordered statistics of the sample $X_1$, $\cdots$, $X_n$ by
$$X_{1,n}\leq \cdots \leq X_{n,n}.$$
Usually, in extreme value theory, we focus on upper extreme and the hypothesis $X>0$ and the log-transform $Y=\log X$ is instrumental in all major results in that field. We denote $G(x)=F(e^x)$, $x \in \mathbb{R}_{+}$. The Renyi representation is also of common use in the following form. The sequence is replaced as follows
$$\{ \{X_{1,n}\leq \cdots \leq X_{n,n}\}, \ n\geq 1\}=_{d} \{ \{F^{-1}(1-U_{n-j+1,n}), 1\leq j\leq n\}, \ n\geq 1\}, \label{renyi}$$
where $=_{d}$ stands for the equality in distribution. Finally, the following Malmquist representation (see [@shwell], also [@ips-wcia-ang], page ...) is also used : for each $n\geq 1$, there exist a finite sequence of standard independent exponential random variables $E_{1,n}$, $\cdots$, $E_{n,n}$ such that
$$\label{malmquist}
\left\{ \left(\frac{U_{i+1,n}}{U_{i,n}}\right)^{i}, \ 1\leq i \leq n \right\}=_{d} \left\{E_{i,n}, 1\leq i \leq n \right\}.$$
Extremes
========
We can directly see that $F$ is the Gumbel distribution $G_{0}$ by three different arguments. First, by using the Von Mises’ argument (see [@dehaan] or [@ips-wciia-ang], Proposition 24, page 184)
$$\begin{aligned}
\lim_{x \rightarrow 0} \frac{f^{\prime}(x)(1-F(x))}{f^2(x)}=-1. \end{aligned}$$
A second argument comes from that $Y=\exp(X)$ has the distribution $G(x)=F(\log x)=\beta^{-1} (\beta + \theta \log x) x^{-\theta x}$. Since
$$\begin{aligned}
\lim_{x \rightarrow 0} \frac{1-G(\lambda x)}{1-G(x)}=\lambda^{-\theta}, \end{aligned}$$
$G \in D(G_{1/\theta})$ and since $F(x)=G(e^x)$ for $x\geq 1$, by Theorem (in Lo), $F \in D(G_0)$.\
A third argument is relation to the development of the quantile function. In the appendix (page ), we give a number of expansions of that quantile that could be used for different purposes. For example we have (see page ),
$$\label{quantile}
F^{-1}(1-u) = \theta^{-1}(\log(1/u) - \log\log(1/u)) + \theta^{-1} K(u)$$
with $K(u)=O(\log 1/u)^{-2})$. By using it we get
$$\frac{F^{-1}(1-\lambda u)-F^{-1}(1-u)}{(1/\theta)} \rightarrow -\log \lambda \ as \ u\rightarrow 0.$$
By the $\pi$-variation criteria of [@dehaan] (See [@ips-wcia-ang], Proposition 11, page 88), we have $F\in D(G_0)$ and $R(x,F)\rightarrow \gamma=1/\theta$ as $x\rightarrow +\infty$. Formula is actually a second-order condition for the quantile function (see [@dehaan]). We apply it right to get a rate of convergence of the maximum observations. Put $\gamma=1/\theta$.
**2. Expansion of the maximum values**.\
By the Renyi representation and by denoting $Z_n=-\log(nU_{1,n})$, we have that $\log(1+Z_n/(\log n)) \rightarrow_{\mathbb{P}} 0$ and since $\log U_{1,n}=O_{\mathbb{P}}(\log n)^{-1}$
$$\begin{aligned}
X_{n,n}-F^{-1}(1-1/n)&=&\gamma Z_n + \gamma \log(1+Z_n/(\log n))\\
&+& O((\log n)^{-2})+O((\log U_{1,n})^{-2}).\end{aligned}$$
and hence
$$\begin{aligned}
\frac{X_{n,n}-F^{-1}(1-1/n)}{\gamma}= Z_n + O_{\mathbb{P}}(\log n)^{-1})=\Gamma + O_{\mathbb{P}}(1) \end{aligned}$$
It is easy to see that $Z_n$ converge to Gumbel law $\Lambda$ with *cdf*
$$G_{0}(x)=\exp(-\exp(-x)), \ x \in \mathbb{R}$$
So we have that $X_{n,n}$ converges to a $\Lambda$ law. But we obtain the random rate of convergence $Z_n/\log n$ since
$$\frac{\log Z_n}{\log n} \left(\frac{X_{n,n}-F^{-1}(1-1/n)}{\gamma}-Z_n \right)=1$$
As well for $k=k(n)\rightarrow +\infty$ such that $k(n)/n\rightarrow 0$, and by taking $T_n=\log (nU_{k,n}/k)$ and $q_n=n/k(n)$ which goes to $+\infty$, we have
$$\begin{aligned}
\frac{X_{n-k,n}-F^{-1}(1-k/n)}{\gamma}= T_n + \log(1+T_n/\log q_n)) + O_{\mathbb{P}}((\log q_n)^{-2}).\end{aligned}$$
**3. Estimating the extreme value index $\gamma=1/\theta$**.\
The [@hill]’s estimator
$$H_n=\frac{1}{k(n)} \sum_{j=1}^{k(n)} j\left(X_{n-j+1,n}-X_{n-j,n}\right),$$
is the most celebrated estimator the extreme value index $\gamma=1/\theta$ of $Z=\exp(X)$. Among a significant number of generalizations, the [@ngomlo2016]’s generalization, called the Double-indexed function Hill estimator, is one the sharpest one. It is defined as
$$H_n(f,s)= \left(\sum_{j=1}^{k(n)} f(j)\left(X_{n-j+1,n}-X_{n-j,n}\right)^s/(a_n(f,s)\right)^{1/s},$$
where $f:\mathbb{N}\setminus \{0\} \rightarrow \mathbb{R}_+\setminus \{0\}$ is a measurable mapping and $s>0$. Let us define
$$a_n(f,s)=\Gamma(s+1) \sum_{j=1}^{k(n)} f(j)j^{-s}, \ C^2(s)=\Gamma(2s+1)-\Gamma(s+1)^2, \ s_n^2(f,s)= C^2(s) \sum_{j=1}^{k(n)} f(j)^2j^{-2s},$$
and
$$B_n(f,s)=\max\{ f(j)j^{-s}/s_n(f,s), \ 1\leq j \leq k(n)\}.$$
We simply notice that the classical Hill’s estimator is $H_n(I_d,1)$ where $I_d$ is the identity function on $\mathbb{N}\setminus \{0\}$. Let us give asymptotic normality for Double-indexed function Hill estimator.\
**(a) Extreme Limit Theorem**.\
We begin with the simple Hill’s estimator.\
\[theoHill\] For $]0, \ n] \ni k(n)\rightarrow +\infty$ such that
$$k(n)^{3/4}/\log n \rightarrow 0. \tag{K1}$$
we have, as $n\rightarrow +\infty$,
$$\sqrt{k(n)} \left(H_n-\gamma\right) \rightsquigarrow \mathcal{N}(0,\gamma^2). \label{hillSN}$$
We want to establish the random rate of convergence associated with the convergence \[hillSN\] in the part (a) of the following corollary. In the part (b), we want to share that we need any other condition on top of $k(n)/n \rightarrow 0$ to have the central limit theorem if $F^{-1}$ is reduced to
$$F_{\ast}^{-1}(1-u)=\gamma \log u - C(\gamma) \log \log(1/u), \ u \in ]0,1[, \ C(\gamma)\geq 1. \label{Freduced}$$
\[theoHillCoro\] We have the following results.\
(a) Here again $F$ is the *cdf* of the Pseudo-Lindley distribution with parameters $\theta>0$ and $\beta>0$ and the notation above. Let $k(n)/\log n\rightarrow 0$.
Let $W(1)$ is a standard Gaussian random variable. Then we have
$$\frac{\log n}{\gamma \sqrt{k(n)}} \biggr(\sqrt{k(n)}(H_n-\gamma)-\gamma W(1)\biggr) \rightarrow_{\mathbb{P}} 1,$$
\(b) If $F^{-1}$ were reduced as in Formula , we have the asymptotic normality
$$\sqrt{k(n)}(H_n-\gamma) \rightarrow \mathcal{N}(0,\gamma^2)$$
whenever $k(n)/n \rightarrow 0$ and
$$\log n \biggr(\sqrt{k(n)}(H_n-\gamma) - \gamma W(1)\biggr)=O_{\mathbb{P}}(1). \ \Diamond$$
**Proof of Theorem \[theoHill\]**. By Malmquist representations (See [@shwell] or [@ips-wcia-ang], Proposition 32, page 135), by Formula , we have for any $1\leq j \leq k$,
$$\begin{aligned}
X_{n-j+1,n}-X_{n-j,n}&=&F^{-1}(1-U_{j,n})-F^{-1}(1-U_{j+1,n}) \notag\\
&=& \gamma j^{-1} E_{j,n} - \gamma \int_{U_{j,n}}^{U_{j+1,n}} \frac{du}{u \log (1/u)} + O_{\mathbb{P}}\left(\left(log n\right)^{-2}\right) \label{spacingF}\end{aligned}$$
and next
$$\begin{aligned}
j \left(X_{n-j+1,n}-X_{n-j,n}\right)&=& \gamma E_{j,n} - \gamma j\int_{U_{j,n}}^{U_{j+1,n}} \frac{du}{u \log (1/u)} + O_{\mathbb{P}}\left(k \left(log n\right)^{-2}\right).\end{aligned}$$
So for $Z_n=\log nU_{1,n}$ (which converges in law to $\Lambda$) and
$$\label{step_01}
\left| \int_{U_{j,n}}^{U_{j+1,n}} \frac{du}{u \log (1/u)} \right|\leq \frac{j^{-1} E_{j,n}}{|\log n - Z_n|}.$$
Hence
$$\frac{1}{k(n)}\left| \sum_{j=1}^{k(n)} j \int_{U_{j,n}}^{U_{j+1,n}} \frac{du}{u \log (1/u)} \right|\leq \frac{S^{\ast}_{k(n)}}{k} O_{\mathbb{P}}((\log n)^{-1}).$$
where $S^{\ast}_{k(n)}=E_{j,n}+\cdots+E_{k,n}$. We finally get
$$\sqrt{k(n)} \left(H_n-\gamma\right)=\gamma \frac{S^{\ast}_{k(n)}-k}{\sqrt{k(n)}} + O_{\mathbb{P}}\left(\frac{1}{\log n}, \frac{k^{3/2}}{(\log n)^2}\right)$$
We conclude that, whenever (K1) holds, we have
$$\sqrt{k(n)} \left(H_n-\gamma\right)=\gamma \frac{S^{\ast}_{k(n)}-k}{\sqrt{n}} + o_{\mathbb{P}}(1). \ \square$$
**Proof of the Corollary \[theoHillCoro\]**. The proof of Part (b) is the conclusion of the proof of Theorem \[theoHill\] up to the formula . If \[Freduced\] holds, further steps are dismissed. And we need only $k(n)/n\rightarrow 0$ to conclude. Let us set
$$Z^{\ast}_n=\frac{1}{\sqrt{k(n)}} \sum_{j=1}^{k(n)} j\int_{U_{j,n}}^{U_{j+1,n}} \frac{du}{u \log (1/u)}, \ n\geq 1.$$
For the first part, we already knew that $Z_n^{\ast}=O_{\mathbb{P}}(1/\log n)$. We denoted by $W(1)$ a standard Gaussian random variable. By the classical Kómlos-Màjor-Tusnàdy (KMT) approximation, we have
$$\left|\frac{S^{\ast}_k(n)-k(n)}{\sqrt{k(n)}}-\gamma W(1)\right|=O_{\mathbb{P}}\left(\frac{\log k(n)}{\sqrt{k(n)}}\right).$$
Straightforward expansions using the different rates of convergence lead to
$$\frac{\sqrt{k(n)}(H_n-\gamma)-\gamma W(1)}{\gamma Z_n^{\ast}} \rightarrow_{\mathbb{P}} 1,$$
whenever $k(n)/n \rightarrow 0$. Now we apply Proposition in [@ips-wciaa-ang], page 22. Since the function $\log (1/u)$ is slowly varying and that $U_{1,n}/U_{k+1,n}$ and $U_{k+1,n}/U_{1,n}$ are both asymptotically bounded in probability, we have
$$t_n=\sup_{1\leq j \leq k(n)} \sup_{s \in [U_{j,n},U_{j+1,n}]} \left|\frac{\log (1/s)}{\log n}-1\right| \rightarrow_{\mathbb{P}} 0.$$
It comes that
$$Z_n^{\ast}=\frac{\sqrt{k(n)}}{\log n} (k_{(n)}^{-1} S^{\ast}_{k(n)})(1 + O(t_n))=\frac{\sqrt{k(n)}}{\log n} (1+o(1)),$$
which gives the desired result. $\blacksquare$\
We have the following convergence of the Double-indexed functional Hill statistics.
\[theoDH\] We have the following two results.\
(a) If the following conditions hold, as $n\rightarrow +\infty$
$$s_n(f,1)/(s_n(f,s) \log n)\rightarrow 0 \ and \ B_n(f,s) \rightarrow 0,$$
then
$$\frac{T_n(f,s)-\gamma^s a_n(f,s)}{s_n(f,s)} \rightsquigarrow \mathcal{N}\left(0,\gamma^{2s}\right).$$
\(b) Furthermore, if $a_n(f,s)/a_n(f,s)\rightarrow +\infty$, then
$$\frac{s_n(f,s)}{a_n(f,s)} \biggr(\left(\frac{T_n(f,s)}{a_n(f,s)}\right)^{1/s}-\gamma\bigg) \rightsquigarrow \mathcal{N}(0,s^{-2}\gamma^2). \blacksquare.$$
**Proof**. Let us exploit the proof of Theorem \[theoHill\]. we have for $j \in \{1,\cdots,k(n)\}$, $s\geq 1$, $$\begin{aligned}
A_{i,n}&=&f(j) \left(X_{n-j+1,n}-X_{n-j,n}\right)^s\\
&=& f(j) \biggr(\gamma j^{-1}E_{j,n} - \gamma \int_{U_{j,n}}^{U_{j+1,n}} \frac{du}{u \log (1/u)} + O_{\mathbb{P}}\left(F_{k(n)} \left(log n\right)^{-2}\right)\biggr)^{s}\\
&=:& f(j) \left(\gamma j^{-1}E_{j,n} - R_{j,n}+C_{j,n}\right)^{s}, \end{aligned}$$
with
$$C_{j,n}= O_{\mathbb{P}}\left(\left(log n\right)^{-2}\right) (uniformly \ in \ j), \left|\gamma \int_{U_{j,n}}^{U_{j+1,n}} \frac{du}{u \log (1/u)} \right| \leq
\frac{\gamma j^{-1} E_{j,n}b(n)}{|\log n - Z_n|}.$$
We get, by the mean value theorem, $j \in \{1,\cdots,k(n)\}$, $s\geq 1$,
$$\begin{aligned}
&&A_{i,n}-\gamma^s f(j)j^{-s} E_{j,n}^{s}\\
&&\leq sf(j) \left|R_{j,n}+C_{j,n}\right| \left(\gamma j^{-1} E_{j,n} +\left|R_{j,n}\right|+\left|C_{j,n}\right|\right)^{s-1}\\
&& \leq \biggr(\frac{s \gamma f(j) j^{-1}E_{j,n}}{|\log n -Z_n|}\biggr)\biggr(\gamma j^{-1} E_{j,n} +\left|R_{j,n}\right|+\left|C_{j,n}\right|\biggr)^{s-1}.\end{aligned}$$
In the lines below, we will bound the term with the power $s-1$. But if $s=1$, there will is nothing to bound. So formulas regarding that term are dismissed for $s=1$ and are used only for $s>2$. For $s\geq 1$, we will use the $C_{s-1}$ inequality ( for $s\leq 2$, with $|a+b|^{s-1} \leq 2^{s-2} |a|^{s-1} + |b|^{s-1}$ $C_{s-1}=2^{s-2}$). For $0<r<1$, it can be easily checked that, for $u>0$ fixed, the function $g(v)=(u+v)^r-u^r-v^r$ of $v\geq 0$ takes the value $g(v)=0$ and has a non-positive derivative function, so that $g(v)\leq g(0)=0$ for any $v\geq 0$, which is equivalent to $(u+v)^r\leq u^r+v^r$. We finally have that $|a+b|^{s-1} \leq D_{s} |a|^{s-1} + |b|^{s-1}$ with $D_s=1$ for $1<s<2$ and $D_s=C_{s-1}$ for $s\geq 2$. Applying that inequality leads, $j \in \{1,\cdots,k(n)\}$, $s\geq 1$, to
$$\begin{aligned}
&&A_{i,n}-\gamma^s f(j)j^{-s} E_{j,n}^{s}\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ (A)\\
&&\leq \biggr(\frac{s \gamma f(j) j^{-1}E_{j,n}}{|\log n -Z_n|}\biggr)\biggr(D_s\gamma^{s-1} j^{s-1}E_{j,n}^{s-1}+\frac{D_{s}^2\gamma^{s-1}j^{s-1}E_{j,n}^{s-1}}{(|\log n - X_n|^{s-1})} + O_{\mathbb{P}}\left(\frac{D_s^2}{(\log n)^{2(s-1)}}\right)\biggr).\end{aligned}$$
Let us denote
$$S_n(f,s)=\sum_{j=1}^{k(n)} f(j) j^{-s} E_{j,n}^s$$
and
$$T_n(f,s)=\sum_{j=1}^{k(n)} f(j) \left(X_{n-j+1,n}-X_{n-j,n}\right)^s,$$
By combining the results above, we arrive at \[B\]
$$\begin{aligned}
&&\biggr|T_n(f,s)-\gamma^s S_n(f,s)\biggr| \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ (B)\\
&&\leq \biggr(\frac{s \gamma S_n(f,1)}{|\log n -Z_n|}\biggr)\biggr(D_s\gamma^{s-1} S_n(Id,s-1)+\frac{D_{s}^2\gamma^{s-1}S_n(Id,s-1)}{(|\log n - Z_n|^{s-1})} +
O_{\mathbb{P}}\left(\frac{D_s^2}{(\log n)^{2(s-1)}}\right)\biggr).
\end{aligned}$$
Let us study $S_n(f,s)$. As a sequence of partial sums of real-value independent random variables indexed by $j \in \{1,\cdots,k(n)\}$ with first and second moments
$$\Gamma(s+1)f(j)j^{-s} \ and \ (\Gamma(2s+1)-\Gamma(s+1)^2) f(j)^2j^{-2s},$$
the asymptotic normality is given by the the theorem of Levy-Feller-Linderberg (See Theorem 20 in [@ips-mfpt-ang] we apply to the centered *rrv*’s $\xi_j=f(j) j^{-s}(E_{j,n}^s-\Gamma(s+1))$, after remarking that
$$\left\{\frac{\mathbb{V}ar(\xi_j) }{\sum_{j=1}^{k(n)} \mathbb{V}ar(\xi_j)}, \ 1\leq j\leq k(n)\right\}=C(s) B_n(f,s).$$
So, as $n\rightarrow +\infty$,
$$\left(\frac{1}{s_n(f,s)} \sum_{j=1}^{k(n)} \left(f(j) j^{-s} (E_{j,n}^s-\Gamma(s+1))\right) \rightsquigarrow \mathcal{N}(0,1)\right) \ and \ B_n(f,s) \rightarrow 0$$
and the Lynderberg condition holds, that is, for any $\varepsilon>0$,
$$g(n,\varepsilon)=\frac{1}{s_n(f,s)} \sum_{j=1}^{k(n)} \int_{(|\xi_j|>\varepsilon s_n(f,s))} \xi_j^2 \ d\mathbb{P} \rightarrow 0.$$
But, for $K^2(s)=\Gamma(4s+1)-4\Gamma(3s+1)\Gamma(s+1)+\Gamma(2s+1)\Gamma(s+1)2-3\Gamma(3s+1)^4$,
$$\mathbb{E} \xi^4 = K(s) f(s)^4 j^{-4s}$$
and, by the Cauchy-Schwarz inequality
$$\begin{aligned}
&&\int_{(|\xi_j|>\varepsilon s_n(f,s))} \xi_j^2 \ d\mathbb{P}\\
&&\leq \left(\int \xi_j^4 \ d\mathbb{P}\right)^{1/2} \left(\int 1_{(|\xi_j|>\varepsilon s_n(f,s))} \ d\mathbb{P}\right)^{1/2}\\
&& =K f(j)^2 j^{2s} \left(\int 1_{(|\xi_j|>\varepsilon s_n(f,s))} \ d\mathbb{P}\right)^{1/2}\\
&& =K f(j)^2 j^{2s} \mathbb{P}\left(|\xi_j|>\varepsilon s_n(f,s)) \right)^{1/2}\\
&& \leq K f(j)^2 j^{2s} \left(\frac{K(s)^2 f(j)^4 j^{-4s}}{\varepsilon^4 s_n^4(f,s)}\right)^{1/2}\\
&& =K(s)^2 \left(f(j)^2 j^{2s}\right)^2 (s_n^{-2}(f,s)^2\\
&& =\frac{C(s)}{K(s)}B_n(f,s) \frac{\mathbb{V}ar(\xi_j)}{s^{2}\left(f,s\right)}\\\end{aligned}$$
So
$$g(n,\varepsilon)= \left(\frac{K(s)}{C(s)}\right)^{2}B_n(f,s) \rightarrow 0.$$
Our hypothesis $B_n(f,s)\rightarrow 0$ makes the Lynderberg hold and the central limit theorem holds for $S_n(f,s)$, that is
$$\frac{S_n(f,s)-\gamma^{s}a_n(f,s)}{s_n(f,s)} \rightsquigarrow \mathcal{N}(0,1)$$
Now, let us return to the approximation (B) at page . We have that for $s=1$, the expression denoted as $C_n$ between the pair of big parentheses should be equal to one as explained before. If $s>1$, we have $\sigma^2(s)=\sum_{j\geq 1} j^{-2(s-1)}<+\infty$, we apply a theorem of Kolmogorov (see [@ips-mfpt-ang], Proposition 25, page 233), $S_n(Id,s-1)$ weakly converges to the random variable $W(s)$ with variance $\sigma^2(s)$. Hence $C_n=O_{\mathbb{P}}(1)$. We arrive at
$$\biggr|\frac{T_n(f,s)-a_n(f,s)}{s_n(f,s)}-\frac{\gamma^{s}(S_n(f,s)-a_n(f,s))}{s_n(f,s)}\biggr|\leq O_{\mathbb{P}}\bigg(\frac{S_n(f,1)}{s_n(f,s)\log n}\biggr).$$
The later bound goes to zero in probability if and only if $s_n(f,1)/(s_n(f,s) \log n)\rightarrow 0$. Now, we have
$$\frac{s_n(f,s)}{a_n(f,s)} \biggr(\frac{T_n(f,s)}{a_n(f,s)}-\gamma^s\biggr)=Z_n+ o_{\mathbb{P}}(1).$$
If $s_n(f,s)/a_n(f,s)\rightarrow +\infty$, we can use the $\delta$-method applied to $g(t)=t^{1/s}$ to get
$$\frac{a_n(f,s)}{s_n(f,s)} \biggr(\left(\frac{T_n(f,s)}{a_n(f,s)}\right)^{1/s}-\gamma\biggr) \rightsquigarrow \mathcal{N}(0,s^{-2}\gamma^2). \blacksquare.$$
**Remark**. In [@ngomlo2016], we gave a direct proof of the asymptotic normality of $S_n(f,s)$ by using the two hypotheses $B_n(f,s)\rightarrow 0$ and $s_n(f,s) \rightarrow +\infty$. Here, it seems that we only used the first one. But that one could not hold if $S_n(f,s)$ contains a sub-sequence converging to a finite and positive number. That remark should be recalled in interpreting the results in [@ngomlo2016].\
Upper records values
====================
The main result is :
If, for each $n\geq 1$, $X^{(n)}$ stands for $n$-th record value, we have as $n\rightarrow +\infty$, $$\frac{X^{(n)}-\gamma n}{\gamma \sqrt{n}} \rightsquigarrow \mathcal{N}(0,1).$$
**Remark**. We refer the reader to [@lo2019ALR] for a simple introduction to records theory.\
**Proof**. We already noticed that $Z=exp(X)$ is the extremal domain of attraction of $G_{\gamma}(x)=\exp(-(1+\gamma x))$, for $\gamma x>-1$. From Part (b) of Theorem 1 in [@lo2019ALR], the $n$-th record $Z^{(n)}=exp(X^{(n)})$ have the representation
$$\left(\frac{\exp(X^{(n)})}{H^{-1}(1-e^{-n}}\right)^{1/\sqrt{n}}=\exp(\gamma S_{n}^{\ast}) +o_{\mathbb{P}}(1)$$
where $S_{n}^{\ast}$ has the same law as $\gamma^{-1}(T_n-n)/\sqrt{n}$ with $T_n$ denoting a $\gamma$ law with parameters $n$ and $1$. Since $H^{-1}(1-u)=\exp(F^{-1}(1-u))$, we have
$$\frac{X^{(n)}-F^{-1}\left(1-e^{-n}\right)}{\gamma \sqrt{n}}=S_{n}^{\ast} + o_{\mathbb{P}}(1)$$
By the central limit theorem, it comes that
$$\frac{X^{(n)}-F^{-1}\left(1-e^{-n}\right)}{\gamma \sqrt{n}}=\mathcal{N}(0,1) + o_{\mathbb{P}}(1).$$
By using Formula , we get
$$\begin{aligned}
&&\frac{X^{(n)}-\gamma n}{\gamma \sqrt{n}}=S^{\ast}_n + o_{\mathbb{P}}(1) \label{records01}\\
&&\frac{X^{(n)}-\gamma n}{\gamma \sqrt{n}}=\mathcal{N}(0,1) + o_{\mathbb{P}}(1). \label{records02}\end{aligned}$$
The proof is over. $\square$
The moment problem
==================
Typically, the moment problem on $\mathbb{R}$(see [@shohat]) is the following. Given a sequences real numbers $(m_n)_{n\geq 1}$, can we find a distribution (not necessarily a *cdf*) $F$ on $\mathbb{R}$ as the unique solution of the moments equations.
$$\forall n\geq 1, \ m_n=\int x^n \ dF(x).$$
This is a nice but difficult mathematical question treated in [@shohat]. But in the context of probability theory on $\mathbb{R}$, we may have a fixed *cdf* $F$ of random variable $X$ having moments
$$\forall n\geq 1, \ \mathbb{E} X^n = m_n \ finite.$$
The moment problem becomes : Is the sequence of moments $(m_n)_{n\geq 1}$ characterize the probability law of $X$. In that regard, we have
\[momProb\] The moments of the pseudo-Lindely probability law are the following
$$\forall n\geq 1, \ m_n=\frac{n! (\beta+n)}{\theta^n \beta}.$$
Any real-valued random variable have the moments $(m_n)_{n\geq 1}$ follows the pseudo-Limdley law.
**Proof**. At the place of a simple proof, we proceed to slight round-up of the moment problem and explain how to find a simple criteria based on Analysis. A possible tool is the characteristic function which characterize its associated probability law. We have the following expansion of any characteristic function of $X$ (see [@loeve] or [@ips-mfpt-ang], Lemma 5, page 255), we have
$$\mathbb{E}e^{iuX}=1+\sum_{k=1}^{n}\frac{(iu)^{k}mk}{k!}+\theta 2^{1-\delta }\mu
^{n+\delta }\frac{\left\vert u\right\vert ^{n+\delta }}{(n+1)!}. \label{expan-ch}$$
By usual analysis tools, the series in Formula converges in the $]-R, R[$ where $R$ is found according the Cauchy rule
$$\limsup_{n\rightarrow +\infty} (m_n)^{1/n} = R > 0.$$
The conclusion is that two random variables have the same moments of all orders have characteristic functions coinciding on $]-R, R[$. Finally, (see [@loeve], page 225, Part B.; see also [@billinsgleypm]) two characteristic function coinciding on an interval $]-R, R[$ coincide everywhere and thus, are associated to the same probability law.\
Let us apply to the pseudo-Lindley law. In [@Zeghoudi3], the moments are given by
$$\forall n\geq 1, \ m_n=\frac{n! (\beta+n)}{\theta^n \beta}.$$
Straightforward computation based on the Stirling formula leads to $R=1/\theta$. This is enough to prove the claim of the theorem.
**Appendix** \[appendix-erZ\]. Let $R=\beta/\theta$. In the computations below, $u \in (0,1)$ and $x\geq 0$ are linked by $u=1-F(x)$. So $u\rightarrow 0$ if and only if $x\rightarrow +\infty$. Also, below, functions of $x$ are functions of $u$ actually. We denote $A(u)=\log(1+R/x)$. We have $A(u)\rightarrow 0$ as $u\rightarrow 0$. By writing
$$\log(\beta +\theta x)=\log(\beta +\theta x)-\log \theta x+\log \theta x= \log \theta x +A(u),$$
we see that $u=1-F(x)$ gives
$$\theta x = \log(1/u) +\log R + \log x + A(u). \label{erZ_01}$$
So, we have
$$\label{erZ_02}
F^{-1}(1-u)=\theta^{-1} \log(1/u) (1+o(1)).$$
and
$$\label{erZ_03}
\log x = \log\log(1/u) (1+o(1)).$$
Now, we wish to develop that asymptotic equivalence with rates of convergence. Let $B(u)=\log R + \log x+ A(u)$. From Formula \[erZ\_01\], we have
$$\label{erZ_04}
\frac{x}{\theta^{-1}\log(1/u)}-1=\frac{B(u)}{\log(1/u)}.$$
By Formula , we notice that
$$\label{erZ_05}
B(u)= \log R + \log x + (R/x) - (R/x)^2/2 + O(\log(1/u)^{-3})=O(\log x)= (\log\log u)(1+o(1)),$$
and hence, for $D(u)=\log R + A(u)$,
$$\label{erZ_06}
\frac{\log(1/u)}{\log x}\left(\frac{x}{\theta^{-1}\log(1/u)}-1\right)= 1 + \frac{D(u)}{\log x}.$$
Also
$$\frac{D(u)}{\log x}=\frac{\log R + (R/x) - (R/x)^2/2 + O(x^{-3}}{\log x}$$
Next, we have
$$\begin{aligned}
\label{erZ_07}
&&\frac{\log x}{-\log R}\left(\frac{\log(1/u)}{\log x}\left(\frac{x}{\theta^{-1}\log(1/u)}-1\right)-1\right)\\
&&=1+\frac{R}{x \log R} -\frac{R^2}{2x^2\log R}+O(x^{-3}) \notag\end{aligned}$$
and finally
$$\begin{aligned}
\label{erZ_08}
&&\frac{x \log R}{R}\left(\frac{\log x}{\log R}\left(\frac{\log(1/u)}{\log x}\left(\frac{x}{\theta^{-1}\log(1/u)}-1\right)-1\right)-1\right)\\
&&=1-\frac{R}{2x}+O(x^{-2}).\notag\end{aligned}$$
Now we want to do the same for $\log x$. From Formula \[erZ\_04\], we have
$$\label{erZ_09}
\log (\theta x)= \log\log(1/u) + \log(1+ B(u)/\log(1/u))$$
from which we get
$$\label{erZ_09}
\log x - \log\log(1/u)=-\log \theta + (B(u)/\log(1/u))+ O\left((B(u)/\log(1/u)^2\right).$$
From Formula , we have
$$\begin{aligned}
&&\frac{\log(1/u)}{\log x}\left(\frac{x}{\theta^{-1}\log(1/u)}-1\right)-\frac{\log(1/u)}{\log\log 1/u}\left(\frac{x}{\theta^{-1}\log(1/u)}-1\right)\\
&&=\left(\frac{x}{\theta^{-1}\log(1/u)}-1\right)\frac{-(\log(1/u))(\log x -\log\log 1/u)}{(\log x)(\log\log 1/u)}\\
&&=(1+D(u)/\log x)\left(\frac{1}{(\log x)(\log\log 1/u)}\left( -\log \theta + (B(u)/\log(1/u))+ O(B(u)/\log(1/u)^2) \right)\right)\\
&&=O((\log\log 1/u)^{2})\end{aligned}$$
Formula becomes
$$\label{erZ_10}
\frac{\log(1/u)}{\log\log 1/u}\left(\frac{x}{\theta^{-1}\log(1/u)}-1\right)= 1 + \frac{D(u)}{\log x}+ O((\log\log 1/u)^{2}).$$
That formula will be used with Formula \[erZ\_09\] and
$$\begin{aligned}
\label{erZ_11}
\frac{B(u)}{\log 1/u}&=&\frac{\log R}{\log 1/u} + \frac{\log\log 1/u}{\log 1/u}(1+o(1))\\
&+& \frac{(R/x)-(R/x)^2/2}{\log 1/u} + O((\log 1/u)^{-4}). \notag\end{aligned}$$
From \[erZ\_01\], and from the following formula we can check by using differentiation methods to establish monotonicity
$$x - x^2/2 \leq \log(1+x) \leq x$$
we have
$$\label{erZ_12a}
(R/x)-R^2/(2x^2) +\log R +\log x \leq \theta x - \log(1/u) \leq (R/x) +\log R +\log x.$$
But we also have
$$x= \log(1/u) \left(1 + \frac{\log \beta^{-1}+ \log x + A(u)}{\log(1/u)}\right)$$
which implies
$$\log x= \log\log(1/u) + \log \left(1 + \frac{\log \beta^{-1}+ \log x + A(u)}{\log(1/u)}\right)$$
By putting
$$H(u)=\frac{\log \beta^{-1}+ \log x + A(u)}{\log(1/u)},$$
we finally get
$$\label{erZ_12b}
H(u)-H(u)^2/2 \leq \log x -\log\log(1/u)\leq H(u).$$
By combining Formulas and , we get
$$\label{erZ_12c}
\left|\theta x - \log(1/u) - \log(1/u) \right|\leq \frac{1}{2}\left(\frac{R^2}{x^2}+H(u)^2\right).$$
Since $(R/x^2)$ and $H(u)^2$ are both $O(\log 1/u)^{-2})$, we have
$$\label{erZ_12d}
F^{-1}(1-u) = \theta^{-1}(\log(1/u) - \log\log(1/u)) + O(\log 1/u)^{-2}).$$
But since the derivative $\log\log(1/u)$ is $(-u \log(1/u))^{-1}$, we have for $d=-\log\log 2$,
$$\forall u \in ]0,1[, \ \log\log(1/u)- =\int_{u}^{1/2} \frac{1}{u \log(1/u} \ du,$$
and finally
$$\label{erZ_12e}
F^{-1}(1-u) = d + \theta^{-1}(\log(1/u) - \int_{u}^{1/2} \frac{1}{u \log(1/u)} \ du + O\left(\left(\log 1/u\right)^{-2}\right).$$
[99]{} Patrick Billingsley (1995). *Probability and Measure*. Wiley. Third Edition.\
Lo G.S. (2019). A series of statistical packages using VB6. Arxiv:
Lo, G.S.(2016). Weak Convergence (IA). Sequences of random vectors. . Saint-Louis, Senegal - Calgary, Canada. Doi : 10.16929/sbs/2016.0001. Arxiv : 1610.05415. ISBN : 978-2-9559183-1-9
Lo G. S.(2010). A simple note on some empirical stochastic process as a tool in uniform L-statistics weak laws. *Afrika Statistika*, **5**, pp. 437-446. (MR2920301)
Lo G. S.(2016). How to use the functional empirical process for deriving asymptotic laws for functions of the sample. Arxiv : 1607.02745 (https://arxiv.org/pdf/1607.02745)
Zeghdoudi H and Nedjar S. (2016). A pseudo Lindley-distribution and its application. *Afrika Statistika* Volume 11 (1), pp. 923-932. DOI: http://dx.doi.org/10.16929/as/2016.923.83
Lo G.S., K. T. A. Ngom M. and Diallo M.(2018). Weak Convergence (IIA) - Functional and Random Aspects of the Univariate Extreme Value Theory. Arxiv : 1810.01625
Ngom and Lo G.S. (2016) A double-indexed functional Hill Process and Applications. Journal of Mathematical Research (e-ISSN 1916-9809). Vol 8 (4) , pp. 144-165, Doi : 105539/jmr/v8n4p144
Shorack G.R. and Wellner J.A. (1986). Empirical Processes with Applications to Statistics, Wiley-Interscience, New-York.\
de Haan, L. (1970). *On regular variation and its application to the weak convergence of sample extremes*. Mathematical Centre Tracts, **32**, Amsterdam. (MR0286156)
de Haan, L. and Ferreira A. (2006). *Extreme value theory: An introduction*. Springer. (MR2234156)
Lo, G.S. and Ahnsanullah M.(2019). Asymptotic Laws of the strong Upper Records Values in the extreme domain of attraction and beyond. arXiv:1905.03380
Shohat J.A. and Tamarkin J.D (1943) The problem of moment, Mathematical Surveys and monographs. Volume I. American Society of Mathematics (Re-edited in 1950, 1963 and 1970).
Hill, B.(1975). A simple general approach to the inference about the tail index of a distribution. *Ann Statist*. 3, 1163-1174.
Loève, Michel.(1997). *Probability Theory I*. Springer-Verlag, 4th Edition.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the space of orthogonally additive $n$-homogeneous polynomials on $C(K)$. There are two natural norms on this space. First, there is the usual supremum norm of uniform convergence on the closed unit ball. As every orthogonally additive $n$-homogeneous polynomial is regular with respect to the Banach lattice structure, there is also the regular norm. These norms are equivalent, but have significantly different geometric properties. We characterise the extreme points of the unit ball for both norms, with different results for even and odd degrees. As an application, we prove a Banach-Stone theorem. We conclude with a classification of the exposed points.'
author:
- 'Christopher Boyd, Raymond A. Ryan and Nina Snigireva'
bibliography:
- 'OAddC(K).bib'
title: ' Geometry of Spaces of Orthogonally Additive Polynomials on $C(K)$'
---
Introduction
============
A real function $f$ on a Banach lattice is said to be *orthogonally additive* if $f(x+y) = f(x)+f(y)$ whenever $x$ and $y$ are disjoint. Non-linear orthogonally additive functions on function spaces often have useful integral representations — see, for example the papers of Chacon, Friedman and Katz [@Chacon; @Friedman1; @Friedman2], Mizel [@Mizel] and Rao [@Rao]. In 1990, Sundaresan [@Sund] initiated the study of orthogonally additive $n$-homogeneous polynomials with particular reference to the spaces $L_p[0,1]$ and $\ell_p$ for $1\le p <\infty$. Building on the work of Mizel, he showed that, for every orthogonally additive $n$-homogeneous polynomial $P$ on $L_p[0,1]$ with $n\le p$, there exists a unique function $\xi\in L_{\tilde{p}}$, where $\tilde{p}=p/(p-n)$, such that $$\label{e:Sund}
P(x) = \int_0^1 \xi x^n\,d\mu$$ for every $x\in L_p[0,1]$. When $n>p$, there are no non-zero orthogonally additive $n$-homogeneous polynomials on $L_p[0,1]$. He went on to show that the Banach space of orthogonally additive $n$-homogeneous polynomials on $L_p[0,1]$ is isometrically isomorphic to $L_{\tilde{p}}$ where the latter space is equipped not with the usual norm, but with the equivalent norm $\|x\|=
\max\{\|x^+\|_{\tilde{p}},\|x^-\|_{\tilde{p}} \}$.
The next significant development was the discovery of an integral representation for orthogonally additive $n$-homogeneous polynomials on $C(K)$ spaces by Pérez and Villanueva [@PV] and by Benyamini, Lassalle and Llavona [@BLL], who proved a representation of the form $$\label{e:IntRep}
P(x) = \int_K x^n\,d\mu$$ where $\mu$ is a regular Borel signed measure on $K$. The integral representations (\[e:Sund\]) and (\[e:IntRep\]) have been extended and generalized in various directions in recent years. See, for example, [@Palazuelos; @Kusraeva11; @Alaminos; @Villena].
Orthogonally additive $n$-homogeneous polynomials are also of interest in the study of multilinear operators on Banach lattices and, more generally, on vector lattices. If $E,F$ are vector lattices, an $n$-linear mapping $A\colon E^n \to F$ is *orthosymmetric* if $A(x_1,\dots,x_n) =0$ whenever $x_i$ and $x_j$ are disjoint for some pair of distinct indices $i,j$. Orthosymmetric multilinear mappings are automatically symmetric [@Boul03]. In [@BuBuskes12], Bu and Buskes prove that an $n$-linear function is orthosymmetric if and only if the associate $n$-homogeneous polynomial is orthogonally additive.
Let $E$ be a Banach lattice. For every positive element $a$ of $E$, we may form the principal ideal $$E_a = \{x\in E: |x|\le na \text{ for some $n\in{\ensuremath{\mathbb{N}}}$} \}$$ with lattice structure inherited from $E$ and the norm defined by $\|x\|_a = \inf\{C>0:|x|\le Ca\}$. With this norm, $E_a$ is a Banach lattice. By virtue of the Kakutani representation theorem [@KakAM], the Banach lattice $E_a$ is canonically Banach lattice isometrically isomorphic to $C(K)$ for some compact Hausdorff topological space $K$, with $a$ being identified with the unit function on $K$. The Banach lattice structure of $E$ is uniquely determined by its principal ideals. It follows that an analysis of the orthogonally additive $n$-homogeneous polynomials on $C(K)$ is central to an understanding of the behaviour of orthogonally additive $n$-homogeneous polynomials on general Banach lattices.
In this paper, we focus on the geometric properties of the spaces ${\EuScript{P\!}_o(^{n}{{C(K)}})}$ of orthogonally additive $n$-homogeneous polynomials on ${C(K)}$. There are two phenomena that are of particular interest. The first is that there are two natural ways to norm the space ${\EuScript{P\!}_o(^{n}{E})}$. The first is the norm of uniform convergence on the unit ball of $E$, given by $\|P\|_\infty= \sup\{|P(x)|: x\in E, \|x\|\le 1\}$. In this norm, ${\EuScript{P\!}_o(^{n}{E})}$ is a Banach space. Now ${\EuScript{P\!}_o(^{n}{E})}$ also has a lattice structure and so another choice of norm is the *regular norm*, defined by $\|P\|_r = \| |P| \|_\infty$, where $|P|$ is the absolute value of $P$. In this norm, ${\EuScript{P\!}_o(^{n}{E})}$ is a Banach lattice. The existence of these two norms was first observed by Bu and Buskes [@BuBuskes12] and is hinted at in the paper of Sundaresan [@Sund]. These norms are equivalent, but we shall see that they have significantly different geometric properties.
The second phenomenon is the influence of the parity of the degree $n$ on the structure of the space ${\EuScript{P\!}_o(^{n}{E})}$ for the two norms. Bu and Buskes [@BuBuskes12] showed that, when $n$ is odd, the supremum and regular norms on ${\EuScript{P\!}_o(^{n}{E})}$ are the same and that they are equivalent when $n$ is even. We sharpen their results, using the strategy of working first on ${\EuScript{P\!}_o(^{n}{{C(K)}})}$ and then extending to general Banach lattices. The integral representation (\[e:IntRep\]) gives a canonical isomorphism between ${\EuScript{P\!}_o(^{n}{{C(K)}})}$ and ${\EuScript{M}(K)}$, the space of regular Borel signed measures on $K$. The regular norm on ${\EuScript{P\!}_o(^{n}{{C(K)}})}$ corresponds to the usual variation norm on ${\EuScript{M}(K)}$, but the supremum norm is identified with a different norm on ${\EuScript{M}(K)}$, given by $\|\mu\|_0 =
\max\{\|\mu^+\|_1,\|\mu^-\|_1\}$. We show that $({\EuScript{M}(K)}, \|\cdot\|_0)$ is isometrically isomorphic to the dual space of ${C(K)}$, where ${C(K)}$ is endowed with the norm ${\|{x}\|_d}=
\|x^+\|_\infty + \|x^-\|_\infty$ and we show that this norm is closely related to the *diameter seminorm* (see, for example, [@Cabello]). We use these identifications to give a complete description of the extreme points of the unit ball of ${\EuScript{P\!}_o(^{n}{{C(K)}})}$ for both norms, extending the results in [@CLZ1]. Our starting point is a characterisation of the extreme points in ${C(K)}$ for the norm ${\|\cdot\|_d}$ and ${\EuScript{M}(K)}$ for the norm $\|\cdot\|_0$. This allows us to prove a Banach-Stone theorem for $({C(K)},{\|\cdot\|_d})$.
We finish with a study of the exposed points of the unit ball of the space ${\EuScript{P\!}_o(^{n}{{C(K)}})}$. The identification of this space with the space of measures ${\EuScript{M}(K)}$, which is a dual space for both norms, allows us to use the theory of [Š]{}mul’yan [@Smulyan1; @Smulyan2]. Using this machinery, we characterise the weak$^*$ exposed and the weak$^*$ strongly exposed points of the unit ball.
Preliminaries {#preliminaries .unnumbered}
-------------
Let $E$ be a real Banach space and let $n$ be a natural number. A function $P\colon E\to {\ensuremath{\mathbb{R}}}$ is an *$n$-homogeneous polynomial* if there exists a necessarily unique, bounded $n$-linear function $A\colon E^n \to {\ensuremath{\mathbb{R}}}$ such that $P(x)= A(x,\dots,x)$ for all $x\in E$. We write $P=\widehat{A}$ if $P$ and $A$ are related in this way. The space ${\EuScript{P}(^{n}{E})}$ of $n$-homogeneous polynomials is a Banach space with the supremum norm, $$\|P\|_\infty= \sup\{|P(x)|:
x\in E, \|x\|\le 1 \} \,.$$ The Banach space $\bigl({\EuScript{P}(^{n}{E})},\|\cdot\|_\infty\bigr)$ is a dual space. We refer to the book by Dineen [@Dineen] for this and other facts about $n$-homogeneous polynomials.
Now assume that $E$ is a Banach lattice. A partial order is defined on ${\EuScript{P}(^{n}{E})}$ by $P=\widehat{A} \le
Q= \widehat{B}$ if $A(x_1,\dots,x_n) \le
B(x_1,\dots,x_n)$ for all $x_1,\dots,x_n\ge 0$. In particular, an $n$-homogeneous polymonial $P$ is said to be *positive* if $P\ge 0$ in the sense of this order and $P$ is *regular* if it is the difference of two positive $n$-homogeneous polynomials. The regular polynomials are precisely those that have an *absolute value*, which is given by the formula $$\label{e:AbsFormula}
|P|(x) = \sup\Bigl\{\, \sum_{i_1,\dots,i_n}
|A(u^1_{i_1},\dots,u^n_{i_n})|: u^1,\dots,u^n \in \Pi(x)\Bigr\}\,,$$ where $\Pi(x)$ denotes the set of partitions of $x$, namely, all finite sets of positive elements of $E$ whose sum is $x$ [@BuBuskes12].
The space ${\EuScript{P}_r(^{n} {E})}$ of regular $n$-homogeneous polynomials on $E$ is a Banach lattice with the *regular norm*, $$\|P\|_r = \| \,|P|\,\|_\infty\,.$$ We have $\|P\|_\infty \le \|P\|_r$ and in general these norms are not equivalent on ${\EuScript{P}_r(^{n} {E})}$. Every regular $n$-homogeneous polynomial $P$ can be decomposed canonically as the difference of two positive $n$-homogeneous polynomials, so that $P=P^+ - P^-$ and $|P|= P^+ + P^-$. We refer to the paper of Bu and Buskes [@BuBuskes12] for further details. For example, they show that $\bigl({\EuScript{P}_r(^{n} {E})},\|\cdot\|_r\bigr)$ is a dual Banach lattice.
Let $K$ be a compact, Hausdorff space. The space ${C(K)}$ of continuous real functions on $K$ is a Banach lattice with the supremum norm, $\|x\|_\infty= \sup\{|x(t)| : t\in K\}$. We denote by ${\EuScript{M}(K)}$ the space of regular Borel signed measures on $K$. Then the Banach lattice dual of ${C(K)}$ can be identified with ${\EuScript{M}(K)}$ under the variation norm, which we denote by $\|\cdot\|_1$. Thus, $$\|\mu\|_1 = |\mu|(K) = \mu^+(K)+\mu^-(K)
=\|\mu^+\|_1 + \|\mu^-\|_1\,,$$ where $\mu^+$, $\mu^-$ are the positive and negative parts of $\mu$.
Orthogonally additive $n$-homogeneous polynomials
=================================================
Let $E$ be a Banach lattice and $n$ a positive integer. A function $P\colon E\to {\ensuremath{\mathbb{R}}}$ is called an *orthogonally additive $n$-homogeneous polynomial* if $P$ is a bounded $n$-homogeneous polynomial with the property that $P(x+y)=P(x)+P(y)$ whenever $x,y\in E$ are disjoint. The space of orthogonally additive $n$-homogeneous polynomials on $E$ is denoted by ${\EuScript{P\!}_o(^{n}{E})}$. It is easy to see that ${\EuScript{P\!}_o(^{n}{E})}$ is a closed subspace of the space $\bigl({\EuScript{P}(^{n}{E})} ,\|\cdot\|_\infty\bigr)$ of bounded $n$-homogeneous polynomials with the supremum norm. Thus ${\EuScript{P\!}_o(^{n}{E})}$, with this norm, is a Banach space. When $n=1$, this space is simply the dual space $E'$, since every bounded linear functional is orthogonally additive.
We have the following integral representation for orthogonally additive $n$-homogeneous polynomials on $C(K)$ spaces, due to Pérez-García and Villanueva [@PV] and Benyami, Lassalle and Llavona [@BLL] (see also [@CLZ1]).
\[Rep\] Let $K$ be a compact, Hausdorff topological space. For every orthogonally additive $n$-homogeneous polynomial $P$ on $C(K)$ there is a regular Borel signed measure $\mu$ on $K$ such that $$P(x) = \int_K x^n\,d\mu$$ for all $x\in C(K)$.
In general, there is no guarantee that a Banach lattice supports any non-trivial orthogonally additive polynomials of degree greater than one. Sundaresan [@Sund] showed that there are no non-zero orthogonally additive $n$-homogeneous polynomials on $L_1[0,1]$ for $n>1$. In the case of $\ell_1$, it is easy to see that an $n$-homogeneous polynomial $P$ is orthogonally additive if and only if there exists a bounded sequence of real numbers, $(a_j)$, such that $$P(x) = \sum_{j=1}^\infty a_j x_j^n$$ for every $x\in \ell_1$, and that $\|P\|_\infty = \sup_j |a_j|$. Thus ${\EuScript{P\!}_o(^{n}{\ell_1})}$ is isometrically isomorphic to $\ell_\infty$ for every $n$.
To put the results of the previous paragraph in a general context, we recall that a Banach lattice $E$ is an AL-space if the norm is additive on the positive cone: $\|x+y\| = \|x\| + \|y\|$ for all $x,y\ge 0$. The Kakutani representation theorem [@KakAL; @Lacey] states that every AL-space $E$ can be decomposed into a disjoint sum of copies of $\ell_1$ and $L_1$ spaces. Accordingly, $E$ is Banach lattice isometrically isomorphic to a space of the form $$\Bigl[\ell_1(\Gamma) \oplus \bigl(
\bigoplus_{\alpha\in A }
L_1[0,1]^{m_\alpha}\bigr )\Bigr]_1$$ In this representation, the unit basis vectors $e_\gamma$ in $\ell_1(\Gamma)$ are in one-to-one correspondence with the atoms in $E$ of unit norm. We recall that a positive element $x$ of $E$ is said to be an *atom* if $0\le y\le x$ implies that $y$ is a scalar multiple of $x$. We can write the second component in this representation as $L_1(\mu)$, where $\mu$ is the product of the Lebesgue measures on the sets $[0,1]^{m_\alpha}$. Thus, we see that $E$ can be represented as the disjoint sum $\ell_1(\Gamma)\oplus_1 L_1(\mu )$, where the measure $\mu$ is nonatomic.
Let $E$ be an AL-space and let $n>1$. There is a non-zero orthogonally additive $n$-homogeneous polynomial on $E$ if and only if $E$ contains at least one atom.
Let $\ell_1(\Gamma)\oplus_1 L_1(\mu )$ be the Kakutani representation of $E$ as described above.
Suppose that $E$ contains an atom. Then the set $\Gamma$ in the Kakutani representation is non-empty. Choose $\gamma_0 \in \Gamma$ and define $P(x) = x_{\gamma_0}^n$ for $x= (x_\gamma)\in
\ell_1(\Gamma)$ and $P(x) = 0$ for $x\in L_1(\mu)$. Then $P$ is a non-zero orthogonally additive $n$-homogeneous polynomial.
Conversely, suppose that $P$ has no atoms. Then the Kakutani representation of $P$ is $L_1(\mu)$ where the measure $\mu$ is nonatomic. The proof in this case can be gleaned from [@Sund], but we can give a direct proof as follows. We treat the case $n=2$ for simplicity. Suppose that $P$ is an orthogonally additive $2$-homogeneous polynomial on $L_1(\mu)$, where $\mu$ is nonatomic. Let $A$ be the bounded, symmetric bilinear form that generates $P$. Then $A$ is orthosymmetric: if $x,y$ are disjoint, then $A(x,y)=0$ [@BuBuskes12 Lemma 4.1]. It follows from the fact that $L_1(\mu)\hat{\otimes}_\pi L_1(\mu)$ is isometrically isomorphic to $L_1(\mu^2)$ that there exists $g\in L_\infty(\mu^2)$ such that $$A(x,y) = \int x(s)y(t) g(s,t)\,d\mu^2(s,t)$$ If we take $x$, $y$ to be the characteristic functions of arbitrary disjoint measurable sets, this integral is zero and so we have $$A(x,y) =
\int_D x(t) y(t) g(t,t)\,d\mu^2$$ for all $x,y\in L_1(\mu)$, where $D$ is the diagonal. However, if $\mu$ has no atoms, then the product measure of the diagonal is zero. Hence $P(x)=0$ for every $x$.
The Banach lattices $L_1(\mu)$, where $\mu$ is nonatomic, do not support any real valued lattice homomorphisms. Our next result indicates that the existence of non-trivial $n$-homogeneous orthogonally additive polynomials on a Banach lattice is closely related to the existence of lattice homomorphisms.
Let $E$ be a Banach lattice, let $\varphi\in E'$ and let $n\ge 2$. The $n$-homogeneous polynomial defined by $P(x)= \varphi(x)^n$ is orthogonally additive if and only if either $\varphi$ or $-\varphi$ is a lattice homomorphism.
Suppose that $\varphi$ or $-\varphi$ is a lattice homomorphism. Then if $x$ and $y$ are disjoint, we have either $\varphi(x)$ or $\varphi(y)=0$ and so $P(x+y)=P(x)+P(y)$.
Conversely, suppose that $P=\varphi^n$ is orthogonally additive. For every $x\in E$, the vectors $x^+$ and $tx^-$ are disjoint for all $t\in {\ensuremath{\mathbb{R}}}$. Therefore $$\varphi(x^+)^n+t^k\varphi(x^-)^n = P(x^+ + tx^-) =
\sum_{j=0}^n \binom{n}{j} \varphi(x^+)^{n-j}\varphi(x^-)^j t^j\,.$$ for every $t\in {\ensuremath{\mathbb{R}}}$. Hence either $\varphi(x^+)=0$ or $\varphi(x^-)=0$. If we can show that $\varphi$ (or $-\varphi$) is positive, then it follows that $\varphi$ (or $-\varphi$) is a lattice homomorphism.
Let $a$ be a positive element of $E$. The principal ideal $E_a$ generated by $a$ is isometrically Banach lattice isomorphic to $C(K)$ for some compact Hausdorff topological space $K$. The functional $\varphi$ is represented by a regular Borel signed measure $\mu$ on $K$ and the fact that $\varphi(x^+)$ or $\varphi(x^-)=0$ for all $x\in E_a$ implies that the support of $\mu$ consists of a single point. It follows that either $\varphi$ or $-\varphi$ is positive on $E_a$. Now $E$ is the union of the principal ideals $E_a$, which are upwards directed by inclusion. Thus, if $\varphi$ (or $-\varphi$) is positive on one $E_a$, then $\varphi$ (or $-\varphi$) is positive on all of $ E$.
A Banach lattice $E$ is an AM-space if the norm has the property that $x\wedge y=0$ implies $\|x\vee y\| = \max\{\|x\|,\|y\|\}$. In contrast with AL-spaces, there is a good supply of orthogonally additive $n$-homogeneous polynomials on every AM-space. The Kakutani representation theorem for AM-spaces [@KakAM] shows that the real valued lattice homomorphisms on an AM-space $E$ separate the points of $E$. It follows that there is a rich supply of orthogonally additive $n$-homogeneous polynomials of every degree on $E$.
We now look at some properties of orthogonally additive polynomials on general Banach lattices. Our starting point is the fact that every orthogonally additive $n$-homogeneous polynomial on a Banach lattice $E$ is regular. This has been shown by Toumi [@ToumiReg Theorem 1]. One may also argue as follows. Let $P$ be an orthogonally additive $n$-homogeneous polynomial on a Banach lattice $E$. As $E$ is the upwards directed union of its principal ideals, it suffices to show that $P$ is regular on each of them. Since each principal ideal is Banach lattice isometrically isomorphic to a $C(K)$, we can use the integral representation in Theorem \[Rep\]. Then the Jordan decomposition of the representing measure gives a decomposition of the polynomial into the difference of two positive orthogonally additive $n$-homogeneous polynomials. Therefore $P$ is regular.
Let $P= \widehat{A}$ be a regular $n$-homogeneous polynomial $P$ on $E$. The absolute value of $P$ is given by [@BuBuskes12; @Loane] $$\label{e:AbsFormula2}
|P|(x) = \sup\Bigl\{\, \sum_{i_1,\dots,i_n}
|A(u^1_{i_1},\dots,u^n_{i_n})|: u^1,\dots,u^n \in \Pi(x)\Bigr\}$$ for $x\ge 0$, where $\Pi(x)$ denotes the set of partitions of $x$, namely, all finite sets of positive vectors whose sum is $x$. In general, we have $$\label{e:Abs}
|P(x)| \le |P|(|x|)$$ for every $x\in E$ and $|P|$ is the smallest positive $n$-homogeneous polynomial, in the sense of the lattice structure of ${\EuScript{P}_r(^{n} {E})}$, with this property. The space ${\EuScript{P}_r(^{n} {E})}$ is a Banach lattice in the *regular norm*, $$\|P\|_r = \| \,|P|\, \|_\infty\,.$$ It follows from (\[e:Abs\]) that $\|P\|_\infty \le \|P\|_r$ for every $P\in {\EuScript{P}_r(^{n} {E})}$. In general, these norms are not equivalent.
Now ${\EuScript{P\!}_o(^{n}{E})}$ is complete in the regular norm; indeed, it is even a dual Banach lattice [@BuBuskes12 Theorem 5.4]. It follows that the supremum and regular norms are equivalent on this space. Thus, there is a sequence $(C_n)$ of positive real numbers such that $\|P\|_r \le C_n\, \|P\|_\infty$ for every $n$ and every $P\in {\EuScript{P\!}_o(^{n}{E})}$. Bu and Buskes [@BuBuskes12] show that the two norms are the same for odd values of $n$. For even values of $n$, they show that $C_n \le n^n/n!$, the polarization constant. We shall show that, in fact, $C_n=2$ for even values of $n$ and that this is sharp. This will follow from estimates we give for the value of $|P|$ at positive points in $E$.
If $\varphi$ is a bounded linear functional on $E$, then [@Meyer-Nieberg] $$|\varphi|(x) = \sup\{|\varphi(y)|: |y|\le x\}$$ for every $x\ge 0$. It would be suprising if there were such a simple formula for $|P|(x)$ when $P$ is a regular $n$-homogeneous polynomial. As a linear functional, $P$ acts on an $n$-fold symmetric tensor power of $E$ and the set of vectors $y$ satisfying $|y|\le x$ is now a set of tensors, rather than elements of $E$. However, if $P$ is orthogonally additive, it is possible to establish a relatively simple estimate for the values of $|P|$.
\[p: basic\] Let $P$ be an orthogonally additive $n$-homogeneous polynomial on the Banach lattice $E$.
1. If $n$ is odd, then $$|P|(x) =
\sup\bigl\{|P(y)|: |y| \le x\bigr\}\,.$$ for every $x\ge 0$ in $E$.
2. If $n$ is even, then $$|P|(x) \le 2\,
\sup\bigl\{|P(y)|: |y| \le x\bigr\}\,.$$ for every $x\ge 0$ in $E$.
Let $x\ge 0$. It follows from (\[e:AbsFormula\]) that the value $|P|(x)$ is unchanged if we consider $P$ as an $n$-homogeneous polynomial on the principal ideal $E_x$ generated by $x$. Now $E_x$ is Banach lattice isomorphic to $C(K)$ for some compact topological space $K$. Since $P$ is orthogonally addive there exists a regular signed Borel measure $\mu$ on $K$ such that $$P(y)= \int_K y^n \,d\mu\,.$$ for every $y\in E_x\cong C(K)$. The symmetric $n$-linear form on $C(K)^n$ that generates $P$ is given by $$A(x_1,\dots,x_n) = \int_K x_1\dots x_n\,d\mu\,.$$ Thus, for $x_1,\dots,x_n \ge 0$, $$|A(x_1,\dots,x_n)| \le \int_K x_1\dots x_n\,d|\mu|$$ and it follows that $$|P|(x) \le \int_K x^n\,d|\mu|$$ for $x\ge 0$.
Now in general, for a nonnegative function $w\in C(K)$ we have $$\int_K w\,d|\mu| =
\sup\Bigl\{\Bigl| \int_K g\,d\mu\Bigr|:
g \in C(K), |g| \le w \Bigr\}\,,$$ Therefore $$|P|(x) \le \sup
\Bigl\{\Bigl|\int_K y\,d\mu\Bigr| : y\in C(K), \;|y| \le x^n \Bigr\}\,.$$ where we are identifying elements of $E$ with continuous functions on $K$. We now consider separately the cases where $n$ is odd and even.
\(a) We first consider the case when $n$ odd.\
If $|y|\le x^n$, let $v= y^{1/n}$. Then $|v|\le x$ and $\int_K y\,d\mu = \int_K v^n\,d\mu$. Therefore $$|P|(x) \le \sup
\Bigl\{\Bigl|\int_K v^n\,d\mu\Bigr| : v\in E, |v| \le x \Bigr\}\,.$$ Thus we have $$|P|(x) \le \sup \{|P(y)| : |y| \le x\}\,.$$ and it is easy to see that the reverse inequality also holds.
\(b) We now consider the case when $n$ even.\
We have $$|P|(x) \le \sup
\Bigl\{\Bigl|\int_K y\,d\mu\Bigr| : |y| \le x^n \Bigr\}\,.$$ Given $v\in E_x\cong C(K)$ satisfying $|v|\le x^n$, we define $v_1,v_2\in C(K)$ by $$v_1(t)= \begin{cases} v(t)^{1/n} &\text{if } v(t)\ge 0\\
0 &\text{if } v(t) <0
\end{cases} \qquad
v_2(t)= \begin{cases} 0 &\text{if } v(t)\ge 0\\
|v(t)|^{1/n} &\text{if } v(t) <0
\end{cases}$$ Then $v= v_1^n - v_2^n$, and so $$\Bigl| \int_K v\,d\mu\Bigr|
\le \Bigl|\int_K v_1^n\,d\mu\Bigr| +
\Bigl| \int_K v_2^n\,d\mu\Bigl| =
\bigl| P(v_1)\bigr| +\bigl|P(v_2)\bigr|\,.$$ It follows from $|v|\le x^n$ that $0\le v_1,v_2 \le x$. Therefore $$\label{e:bound}
|P|(x) \le 2\,
\sup\bigl\{|P(y)|: 0\le y \le x\bigr\} =
2\,\sup\bigl\{|P(y)|: |y| \le x\bigr\}\,,$$ since $n$ is even.
To see that the bound in (\[e:bound\]) for even values of $n$ is sharp, consider the example $P(x)=x_1^n-x_2^n$ on ${\ensuremath{\mathbb{R}}}^2$ with any Banach lattice norm. The bound is attained for the vector $x=(1,1)$.
\[c: basic\] Let $P$ is an orthogonally additive $n$-homogeneous polynomial on a Banach lattice $E$. Then $\|P\|_r = \|P\|_\infty$ if $n$ is odd and $\|P\|_\infty \le \|P\|_r \le 2\,\|P\|_\infty$ if $n$ is even. These inequalities are sharp.
Orthogonally additive polynomials on $C(K)$
===========================================
In this section, we study the supremum and regular norms on the spaces of orthogonally additive $n$-homogeneous polynomials on $C(K)$.
The integral representation for orthogonally additive $n$-homogeneous polynomials on $C(K)$ allows us to identify the vector space ${\EuScript{P}_r(^{n} {C(K)})}$ with ${\EuScript{M}(K)}$, the space of regular Borel signed measures on $K$. The natural norm on ${\EuScript{M}(K)}\cong {C(K)}' $ is the dual norm. This is the *variation norm* for measures: $\|\mu\|_1 = |\mu|(K)$. We shall see that this norm corresponds to the regular norm on the spaces of orthogonally additive $n$-homogeneous polynomials. However, the supremum norm on ${\EuScript{P\!}_o(^{n}{K})}$ corresponds to a different, but equivalent norm on the space of regular Borel signed measures.
The space ${\EuScript{P}_r(^{n} {C(K)})}$ is a Banach lattice with the regular norm, as is the dual Banach lattice ${\EuScript{M}(K)}$ with the variation norm. We shall see that the lattice structures of these two Banach lattices are the same. We note that the lattice structure of ${\EuScript{M}(K)}$ as the dual of $C(K)$ is the same as the lattices structure of ${\EuScript{M}(K)}$ considered as a sublattice of the lattice of Borel signed measures on $K$. In other words, a measure $\mu\in {\EuScript{M}(K)}$ is positive, in the sense that $\int_K f\,d\mu \ge 0$ for every nonnegative $x\in C(K)$, if and only if $\mu(E)\ge 0$ for every Borel subset $E$ of $K$ [@Rudin Theorem 2.18].
\[l:|P|\] Let $K$ be a compact Hausdorff topological space and let $P$ be an orthogonally additive $n$-homogeneous polynomial on $C(K)$, given by $$P(x) = \int_K x^n\,d\mu\,.$$ Then the absolute value of $P$ is given by $$|P|(x) = \int_K x^n \,d|\mu|\,.$$
We have seen in the proof of Theorem \[p: basic\] that $$|P|(x) \le \int_K x^n \,d|\mu|$$ for every $x\ge 0$.
To prove the reverse inequality, we start with the definition of the absolute value: $$|P|(x) = \sup\Bigl\{\, \sum_{i_1,\dots,i_n}
|A(u^1_{i_1},\dots,u^n_{i_n})|: u^1,\dots,u^n \in \Pi(x)\Bigr\}$$ for $x\ge 0$ Taking each of $u^2,\dots,u^n$ to be the trivial partition $\{x\}$ gives $$\begin{aligned}
|P|(x) &\ge
\sup\Bigl\{\, \sum_{i}
|A(u^1_{i},x,\dots,x)|: u^1 \in \Pi(x)\Bigr\}\\
& =
\sup\Bigl\{\, \sum_{i}
\Bigl|\int_K u^1_i x^{n-1}\,d\mu\Bigr|: u^1 \in \Pi(x)\Bigr\} = \int_K x^n\,d|\mu|\,,
\end{aligned}$$ applying the partition form of the Riesz-Kantorovich formula for the absolute value of a linear functional [@Aliprantis Theorem 1.16] to the measure $d\lambda = x^{n-1}d\mu$ in ${\EuScript{M}(K)}$ and using the fact that $d|\lambda| = x^{n-1}d|\mu|$.
\[p:isomorphism\] Let $K$ be a compact, Hausdorff space. Let $J_n \colon {\EuScript{M}(K)}\to {\EuScript{P\!}_o(^{n}{C(K)})}$ be given by $$(J_n\mu)(x) = \int_K x^n \,d\mu\,.$$
- For every $n$, $J_n$ is a Banach lattice isometric isomorphism from $\bigl({\EuScript{M}(K)},\|\cdot\|_1\bigr)$ onto $\bigl({\EuScript{P\!}_o(^{n}{C(K)})},\|\cdot\|_r\bigr)$.
- If $n$ is odd, then the regular and supremum norms coincide on ${\EuScript{P\!}_o(^{n}{C(K)})}$ and so $J_n$ is an isometric isomorphism for the supremum norm on ${\EuScript{P}_o(^{n}{C(K)})}$.
- If $n$ is even, then $J_n$ is an isometric isomorphism for the norm on ${\EuScript{M}(K)}$ defined by $$\|\mu\|_0 := \max\{\|\mu^+\|_1,\|\mu^-\|_1\}$$ and the supremum norm on ${\EuScript{P}_o(^{n}{C(K)})}$.
Cleary, $J_n$ is linear and surjective. To see that it is injective, suppose that the $n$-homogeneous polynomial $P(x) = \int_K x^n \,d\mu$ is zero. The associated symmetric $n$-linear form is $$A(x_1,\dots,x_n) = \int_K x_1\dots x_n\,d\mu$$ and so $A(x_1,\dots,x_n)=0$ for all $x_1,\dots,x_n \in
C(K)$. Taking $x_2= \dots = x_n = 1$, we have $\int_K x\,d\mu=0$ for every $x\in C(K)$ and so $\mu=0$.
\(a) Clearly, $J_n$ is positive. If we show that $J_n^{-1}$ is also positive, then it will follow that $J_n$ is a lattice homomorphism [@Aliprantis Theorem 7.3]. Let $P=\widehat{A}$ be a positive element of ${\EuScript{P}_r(^{n} {C(K)})}$, with $\mu\in {\EuScript{M}(K)}$ satisfying $J_n\mu = P$. Then, for every nonnegative $x\in C(K)$, we have $\int_K x\,d\mu = A(x,1,\dots,1) \ge 0$ and so $\mu$ is positive. Therefore $J_n$ is a lattice isomorphism for every $n$.
By Proposition \[l:|P|\], the regular norm of $P=J_n\mu$ is $\| P\|_r =
\|\,|P|\,\|_\infty =
|P|(1) = |\mu|(K) = \|\mu\|_1$, since $|P|$ is increasing on the positive cone of $C(K)$. Therefore $J_n$ is both a lattice isomorphism and an isometry.
\(b) This has already been proved in Corollary \[c: basic\].
\(c) Let $\mu \in {\EuScript{M}(K)}$ and let $P= J_n\mu$. It follows from (a) that $P^+= J_n \mu^+$ and $P^- = J_n \mu^-$. We have $$P(x) = \int_K x^n \,d\mu^+ - \int_K x^n d\mu^-$$ for every $x\in {C(K)}$. As $|a-b| \le \max\{|a|,|b|\}$ for $a,b\in {\ensuremath{\mathbb{R}}}^+$ and $n$ is even, it follows that $\|P\|_\infty \le \max\{\|\mu^+\|_1, \|\mu^-\|_1 \}$.
Now let $\{A,B\}$ be a Hahn decomposition of $\mu$, with $\mu$ positive on $A$ and negative on $B$. If $F \subset A$ is compact, then by a standard argument using Urysohn’s lemma (see, for example, [@Hewitt Theorem 12.41]) there is a decreasing sequence $(x_k)$ of continuous functions on $K$ with values in $[0,1]$ that converges almost everywhere with respect to $|\mu|$ to $1_F$, the characteristic function of $F$. Then, by the bounded convergence theorem, $$\|P\|_\infty \ge \lim_{k\to \infty}
\biggl| \int_K x_k^n\,d\mu\biggr|
= \biggl| \int_K 1_F \,d\mu \biggr|
= \mu^+(F)\,.$$ It follows from the regularity of $\mu^+$ that $\|P\|_\infty \ge \mu^+(A) = \|\mu^+\|_1$. Similarly, $\|P\|_\infty \ge \|\mu^-\|_1$. Therefore $\|P\|_\infty = \|\mu\|_0$.
We summarize the identifications of the various norms, bearing in mind that the supremum and regular norms coincide for positive polynomials.
Let $P$ be an orthogonally additive $n$-homogeneous polynomial on $C(K)$, with corresponding measure $\mu\in {\EuScript{M}(K)}$. Then
- $\|P\|_r =
\|P^+\|_r + \|P^-\|_r= \|\mu^+\|_1 + \|\mu^-\|_1 = \|\mu\|_1$.
- If $n$ is odd, then $\|P\|_\infty = \|P\|_r$.
- If $n$ is even, then $\|P\|_\infty =
\max\{ \|P^+\|_r, \|P^-\|_r \} =
\max\{\|\mu^+\|_1, \|\mu^-\|_1\} = \|\mu\|_0$.
We note that the norm $\|\cdot\|_0$ is easily seen to be equivalent to the dual (variation) norm on ${\EuScript{M}(K)}$. In fact, we have $$\|\mu\|_0 \le \|\mu\|_1 \le 2\,\|\mu\|_0$$ for every $\mu\in {\EuScript{M}(K)}$.
It will be useful to have an alternative expression for the norm $\|\cdot\|_0$ on ${\EuScript{M}(K)}$. Using the identity $\max\{a,b\} = \frac{1}{2} \bigl(|a+b| + |a-b|\bigr)$ for non-negative real numbers, we have $$\begin{aligned}
\|\mu\|_0 & =
\frac{1}{2}\bigl( \bigl|\|\mu^+\|_1+
\| \mu^-\|_1\bigr| +
\bigl| \|\mu^+\|_1 - \|\mu^-\|_1 \bigr|
\bigr) \\
&= \frac{1}{2}\bigl( \|\mu\|_1 +
\bigl| \mu^+(K)-\mu^-(K)\bigr| \bigr)
= \frac{1}{2}\bigl( \|\mu\|_1 + \bigl| \mu(K)\bigl| \bigr)
\label{e:mu0}\end{aligned}$$
Thus, we have $$\label{e:mu0}
\|\mu\|_0 = \max\bigl\{\|\mu^+\|_1, \|\mu^-\|_1\bigr\}
= \frac{1}{2}\Bigl( \|\mu\|_1 + \bigl| \mu(K)\bigl| \Bigr)$$ These results clarify the geometric properties of the spaces ${\EuScript{P}_o(^{n}{C(K)})}$; for the regular norm, these spaces are all essentially the same as the dual space ${\EuScript{M}(K)}$ with the variation norm. The case of ${\EuScript{P}_o(^{n}{C(K)})}$ with the supremum norm and $n$ even is substantially different. To understand this, we must study the extreme point structure of the unit ball of ${\EuScript{M}(K)}$ for the norm $\|\cdot \|_0$.
Extreme points in ${\EuScript{P\!}_o(^{n}{C(K)})}$
==================================================
In this section, we study the extreme points of the unit ball of the space ${\EuScript{P\!}_o(^{n}{C(K)})}$. We begin with the regular norm. We have seen in Proposition \[p:isomorphism\] that there is an isometric isomorphism $$\bigl({\EuScript{P\!}_o(^{n}{C(K)})}, \| \cdot\|_r\bigr)
\cong \bigl({\EuScript{M}(K)}, \|\cdot\|_1\bigr)$$ where $\|\cdot\|_1$ denotes the variation norm on ${\EuScript{M}(K)}$, the space of regular Borel signed measures on $K$. Furthermore, when the degree $n$ is odd, the supremum and regular norms on ${\EuScript{P\!}_o(^{n}{C(K)})}$ coincide.
It is a classical result that the extreme points of the unit ball of ${\EuScript{M}(K)}$ for the variation norm are the measures of the form $\pm \delta_t$, where $t\in K$ (see, for example, [@DS V.8.6]). The isomorphism between ${\EuScript{P\!}_o(^{n}{C(K)})}$ and ${\EuScript{M}(K)}$ associates the polynomial $P(x)= x(t)^n$ with the measure $\delta_t$. Thus, we have
\[p:extremeRegular\] Let $K$ be a compact Hausdorff topological space. The extreme points of the closed unit ball of the space $\bigl({\EuScript{P\!}_o(^{n}{C(K)})},\|\cdot\|_r\bigr)$ are the $n$-homogeneous polynomials $\pm \delta_t^n$, where $t\in K$ and $\delta_t^n(x)= x(t)^n$.
This result is given in [@CLZ1] for the supremum norm, but the proof given there is not valid for polynomials of even degree. However, this does not affect the results that follow in [@CLZ1]. In particular, their elegant proof of the integral representation still stands. Essentially, all that is required for their arguments to work is that ${\EuScript{P}_o(^{n}{C(K)})}$ is a dual space and that the extreme points of the unit ball are as described above.
We now turn to the geometry of ${\EuScript{P}_o(^{n}{C(K)})}$ for the supremum norm, where the degree $n$ is even. We have the isometric isomorphism $$\label{e: iso}
\bigl({\EuScript{P\!}_o(^{n}{C(K)})}, \|\cdot\|_\infty\bigr)
\cong \bigl({\EuScript{M}(K)}, \|\cdot\|_0\bigr)$$ where $\|\mu\|_0 = \max\{\|\mu^+\|_1, \|\mu^-\|_1\}$. We will show that $\|\cdot\|_0$ is the dual of a norm on ${C(K)}$ that is equivalent to the supremum norm.
The norm we seek is related to the *diameter seminorm* on ${C(K)}$, which is defined by $$\rho(x) = \operatorname{diam}(x) = \sup\{|x(s)-x(t)|: s,t\in K\}\,.$$ It is easy to see that we also have $$\rho(x) = 2\,\inf\bigl\{\|x-\alpha 1_K\|_\infty:
\alpha\in {\ensuremath{\mathbb{R}}}\bigr\}$$ The kernel of $\rho$ is the one dimensional subspace of constant functions. As in [@Cabello], we use $C_\rho(K)$ to denote the quotient space ${C(K)}/ \ker\rho$. It is a Banach space under the norm $$\|\pi(x)\|_\rho = \rho(x)$$ where $\pi\colon {C(K)}\to {C(K)}/\ker \rho$ is the quotient map. Following Cabello-Sanchez [@Cabello], we note that this means that $\bigl(C_\rho(K),\|\cdot\|_\rho\bigr)$ is isometrically isomorphic, up to a a constant factor $2$, to the quotient space of $\bigl({C(K)}, \|\cdot\|_\infty\bigr)$ by the subspace of constant functions. Therefore the dual space $\bigl(C_\rho(K),\|\cdot\|_\rho\bigr)'$ is isometrically isomorphic, up to a constant factor $1/2$, to a subspace of $\bigl({C(K)}, \|\cdot\|_\infty\bigr)'$, the space of regular Borel signed measures with the variation norm. This subspace is the space of measures $\mu$ satisfying $\mu(K)=0$ and on it we have [@Cabello] $$\|\mu\|_1 = 2\|\mu\|_{\bigl(C_\rho(K),\|\cdot\|_\rho\bigr)'}
\,.$$
\[p:Cabello\] Let $K$ be a compact Hausdorff topological space. A regular Borel signed measure $\mu$ is an extreme point of the unit ball of the dual space $\bigl(C_\rho(K),\|\cdot\|_\rho\bigr)'$ if and only if $\mu = \delta_s - \delta_t$, where $s$ and $t$ are distinct points of $K$.
In order to apply this result, we first need to identify the predual of the norm $\|\cdot\|_1$ on ${\EuScript{M}(K)}\cong {C(K)}'$.
\[p: predual\] Let $K$ be a compact Hausdorff topological space. Let ${\|\cdot\|_d}$ be the norm on $C(K)$ defined by $$\label{e:predual}
{\|{x}\|_d} := \|x^+\|_\infty + \|x^-\|_\infty
= \max\bigl\{\|x\|_\infty,\rho(x) \bigr\}$$ where $\rho$ is the diameter seminorm. Then the dual space of $\bigl(C(K), {\|\cdot\|_d}\bigr)$ is isometrically isomorphic to the space of regular Borel signed measures on $K$ with the norm $\|\mu\|_0
= \max\{\|\mu^+\|_1, \|\mu^-\|_1\}$.
A routine calculation shows that the formula ${\|{x}\|_d} = \|x^+\|_\infty + \|x^-\|_\infty$ defines a norm on ${C(K)}$. To establish the second equality in (\[e:predual\]), we consider two cases.
1. Suppose the function $x$ has constant sign. Then $\rho(x) \le \|x\|_\infty$ and one of $\|x^+\|_\infty$, $\|x^-\|_\infty$ is zero. Therefore ${\|{x}\|_d}
=\|x\|_\infty$.
2. If $x$ changes sign, then $\|x\|_\infty \le \|x^+\|_\infty + \|x^-\|_\infty = \rho(x)$.
Therefore ${\|{x}\|_d} =
\max\bigl\{\|x\|_\infty,\rho(x) \bigr\}$ for every $x\in {C(K)}$.
Let us denote the dual norm of $\|\cdot\|_1$ by $\|\cdot\|_1'$. If $x\in {C(K)}$ and $\mu\in{\EuScript{M}(K)}$, then $$\int_K x\,d\mu =
\Bigl( \int_K x^+\,d\mu^+ + \int_K x^-\,d\mu^-\Bigr)
-\Bigl(\int_K x^+\,d\mu^- + \int_K x^-\,d\mu^+\Bigr)\,.$$ Now $$0\le \int_K x^+\,d\mu^+ + \int_K x^-\,d\mu^-
\le \|x^+\|_\infty \|\mu^+\|_1 + \|x^-\|_\infty \|\mu^-\|_1
\le {\|{x}\|_d} \|\mu\|_0$$ and similarly $$0\le \int_K x^+\,d\mu^- + \int_K x^-\,d\mu^+
\le {\|{x}\|_d} \|\mu\|_0 \,.$$ Therefore $$\Bigl| \int_K x\,d\mu \Bigr| \le {\|{x}\|_d} \|\mu\|_0$$ and so $\|\mu\|_1' \le \|\mu\|_0$.
Fix $\mu\in {\EuScript{M}(K)}$ and let $\varepsilon>0$. Let $\{A,B\}$ be a Hahn decomposition for $\mu$, where $A$ is a positive set and $B$ a negative set. Since $\mu$ is regular, there exist compact sets $C\subset A$ and $D\subset B$ such that $|\mu|(A\setminus C), |\mu|(B\setminus D)
<\varepsilon$. By Urysohn’s lemma, there is a continuous function $y\colon K \to [0,1]$ that takes the values $1$ and $0$ on the sets $C$ and $D$ respectively. Then $\|y\|_1 = 1$ and $$\int_K y\,d\mu = \int_C y\,d\mu
+ \int_{A\setminus C} y\,d\mu
+ \int_{B\setminus D} y\,d\mu$$ It follows that $$\Bigl|\int_K y\,d\mu\Bigl| \ge
\mu^+(C)-2\varepsilon
\ge \mu^+(A) -3\varepsilon = \|\mu^+\|_1 -3\varepsilon\,.$$ Similarly, $$\Bigl|\int_K y\,d\mu\Bigr| \ge \|\mu^-\|_1 - 3\varepsilon$$ Thus, $\|\mu\|_1' \ge \|\mu\|_0 - 3\varepsilon$ for every $\varepsilon>0$
Therefore $\|\mu\|_1' = \|\mu\|_0$ for every $\mu\in {\EuScript{M}(K)}$.
The extreme points of the unit ball of $\bigl({C(K)},{\|\cdot\|_d}\bigr)$
-------------------------------------------------------------------------
The extreme points of the closed unit ball of ${C(K)}$ with the supremum norm are the constant functions $\pm 1$. Our next result shows that changing to the equivalent norm given in the preceding proposition leads to a different set of extreme points.
A function $x$ is an extreme point of the closed unit ball of $(C(K),{\|\cdot\|_d})$ if and only if either\
(i) $x(t)= 1$ or $0$ for every $t\in K$, or\
(ii) $x(t)= -1$ or $0$ for every $t\in K$\
(and $\{t: x(t)\neq 0 \}\neq \varnothing$ in each case.)
To show that every such function is extreme, let ${\|{x}\|_d}=1$, with $x(t)=1$ for $t\in A$ and $x(t)=0$ for $t\in A^c$, where $A$ is a nonempty subset of $K$. Suppose that $$x = a y + bz\,,$$ where $a,b\in (0,1)$ with $a+b=1$ and ${\|{y}\|_d} = {\|{z}\|_d} = 1$. Then, for $t\in A$, $ay(t)+bz(t) =1$. But $|y(t)|, |z(t)| \le 1$ and it follows that $y(t)= z(t)=1$ for every $t\in A$.
Now, if $t\in A^c$, then $ay(t)+bz(t) = 0$. But $\operatorname{diam}(y), \operatorname{diam}(z) \le 1$ and $\|y\|_\infty,
\|z\|_\infty =1$ imply that $0\le y(t), z(t)
\le 1$ for every $t\in K$ and hence $y(t)= z(t) = 0$ for every $t\in A^c$. Therefore $y(t)=z(t)=x(t)$ for every $t\in K$ and so $x$ is an extreme point. The case in which $x$ takes values $-1$ and $0$ is done in exactly the same way.
We now show that every extreme point is of this type. Let $x$ be an extreme point. Since ${\|{x}\|_d} =
\max\bigl\{\|x\|_\infty, \operatorname{diam}(x) \bigr\} =1$, there are two cases to consider.
Case 1: $\|x\|_\infty =1$ and $\operatorname{diam}(x) \le 1$. Then $x$ takes its values either in $[-1,0]$ or $[0,1]$. Suppose it is the latter. Then there is at least one point at which $x(t)=1$. Suppose there is a point $s\in K$ for which $0<x(s) <1$. Then, by a standard argument, there is a function $y\in C(K)$ with values in $[0,1]$ and supported by a neighbourhood of $s$, such that $\|x\pm y\|_\infty \le 1$. Clearly, we also have $\operatorname{diam}(x\pm y)\le 1$. This implies that $x$ is not extreme and so we have a contradiction. Therefore $x$ can only have values $0$ or $1$.
Case 2: $\operatorname{diam}(x) =1$ and $\|x\|_\infty < 1$. There exist points $s,t$ in $K$ such that $|x(t)-x(s)|= \operatorname{diam}(x)=1$. Without loss of generality we may assume that $x(t)>x(s)$. Then $x$ takes its values in the interval $[x(s),x(t)]$. If there exists $u\in K$ such that $x(s)< x(u)<x(t)$, then, using the same perturbation argument as in the proof of Case 1, it follows that $x$ is not extreme. Therefore $x$ has precisely two distinct values, $x(s)$ and $x(t)$.
Suppose that $-1<x(s)<x(t)<1$. Then, for sufficiently small $\varepsilon>0$, we have ${\|{x\pm \varepsilon 1_K}\|_d} = 1$, which implies that $x$ is not extreme. Therefore either $x(t)=1$ and $x(s)=0$, or $x(s)=-1$ and $x(t)=0$.
Note that if $K$ is connected, then the closed unit ball of $C(K)$ for both the supremum norm and the norm ${\|\cdot\|_d}$ has the same extreme points — the constant functions $1$ and $-1$. However, if $K$ has more than one connected component, then there are functions that are extreme for $\|\cdot\|_\infty$ but not ${\|\cdot\|_d}$, and vice versa.
The extreme points of the unit ball of $\bigl({\EuScript{M}(K)}, \|\cdot\|_0\bigr)$
-----------------------------------------------------------------------------------
The isometric isomorphism $$\bigl(C(K),{\|\cdot\|_d})' \cong
\bigl({\EuScript{M}(K)}, \|\cdot\|_0\bigr)\,.$$ now enables us to identify the extreme points of the unit ball of $\bigl({\EuScript{M}(K)},\|\cdot\|_0\bigr)$.
\[p:extremeMK\] Let $K$ be a compact Hausdorff topological space. A regular Borel signed measure $\mu$ on $K$ is an extreme point of the unit ball of $\bigl({\EuScript{M}(K)},\|\cdot\|_0\bigr)$ if and only if it is one of the following:
- $\mu = \pm \,\delta_t$, where $t\in K$;
- $\mu = \delta_s-\delta_t$, where $s$, $t$ are distinct points in $K$.
\
Step 1. We show that every extreme point must be one of the types described in the statement. Let $\tilde{K}$ be the space $K \cup K^2$, with the sum topology, where $K^2$ carries the product topology. For $x\in {C(K)}$, let $\tilde{x}$ be the continuous function on $\tilde{K}$ defined by $\tilde{x}(u) = x(u)$ for $u\in K$ and $\tilde{x}(s,t) = x(s)-x(t)$ for $(s,t)\in K^2$. The fact that $$\|\tilde{x}\|_\infty =
\max\bigl\{\sup\{|x(u)|:u\in K\},
\sup\{|x(s,t)| : s,t\in K \}\bigr\} = {\|{x}\|_d}$$ shows that the mapping $x\mapsto \tilde{x}$ is an isometric embedding of $\bigl(C(K),{\|\cdot\|_d}\bigr)$ into a closed subspace of $\bigl(C(\tilde{K}),\|\cdot\|_\infty \bigr) $. It follows from [@DS V.8.6] that every extreme point of the unit ball of $\bigl(C(K),{\|\cdot\|_d})' \cong
\bigl({\EuScript{M}(K)},\|\cdot\|_0\bigr)$ is either $\pm\delta_u$ for some $u\in K$, or $\delta_s-\delta_t$ for some $s,t\in K$.
Step 2. $\pm\,\delta_t$ are extreme points: Suppose that $$\delta_t = a \mu_1 + b\mu_2 \,,$$ where $\mu_1, \mu_2\in {\EuScript{M}(K)}$, $\|\mu_1\|_0 = \|\mu_2\|_0=1$, $a,b\in (0,1)$ and $a+b=1$. Applying $\delta_t$ to the function $1_K$, we have $$a\mu_1(K) + b\mu_2(K) = 1\,.$$ On the other hand, ${\|{1_K}\|_d} = 1$ implies that $|\mu_i(K)| \le 1$ for $i=1,2$. Therefore $\mu_1(K)= \mu_2(K) = 1$ and it follows from $\|\mu_i\|_0 =
\frac{1}{2} \bigl(\|\mu_i\|_1+ |\mu_i(K)|\bigr)
=1$ that $\|\mu_1\|_1 = \|\mu_2\|_1 = 1$. Since $\delta_t$ is an extreme point of the unit ball of ${\EuScript{M}(K)}$ for the variation norm, it follows that $\mu_1=\mu_2 = \delta_t$. Therefore $\delta_t$ is an extreme point of the unit ball of $\bigl({\EuScript{M}(K)},\|\cdot\|_0\bigr)$.
Step 3. $\delta_s-\delta_t$ is extreme for every pair of distinct points $s,t\in K$: Suppose that $$\delta_s - \delta_t = a \mu_1 + b\mu_2 \,,$$ where $\mu_1, \mu_2\in {\EuScript{M}(K)}$, $\|\mu_1\|_0 = \|\mu_2\|_0=1$, $a,b\in (0,1)$ and $a+b=1$. Without loss of generality, we may assume that $a=b= \frac{1}{2}$. As $\|\delta_s - \delta_t\|_1=2$, we have $4 \le \|\mu_1\|_1 + \|\mu_2\|_1$. On the other hand, $$\|\mu_i\|_0 = \frac{1}{2}\bigl( \|\mu_i\|_1 +
|\mu_i(K)| \bigr) = 1\quad\text{for $i=1,2$}$$ and it follows that $\|\mu_i\|_1 =2$ and $\mu_i(K) = 0$ for $i=1,2$. Therefore $\delta_s - \delta_t$, $\mu_1$ and $\mu_2$ all lie in $\bigl(C_\rho(K),\|\cdot\|_\rho\bigr)'$, the space of regular Borel signed measures on $K$ that are zero on $K$. Furthermore, these measures are all unit vectors in this space, since the variation norm is exactly twice the dual norm in $\bigl(C_\rho(K),\|\cdot\|_\rho\bigr)'$. It follows from the result of Cabello-Sanchez (Theorem \[p:Cabello\] above) that $\mu_1 = \mu_2 = \delta_s - \delta_t$. Therefore $\delta_s - \delta_t$ is an extreme point of the unit ball of $\bigl({\EuScript{M}(K)},\|\cdot\|_0\bigr)$.
We can now describe the extreme points of the unit ball of ${\EuScript{P}_o(^{n}{C(K)})}$ for the supremum norm. Recall that, when $n$ is odd, the supremum and regular norms coincide. Thus, by Propositions \[p:isomorphism\], \[p:extremeRegular\] and Theorem \[p:extremeMK\] we have the following result.
\[p:extremeSup\] Let $K$ be a compact, Hausdorff space.
- If $n$ is odd, then $P\in {\EuScript{P}_o(^{n}{C(K)})}$ is an extreme point of the closed unit ball of the space $\bigl({\EuScript{P\!}_o(^{n}{C(K)})},\|\cdot\|_\infty)$ if and only if $P= \pm \delta_t^n$, for some $t\in K$, where $$\delta_t^n(x)= x(t)^n\,.$$
- If $n$ is even, then $P\in {\EuScript{P}_o(^{n}{C(K)})}$ is an extreme point of the closed unit ball of the space $\bigl({\EuScript{P\!}_o(^{n}{C(K)})},\|\cdot\|_\infty)$ if and only if $P$ is one of the following:
- $P = \pm \delta_t^n$ for some $t\in K$, where $\delta_t^n(x)= x(t)^n$;
- $P= \delta_s^n - \delta_t^n$, where $s,t$ are distinct points in $K$, and $$(\delta_s^n - \delta_t^n)(x) = x(s)^n - x(t)^n \,.$$
Suppose that the compact, Hausdorff space $K$ has just two points, $\alpha, \beta$. Then the vector lattice ${C(K)}$ can be identified with ${\ensuremath{\mathbb{R}}}^2$, where $(x_1,x_2)\in {\ensuremath{\mathbb{R}}}^2$ corresponds to the function $\alpha \mapsto x_1$, $\beta\mapsto x_2$. The supremum norm on ${C(K)}$ is identified with the supremum norm on ${\ensuremath{\mathbb{R}}}^2$. The orthogonally additive $n$-homogeneous polynomials on ${\ensuremath{\mathbb{R}}}^2$ have the form $P(x) = a_1 x_1^n + a_2 x_2^n$. The regular and supremum norms are $$\begin{aligned}
\|P\|_r &= |a_1| + |a_2| \,,\\
\|P\|_\infty & =
\begin{cases}
|a_1| + |a_2|\,, \quad\text{if $n$ is odd,}\\
\max\{|a_1|, |a_2|, |a_1+a_2| \} \,,
\quad\text{if $n$ is even.}
\end{cases}
\end{aligned}$$ The diagrams below show the unit balls for both norms.
plot [-x+1]{}; coordinates [(1,0)]{}; coordinates [(1.2,0.2)]{} node; plot [x+1]{}; coordinates [(0.3,1)]{} node; coordinates [(0,1)]{}; plot [-x-1]{}; coordinates [(0,-1)]{}; coordinates [(0.5,-1)]{} node; plot [x-1]{}; coordinates [(-1,0)]{}; coordinates [(-1.4,0.2)]{} node;
plot [-x+1]{}; coordinates [(1,0)]{}; coordinates [(1.2,0.3)]{} node[[$x_1^n$]{}]{}; plot [x+1]{}; coordinates [(0,1)]{}; coordinates [(0.3,1.3)]{} node[[$x_2^n$]{}]{}; plot [-x-1]{}; coordinates [(0,-1)]{}; coordinates [(-0.4,-1.4)]{} node[[$-x_2^n$]{}]{}; plot [x-1]{}; coordinates [(-1,0)]{}; plot [1]{} ; coordinates [(-1,1)]{}; coordinates [(-1, 0) (-1, 1)]{}; coordinates [(1,-1)]{}; coordinates [(-1.1, 1.3)]{} node[[$-x_1^n+x_2^n$]{}]{}; plot [-1]{}; coordinates [(1, -1) (1, 0)]{} ; coordinates [(0.9,-1.4)]{} node[[$x_1^n-x_2^n$]{}]{}; coordinates [(-1.4,0.3)]{} node[[$-x_1^n$]{}]{};
[Unit ball of $\bigl({\EuScript{P}_o(^{n}{C(K)})},\|\cdot\|_r\bigr)$ for any $n$\
and $({\EuScript{P\!}_o(^{n}{{C(K)}})},\|\cdot\|_\infty)$ for $n$ odd.]{}
[Unit ball of $\bigl({\EuScript{P}_o(^{n}{C(K)})},\|\cdot\|_\infty\bigr)$\
for $n$ even.]{}
The isometries of $(C(K), {\|\cdot\|_d})$
-----------------------------------------
We would like next to determine the isometries of the spaces ${\EuScript{P}_o(^{n}{C(K)})}$, both for the regular and the supremum norms. Our results show that this reduces to the problem of finding the isometries between the spaces ${\EuScript{M}(K)}$ for the variation norm and the equivalent norm $\|\cdot\|_0$.
The Banach-Stone theorem [@DS V.8.8] uses the classification of the extreme points of the space of regular Borel signed measures to determine the isometries of $C(K)$ spaces with the supremum norm. We recall the statement of this theorem: if $T$ is an isometric isomorphism between $C(K)$ and $C(L)$, then there exists a homeomorhism $\varphi\colon L\to K$ and a function $\alpha\in C(L)$ with values $\pm 1$, such that $$\label{e:BS}
\bigl(Tx\bigr)(s) = \alpha(s) x(\varphi(s))$$ for all $x\in {C(K)}$, $s\in L$. We shall say that an linear bijection, $T$, from $C(K)$ to $C(L)$ is *canonical* if it has this form. In other words, $$Tx = \alpha\,x\circ\varphi \,,$$ where $\alpha$, $\varphi$ are as described above.
Consider the space $\bigl({\EuScript{M}(K)},\|\cdot\|_0\bigr) \cong \bigl(C(K),{\|\cdot\|_d}\bigr)'$. By Theorem \[p:extremeMK\], the set of extreme points of the unit ball of $({\cal M}(K),\|\cdot\|_0)$ is $\{\pm\delta_u,\delta_t-\delta_s
:u,t,s\in K, t\neq s\}$. The crucial step in showing that an isometry $T$ of from $(C(K),{\|\cdot\|_d})$ to $(C(L),{\|\cdot\|_d})$ is canonical is to establish that $T^t$, the transpose of $T$, maps each $\delta_t$ to $\pm\delta_s$ for some $s$ in $K$. This leads to the following proposition.
\[morethan2\] Let $K$ and $L$ be compact Hausdorff topological spaces and let $T\colon \bigl(C(K),{\|\cdot\|_d}\bigr)
\to \bigl(C(L),{\|\cdot\|_d}\bigr)$ be an isometric isomorphism. Let $S_L=\{t\in L: T^t(\delta_t)=\pm\delta_s,\mbox{ for some } s\in K\}$. If $S_L$ contains more than one point, then $T$ is canonical. Moreover, in addition, $\alpha$ will either take the constant value $1$ or $-1$ on $L$.
Assume that $|S_L|\ge 2$ and $S_L^c$ is non-empty. Choose $r\in S_L^c$. Then we have that $T^t(\delta_r)=\delta_u-
\delta_v$ for some $u$ and $v$ in $K$. Since $|S_L|\ge 2$, there are $t$ and $s$ in $L$ with $t\neq s$ so that $T^t(\delta_t)=\pm\delta_w$ and $T^t(\delta_s)=\pm\delta_p$ for some $w$ and $p$ in $K$ with $w\neq p$. We now claim that $\{w,p\}\not=\{u,v\}$ and so there $t$ in $L$ so that $T^t(\delta_t)=\pm\delta_w$ with $w\not=u,v$. Without loss of generality suppose that $w=u$ and $p=v$. Then we have $T^t(\delta_t)=\pm\delta_u$ and $T^t(
\delta_s)=\pm\delta_v$. Hence $$T^t(\delta_t-\delta_s)=\pm\delta_u\mp\delta_v$$ and therefore $(T^t)^{-1}(\delta_u-\delta_v)=\pm (\delta_t-\delta_s)$. Since $(T^t)^{-1}$ is a bijection we have $\delta_r=\pm (\delta_t-\delta_s)$ which is impossible. Let $t$ in $L$ be such that $T^t(\delta_t)=\pm\delta_w$ with $w\not=u,v$. Then $\delta_r-\delta_t$ is an extreme point of the unit ball of $({\cal M}(L),\|\cdot\|_0)$. However, $$T^t(\delta_r-\delta_t)=\delta_u-\delta_v\pm\delta_w\,.$$ Since $w\not= u,v$, $\delta_u-\delta_v\pm\delta_s$, is not an extreme point of the unit ball of $({\cal M}(K),\|\cdot\|_0)$. This is a contradiction. Hence, if $|S_L|\ge 2$, then $S_L=L$.
Note that $T\colon (C(K),\|\cdot\|_\infty) \to (C(L),\|\cdot
\|_\infty)$ is an isomorphism since the norms $\|\cdot\|_\infty$ and ${\|\cdot\|_d}$ are equivalent. Further, since $S_L=L$, we have that for every $t$ in $L$ there is $s$ in $K$ such that $T^t(\delta_t)=\pm\delta_s$. Hence $T^t$ maps extreme points of the unit ball of $({\cal M}(L),\|\cdot\|_1)$ to the extreme points of the unit ball of $({\cal M}(K),\|\cdot\|_1)$ in one to one manner. Hence $T^t(B_{{\cal M}(L)})\subseteq B_{{\cal M}(K)}$ and $(T^t)^{-1}(B_{{\cal M}(K)})
\subseteq B_{{\cal M}(L)}$. This gives us that $T: (C(K),\|\cdot\|_\infty) \to (C(L
),\|\cdot
\|_\infty)$ is an isometric isomorphism. Hence we can now apply Banach-Stone theorem to find a homeomorphism $\varphi$ from $L$ to $K$ and a function $\alpha\in C(K)$ with $\alpha(t)=\pm 1$ for all $t\in K$ such that
$$T(x)=\alpha\, x\circ\varphi.$$
Now let us see that $\alpha$ is constant on $L$. To see this suppose that $$S_L^+=\{t\in L:T^t(\delta_t)=\delta_s \hbox{ for some } s\in K\}$$ and $$S_L^-=\{t\in L:T^t(\delta_t)=-\delta_s \hbox{ for some } s\in K\}$$ are both non empty. Choose $t$ in $S_L^+$ and $r$ in $S_L^-$. Suppose that $T^t(\delta_t)=\delta_u$ and that $T^t(\delta_r)=-\delta_v$. Then $\delta_t-
\delta_r$ is an extreme point of the unit ball of $({\cal M}(L),\|\cdot\|_0)$ yet $T^t(\delta_t-\delta_r)=\delta_u+\delta_v$ is not extreme point of the unit ball of $({\cal M}(K),\|\cdot\|_0)$. The result now follows and we get that $$T(x)=\pm\, x\circ\varphi.$$
Let us now consider the case when $|S_L|=1$ and show that we can construct a non canonical isometry in this case. To help understand this result, we first consider the following example.
Let $K=\{a,b\}$ and $L=\{\alpha,\beta\}$. We observe that we can identify both $(C(K),{\|\cdot\|_d})$ and $(C(L),{\|\cdot\|_d})$ with $\mathbb{R}^2$. Let $x$ in $C(K)$ and set $x_1=x(a)$ and $x_2=x(b)$. Then $(x_1,x_1) \in \mathbb R^2$ and the norm of $(x_1,x_1)$ is given by $${\|{(x_1,x_2)}\|_d}=\max\{|x_1|,|x_2|,|x_1-x_2|\}.$$
Now define $T: (\mathbb{R}^2,{\|\cdot\|_d}) \to (\mathbb{R}^2,{\|\cdot\|_d})$ by
$$T(x_1,x_2)=(x_1,x_1-x_2)$$
Clearly, $T$ is a continuous linear bijection. We can also show that $$\begin{aligned}
T^t(\delta_{\alpha})=&\delta_a,\\
T^t(\delta_{\beta})=&\delta_a-\delta_b.
\end{aligned}$$
We have that $${\|{T(x_1,x_2)}\|_d}=\max\{|x_1|,|x_1-x_2|,|x_2|\}=
{\|{(x_1,x_2)}\|_d}$$ and hence $T$ is an isometry. However, $T$ is not canonical since
$$\begin{aligned}
(Tx)(\alpha) & = x(a)\,,\\
(Tx)(\beta) & = x(a)-x(b)\,.
\end{aligned}$$
Guided by Proposition \[morethan2\] and the above example, we now have the following result.
Let $K$ and $L$ be compact Hausdorff topological spaces.
1. Suppose that $K$ and $L$ do not contain isolated points. Then every isometric isomorphism $T$ from $(C(K),{\|\cdot\|_d})$ onto $(C(L),{\|\cdot\|_d})$ has the form $$T(x)=\pm\, x\circ\varphi.$$ for some homeomorphism $\varphi\colon L\to K$.
2. Suppose that either $K$ or $L$ contains an isolated point. Let $T:(C(K),{\|\cdot\|_d}) \to (C(L),{\|\cdot\|_d})$ be an isometric isomorphism. Then $T$ is one of the following types.
1. $$T(x)=\pm\, x\circ\varphi.$$ for some homeomorphism $\varphi\colon L\to K$.
2. There exist $p$ in $K$ and $t$ in $L$ and a homeomorphism $\varphi\colon L\setminus\{t\}\to K\setminus\{p\}
$ such that $T=\pm T_1$, where $$\begin{aligned}
(T_1x)(t)&=x(p)\\
(T_1 x)(s)&=x(p)-x(\varphi(s))
\qquad\text{for $s\neq t$.}
\end{aligned}$$
\(a) Note that $L=S_L\cup S^c_L$. We claim that, if $|S_L|=1$, then $L$ contains an isolated point. Suppose that $S_L=\{t\} $ and, without loss of generality, $T^t\delta_t= \delta_s$. Then $$(T 1_K) (t) =\delta_t(T 1_K) = (T^t\delta_t)(1_K)=
\delta_s(1_K) =1 \,.$$ For any $r\in S_L^c$, a similar calculation shows that $(T 1_K)(r) = 0$. As $S_L= (T 1_K)^{-1}(1)$ and $S_L^c= (T 1_K)^{-1}(0)$ and $T1_K$ is continuous, it follows that $S_L$ and $S_L^c$ are disjoint closed sets. Therefore $S_L=\{t\}$ is an isolated point of $L$. Therefore, if $L$ does not contain isolated points then $|S_L|\geq 2$ and Proposition \[morethan2\] gives us that $T$ is canonical.
\(b) We only need to consider the case $|S_L|=1$ as otherwise Proposition \[morethan2\] gives us that $T$ is canonical. Suppose $L$ contains an isolated point $t$ and $K$ an isolated point $p$. For each $x$ in $C(K)$ the function $Tx$ as defined in (b) is continuous and the mapping $x\to Tx$ is easily seen to be an isometry.
Let us see that if $T$ is not canonical then this is the form that an isometry can take. By definition and the fact that $T$ is invertible we have that $|S_K|=1$, where $S_K$ is the set of points $q$ in $K$ for which $T^t (\delta_t) = \pm \delta_q$ for some $t\in L$. Let $S_L=\{t\}$ and $S_K=\{p\}$. Then $T^t(\delta_t)=\pm\delta_p$. Let us suppose that $T^t(\delta_t)=\delta_p$. We claim that for each $s$ in $S_L^c$ we have $T^t(\delta_s)=\delta_p-\delta_q$ for some $q$ in $K\setminus\{p
\}$. Otherwise we have that $\delta_t-\delta_s$ is extreme but $T^t(\delta_t-
\delta_s)=\delta_p-\delta_u+\delta_v$ is not. The mapping $T^t(\delta_s)=
\delta_p-\delta_q$ now induces a bijection $\varphi\colon L\setminus\{t\}\to K
\setminus\{p\}$ so that $T^t(\delta_s)=\delta_p-\delta_{\varphi(s)}$. Since the mapping $L\setminus\{p\}\to ({\cal M}(K),\sigma({\cal M}(K),{\cal C}(K)))$, $s\mapsto \delta_p-\delta_{\varphi(s)}$, is continuous, $\varphi$ will be continuous. As $\varphi$ is a continuous bijection from the compact space $L\setminus \{t\}$ to the Hausdorff space $K\setminus \{p\}$ it is a homeomorphism. Rewriting $s\mapsto \delta_p -
\delta_{\varphi(s)}$ in terms of $x$, we see that $(Tx)(s)=x(p)-x(\varphi(s))$. When $T^t(\delta_t)= -\delta_s$, we obtain $(Tx)(s)=x(\varphi(s))-x(p)$.
Our characterisation of the isometries of $(C(L),{\|\cdot\|_d})$ onto $(C(K),{\|\cdot\|_d})$ allows us to construct isometries of $({\EuScript{P\!}_o(^{n}{C(K)})}, \|\cdot\|_\infty)$ onto $({\EuScript{P\!}_o(^{n}{C(L)})},\|\cdot\|_\infty)$. Given a homeomorphism $\varphi\colon K\to L$ we use $C_\varphi$ to denote the composition operator $C_\varphi\colon C(L)\to C(K)$ defined by $C_\varphi(f)=f\circ\varphi$ for each $f$ in $C(L)$. The transpose of the canonical isometry of $(C(K),{\|\cdot\|_d})$ onto $(C(L),{\|\cdot\|_d})$ determined by $\varphi$ now gives rise to the isometry $T\colon ({\EuScript{P\!}_o(^{n}{C(K)})},\|\cdot\|_\infty)
\to ({\EuScript{P\!}_o(^{n}{C(L),\|\cdot\|_\infty})})$ given by $T(P)=P\circ C_\varphi$.
To understand the isometries from $({\EuScript{P\!}_o(^{n}{C(K)})}, \|\cdot\|_\infty)$ to $({\EuScript{P\!}_o(^{n}{C(L)})},\|\cdot\|_\infty)$ induced by non canonical isometries of $(C(L),{\|\cdot\|_d})$ onto $(C(K),{\|\cdot\|_d})$ we note that if $K$ and $L$ have isolated points $t$ and $p$ respectively then we have that $(C(K),\|\cdot
\|_\infty)$ is isometrically isomorphic to $(C(\{t\}),\|\cdot\|_\infty)\oplus_
\infty(C(K\setminus\{t\}),\|\cdot\|_\infty)$ while $(C(L),\|\cdot\|_\infty)$ is isometrically isomorphic to $(C(\{p\}),\|\cdot\|_\infty)\oplus_\infty(C(L
\setminus\{p\}),\|\cdot\|_\infty)$. Hence, if $P$ is an $n$-homogeneous orthogonally additive polynomial on $(C(K),\|\cdot\|_\infty\|)$ then we can write $P$ as $P=\lambda \delta_t^n+P_2$ where $P_2=P|_{C(K\setminus\{t\})}$. It follows that the transpose of each non canonical isometry from $(C(K),{\|\cdot\|_d})$ onto $(C(L),{\|\cdot\|_d})$ gives an isometry from $({\EuScript{P\!}_o(^{n}{C(K)})},\|\cdot
\|_\infty)$ onto $({\EuScript{P\!}_o(^{n}{C(L)})},\|\cdot\|_\infty)$ of the form $$T(P)=P(1)\,\delta_p^n-P_2\circ C_\varphi$$ where $\varphi$ is a homeomorphism of $K\setminus
\{t\}$ to $L\setminus\{p\}$.
In a similar manner, we can construct canonical isometries from $({\EuScript{P\!}_o(^{n}{C(K)})},\|\cdot\|_r)$ onto $({\EuScript{P\!}_o(^{n}{C(L)})},\|\cdot\|_r)$.
Exposed points in ${\EuScript{P}_o(^{n}{C(K)})}$
=================================================
In this section we shall characterise the weak${}^*$ exposed and weak${}^*$ strongly exposed point of the unit ball of $(C(K),{\|\cdot\|_d})'$. We have an upper bound for this set. We know that it is contained in the set of extreme points of the unit ball of $(C(K),
{\|\cdot\|_d})'\cong ({\cal M}(K),\|\cdot\|_0)$ and that the set of extreme points of this set is equal to $\{\pm\delta_p,\delta_t-\delta_s:p, t,s\in K, t\neq s\}$.
Let us begin with some definitions.
Let $E$ be a Banach space. A point $x$ in the closed unit ball of $E$ is said to be an *exposed point* if there exists $\varphi \in E'$ with $\|\varphi\|=
1$ such that $$\varphi(x) = 1 \text{ and } \varphi (y) < 1 \text{ for } y \in
\overline{B}_E \backslash \{x\}.$$ If this is the case then we say that $\varphi$ *exposes* $x$.
We say that $x$ is a *strongly exposed point* of the closed unit ball of $E$ if there exists $\varphi \in E'$ such that $$\varphi(x) = 1$$ and whenever $(x_n)_n$ is a sequence in $\overline{B}_E$ with $
\lim_{n \rightarrow \infty} \varphi(x_n)=1
$ then $(x_n)_n$ converges to $x$ in norm. We will say that $\varphi$ *strongly exposes* $x$.
If $E=F'$ is a dual Banach space and the point $x\in E$ is exposed (respectively, strongly exposed) by $\varphi$ in $F$ we say that $x$ is a *weak${}^*$ exposed* (respectively, *weak${}^*$ strongly exposed*) point of $E$ and that $\varphi$ *weak${}^*$ exposes* (respectively, *weak$^*$-strongly exposes*) the unit ball of $E$ at $x$.
We also observe that if each $\delta_t$, $t\in K$ and each $\delta_t-\delta_s$, $t,s\in K$ with $t\not=s$ are of norm $1$ in $(C(K),{\|\cdot\|_d})'$. Hence, if $\delta_t$ is exposed by $x$ then we must have ${\rm diam}(x)<1$. Conversely, if $\delta_t-\delta_s$ is exposed by $x$ then we must have $\|x\|_\infty<1$.
We note that if $K$ is a compact Hausdorff topological space then a net $(t_\alpha)_\alpha$ converges to $t$ in $K$ if and only if $y(t_\alpha)$ converges to $y(t)$ for every $y$ in $C(K)$.
Gâteaux differentiability of the norm
-------------------------------------
We start with a characterisation of Gâteaux differentiability of the norm on $(C(K),{\|\cdot\|_d})$.
\[gatsingle\] Let $K$ be a compact Hausdorff topological space. Let $t\in K$, $x\in C(K)$ with ${\|{x}\|_d}=1$. Then the following are equivalent
1. The norm of $(C(K),{\|\cdot\|_d})$ is Gâteaux differentiable at $x$ with derivative $\delta_t$.
2. 1. ${\|{x}\|_d}=x(t)=1$ and ${\rm diam}(x)<1$.
2. If $(t_n)_n$ is a sequence of points in $K$ such that $\lim_{n\to
\infty}x(t_n)=1$ then $(t_n)_n$ has a subnet, $(t_\alpha)_\alpha$ such that $(t_
\alpha)_\alpha$ converges to $t$.
3. $t$ is the unique point in $K$ with $x(t)=1$ and ${\rm diam}(x)<1$.
First observe that [Š]{}mul’yan [@Smulyan1; @Smulyan2] (see also [@DGZ]) showed that a point $x$ in $B_{C(K)}$ weak${}^*$ exposes the unit ball of $(C(K),{\|\cdot\|_d})'$ at $\delta_t$ if and only if the norm of $C(K)$ is Gâteaux differentiable at $x$ with derivative $\delta_t$. Hence we have that (a) implies (c).
Let us see that (c) implies (b). Clearly we have that (c) implies (b) (i).
Suppose that (c) is true and that (b) part (ii) fails. Then there is a sequence $(t_n)_n$ in $K$ with $\lim_{n\to\infty}x(t_n)=x(t)=1$ but that for all subnets $(t_\alpha)_\alpha$ of $(t_n)_n$ there is $y$ in $C(K)$ such that $y(t_\alpha)\not
\to y(t)$. As $K$ is compact, we can choose a subnet $(t_\alpha)_\alpha$ of $(t_n)_n$ and $s$ in $K$ so that $\lim_{\alpha\to\infty}t_\alpha=s$. We claim that $t\not=s$. Suppose $t=s$. Then for every $y$ in ${C(K)}$ we have that $\lim_\beta
y(t_\beta)=y(t)$ contrary to what we have assumed. As $s\not=t$ and $x$ is continuous we have $x(s)=\lim_\beta x(t_\beta)=1$ which contradicts (c). Hence, we see that (c) implies (b).
Next suppose that (b) is true and that (a) is false. Then we can find $y$ in $C(K)$, $\varepsilon>0$ and a sequence of positive numbers $(\lambda_n)_n$ converging to $0$ so that $$\bigl|{\|{x+\lambda_ny}\|_d}-{\|{x}\|_d}-\lambda_n y(t)\bigr|\ge \varepsilon\lambda_n$$ for every positive integer $n$.
Note that as ${\|{x+\lambda_n y}\|_d}\ge(x+\lambda_n y)(t)$ we actually have that ${\|{x+\lambda_ny}\|_d}-{\|{x}\|_d}-\lambda_n y(t)$ is non-negative and therefore we have $${\|{x+\lambda_ny}\|_d}-{\|{x}\|_d}-\lambda_n y(t)
\ge \lambda_n\varepsilon$$ for every positive integer $n$.
Each of the functions $x+\lambda_n y$ attains its norm either at a point of the form $\delta_t$ or at a point $\delta_u-\delta_v$. As ${\rm diam}(x)<1$ choosing $n$ sufficiently large we can assume that $x+\lambda_n y$ attains its norm at a point of the first type. Hence, for each $n$ in $\mathbb{N}$, we can find $t_n$ in $K$ and $\beta_n=\pm 1$ so that $$\beta_n(x+\lambda_n y)(t_n)={\|{x+\lambda_n y}\|_d}\,.$$ Then we have $$\begin{aligned}
1&= {\|{x}\|_d} \ge \beta_n x(t_n)=\beta_n(x+\lambda_n y)(t_n)-\beta_n\lambda_n y(t_n)\\
&\ge {\|{x+\lambda_n y}\|_d}-|\lambda_n|{\|{y}\|_d}.\end{aligned}$$ As $(\lambda_n)_n$ is a null sequence we have that ${\|{x+\lambda_n y}\|_d}-
|\lambda_n|{\|{y}\|_d}$ converges to ${\|{x}\|_d}$ as $n$ tends to $\infty$. Hence we have that $\beta_nx(t_n)\to 1$. However, as ${\rm diam}(x)<1$, we have $x(t_n)>0$ for all $n$. Hence, without loss of generality, we may assume that $\beta_n=1$ for all $n$ and therefore we have $\lim_{n\to\infty}x(t_n)=1$.
Then we have that $$\begin{aligned}
\varepsilon \lambda_n\le &{\|{x+\lambda_n y}\|_d}-{\|{x}\|_d}-\lambda_n y(t)\\
= &(x+\lambda_n y)(t_n)-x(t)-\lambda_n y(t)\\
= & x(t_n)-x(t)+\lambda_n\left( y(t_n)-y(t)\right)\\
\le & \lambda_n\left(y(t_n)-y(t)\right)\\\end{aligned}$$ However this means that there is no subnet $(t_\alpha)$ of $(t_n)_n$ so that $y(t_\alpha)$ converges to $y(t)$ and so (b) (ii) is false.
We recall that a function $x\in {C(K)}$ is said to peak at a point $t\in K$ if $t$ is the unique point at which $x$ attains its maximum.
\[l:peak\] Let $K$ be a compact Hausdorff topological space and $t\in K$. Then there is $x$ in $C(K)$ which peaks at $t$ if and only if $\{t\}$ is a $G_\delta$ subset of $K$.
We first suppose that $\{t\}$ is a $G_\delta$ subset of $K$. Then we can find a sequence of open sets $(U_n)_n$ so that $\{t\}=\bigcap_
{n=1}^\infty U_n$. As $K$ is compact and Hausdorff it is completely regular. Hence, for each $n\in \mathbb{N}$ we can find a continuous function $x_n\colon
K\to [0,1]$ such that $x_n(t)=1$ and $x_n(U_n^c)=0$. Now let $x\colon K\to
[0,1]$ be defined by $$x(t)=\frac{6}{\pi^2}\sum_{n=1}^\infty \frac{1}{n^2}x_n(t).$$ Then we have $x(t)=1$ and $x(s)<1$ for $s\in K$, $s\not=t$. So $x$ peaks at $t$.
Conversely, if there is $x$ in $C(K)$ which peaks at $t$, for each $n\in \mathbb
{N}$ let $U_n=\{s\in K:x(s)>1-\frac{1}{n}$. Then $\{t\}=\bigcap_{n=1}^\infty U_n
$. As each $U_n$ is open, $\{t\}$ is a $G_\delta$ set.
The weak${}^*$ exposed points of the ball of the form $\delta_t$ are characterised by the following proposition.
\[peak\] Let $K$ be a compact Hausdorff topological space. Then $\{\pm \delta_t:t\in K\}$ is contained in the set of weak${}^*$ exposed points of the unit ball of $({\cal M}(K),\|\cdot\|_0)$ if and only if $K$ is first countable.
Just as we have characterised the weak${}^*$ exposed points of the ball of the form $\delta_t$ we now characterise weak${}^*$ exposed points of the form $\delta_t-\delta_s$. Replacing $\delta_t$ with $\delta_t-\delta_s$ in Theorem \[gatsingle\] we obtain the following result.
Let $K$ be a compact Hausdorff topological space. Let $t,s\in K$, $x\in C(K)$ with ${\|{x}\|_d}=1$. Then the following are equivalent
1. The norm of $(C(K),{\|\cdot\|_d})$ is Gâteaux differentiable at $x$ with derivative $\delta_t-\delta_s$.
2. 1. ${\|{x}\|_d}=x(t)-x(s)=1$ and $\|x\|_\infty<1$.
2. If $(t_n)_n$ and $(s_n)_n$ are sequences of points in $K$ such that $\lim_{n\to \infty}x(t_n)-x(s_n)=1$ then $(t_n)_n$ and $(s_n)$ have subnets $(t_\alpha)_\alpha$ and $(s_\alpha)_\alpha$ which converge to $t$ and $s$ respectively.
3. $t,s$ is the unique pair of points in $K$ with $x(t)-x(s)=1$ and $\|x\|_\infty<1$.
As the proof of the following lemma is similar to that of Lemma \[l:peak\] we omit it.
Let $K$ be a compact Hausdorff topological space and $t,s\in K$ with $t\not=s$. Then there is $x$ in $C(K)$ such that $x(t)=\frac{1}{2}$, $x(s)=-\frac{1}{2}$ and $-\frac{1}{2}<x(u)<\frac{1}{2}$ for $u\in K\setminus\{t,s\}$ if and only if $\{t\}$ and $\{s\}$ are $G_\delta$ sets.
The weak${}^*$ exposed points of the ball of the form $\delta_t-\delta_s$ are now characterised by the following proposition.
\[2peak\] Let $K$ be a compact Hausdorff topological space and $n$ be an even integer. Then $\{\delta_t-\delta_s:t,s
\in K, t\neq s\}$ is contained in the set of weak${}^*$ exposed points of the unit ball of $({\cal M}(K),\|\cdot\|_0)$ if and only if $K$ is first countable.
Propositions \[peak\] and \[2peak\] can be rephrased in terms of spaces of orthogonally additive polynomials. Since we have canonically identified the space ${\EuScript{P\!}_o(^{n}{{C(K)}})}$ with the space ${\EuScript{M}(K)}$, we may transfer the weak$^*$ topology on ${\EuScript{M}(K)}= {C(K)}'$ to the space ${\EuScript{P\!}_o(^{n}{{C(K)}})}$. References to the weak$^*$ topology on ${\EuScript{P\!}_o(^{n}{{C(K)}})}$ should be understood in this sense. It is easy to see that this is the topology of pointwise convergence on ${\EuScript{P\!}_o(^{n}{{C(K)}})}$.
\[peakpoly\] Let $K$ be a compact Hausdorff topological space and $n$ be an even integer. Then $\{\pm\delta_p^n,\delta_t^n-\delta_s^n:p, t,s\in K, t\neq s\}$ is equal to the set of weak$^*$ exposed points of the unit ball of $({\EuScript{P\!}_o(^{n}{{C(K)}})},\|\cdot\|_\infty)$ if and only if $K$ is first countable.
Fréchet differentibility of the norm
------------------------------------
We now characterise Fréchet differentiability of the norm on $(C(K),{\|\cdot\|_d})$.
\[Frechet\] Let $K$ be a compact Hausdorff topological space. Let $t\in K$, $x\in C(K)$ with ${\|{x}\|_d}=1$. Then the following are equivalent.
1. The norm of $(C(K),{\|\cdot\|_d})$ is Fréchet differentiable at $x$ with derivative $\delta_t$.
2. 1. ${\|{x}\|_d}=x(t)=1$ and ${\rm diam}(x)<1$.
2. If $(t_n)_n$ is a sequence of points in $K$ such that $\lim_{n\to
\infty}x(t_n)=1$ then $(t_n)_n$ is eventually equal to $t$.
3. $x$ weak${}^*$ strongly exposes the unit ball of $(C(K),{\|\cdot\|_d})'$ at $\delta_t$.
First observe that [Š]{}mul’yan [@Smulyan1; @Smulyan2] (see also [@DGZ]) showed that a point $x$ in $B_{C(K)}$ weak${}^*$ strongly exposes the unit ball of $(C(K),
{\|\cdot\|_d})'$ at $\delta_t$ if and only if the norm of $C(K)$ is Fréchet differentiable at $x$ with derivative $\delta_t$. Thus (a) and (c) are equivalent.
If the norm of $(C(K),{\|\cdot\|_d})$ is Fréchet differentiable at $x$ with derivative $\delta_t$ then it is Gâteaux differentiable at $x$ with derivative $\delta_t$. Theorem \[gatsingle\] now implies that (b) (i) holds.
Suppose that (c) is true. Then $x$ in $B_{C(K)}$ weak${}^*$ strongly exposes the unit ball of $(C(K),{\|\cdot\|_d})'$ at $\delta_t$. If $\lim_{n\to \infty}x(t_n)
=1$ then $\lim_{n\to\infty}\delta_{t_n}(x)=\delta_t(x)=1$. As $f$ weak${}^*$-strongly exposes the unit ball of $(C(K),{\|\cdot\|_d})'$ at $\delta_t$ we have that $\lim_{n\to\infty}\delta_{t_n}=\delta_t$ in norm. However, as $\|\delta_u-\delta_v\|_0=1$ whenever $u\not= v$ we see that only way we can have $(\delta_{t_n})_n$ converge to $\delta_t$ is that the sequence $(t_n)_n$ is eventually equal to $t$.
The implication (b) implies (a) is similar to the corresponding part of the proof of Theorem \[gatsingle\] where instead of using the fact that $(t_n)_n$ has a subsequence that converges to $t$ we use the fact that $(t_n)_n$ has a subsequence so that it is eventually equal to $t$.
Let $K$ be a compact Hausdorff topological space and $t\in K$. Then $\delta_t$ is a weak${}^*$ strongly exposed point of the unit ball of $(C(K),{\|\cdot\|_d})'$ if and only if $t$ is an isolated point of $K$.
If $t$ is an isolated point of $K$ then the function given by $$x(s)=\begin{cases} 1, & s=t\\
1/2 &\hbox{otherwise}\\
\end{cases}$$ is continuous on $K$. Moreover, if $x(t_n)\to 1$ then $(t_n)_n$ is eventually equal to $t$.
Conversely, if $t$ is not an isolated point of $K$. Choose a sequence of points $(t_n)_n$ with $t_n\not=t$, all $n$, so that $t_n$ converges to $t$. Let $x$ be any function in $C(K)$ with ${\|{x}\|_d}=1$ and $x(t)=1$. Then we have that $x(t_n)\to x(t)=1$. However, as $(t_n)_n$ is not eventually equal to $t$ we see that condition (b) (ii) of Theorem \[Frechet\] is not satisfied and therefore no $x$ in $C(K)$ with ${\|{x}\|_d}=1$ can expose the unit ball $(C(K),{\|\cdot\|_d})'$ at $\delta_t$.
Let $K$ be a compact Hausdorff topological space. Let $t,s\in K$, $x\in C(K)$ with ${\|{x}\|_d}=1$. Then the following are equivalent
1. The norm of $(C(K),{\|\cdot\|_d})$ is Fréchet differentiable to $x$ with differential $\delta_t-\delta_s$.
2. 1. ${\|{x}\|_d}=x(t)-x(s)=1$ and $\|x\|_\infty<1$.
2. If $(t_n)_n$ and $(s_n)_n$ are sequences of points in $K$ such that $\lim_{n\to \infty}x(t_n)-x(s_n)=1$ then $(t_{n})_n$ and $(s_{n})_n$ are eventually the constant sequences $t$ and $s$ respectively.
3. $x$ weak${}^*$ strongly exposes the unit ball of $(C(K),{\|\cdot\|_d})'$ at $\delta_t-\delta_s$.
The proof is similar to Theorem \[Frechet\] and therefore omitted.
Let $K$ be a compact Hausdorff topological space and $t,s\in K$ with $t\not=s$. Then $\delta_t-\delta_s$ is a weak${}^*$ strongly exposed point of the unit ball of $(C(K),{\|\cdot\|_d})'$ if and only if $t$ and $s$ are isolated points of $K$.
We can rephrase these results in terms of spaces of orthogonally additive polynomials as follows.
Let $K$ be a compact Hausdorff topological space, let $n$ be an even integer and let $s,t$ be distinct points in $K$.
- $\delta_t^n$ is a weak${}^*$ strongly exposed point of the unit ball of $({\EuScript{P\!}_o(^{n}{{C(K)}})},\|\cdot\|_\infty)$ if and only if $t$ is an isolated point of $K$.
- $\delta_t^n-\delta_s^n$ ($s\neq t$) is a weak${}^*$ strongly exposed point of the unit ball of $({\EuScript{P\!}_o(^{n}{{C(K)}})},\|\cdot\|_\infty)$ if and only if $t$ and $s$ are isolated points of $K$.
In particular, we see that if $K$ has no isolated points, then the unit ball of ${\EuScript{P\!}_o(^{n}{{C(K)}})}$ does not contain any weak$^*$ strongly exposed points.
**Acknowledgements**
We thank Dirk Werner and Tony Wickstead for helpful discussions.
Christopher Boyd, School of Mathematics & Statistics, University College Dublin, Belfield, Dublin 4, Ireland.\
e-mail: [email protected]
Raymond A. Ryan, School of Mathematics, Statistics and Applied Mathematics, National University of Ireland Galway, Ireland.\
e-mail: [email protected]
Nina Snigireva, School of Mathematics, Statistics and Applied Mathematics, National University of Ireland Galway, Ireland.\
e-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A detailed study is presented of the decaying solar active region NOAA 10103 observed with the Coronal Diagnostic Spectrometer (CDS), the Michelson Doppler Imager (MDI) and the Extreme-ultraviolet Imaging Telescope (EIT) onboard the [*Solar and Heliospheric Observatory (SOHO)*]{}. Electron density maps formed using <span style="font-variant:small-caps;">Si x</span>(356.03 Å/347.41 Å) show that the density varies from $\sim$10$^{10}$ cm$^{-3}$ in the active region core, to $\sim$7$\times$10$^{8}$ cm$^{-3}$ at the region boundaries. Over the five days of observations, the average electron density fell by $\sim$30 per cent. Temperature maps formed using <span style="font-variant:small-caps;">Fe xvi</span>(335.41 Å)/<span style="font-variant:small-caps;">Fe xiv</span>(334.18 Å) show electron temperatures of $\sim$2.34$\times$10$^{6}$ K in the active region core, and $\sim$2.10$\times$10$^{6}$ K at the region boundaries. Similarly to the electron density, there was a small decrease in the average electron temperature over the five day period. The radiative, conductive, and mass flow losses were calculated and used to determine the resultant heating rate ($P_{H}$). Radiative losses were found to dominate the active region cooling process. As the region decayed, the heating rate decreased by almost a factor of five between the first and last day of observations. The heating rate was then compared to the total unsigned magnetic flux ($\Phi_{tot} = \int dA |B_z|$), yielding a power-law of the form $P_{H}\sim\Phi_{tot}^{0.81\pm0.32}$. This result suggests that waves rather than nanoflares may be the dominant heating mechanism in this active region.'
author:
- |
R. O. Milligan,$^{1,3}$ [^1] P. T. Gallagher,$^{2,3,4}$ M. Mathioudakis,$^1$ F. P. Keenan,$^1$ and D. S. Bloomfield,$^1$\
$^1$Department of Physics and Astronomy, Queen’s University Belfast, University Road, Belfast, BT7 1NN, Northern Ireland\
$^2$Department of Experimental Physics, University College Dublin, Belfield, Dublin 4, Ireland\
$^3$Laboratory for Astronomy and Solar Physics, NASA Goddard Space Flight Centre, Greenbelt, MD 20771, U.S.A.\
$^4$L-3 Communications GSI\
date: Released 2005 Xxxxx XX
title: Plasma Diagnostics of Active Region Evolution and Implications for Coronal Heating
---
\[firstpage\]
Sun: activity – Sun: corona – Sun: evolution – Sun: UV radiation
INTRODUCTION {#intro}
============
Since the discovery of highly ionized species of iron in the solar corona in the 1930’s, physicists have been puzzled by the high temperatures observed in the outer solar atmosphere [@edlen37]. It is widely accepted that the magnetic field plays a fundamental role in the heating process, but precise measurements of the coronal magnetic field are currently impossible. Indirect methods are therefore adopted which rely on measurable quantities such as the electron temperature, density and photospheric magnetic flux. These measurements can then be compared to theoretical predictions.
Models for coronal heating typically belong to one of two broad categories. In wave (AC) heating, the large-scale magnetic field essentially acts as a conduit for small-scale, high-frequency Alfvèn waves propagating into the corona. For constant Alfvèn wave amplitude $\langle v^{2} \rangle$, the total power dissipated in an active region is, $$\label{ac_heating}
P_H = \sqrt{\frac{\rho}{4\pi}} \langle v^2 \rangle \Phi_{tot}~~{\rm~ergs~s^{-1}},$$ where $\rho$ is the mass density, and $\Phi_{tot}$ is the total unsigned magnetic flux, $$\label{phi_tot}
\Phi_{tot} = \int dA |B_z|~~{\rm~Mx},$$ where $B_z$ is the longitudinal component of the magnetic field.
{height="16.5cm"}
In stress (DC) heating, the coronal magnetic field stores energy in the form of electric currents until it can be dissipated, e.g., by nanoflares [@park88]. The total power can be estimated by, $$\label{dc_heating}
P_H \sim |v| \Phi_{tot}^{2}~~{\rm~ergs~s^{-1}},$$ and the constant of proportionality describes the efficiency of magnetic dissipation, which might involve the random footpoint velocity, $v$ [@park83], or simply the geometry [@brow86; @fish98].
Several authors have linked the photospheric magnetic flux to EUV and X-ray line intensity. @gurm74 found that the line intensity of <span style="font-variant:small-caps;">Mg x</span> (624.94 Å) was proportional to the magnetic flux density. @schr87 related the integrated intensities of chromospheric, transition region, and coronal lines to the total magnetic flux by a power law, the index of which was dependent on the scale height. This result was later confirmed by @fish98, who showed that X-ray luminosity is highly correlated with the total unsigned magnetic flux. Van Driel-Gesztelyi et al. (2003) also showed a power-law relationship between the mean X-ray flux, temperature, and emission measure, and the mean magnetic field by studying the long-term evolution of an active region over several rotations, at times when there were no significant brightenings.
In this paper, we study the evolution of a decaying active region using the diagnostic capabilities of the Coronal Diagnostic Spectrometer (CDS; Harrison et al. 1995) and Michelson Doppler Imager (MDI; Scherrer et al. 1995) onboard [*SOHO*]{}. Using the temperatures, densities, and dimensions of the active region, the heating rate is calculated and compared to the total unsigned magnetic flux. These results can be put in the context of theoretical models. Section \[obs\] gives a brief overview of the active region, a summary of the instruments involved, and a description of the data analysis techniques. Our results are given in Section \[results\], and discussion and conclusions in Section \[conc\].
OBSERVATIONS AND DATA ANALYSIS {#obs}
==============================
NOAA 10103 was observed by CDS, EIT, and MDI for five consecutive days during 2002 September 10–14. Fig. \[eit\_mdi\] shows the general evolution of the region over that period.
{width="16.5cm"}
The Coronal Diagnostic Spectrometer {#CDS}
-----------------------------------
EUV spectra were obtained with the CDS instrument, which is a dual spectrometer that can be used to obtain images with a spatial resolution of $\sim$8 arcsec. The Normal Incidence Spectrometer (NIS), used in this study, is a stigmatic spectrometer which forms images by moving the solar image across the slit using a scan mirror. The spectral ranges of NIS (308–381 Å and 513–633 Å) include emission lines formed over a wide range of temperatures, from $\sim$10$^{4}$ K at the upper chromosphere, through the transition region, to $\sim$10$^{6}$ K at the corona. The details of the <span style="font-variant:small-caps;">az\_ddep1</span> observing sequence used in this study can be found in Table \[CDS\_table\].
[lc]{} &\
Date & 2002 September 10–14\
Region name & NOAA 10103\
Instrument & CDS/NIS1\
Wavelength range (Å) & 332–368\
Slit size (arcsec$^{2}$) & 4.064$\times$240\
Area imaged (arcsec$^{2}$) & 243.84$\times$240\
Exposure time (s) & 50\
Number of slit positions & 60\
The raw CDS data were cleaned to remove cosmic rays, and calibrated to remove the CCD readout bias and convert the data into physical units of photons cm$^{-2}$ s$^{-1}$ arcsec$^{-2}$. Due to the broadened nature of post-recovery CDS spectra, the emission lines were fitted with modified Gaussian profiles as described by @thom99. The Gaussian term was defined as, $$G(\lambda)=exp\left[-\frac{1}{2}\left(\frac{\lambda-\lambda_0}{\sigma}\right)^2\right],$$ and the wings by, $$W(\lambda)=\frac{1}{\left(\frac{\lambda-\lambda_0}{\sigma'}\right)^2+1},$$ where $\lambda$ is the wavelength, $\lambda_0$ is the central wavelength of the line, $\sigma$ is the Gaussian width, and $\sigma'$ is the FWHM.
The combined function describing the line profile can be expressed as, $$B(\lambda)=I[(1-a)G(\lambda)+a W(\lambda)],$$ where $I$ is the amplitude of the line profile and $a$ can be the amplitude of the red or the blue wing. The line flux is then given by, $$\begin{aligned}
\label{flux_eqn}
\int B(\lambda) d\lambda & = & I \sigma \left[\left(1-\frac{a_{red}}{2}
\left(1+\frac{a_{blue}}{a_{red}}\right) \right) \sqrt{2\pi}\right.\nonumber \\
& + & \left.a_{red} \left(1+\frac{a_{blue}}{a_{red}}\right) \pi
\sqrt{2\ln(2)} \right].\end{aligned}$$
These broadened Gaussian profiles were then fitted to emission lines in three wavelength intervals using the <span style="font-variant:small-caps;">xcfit</span> routine in the CDS branch of the SolarSoftWare tree (SSW; Freeland & Handy 1998; see Fig. \[line\_fits\]). Two of the intervals were centreed on each of the density sensitive <span style="font-variant:small-caps;">Si x</span> (347.41 Å) and <span style="font-variant:small-caps;">Si x</span> (356.03 Å) lines, while the third contained the temperature sensitive <span style="font-variant:small-caps;">Fe xiv</span> (334.18 Å) and <span style="font-variant:small-caps;">Fe xvi</span> (335.41 Å) pair. Due to the relatively low intensity of <span style="font-variant:small-caps;">Fe xiv</span> (334.18 Å) compared to that of the adjacent <span style="font-variant:small-caps;">Fe xvi</span> (335.41 Å) transition, an upper constraint on the width of $\sigma$ = 0.4 Å (FWHM = 0.17 Å) was placed on the <span style="font-variant:small-caps;">Fe xiv</span> line. The primary lines, their formation temperatures, transitions, and rest wavelengths are given in Table \[line\_data\].
[lclc]{} &Log T$_{e}$ &Transition &$\lambda$/Å\
<span style="font-variant:small-caps;">Si x</span> &6.1 &2s$^{2}$2p$^{2}$ $^{3}$P$_{1/2}$–2s2p$^{2}$ $^{2}$D$_{3/2}$ &347.41\
<span style="font-variant:small-caps;">Si x</span> &6.1 &2s$^{2}$2p $^{3}$P$_{3/2}$–2s2p$^{2}$ $^{2}$D$_{3/2,5/2}$ &356.03\
<span style="font-variant:small-caps;">Fe xiv</span> &6.3 &3s$^{2}$3p $^{2}$P$_{1/2}$–3s3p$^{2}$ $^{2}$D$_{3/2}$ &334.18\
<span style="font-variant:small-caps;">Fe xvi</span> &6.4 &3s $^{2}$S$_{1/2}$–3p $^{2}$P$_{3/2}$ &335.41\
<span style="font-variant:small-caps;">Fe xvii</span>&6.7 &2p$^{5}$ 3s $^{3}$P$_{1}$–2p$^{5}$ 3p $^{1}$D$_{2}$ &347.85\
The Michelson Doppler Interferometer {#MDI}
------------------------------------
Magnetic field measurements were taken by the MDI instrument, which images the Sun on a 1024 $\times$ 1024 pixel CCD camera through a series of increasingly narrow filters. The final elements, a pair of tunable Michelson interferometers, enable MDI to record filtergrams with a FWHM bandwidth of 94 mÅ. Several times each day, polarizers are inserted to measure the line-of-sight magnetic field. In this paper, 5-minute-averaged magnetograms of the full disk were used, with a 96 minute cadence and a pixel size of 2 arcsec.
@berg03 analyzed Advanced Stokes Polarimeter (ASP) and MDI magnetograms, and found that MDI underestimates the magnetic flux densities by a factor of 1.45 for values below 1200 G. For flux densities higher than 1200 G, this underestimation becomes nonlinear, with the MDI fluxes saturating at $\sim$1300 G. Values below 1200 G were therefore corrected by multiplying by 1.45, and values above 1200 G were approximately corrected by multiplying by a factor of 1.9 [@gree03].
Before calculating the total unsigned magnetic flux, two final corrections were applied to the data. The first results from the fact that the measured line-of-sight flux deviates more and more from a radial measurement as one approaches the limb. For simplicity, we assume that magnetic fields in the photosphere are predominantly radial. Then, the radial field strength becomes equal to the line-of-sight field strength times $1/\cos{\theta}$, where $\theta$ is the heliocentric distance of the region from Sun centre.
Active region areas, $A$, were calculated by counting all pixels above 500 G and multiplying by the appropriate factor to obtain the active region area in cm$^2$. This threshold value for the magnetic field was found to adequately separate active region structures from neighbouring areas of quiet-Sun and plage. As the region approached the limb, the effects of foreshortening became significant. Measured areas were therefore corrected by dividing by $\cos{\theta}$. The resulting total unsigned flux was then calculated using Equation (\[phi\_tot\]).
Modern high resolution space based instruments, such as the [*Transition Region and Coronal Explorer (TRACE)*]{}, have been used by [@asch01b] to show that coronal loops observed in different bandpasses are not necessarily cospatial. Due to the coarse resolution of CDS, we make the assumption that active regions occupy a hemispherical volume, $V = 2/3~\pi^{-1/2}A^{3/2}$, with a mean loop length of $L = \sqrt{ \pi A}$.
![The cosine-corrected active region area, total magnetic flux density, and total magnetic flux for September 10–14.[]{data-label="mag_plot"}](f3.eps){width="8.0cm"}
RESULTS
=======
Morphology and Magnetic Field
-----------------------------
The top row of Fig. \[eit\_mdi\] shows a series of 360 $\times$360 arcsec EIT images obtained in the 195 Å bandpass. The first three images show a number of loops to the south and north of the region, which are not visible from 2002 September 13 onwards. On September 10, the MDI magnetogram shows a simple bipolar $\beta$ region, which is then classified as a $\beta\gamma$ on the following day. The region subsequently decreased in both size and complexity as it approached the west limb. The overall decay of the active region is clearest in the MDI continuum images in the bottom row of Fig. \[eit\_mdi\].
Table \[res\_tab\] and Fig. \[mag\_plot\] show the decay of the region in terms of the cosine-corrected area and magnetic flux for September 10–14. The region was observed to have an initial area of 1.26$\times$10$^{19}$ cm$^2$, which fell to 6.54$\times$10$^{18}$ cm$^2$ by September 14. The total unsigned magnetic flux density also shows a similar trend, falling from close to 4.80$\times$10$^{5}$ Mx cm$^2$ to 1.40$\times$10$^{5}$ Mx cm$^2$. The product of the region area times the total magnetic flux density is then given in the bottom panel of Fig. \[mag\_plot\]. As the region decays, the total unsigned magnetic flux falls off by a factor of 5–6 over the five days from September 10 to 14.
Electron Densities {#den_map}
------------------
Electron density maps were generated using the <span style="font-variant:small-caps;">Si x</span> (356.03 Å/347.41 Å) ratio in conjunction with theoretical data from the <span style="font-variant:small-caps;">chianti</span> v4.2 atomic database [@dere97]. Fig. \[chianti\] shows a plot of the density sensitive <span style="font-variant:small-caps;">Si x</span> (356.03 Å/347.40 Å) ratio together with the expected range of densities in an active region of this size and class.
The density maps, presented in the second row of Fig. \[maps\_fig\], show values of $\sim$10$^{10}$ cm$^{-3}$ in the active region core, and $\sim$7$\times$10$^{8}$ cm$^{-3}$ in the region boundaries. These are in good agreement with previous active region density measurements (Gallagher et al. 2001; Warren & Winebarger 2003). Densities in the core were observed to fall from $\sim$10$^{10}$ on September 10 to 3.9$\times$10$^{9}$ cm$^{-3}$ on September 14, excluding the high density limit ($\sim$2.5$\times$10$^{10}$ cm$^{-3}$) that was reached during a C-class flare on September 13. This behaviour is more evident in the top panel of Fig. \[ne\_te\_pe\_plot\], which clearly shows that the average electron density changed by $\sim$30 per cent over the five days. Subsequently, plasma that was found to reach the high density limit was excluded from any further calculations. The average electron densities are listed in Table \[res\_tab\].
![Theoretical line ratio as a function of density from <span style="font-variant:small-caps;">chianti</span> for <span style="font-variant:small-caps;">Si x</span> (356.03 Å/347.41 Å). The vertical dotted lines denote the typical range of densities expected to be found in an active region.[]{data-label="chianti"}](f4.eps){height="8.0cm"}
{width="16.5cm"}
[lccccccccccc]{} & $\langle N_e \rangle$ & $\langle T_e \rangle$ & $\langle P_e \rangle$ & $A$ & $V$ & $L$\
Date & 10$^{9}$cm$^{-3}$ & 10$^{6}$K & 10$^{15}$K cm$^{-3}$ & 10$^{18}$cm$^{2}$ & 10$^{28}$cm$^{3}$ & 10$^{9}$cm\
2002 September 10& 2.46$\pm$0.31 & 2.26$\pm$0.03 & 3.10$\pm$0.60 & 12.66$\pm$0.75 & 1.69$\pm$0.15 & 6.31$\pm$0.37\
2002 September 11 & 2.71$\pm$0.53 & 2.24$\pm$0.03 & 3.41$\pm$0.84 & 10.23$\pm$0.64 & 1.23$\pm$0.11 & 5.66$\pm$0.35\
2002 September 12 & 3.04$\pm$0.79 & 2.26$\pm$0.03 & 3.83$\pm$1.16 & 10.58$\pm$0.76 & 1.29$\pm$0.14 & 5.76$\pm$0.41\
2002 September 13 & 2.46$\pm$0.52 & 2.24$\pm$0.02 & 3.10$\pm$0.65 & 9.89$\pm$1.03 & 1.17$\pm$0.18 & 5.57$\pm$0.58\
2002 September 14 & 1.87$\pm$0.20 & 2.16$\pm$0.02 & 2.35$\pm$0.43 & 6.54$\pm$1.22 & 0.62$\pm$0.17 & 4.53$\pm$0.84\
& $P_R$ & $P_C$ & $P_F$ & $P_H$ & $\Phi_{tot}$\
Date & 10$^{25}$ergs s$^{-1}$ & 10$^{25}$erg s$^{-1}$ & 10$^{25}$erg s$^{-1}$ & 10$^{25}$ergs s$^{-1}$ & 10$^{24}$Mx\
2002 September 10 & 3.37$\pm$0.90 & 0.28$\pm$0.07 & 0.25$\pm$0.04 & 3.90$\pm$1.57 & 5.97$\pm$0.35\
2002 September 11 & 2.97$\pm$1.12 & 0.19$\pm$0.05 & 0.37$\pm$0.08 & 3.53$\pm$1.92 & 4.01$\pm$0.25\
2002 September 12 & 3.92$\pm$2.07 & 0.21$\pm$0.06 & 0.43$\pm$0.12 & 4.56$\pm$3.10 & 3.97$\pm$0.28\
2002 September 13 & 2.33$\pm$1.05 & 0.30$\pm$0.07 & 0.42$\pm$0.14 & 2.59$\pm$1.40 & 2.78$\pm$0.29\
2002 September 14 & 0.71$\pm$0.24 & 0.07$\pm$0.03 & 0.18$\pm$0.06 & 0.97$\pm$0.68 & 0.91$\pm$0.17\
![Variations in electron density ($N_e$), temperature ($T_e$), and pressure ($P_e$) for September 10–14.[]{data-label="ne_te_pe_plot"}](f6.eps){width="8.0cm"}
Electron Temperatures {#temp_map}
---------------------
Electron temperature maps were created using the method of @bros96, which employs the ratio of the fluxes of various ionization stages of iron. This method obtains a polynomial fit to the logarithm of the temperature as a function of the logarithm of the emissivity ratio for selected line pairs, under the assumption of an isothermal plasma, with the form, $$\label{brosius_eqn}
log~T_{e}=a_{0}+a_{1}(log~R)+a_{2}(log~R)^{2}+a_{3}(log~R)^{3},$$ where the parameters, $a_{0}$, $a_{1}$, $a_{2}$, and $a_{3}$ were initially tabulated by @bros96, and $R$ is the intensity ratio <span style="font-variant:small-caps;">Fe xvi</span> (335.41 Å)/<span style="font-variant:small-caps;">Fe xiv</span> (334.18 Å).
Using the most recent theoretical atomic data from <span style="font-variant:small-caps;">chianti</span> v4.2, the values of $a_{0}$,..., $a_{3}$ were found to change somewhat and resulted in a slightly higher temperature ($\sim$5 per cent) than those predicted using [@bros96] values. Both previous and updated parameters have been included in Table \[bros\_chianti\], while the resulting temperatures are presented in Fig. \[bros\_ch\_fig\]. In addition, the <span style="font-variant:small-caps;">Fe xiv</span> (334.18 Å) line is density sensitive, and was accounted for by determining the temperature across the active region at the corresponding density.
The temperature maps are displayed in the third row of Fig. \[maps\_fig\]. As expected, the temperature maps show a close spatial correlation with the intensity and density maps. As the region evolves, the temperature remains constant at $\sim$2.25$\times$10$^{6}$ K, with just a slight decrease (by $\sim$4 per cent) on the final day of observation. More statistically significant is the variation of electron temperature across the region, which ranges from $\sim$2.10$\times$10$^{6}$ K in the active region boundary to $\sim$2.34$\times$10$^{6}$ K in the core. The values derived for the temperature are heavily dependent on the $a_{0}$ coefficient in Equation (\[brosius\_eqn\]), which accounts for the restricted temperature sensitivity of this method. The temperature is also affected by the brightening on September 13 as can be seen from the corresponding map. Average temperature values are listed in Table \[res\_tab\], and are plotted in Fig. \[ne\_te\_pe\_plot\]. Again, these results are in good agreement with those found in [@gall01] using similar methods.
The C-class flare on September 13 is not only evident in the intensity and density maps of September 13, but was also identified due to the presence of the <span style="font-variant:small-caps;">Fe xvii</span> (347.85 Å) line which has a formation temperature of $\sim$5$\times$10$^{6}$ K. A portion of the spectrum for the flare is shown in Fig. \[fe17\_plot\].
![The temperature sensitive curves for <span style="font-variant:small-caps;">Fe xvi</span> (335.41 Å)/<span style="font-variant:small-caps;">Fe xiv</span> (334.18 Å) for the @bros96 method and the updated <span style="font-variant:small-caps;">chianti</span> v4.2 data, both plotted for plasma densities of 10$^{9}$ and 10$^{10}$ cm$^{-3}$.[]{data-label="bros_ch_fig"}](f7.eps){height="8.0cm"}
![Portion of the spectrum from the compact brightening on September 13 showing the <span style="font-variant:small-caps;">Fe xvii</span> (347.85 Å) emission line and neighbouring lines. <span style="font-variant:small-caps;">Fe xvii</span> has a formation temperature of $\sim~5~\times~10^{6}~K$.[]{data-label="fe17_plot"}](f8.eps){height="8.0cm"}
[lccccc]{} &Log N$_{e}$ &$a_{0}$ &$a_{1}$ &$a_{2}$ &$a_{3}$\
[@bros96] &9.0 &6.237 &0.111 &0.0087 &0.0016\
&10.0 &6.217 &0.108 &0.0076 &0.0018\
Updated values &9.0 &6.259 &0.111 &0.0105 &0.0015\
&10.0 &6.240 &0.108 &0.0097 &0.0015\
Electron Pressures {#pres_map}
------------------
The pressure maps in the bottom row of Fig. \[maps\_fig\] follow a similiar behaviour to that seen in the density. Values for the pressure in the high density core of the active region remain $\sim$10$^{16}$ K cm$^{-3}$ from September 10–13 and drop to 4$\times$10$^{15}$ K cm$^{-3}$ on September 14, again not taking into account the effects of the brightening. Similar to the average electron density, the average pressure varied by $\sim$30 per cent between September 10–14 (see the values presented in Table \[res\_tab\] and Fig. \[ne\_te\_pe\_plot\]).
Power Balance {#energy_bal}
-------------
The steady-state energetics of a coronal loop can be expressed as, $$\label{energy_bal_eqn}
E_H = E_R + E_C + E_F~~{\rm~ergs~cm^{-3}~s^{-1}},$$ where $E_R$ is the total radiative losses of the plasma, $E_C$ is the thermal conductive flux, $E_F$ is the energy lost due to mass flows, and $E_H$ is the energy required to balance these losses (e.g. Antiochos & Sturrock 1982, Bradshaw & Mason 2003). The radiative loss term, $E_R$, in Equation (\[energy\_bal\_eqn\]) can be written in terms of the electron density, $N_e$, and the radiative loss function, $\Lambda(T_e)$, $$\label{rad_eqn}
E_R = -N_e^2 \Lambda(T_e)~~{\rm~ergs~cm^{-3}~s^{-1}}.$$ The radiative loss function is usually approximated by analytical expressions of the form $N_e^2 \chi T_e^\alpha$ [@cook89]. While this method is simple to implement computationally, it does not capture accurately the fine-scale structure of the radiative loss function. A more appropriate technique is used here, which relies on interpolating the values obtained from <span style="font-variant:small-caps;">chianti</span>, using the coronal abundances of [@feld92] and the ionization balance calculations of @mazz98. The choice of coronal abundances was motivated by recent work by [@delz03] who show that low FIP elements in active regions have abundances which are coronal.
Assuming classical heat conduction along the magnetic field, the conductive flux can be expressed as, $$\label{conduct_eqn}
E_C = \frac{d}{ds} \left [ -\kappa T_e^{5/2}
\frac{dT_e}{ds} \right ]~~{\rm~ergs~cm^{-3}~s^{-1}},$$ where the thermal conductivity is $\kappa = 0.92 \times 10^{-6}~~{\rm
ergs~s^{-1}~cm^{-1}~K^{-7/2}}$ [@spit62], $T_e$ is the electron temperature. Various approximations for Equation (\[conduct\_eqn\]) have been used in previous studies: @asch99 used an expression that was heavily dependent upon the temperature gradient, $(dT/ds)^2$, in the study of active region loops, whereas [@vara00] used an approximation strongly dependent upon temperature, $T^{7/2}$, for post-flare loops. Here, we assume a semi-circular loop geometry and approximate Equation (\[conduct\_eqn\]) by, $$\label{conduct_approx}
E_C = -\kappa T_e^{5/2} \frac{\Delta T_e}{(L/2)^{2}}~~{\rm~ergs~cm^{-3}~s^{-1}},$$ where $\Delta T_{e}$ is the difference in the maximum and minimum temperatures given in Section \[temp\_map\], and $ds$ is approximated by $L/2$. This approximation is therefore not overly dependent on temperature or the temperature gradient, but acts as a balance between the approximations of [@asch99] and [@vara00].
The energy losses due to mass motions of the plasma, $E_F$, can be expressed as a sum of the kinetic energy and the internal energy of the plasma, $$\label{flow_eqn}
E_F = \frac{d}{ds} \left [ v \left ( \frac{1}{2} \rho v^2 + \frac{5}{2} N_e k_B T_e \right ) \right ]~{\rm~ergs~cm^{-3}~s^{-1}},$$ where $v$ is the flow velocity of the plasma, and $ds$ is again approximated to be the loop half-length ($L/2$). Line-of-sight velocities at the core of the active region were found to be $\sim$10 km s$^{-1}$ which is consistant with @bryn99 who detected flow speeds of $\sim$15 km s$^{-1}$ also using CDS. Using these velocity values, as well as typical parameters from Table \[res\_tab\], the mass loss term was found to be $\sim$10$^{24}$ ergs s$^{-1}$. Energy losses due to mass motions can therefore be considered negligible compared to radiative losses.
![The radiative, conductive, and mass losses, and the heating rate for NOAA 10103 for September 10–14.[]{data-label="plot_loss"}](f9.eps){width="8.0cm"}
In order to compare the work described in this paper with that of @fish98, @park83, and others, Equation (\[energy\_bal\_eqn\]) has been integrated over the volume of the active region, and expressed as a power balance equation: $$\label{power_eqn}
P_H = P_R + P_C + P_F ~~{\rm~ergs~s^{-1},}$$ where $P_R$, $P_C$, and $P_F$ are the power lost due to radiation, conduction, and mass flows, respectively. $P_H$ is then the heating rate required to balance these losses.
Using the active region properties detailed in Table \[res\_tab\], the radiative, conductive, and mass flow losses were then calculated using Equations (\[rad\_eqn\]), (\[conduct\_approx\]), and (\[flow\_eqn\]). These results, together with the heating rate calculated using Equation (\[power\_eqn\]), are presented in Table \[res\_tab\] and Fig. \[plot\_loss\]. Due to the difficulty in correlating emission seen in CDS with a particular magnetic flux concentration, the region’s average properties were used for comparision with MDI. The second panel down of Fig. \[plot\_loss\] also shows how the conductive flux values depend on how Equation (\[conduct\_eqn\]) is approximated. In each case, the conductive losses are found to be much less than the radiative losses. Indeed, the average radiative losses ($\langle P_R \rangle \sim 3 \times 10^{25}$ ergs s$^{-1}$) are found to exceed both the conductive losses ($\langle P_C \rangle \sim 2
\times 10^{24}$ ergs s$^{-1}$) and the mass flow losses ($\langle P_F \rangle
\sim 4 \times 10^{24}$ ergs s$^{-1}$) by approximately an order of magnitude. [@asch00] also found an order of magnitude difference between the radiative and conductive losses in coronal loops despite the actual values being somewhat lower than those found here due to the insensitivity of EIT filter ratio techniques. As can be seen from Fig. \[plot\_loss\], the heating rate falls by close to a factor of 5 between September 10 and 14 and is not significantly affected by whichever conductive loss equation is used.
The top panel of Fig. \[phi\_power\] shows the heating rate as a function of the total unsigned magnetic flux. A least-squares fit to the non-flaring data (i.e., neglecting the high density plasma of September 13) yielded a power-law of the form $P_H \sim \Phi_{tot}^{\gamma}$, where $\gamma = 0.81\pm0.32$. Power-laws with slopes of 1 and 2 from Equations (\[ac\_heating\]) and (\[dc\_heating\]) are also shown for comparison. The bottom panel of Fig. \[phi\_power\] shows how this relationship varies when densities from the core and the boundaries of the active region as well as from the quite solar corona (see Fig. \[ne\_te\_pe\_plot\]) are used instead of average values. This wide range of $N_{e}$ shows that ${\gamma}$ varies from 0.2 in the quiet-Sun to 1.5 at the core of the region. It can therefore be concluded that average values of these parameters are a reasonable representation of the entire active region.
![[*Top Panel*]{}: Plot of $P_H$ versus $\Phi_{tot}$ for five days of MDI and CDS observations. Also shown is a least-squares fit to the non-flaring data together with power-laws with slopes of 1 and 2 from the theoretical predictions. [*Bottom Panel*]{}: Plot of $P_H$ versus $\Phi_{tot}$, where $P_H$ has been calculated using non-flaring core ([*dotted line*]{}), boundary ([*dashed line*]{}), and quiet-Sun ([*dot-dashed line*]{}) parameters from Fig. \[ne\_te\_pe\_plot\].[]{data-label="phi_power"}](f10a.eps "fig:"){height="8.0cm"} ![[*Top Panel*]{}: Plot of $P_H$ versus $\Phi_{tot}$ for five days of MDI and CDS observations. Also shown is a least-squares fit to the non-flaring data together with power-laws with slopes of 1 and 2 from the theoretical predictions. [*Bottom Panel*]{}: Plot of $P_H$ versus $\Phi_{tot}$, where $P_H$ has been calculated using non-flaring core ([*dotted line*]{}), boundary ([*dashed line*]{}), and quiet-Sun ([*dot-dashed line*]{}) parameters from Fig. \[ne\_te\_pe\_plot\].[]{data-label="phi_power"}](f10b.eps "fig:"){height="8.0cm"}
DISCUSSION AND CONCLUSIONS {#conc}
==========================
A detailed study of an evolving active region has been described using measurements from several instruments onboard [*SOHO*]{}. The region was observed to decay in size and complexity as it passed from close to the central meridian on September 10, to close to the west limb on September 14. In the photosphere, the total sunspot area fell by close to a factor of 2, while the total magnetic flux fell by approximately a factor of 6. In the corona, the average electron density, temperature, and pressure all showed similar decreases in value, which is to be expected considering that the plasma in the corona traces out field lines which are ultimately rooted in the photosphere. These results are of consequence to efforts in understanding active region evolution and the relationship between the photosphere and corona of solar active regions (e.g., Abbett & Fisher 2003; Ryutova & Shine 2004).
In addition to studying active region evolution, CDS and MDI were used to investigate the power-law relationships predicted by theoretical models of the corona. @mand00 also investigated theoretical scaling laws using magnetic field extrapolations from both vector and longitudinal magnetograms, in light of the work of @klim95. They concluded that models involving the gradual stressing of the coronal magnetic field are in better agreement with the observational contrains than are wave heating models. This same general conclusion was also reached by [@demoulin03] using photospheric and coronal measurements from MDI and [*Yohkoh*]{}. Unfortunately these studies relied on broad-band filter ratios to determine electron temperatures and densities; the difficulty associated with making such measurements is clear from the contradictory results of @priest98, @asch01, and @reale02, who also investigated coronal loop heating models using filter ratios from [*Yohkoh*]{}/SXT.
The analysis presented in this paper, on the other hand, is based on well understood line ratio techniques, which offer a less ambiguous determination of plasma properties. With this in mind, the power-law relationship between the total heating rate and the total unsigned magnetic flux ($\Phi_{tot}$) was determined, finding a relationship of the form $P_H \sim
\Phi_{tot}^{0.81\pm0.32}$. [@fish98] compared the X-ray luminosity (which was assumed to be some fraction of the total heating power) to active region vector magnetograms, finding a similar power-law relation of $L_{X}
\sim \Phi_{tot}^{1.19}$. The result of [@fish98] suggests a “universal” relationship between magnetic flux and the amount of coronal heating, regardless of the age or complexity of the active region. A similar relationship of $L_X \sim
\Phi_{tot}^{0.9}$ was found using statistical samples of late-type stars [@schr87]. The results of this paper therefore lend further observational evidence that active regions are heated by magnetically-associated waves, rather than multiple nanoflare-type events.
Acknowledgments {#acknowledgments .unnumbered}
===============
[*SOHO*]{} is a project of international collaboration between ESA and NASA. This work has been supported by a Cooperative Award in Science and Technology (CAST) studentship from Queen’s University Belfast and the NASA GSFC [*SOHO*]{} project. ROM would like to thank R. T. J. McAteer for useful comments. FPK is grateful to AWE Aldermaston for the award of a William Penny Fellowship.
Abbett, W. P., Fisher, G. H. 2003, ApJ, 582, 475.
Antiochos, S. K., & Sturrock, P. A. 1982, ApJ, 254, 343
Aschwanden, M. J., Newmark, J. S., Delaboudinière, J. P., Neupert, W. M., Klimchuk, J. A., Gary, G. A., Portier-Fozzani, F., & Zucker, A. 1999, ApJ, 515, 842
Aschwanden, M. J., Alexander, D., Hurlburt, N., Newmark, J. S., Neupert, W. M., Klimchuk, J. A., & Gary, G. A. 2000, ApJ, 531, 1129A
Aschwanden, M. J. 2001, ApJ, 559, L171
Aschwanden, M. J., Schrijver, C. J., & Alexander, D. 2001, ApJ, 550, 1036
Berger, T. E., & Lites, B. W. 2003, Sol. Phys., 213, 213
Bradshaw, S. J., & Mason, H. E. 2003, A&A, 401, 699
Brosius, J. W., Davila, J. M., Thomas, R. J., & Monsignori-Fossi, B. C. 1996, ApJ, 106, 143
Browning, P. K., Sakuria, T., & Priest, E. R. 1986, A&A, 158, 217
Brynildsen, N., Maltby, P., Brekke, P., Haugan, S. V. H., & Kjeldseth-Moe, O. 1999, Sol. Phys., 186, 141
Cook, J. W., Cheng, C.-C., Jacobs, V. L., et al. 1989, ApJ, 338, 1176
Dere, K. P., Landi, E., Mason, H. E., Monsignori-Fossi, B. C., & Young, P. R. 1997, A&AS, 125, 149
Del Zanna, G., & Mason, H. E. 2003, A&A, 406, 1089
Démoulin, P., van Driel-Gesztelyi, L., Mandrini, C. H., Klimchuk, J. A., Harra, L. 2003, ApJ, 454, 499
Edlen, B. 1937, Z. Phys., 104, 407
Feldman, U. 1992, Phys. Scripta, 46, 2002
Fisher, G. H., Longcope. D. W., Metcalf, T. R., & Pevtsov, A. A. 1998, ApJ, 508, 885
Freeland, S. L., & Handy, B. N. 1998, Sol. Phys., 182, 497
Gallagher, P. T., Phillips, K. J. H., Lee, J., Keenan, F. P., & Pinfield, D. J. 2001, ApJ, 558, 411
Green, L. M., Démoulin, P., Mandrini, C. H., & van Driel-Gesztelyi, L. 2003, Sol. Phys., 215, 307
Gurman, J. B., Withbroe, G. L. , & Harvey, J. W. 1974, Sol. Phys., 34, 105
Harrison, R. A., et al. 1995, Sol. Phys., 162, 233
Klimchuk, J. A., & Porter, L. J. 1995, Nature, 377, 131
Mandrini, c. H., Démoulin, P., & Klimchuk, J. A. 2000, ApJ, 530, 999
Mazzotta, P., Mazzitelli, G., Colafrancesco, S., & Vittorio, N. 1998, A&AS, 133, 402
Parker, E. N. 1983, ApJ, 264, 642
Parker, E. N. 1988, ApJ, 330, 474
Priest, E. R., Foley, C. R., Heyvaerts, J., Arber, T. D., Culhane, J. L., & Acton, L. W. 1998, Nature, 393, 545
Reale, F. 2002, ApJ, 580, 566
Ryutova, M., & Shine, R. 2004, ApJ, 606, 571R
Scherrer, P. H., et al. 1995, Sol. Phys., 162, 129
Schrijver, C. J. 1987, A&A, 180, 241
Spitzer, L. 1962, Physics of Fully Ionized Gasses (New York: Interscience)
Thompson, W. T. 1999, CDS Software Note No. 53
van Driel-Gesztelyi, L., Démoulin, P., Mandrini, C. H., Harra, L., & Klimchuk, J. A. 2003, ApJ, 586, 579
Varady, M., Fludra, A., & Heinzel, P. 2000, A&A, 355, 769
Warren, H. P., & Winebarger, A. R. 2003, ApJ, 596, L113
[^1]: E-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A two-dimensional periodic array of scatterers has been introduced to a single layer of graphene in the presence of an external magnetic field perpendicular to the graphene layer. The eigenvalue equation for such a system has been solved numerically to display the structure of split Landau subbands as functions of both wave number and magnetic flux. The effects of pseudo-spin coupling and Landau subbands mixing by a strong scattering potential have been demonstrated. Additionally, we investigated the square barrier tunneling problem when magnetic field is present, as well as demonstrate the crucial difference in the modulated band structure between graphene and the two-dimensional electron gas. The low-magnetic field regime is particularly interesting for Dirac fermions and has been discussed. Tunneling of Dirac electrons through a magnetic potential barrier has been investigated to complement the reported results on electrostatic potential scattering in the presence of an ambient magnetic field.'
author:
- 'Godfrey Gumbs$^{1,2}$, Andrii Iurov$^{1}$, Danhong Huang$^{3}$, Paula Fekete$^{4}$ and Liubov Zhemchuzhna$^{5}$'
title: Effects of periodic scattering potential on Landau quantization and ballistic transport of electrons in graphene
---
Dating back to the early work of Azbel [@azbel] and Hofstadter[@hofstadter], the single-particle spectrum of a two-dimensional structure in the presence of both a periodic potential and a uniform ambient perpendicular magnetic field has captivated researchers for many years [@GG]$^{-}$[@cui1]. The paper by Hofstadter [@hofstadter] was for the energy spectrum of a periodic square lattice in the tight-binding approximation and subject to a perpendicular magnetic field. Ever since that time, there have been complementary calculations for the hexagonal lattice [@GG],the two-dimensional electron gas (2DEG) with an electrostatic periodic modulation potential [@pfannkuche; @pfannkuche2; @cui1] and even bilayer graphene where different stacking of the two types of atoms forming the sublattices was considered [@bilayer]. It has been claimed that one may be able to observe evidence of the existence of Hofstadter’s butterfly in such experimentally measured quantities as density-of-states and conductivity [@ye; @nat1; @nat2].
The challenge facing experimentalists so far has been to carry out experiments on 2D structures at achievable magnetic fields where the Hofstadter butterfly spectrum is predicted. One may follow the calculation of Hofstadter by using Harper’s equation which may be regarded as a tight-binding approximation of the Schrodinger equation. Then, assuming that the magnetic flux through unit cell of the periodic lattice is a rational fraction $p/q$ of the flux quantum in conjunction with the Bloch condition for the wave function, one obtains a $p\times p$ Hamiltonian matrix to determine the energy eigenvalues since one only needs to solve the problem in a unit cell. Hofstadter himself was concerned about ever reaching magnetic fields where the rich self-similar structure of the butterfly would be experimentally observed due to the estimated high fields required to achieve this.
In this paper, we propose applying an electrostatic modulation potential to a flat sheet of graphene in the presence of a perpendicular magnetic field to produce the energy spectrum with self-similarity at reasonably low magnetic fields. For a review, see [@sokoloff]. We also compare our results with those for a modulated two-dimensional electron gas and discuss the difference. For completeness, we briefly review tunneling of Dirac electrons through a magnetic barrier.
Model and Theory
================
In the presence of a periodic two-dimensional scattering potential modulation[@cui1]
$$V(x,\,y)=V_0\left[\cos\left(\frac{2\pi x}{d_x}\right)\cos\left(\frac{2\pi y}{d_y}\right)\right]^{2N}\ ,$$
where $N=1,\,2,\,\cdots$ is an integer, $V_0$ is the modulation amplitude, and $d_x,\,d_y$ are the modulation periods in the $x$ and $y$ directions, respectively, we write the Hamiltonian operator as $$\hat{{\mathcal}H} = v_F \left[{
\begin{array}{cccc}
V(x,\,y) & {\hat{p}_x+eB_0y\hat{x}_0 + i\hat{p}_y } & 0 & 0 \\
{\hat{p}_x+eB_0y\hat{x}_0 - i\hat{p}_y } & V(x,\,y) & 0 & 0 \\
0 & 0 & V(x,\,y) & {\hat{p}_x+eB_0\hat{x}_0 - i\hat{p}_y } \\
0 & 0 & {\hat{p}_x+eB_0\hat{x}_0 + i\hat{p}_y } & V(x,\,y)
\end{array}
}\right]$$ $$\hat{\cal H} = v_F \left[{
\begin{array}{cc}
\hat{\cal H}_K+V(x,\,y)\hat{I} & 0 \\
0 & \hat{\cal H}_{K'}+V(x,\,y)\hat{I}
\end{array}
}\right]\ .$$ Here, $v_F$ is the Fermi velocity, $\hat{I}$ is a $2\times 2$ identity matrix. In this system, the magnetic flux through unit cell is $\Phi=B_0(d_xd_y)$, which is assumed to be a rational multiple of the flux quantum $\Phi_0=h/e$, i.e., $\beta\equiv\Phi/\Phi_0=p/q$ is an irreducible fraction and $p$ and $q$ are integers. Furthermore, we choose the first Brillouin zone defined by $ \vert k_x \vert \leq \pi/d_x$ and $|k_y|\leq \pi/(qd_y)$. By using the Bloch-Peierls condition, the wave function of this modulated system may be expanded as
$$\Phi^{\pm}_{\ell;\,n,{\vec k}_{||}} \left( {x,\,y} \right)=\frac{1}{\sqrt{2{\cal N}_y}}\sum\limits_{s=-\infty}^{\infty}\left\{{\rm e}^{ik_y\ell_B^2(sp+\ell)K_1}\right.
\times\left.\left[\Psi_{n,k_x-(sp+\ell)K_1}^{K,\pm} \left( {x,\,y} \right)+\Psi_{n,k_x-(sp+\ell)K_1}^{K',\pm} \left( {x,\,y} \right)\right]\right\}\ ,$$
where ${\vec k}_{||}=(k_x,\,k_y)$, ${\cal N}_y=L_y/(qd_y)$ is the number of unit cells, which are spanned by $b_1=(d_x,\,0)$ and $b_2=(0,\,qd_y)$, in the $y$ direction, $L_y$ ($\to\infty$) is the sample length in the $y$ direction, $K_1=2\pi/d_x$ is the reciprocal lattice vector in the $x$ direction and $\ell=1,\,2,\,\cdots,\,p$ is a new quantum number for labeling split $p$ subbands from a $k_x$-degenerated landau level in the absence of modulation. The above wave function satisfies the usual Bloch condition, i.e., $$\Phi^{\pm}_{\ell;\,n,{\vec k}_{||}} \left( {x+d_x,\,y+qd_y} \right)={\rm e}^{ik_xd_x}\,{\rm e}^{ik_yqd_y}\,\Phi^{\pm}_{\ell;\,n,k_{||}} \left( {x,\,y}\right)\ .$$ Since the wave functions at the $K$ and $K^\prime$ points are decoupled from each other for single-layer graphene, which is different from bilayer graphene [@petters], we may write out explicitly the full expression for the wave function at these two points, i.e.,
$$\Phi^{K,\pm}_{\ell;\,n,{\vec k}_{||}} \left( {x,\,y} \right)=\frac{C_n}{\sqrt{{\cal N}_yL_x}}\sum\limits_{s=-\infty}^{\infty} {\rm e}^{i[k_x+K-(sp+\ell)K_1] x}
{\rm e}^{ik_y\ell_B^2(sp+\ell)K_1} \times$$ $$\times
\left[{
\begin{array}{c}
\alpha_n^{\pm}\phi_{n-1,k_x+K-(sp+\ell)K_1} \left( y \right) \\
\phi_{n,k_x+K-(sp+\ell)K_1} \left(y \right) \\
0 \\
0
\end{array}
}\right] \ ,$$ $$\Phi^{K',\pm}_{\ell;\,n,{\vec k}_{||}} \left( {x,\,y} \right)=\frac{C_n}{\sqrt{{\cal N}_yL_x}}\sum\limits_{s=-\infty}^{\infty}
{\rm e}^{ik_y\ell_B^2(sp+\ell)K_1} \times{\rm e}^{i[k_x+K'-(sp+\ell)K_1] x} \times$$ $$\times
\left[{
\begin{array}{c}
0 \\
0 \\
\phi_{n,k_x+K'-(sp+\ell)K_1} \left( y \right) \\
\alpha_n^{\pm}\phi_{n-1, k_x+K'-(sp+\ell)K_1} \left( y \right)
\end{array}
}\right]\ .$$ From here onwards, we will concentrate on the $K$ point by setting $K=0$ for convenience and omit the label for the $K$ point thereafter, i.e.,
$$\begin{aligned}
\nonumber
&& \psi^{A,\pm}_{\ell,n,{\vec k}_{||}} \left( {x,\,y} \right)\equiv\langle {\vec r}_{||}|\ell,n,{\vec k}_{||};\,A,\pm\rangle
=\frac{\alpha_n^{\pm}C_n}{\sqrt{{\cal N}_yL_x}}\sum\limits_{s=-\infty}^{\infty}
{\rm e}^{ik_y\ell_B^2(sp+\ell)K_1} \times{\rm e}^{i[k_x+K-(sp+\ell)K_1] x}\,\phi_{n-1,k_x+K-(sp+\ell)K_1} \left( y \right) \ , \\
\\
\nonumber
&& \psi^{B,\pm}_{\ell,n,{\vec k}_{||}} \left( {x,\,y} \right)\equiv\langle {\vec r}_{||}|\ell,n,{\vec k}_{||};\,B,\pm\rangle
=\frac{C_n}{\sqrt{{\cal N}_y L_x}}\sum\limits_{s=-\infty}^{\infty}
{\rm e}^{ik_y\ell_B^2(sp+\ell)K_1} \times{\rm e}^{i[k_x+K-(sp+\ell)K_1] x}\,\phi_{n,k_x+K-(sp+\ell)K_1} \left( y \right) \ , \\\end{aligned}$$
where ${\vec r}_{||}=(x,\,y)$. A tedious but straightforward calculation gives rise to an explicit expression for the matrix elements for the potential $V(x,\,y)$ as
$$\begin{aligned}
&& V^{\ell',n',\mu'}_{\ell,n,\mu}(\vec{k}_{||})\equiv\sum\limits_{{\vec k}'_{||}}
\int\int\,dxdy\,\left[\Phi^{\mu'}_{\ell';\,n',{\vec k}'_{||}} \left( {x,\,y} \right)\right]^\dag\,V(x,\,y)\,\Phi^{\mu}_{\ell;\,n,{\vec k}_{||}} \left( {x,\,y} \right) \\
\nonumber
&& =\frac{V_0C_{n'}C_n}{4^{2N}}\left\{{\rm e}^{ik_y\ell_B^2K_1(\ell-\ell')}\,\sum\limits_{i=0}^{N-1}\sum\limits_{j=0}^{N-1}\,\left[{\cal F}^{(B)}_{ij}+\alpha_{n'}^{\mu'}\alpha_n^{\mu}\,{\cal F}^{(A)}_{ij}\right]\right.
\left.+\delta_{\ell,\ell'}\delta_{n,n'}\,\left(1+\alpha_{n'}^{\mu'}\alpha_n^{\mu}\right)
\left[\frac{(2N)!}{(N!)^2}\right]^2\right\}\ , \\\end{aligned}$$
where $$\begin{aligned}
\nonumber
&& {\cal F}_{ij}^{(B,A)}=\left({
\begin{array}{c}
2 N \\
i
\end{array}
}\right)
\left({
\begin{array}{c}
2N \\
N
\end{array}
}\right)
{\mathcal}{T}_1
+
\left({
\begin{array}{c}
2N \\
j
\end{array}
}\right)
\left({
\begin{array}{c}
2N \\
N
\end{array}
}\right){\mathcal}{T}_2 + 2
\left({
\begin{array}{c}
2N \\
i
\end{array}
}\right)
\left({
\begin{array}{c}
2 N \\
j
\end{array}
}\right) {\mathcal}{T}_3 \ , \\
\nonumber
&& {\mathcal}{T}_1 = A^{(B,A)}_1(0,\,N-i) \ , \\
\nonumber
&& {\mathcal}{T}_2 = A^{(B,A)}_2(N-j,\,0) \ , \\
\nonumber
&& {\mathcal}{T}_3 = A^{(B,A)}_3(N-j,\,N-i) \ , \end{aligned}$$
as well as the binomial expansion coefficient with $m \geq n$ $$\left({
\begin{array}{c}
m \\
n
\end{array}
}\right) \equiv \frac{m!}{n!\,(m-n)!} \ .$$ In the above expressions, we have also introduced the following three self-defined functions
$$A_1^{(B,A)}(r,\,s)=D_{n',n}^{rs{(B,A)}}\,T_{\ell}^s\,\delta_{\ell,\ell'}\ ,$$
$$\begin{aligned}
&& A^{(B,A)}_2(r,\,s)=D_{n',n}^{rs{(B,A)}}\left\{\delta_{\ell-\ell',r}\left[{\rm sgn}(n'-n)\right]^{\xi}+\delta_{\ell'-\ell,r}\left[{\rm sgn}(n-n')\right]^{\xi}\right\}\ , \\
\nonumber
&& A^{(B,A)}_3(r,\,s)=D_{n',n}^{rs{(B,A)}}\left\{\delta_{\ell-\ell',r}\left[{\rm sgn}(n'-n)\right]^{\xi}\cos[\Theta_{rs}^{\ell'}(n',\,n)]\right. \left.+\delta_{\ell'-\ell,r}\left[{\rm sgn}(n-n')\right]^{\xi} \cos[\Theta_{rs}^{\ell}(n,\,n')]\right\}\ ,\end{aligned}$$
where $\xi=|n-n^\prime |$, $$D_{n',n}^{rs(B)}=\sqrt{\frac{n_1!}{n_2!}}\,{\rm e}^{-W_{rs}/(2\beta)}\left(\frac{W_{rs}}{\beta}\right)^{\xi/2}L_{n_1}^{(\xi)}\left(\frac{W_{rs}}{\beta}\right)\ ,$$ $n_1={\rm min}(n,\,n')$, $n_2={\rm max}(n,\,n')$, $L_n^{(m)}(x)$ is the associated Laguerre polynomial, $$W_{rs}=\frac{\pi \left( s^2 K_2^2 + r^2 K_1^2 \right)}{K_1 K_2)} \ , \hspace{0.2 in} K_2=\frac{2 \pi}{d_y} \ , \hspace{0.2 in}
D_{n',n}^{rs(A)}=D_{n'-1,n-1}^{rs(B)} \ ,$$ with $T_\ell^s$: $$\begin{aligned}
&& T_\ell^s = \pm 2\cos\left[\frac{s (k_xd_x-2\ell\pi)}{\beta}\right] \,\, \delta_{\xi,\{4 N, \, 4 N+2\}}\ , \\
\nonumber
&& T_\ell^s = \pm 2\sin\left[\frac{s(k_xd_x-2\ell\pi)}{\beta}\right] \,\, \delta_{\xi,\{4 N+1, \, 4 N+3\}}\ ,\end{aligned}$$ and $$\Theta^\ell_{rs}(n',\,n)= \frac{s}{\beta} \left[k_xd_x-2\pi\left(\ell+r/2\right)\right] -
{\rm sgn}(n'-n)\,\xi\tan^{-1}\left(\frac{sd_x}{rd_y}\right) \ .$$ The magnetic band structure for this modulated system is a solution of the eigenvector problem ${\cal M} \bigotimes\vec{\cal A}(\vec{k}_{||})=0$ with the coefficient matrix $\tensor{\cal M}$ given by
$$\{{\cal M}\}_{j,\,j'}=\left[E^{\mu}_n-\varepsilon({\vec k}_{||})\right]
\delta_{n,n'}\delta_{\ell,\ell'}\delta^{(n)}_{\mu,\mu'} + V^{\ell',n',\mu'}_{\ell,n,\mu}(\vec{k}_{||})\ ,
\label{MM}$$
where $\delta^{(n)}_{\mu,\mu'}=1$ for $n=0$ and $\delta^{(n)}_{\mu,\mu'}=\delta_{\mu,\mu'}$ for $n>0$, $j=\{n,\,\ell,\,\mu\}$ is the composite index, and $\{\vec{\cal A}(\vec{k}_{||})\}_j={\cal A}^\mu_{n,\ell}(\vec{k}_{||})$ is the eigenvector. The eigenvalue $\varepsilon_\nu({\vec k}_{||})$ of the system is determined by ${\rm Det}\{\tensor{\cal M}\}=0$.
Numerical Results for Band Structure
------------------------------------
![(Color online). Energy dispersion as functions of $k_xd_x$ for chosen values of $V_0$ and magnetic flux $p/q$ in units of the flux quantum. The energy is scaled in terms of $v_F\sqrt{eB\hbar}$.[]{data-label="FIG:1"}](fig1){width="60.00000%" height="30.00000%"}
In Fig. \[FIG:1\], we present the dispersion curves as a function of $k_x d_x$ for chosen value of $V_0$ and two pairs of values of $p$ and $q$ corresponding to two different magnetic field strengths. In each case, there are $p$ Landau subbands, $q/p$ determines the number of oscillation periods in the first Brillouin zone for each of these subbands. Both the valence and conduction subbands are shifted upward but the conduction subbands are shifted more than the valence subbands for each corresponding Landau label for the unmodulated structure. This shift is increased when the modulation amplitude is increased. The original zero-energy Landau level is on;y slightly broadened and is the least affected by $V_0$. If the sign of the modulation amplitude is reversed to correspond to an array of quantum dots, then the subbands are all shifted downward. from their positions for an unmodulated monolayer graphene.
Figure \[FIG:2\] shows the results of our calculations for the energy eigenvalues of modulated graphene as a function of magnetic flux. We included the $n=0,\pm 1,\pm2,
\pm3,\pm 4$ as we did in obtaining Fig. \[FIG:1\]. For weak magnetic fields, the Landau levels in both valence and conduction bands are slightly broadened into narrow subbands but shifted upward by the perturbing potential $V_0$. Another effect due to modulation is to cause these Landau bands to have negative slope at weak magnetic fields which then broaden enough at higher magnetic fields to produce Landau orbit mixing, reflecting the commensurability between the magnetic and lattice Brillouin zones.
The lowest perturbed Landau subband which originates from the unperturbed $n=0$ Landau level merges with the resulting butterfly spectrum at the highest magnetic field compared to the $n=1,2,3,4$ Landau levels in the conduction band. The onset of the butterfly takes place around $p/q=1/5$ which would correspond to a magnetic field $B\approx 2$ T for $d_x=d_y=10$ nm. Furthermore, our calculations have shown that the symmetry between the valence and conduction bands is destroyed by modulation. There is always mixing of the subbands regardless of the value for $V_0$. This is in contrast with modulated 2DEG where for weak $V_0$, the Landau subbands do not overlap as shown in Fig. \[FIG:3\]. The lowest subband is shifted upward like the other subbands but is not widened as much as the higher subbands. The feature of self-similarity is also apparent in the excited subbands at intermediate magnetic fields. There is only a shift and broadening of the subbands in the low and high magnetic field regimes for modulated 2DEG.
Electron tunneling in the presence of magnetic field
====================================================
A goal of this paper is to investigate single-particle properties of graphene in the presence of both electric and magnetic fields. Consequently, to complement our derived results in the preceding calculations, we now turn our attention to the standard Klein tunneling problem through a square potential barrier [@Kats; @mine1; @mine2] in the presence of the uniform perpendicular magnetic field.
Magnetic barrier and confinement potential for Dirac-Weyl quasiparticles in graphene were addressed in [@Martino]. The reported results support the result that it is not possible for an electron to tunnel in the presence of a uniform magnetic field. This statement follows from the fact that the two substantially different Landau gauge $({\bf A} = - B x \hat{e}_y)$ and the Symmetric gauge $({\bf A} = \frac{1}{2} {\bf B} \times {\bf r})$ are expected to result in invariant observables such as current density or electron momentum.
The electrostatic potential barrier is specified by
$$U(x) = U_0 \left[{\theta(w+x)+\theta(w-x) - 1}\right] \\$$
and the vector potential ${\bf A}(x)$ due to the magnetic field ${ \bf B}(x) = B_0 \Theta(d^2-x^2) \hat{e}_z$ is has the following form $$\begin{aligned}
&& {\bf A}(x) = A(x) \hat{e}_y \, , \\
\nonumber
&& A(x) = \frac{1}{e \ell_B^2} \left[ {- w \theta(w+x)+ x \theta(x^2-d^2) + w \theta(w-x)}\right] \, .\end{aligned}$$ For this, the eigenvalue equations the spinor wave function with components $\Psi_a (x)$ and $\Psi_b (x)$ and energy $\epsilon$ are
$$\begin{aligned}
&& \hbar \left({\frac{\partial}{\partial x} + p_y + e A_y (x) }\right) \Psi_b (x) + i \; U(x) \Psi_a (x) = i \epsilon \Psi_a (x) \, ,
\nonumber\\
\label{a1}
&& \hbar \left({\frac{\partial}{\partial x} - p_y - e A_y (x) }\right) \Psi_a (x) + i \; U(x) \Psi_b (x)= i \epsilon \Psi_b (x) \, .\end{aligned}$$
where $U(x)=U_0$ is the electrostatic potential in the barrier region.
First, we notice that due to the specific spatial dependence of the potentials for both ${{\bf A}(x)}$ and $V(x)$, the transverse component of the electron momentum is conserved. This leads to the form of the electron wave function in each region. The system \[\[a1\]\], determining the wave function components results in the following equations: $$\begin{aligned}
&& \hbar^2 \frac{d^2 \psi_{a/b}(x)}{d x^2} - \left({ \pm e \frac{dA(x)}{dx}+( \hbar k_y + e A(x))}\right)\psi_{a/b}(x) = (\epsilon - U(x))^2 \psi_{a/b}(x) \ .\end{aligned}$$ In the barrier region ,this equation has solutions in the form of parabolic cylinder functions. The incoming particle, incident at angle $\phi$ with the normal to the barrier, has wave function on the left of the barrier given by $$\begin{aligned}
&& \Psi_{in}(x) = \frac{1}{\sqrt{2}} \left({
\begin{array}{c}
1 \\
\frac{p_x + i\left({p_y - \hbar d/ \ell_B^2 }\right)}{\vert \mathbf{p} \vert}
\end{array}
}\right) \texttt{e}^{i \hbar p_x x} \ ,\end{aligned}$$ where $ p_x = \epsilon \cos \phi$, $p_y = \epsilon \sin \phi + \hbar d / \ell_B^2$ and $\texttt{s} =[\epsilon - U(x)]$. The solution within the barrier region is given as follows $$\Psi_{b}(x) = \mathcal{N} \left({
\begin{array}{c}
\mathcal{D}_{\eta-1} (\pm \sqrt{2} \zeta) \\
\texttt{s} \sqrt{1/\eta} \; \mathcal{D}_{\eta} (\pm \sqrt{2} \zeta)
\end{array}
}\right) \texttt{e}^{i \hbar p_x x}$$ with $\zeta=x/\ell_B + p_y \ell_B$ and $\eta=((\epsilon -U(x)) \ell_B)^2 / 2$.
In contrast to the electrostatic potential barrier, the magnetic barrier leads to confinement of Dirac electrons in graphene.
In agreement with [@Martino], we find that for chosen incoming energy and angle of incidence, there is a maximum width of the barrier for finite transmission probability. When we also include an electrostatic potential barrier and take into account electron-hole transition, the energy is renormalized in the barrier region and may be arbitrarily small, which makes the tunneling impossible even for low-width barriers. However, for the case when $U_0 \gg \epsilon$, the transmission probability is restored and may be equal or close to unity, depending on the barrier width. As far as the angular dependence of the transmission is concerned, the largest tunneling probability does not correspond to head-on collision $(\phi = 0)$, but at a finite angle of incidence, revealing the asymmetry due to the applied magnetic field. A sharp drop in the transmission coefficient close to the critical thickness of the barrier corresponds in the classical limit to circular trajectories of the electrons.
The transmission coefficient is obtained by matching the wave function at the boundary of each region, keeping in mind that the longitudinal momentum does not remain constant in the barrier region, i.e.,
$$\begin{aligned}
&& T=\frac{p_{x,3}}{p_{x,1}} \vert t \vert^2 \, ,\\
\nonumber
&& T=\frac{\sqrt{2} \texttt{e}^{- i \hbar w (p_{x,1}+p_{x,3})} \left({1+\texttt{e}^{2 i \phi}}\right) \ell_B (\epsilon - U_0)\left\{{\Lambda^{2,+} \Xi^{2,+} + \Lambda^{2,-} \Xi^{2,+} }\right\}}{2 i \texttt{e}^{i \phi} \Upsilon_1 + i \texttt{e}^{i \phi_2} (U_0 - \epsilon) \epsilon \Upsilon_2 +\sqrt{2} \texttt{e}^{i(\phi+\phi_2)} \epsilon \Upsilon_3 -\sqrt{2}(U_0 - \epsilon) \Upsilon_4} \ . \\
\nonumber\end{aligned}$$
In this notation, $\Upsilon_1 = \Xi^{1,+} \Lambda^{2,-} - \Xi^{1,+}\Lambda^{2,-}$, $\Upsilon_2 = \Lambda^{1,+} \Lambda^{2,-} - \Lambda^{1,-} \Lambda^{2,+}$, $\Upsilon_3 = \Lambda^{1,+}
\Xi^{2,-} + \Lambda^{2,+} \Xi^{2,+}$, $\Upsilon_4 = \Lambda^{2,+} \Xi^{1,-} + \Xi^{2,-} \Lambda^{1,+}$ and $$\begin{aligned}
\nonumber
&& \Lambda^{1, \pm}(\epsilon, U_0, \phi, w) = \mathcal{D}_{\eta - 1}( \pm \sqrt{2} (-\frac{w}{\ell_B} + p_y \ell_B / \hbar)) \, ,\\
\nonumber
&& \Xi^{1, \pm}(\epsilon, U_0, \phi, w) = \mathcal{D}_{\eta}( \pm \sqrt{2} (-\frac{w}{\ell_B} + p_y \ell_B / \hbar)) \, ,\\
\nonumber
&& \Lambda^{2, \pm}(\epsilon, U_0, \phi, w) = \mathcal{D}_{\eta - 1}( \pm \sqrt{2} (\frac{w}{\ell_B} + p_y \ell_B / \hbar)) \, , \\
&& \Xi^{2, \pm}(\epsilon, U_0, \phi, w) = \mathcal{D}_{\eta}( \pm \sqrt{2} (\frac{w}{\ell_B} + p_y \ell_B / \hbar)) \, .\end{aligned}$$
Numerical results for electron transmission are presented in Fig.\[\[FIG:5\]\] and Fig.\[ \[FIG:6\]\]. Our general conclusion is that the transmission incorporates properties of both electric and magnetic potential barriers. Equivalent transmission resonances may be observed for $\epsilon < U_0$, which corresponds to electron-hole transition at the boundary of the potential region.
We note that the *electrostatic potential* $U_0$ increases and drops sharply at the boundaries, whereas the *magnetic vector potential* is continuous. However the electrostatic potential is uniform inside the barrier region. This significant difference in the spatial dependence suggests that we may consider the effect due to each potential separately as an adequate approximation. As a matter of fact, we take into account the refraction due to the electrostatic potential barrier first, and then deal with the transmission of the new state in the magnetic barrier. Consequently, the *energy - width* relationship, determining the condition of complete reflection, must now include the energy in the barrier region $\epsilon - U_0$. As a result, the transmission drops to zero in the vicinity of $\epsilon = U_0$, as we can clearly see from Fig.\[\[FIG:4\]\].
We define the transmission resonances as asymmetric the peaks of the transmission for both electron and hole states in the barrier region. The term comes from the theory of the *Klein paradox* and is employed to separate the head-on complete transmission with its peaks corresponding to the different angles of incidence. In contrast to the Klein paradox which persists for the barriers of arbitrary width and height, the resonances occur for different angles of incidence and the particle energies which depend on the above mentioned parameters. As far as the transmission resonances in the presence of magnetic field are concerned, they exhibit similar properties to the case of pure electric barrier.
We also conclude that the transmission resonances disappear in the limit when $U_0 \to 0$, i.e., when only the magnetic barrier is present. It is interesting to note that even for a finite value of $U_0$, there are no peaks for $\epsilon < U_0$. It could be explained by the fact that for incoming particle energies close to the barrier height, the transmission is suppressed. Analyzing the transmission probability dependence on the barrier width, we confirm that there exists a critical width $w_{\rm cr}$ , such that for any barrier whose width exceeds this critical value, the transmission is completely suppressed. This corresponds to a sharp drop of transmission next to the critical value of the barrier width. This critical value decreases with increasing electron energy. For $w < w_{\rm cr}$, the transmission exhibits peaks and somewhat oscillatory but not periodic behavior.
Angular dependence of the electron transmission probability, presented in Fig.\[\[FIG:5\]\], also demonstrates several interesting features. The transmission maximum no longer corresponds to head-on collision, but is shifted to a finite angle which is determined by the incoming electron energy. Transmission dependence shows the decrease with drastically different properties, corresponding to the larger and smaller angles compared to the angle with the largest transmission probability. For a larger angle of incidence, the decrease is moderate. However, in the opposite direction, we observe a sharp drop. For the barrier width dependence, we once again see oscillatory behavior (see the density plot, panel $(a)$ of Fig.\[\[FIG:5\]\]). There are also angular transmission resonances present.
Concluding Remarks
==================
In this paper, we Presented a formalism for calculating the energy band structure for an electrostatically modulated single-layer of graphene in the presence of a uniform perpendicular magnetic field. At low and high magnetic fields, the Landau levels are broadened into non-overlapping minibands. However, at intermediate fields, the commensurability relations between the modulation period and the cyclotron radius cause a mixing of the Landau orbits and overlapping of the subbands. The resulting picture is that of a Hofstadter butterfly for the conduction and valence bands. The onset of the self-similarity in the energy band structure depends on the modulation amplitude. The minibands for graphene always overlap in the presence of an electrostatic modulation, unlike the 2DEG. We also obtained the energy dispersion for graphene. The asymmetry of these curves in the valence and conduction subbands is due to modulation as may be seen from our eigenvalue equation ${\rm Det}\{\tensor{\cal M}\}=0$ where the matrix $\tensor{\cal M}$ is defined in Eq. (\[MM\]).
Also, we addressed the electron tunneling problem in the presence of both electrostatic and magnetic potential barriers within the same region. We have found that transmission is suppressed for incoming electron energies close to the barrier height and for barrier widths exceeding critical values. In contrast, transmission resonances exist for the case of electron-hole transmission at the boundary of the potential region. The angular dependence of the electron transmission probability demonstrates the shift of the transmission maximum to a finite angle of incidence, compared to Klein paradox for purely electrostatic potential barriers.
This research was supported by contract \# FA 9453-07-C-0207 of AFRL.
[9]{}
M. Ya. Azbel, Sov. Phys. JETP, [**19**]{}, 634 (1964).
D. R. Hofstadter, Phys. Rev.B [**14**]{}, 2239 (1976).
Godfrey Gumbs and Paula Fekete, Phys. Rev. B [**56**]{}, 3787 (1997).
G. H. Wannier, Phys. Stat. Sol. b [**100**]{}, 163 (1980).
A. Rauh, G. H. Wannier, and G. Obermair, Phys. Stat. Solidi b [**63**]{}, 215 (1974).
A. Rauh, Phys. Stat. Sol. b [**69**]{}, K9 (1975).
H. W. Neumann and A. Rauh, Phys. Stat. Solidi b [**96**]{}, 233 (1979).
Y. Hasegawa, Y. Hatsugai, M. Kohmoto, and G. Montambaux, Phys. Rev. B [**41**]{}, 9174 (1990).
Y. Hatsugai and M. Kohmoto, Phys. Rev. B [**42**]{}, 8282 (1990).
D. Pfannkuche and R. R. Gerhardts, Phys. Rev. B [**46**]{}, 12606 (1992).
D. Pfannkuche and R. R. Gerhardts, Surf. Sci. [**263**]{}, 324 (1992).
D. J. Thouless, Phys. Rev. B [**28**]{}, 4272 (1983).
F. Claro, Phys. Stat. Sol. (b) [**104**]{}, K31 (1981).
H. J. Schellnhuber and G. M. Obermair, Phys. Rev. Lett. [**45**]{}, 276 (1980).
T. Perschel and T. Geisel. Phys. Rev. Lett. [**71**]{}, 239 (1993).
X. Wu and S. E. Ulloa, Phys. Rev. B [**47**]{}, 10028 (1993).
H. Silbernauer, J. Phys.: Condens. Matter [**4**]{}, 7355 (1992).
J, B. Sokoloff, Physics Reports [**126**]{}, 189 (1985).
O. Kühn, V. Fessatidis, H. L. Cui, P. E. Selbmann, and N.Horing, Phys. Rev.B [**47**]{}, 19, 13019 (1993).
N. Nemec and G. Cunibert, Phys.Rev.B [**75**]{}, 201404(R) (2007).
P. D. Ye, D. Weiss, R. R. Gerhardts, M. Seeger, K. von Klitzing, K. Eberl, and H. Nickel, Phys. Rev. Lett. [**74**]{}, 3013 (1995).
C. R. Dean et al., Nature [**497**]{}, 598 - 602 (30 May 2013).
L. A. Ponomarenko et al., Nature [**497**]{}, 594 - 597 (30 May 2013).
M. Zarenia, P. Vasilopoulos, and F. M. Peeters, Phys.Rev. B, [**85**]{}, 245426 (2012).
M. I. Katsnelson, K. S. Novoselov, and A. K. Geim, Nature Physics, [**2**]{}, 620 (2006).
Andrii Iurov, Godfrey Gumbs, Oleksiy Roslyak and Danhong Huang, Journal of Physics.: Condensed Matter, [**24**]{}, 1, 015303 (2012)
Andrii Iurov, Godfrey Gumbs, Oleksiy Roslyak and Danhong Huang, Journal of Physics.: Condensed Matter, [**25**]{}, 13, 135502 (2013)
A. De Martino, L. Dell’Anna and R.Egger, Solid State Communications, [**144**]{}, 12 (2007).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '[**Abstract**]{} 0.1cm [We show from exact calculations that a simple tight-binding Hamiltonian with diagonal disorder and long-range hopping integrals, falling off as a power $\mu$ of the inter-site separation, correctly describes the experimentally observed amplitude (close to the value of an ordered ring) and flux-periodicity ($hc/e$) of persistent currents in single-isolated-diffusive normal metal rings of mesoscopic size. Long-range hopping integrals tend to delocalize the electrons even in the presence of disorder resulting orders of magnitude enhancement of persistent current relative to earlier predictions.]{} 0.1cm [*Keywords:*]{} [Model Calculations, Magnetotransport]{}'
author:
- '[Santanu K. Maiti]{} [^1]'
- 'J. Chowdhury'
- 'S. N. Karmakar'
title: 'Persistent currents with long-range hopping in $1$D single-isolated-diffusive rings'
---
The phenomenon of persistent current in mesoscopic normal metal rings has generated a lot of excitement as well as controversy over the past years. In a pioneering work, Büttiker, Imry and Landauer [@butt] predicted that, even in the presence of disorder, an isolated $1$D metallic ring threaded by magnetic flux $\phi$ can support an equilibrium persistent current with periodicity $\phi_0=ch/e$, the flux quantum. Later, experimental observations confirm the existence of persistent currents in isolated mesoscopic rings. However, these experiments yield many results that are not well-understood theoretically even today [@cheu1; @cheu2; @mont; @bouc; @alts; @von; @schm; @ambe; @abra; @bouz; @giam; @burme]. The results of the single loop experiments are significantly different from those for the ensemble of isolated loops. Persistent currents with expected $\phi_0$ periodicity have been observed in isolated single Au rings [@chand] and in a GaAs-AlGaAs ring [@maily]. Levy [*et al.*]{} [@levy] found oscillations with period $\phi_0/2$ rather than $\phi_0$ in an ensemble of $10^7$ independent Cu rings. Similar $\phi_0/2$ oscillations were also reported for an ensemble of disconnected $10^5$ Ag rings [@deb] as well as for an array of $10^5$ isolated GaAs-AlGaAs rings [@reul]. In a recent experiment, Jariwala [*et al.*]{} [@jari] obtained both $\phi_0$ and $\phi_0/2$ periodic persistent currents in an array of thirty diffusive mesoscopic Au rings. Except for the case of the nearly ballistic GaAs-AlGaAs ring [@maily], all the measured currents are in general one or two orders of magnitude larger than those expected from the theory [@cheu1; @cheu2; @bouc; @mont; @von; @alts; @schm; @ambe; @abra; @bouz; @giam]. The diamagnetic response of the measured $\phi_0/2$ oscillations of ensemble-averaged persistent currents near zero magnetic field also contrasts with most predictions [@schm; @ambe].
Free electron theory predicts that at $T=0$, an ordered $1$D metallic ring threaded by magnetic flux $\phi$ supports persistent current with maximum amplitude $I_0=ev_F/L$, where $v_F$ is the Fermi velocity and $L$ is the circumference of the ring. Metals are intrinsically disordered which tends to decrease the persistent current, and the calculations show that the disorder-averaged current $<I>$ crucially depends on the choice of the ensemble [@cheu2; @mont; @bouc]. The magnitude of the current $<I^2>^{1/2}$ is however insensitive to the averaging issues, and is of the order of $I_0l/L$, $l$ being the elastic mean free path of the electrons. This expression remains valid even if one takes into account the finite width of the ring by adding contributions from the transverse channels, since disorder leads to a compensation between the channels [@cheu2; @mont]. However, the measurements on an ensemble of $10^7$ Cu rings [@levy] reported a diamagnetic persistent current of average amplitude $3\times
10^{-3} ev_F/L$ with half a flux-quantum periodicity. Such $\phi_0/2$ oscillations with diamagnetic response were also found in other persistent-current experiments consisting of ensemble of isolated rings [@deb; @reul].
Measurements on single isolated mesoscopic rings on the other hand detected $\phi_0$-periodic persistent currents with amplitudes of the order of $I_0\sim ev_F/L$, (closed to the value for an ordered ring). Theory and experiment [@maily] seem to agree only when disorder is weak. However, the amplitudes of the currents in single-isolated-diffusive gold rings [@chand] were two orders of magnitude larger than the theoretical estimates. This discrepancy initiated intense theoretical activity, and it is generally believed that the electron-electron correlation plays an important role in the disordered diffusive rings [@abra; @bouz; @giam], though the physical origin behind this enhancement of persistent current is still unclear.
In this letter we will address the problem of enhancement of persistent current in single-isolated-diffusive (SID) mesoscopic rings. The large amplitude of the observed currents in SID mesoscopic rings strongly challenges the conventional theories of persistent current. It indicates that in all the previous models some fundamental mechanism is missing which could compensate the effect of impurities, and thus prevents reduction of the current due to disorder. The existing theories are basically within the framework of the Anderson model where the transport properties of the electrons are dominated by the localization phenomenon that essentially reduces the persistent current. In a recent work, Balagurov [*et al.*]{} [@bala] have shown that the electrons become delocalized if one includes long-range hopping (LRH) integrals in the Anderson model.
We describe a $N$-site ring enclosing a magnetic flux $\phi$ (in units of the elementary flux quantum $\phi_0$) by the following Hamiltonian in the Wannier basis $$H=\sum_i\epsilon_i c_{i}^{\dagger}c_{i}
+\sum_{i\ne j} v_{i,j}\left[e^{-i \theta} c_{i}^
{\dagger} c_{j}+ h.c. \right]
\label{hamil}$$ where $\epsilon_i$’s are the site energies, $v_{i,j}$’s are the hopping integrals, and $\theta=\frac{2\pi\phi}{N}(|i-j|)$. The $\epsilon_i$’s are uncorrelated random variables drawn from some distributive function $P(\epsilon)$, and, the non-random LRH integrals are taken as $v_{i,j}=
v/|i-j|^{\mu}$, $v$ being a constant representing the nearest-neighbor hopping (NNH) integrals. It should be noted that the physical domain of the exponent $\mu$ is determined by the boundness of the spectrum [@bala]. The choice of the distribution function $P(\epsilon)$ which we shall use are the “box" distribution $$P(\epsilon)=\frac{\Theta\left(\frac{W}{2}-|\epsilon|\right)}{W}
\label{site1}$$ of width $W$, and the “binary-alloy" distribution $$P(\epsilon)=c\delta\left(\epsilon-\epsilon_A\right)+(1-c)\delta
\left(\epsilon-\epsilon_B\right)
\label{site2}$$ where $c$ and $(1-c)$ are respectively the concentrations of two types of atoms with site energies $\epsilon_A$ and $\epsilon_B$. In the following we use the units $h=e=1$.
For an ordered ring, setting $\epsilon_i=0$ for all $i$, the energy of the $n$th eigenstate can be expressed as $$E_{n}(\phi)=\sum_{m=1}^{N-1}\frac{2v}{m^{\mu}}\cos\left[\frac{2\pi m}{N}
\left(n+\phi \right)\right]
\label{energy}$$ where $m$ is an integer. The current carried by this eigenstate is given by $$I_{n}(\phi)=\left(\frac{4\pi v}{N}\right)\sum_{m=1}^{N-1}m^{(1-\mu)}
\sin\left[\frac{2\pi m}{N}\left(n+\phi \right)\right].
\label{current}$$
For spinless electrons, we can express the total current at $T=0$ in the following form $$I(\phi)=\sum_{n} I_n(\phi)
\label{spinlesscurrent}$$ where $N_e$ is the number of electrons and $-\lfloor N_e/2\rfloor \leq n < \lfloor N_e/2 \rfloor$ ($\lfloor z\rfloor$ denotes the integer part of $z$). In the above expression we restrict $\phi$ in the domains $-0.5 \leq \phi <0.5$ and $0 \leq \phi < 1$ for the systems respectively with odd and even number of electrons. The $I(\phi)$ versus $\phi$ curves for some representative impurity-free systems are plotted in Fig. \[spinless\]. The persistent current as a function of $\phi$ always exhibits discontinuity at certain points as long as there is no impurity in the system. The ground states are degenerate at these points of discontinuity due to the crossing of the energy levels. In the presence of all LRH integrals it is found that the amplitude of the current initially increases as we increase the system size, but eventually it falls when the system becomes larger. This is due to the fact that as we increase the number of sites, the Hamiltonian Eq. (\[hamil\]) includes some additional higher order hopping integrals which causes an increase in the net velocity of the electrons, but after certain system size this increment in velocity drops to zero because the additional hopping integrals are then between far enough sites giving negligible contributions.
If we take into account the spin of the electrons, then the total persistent current in an ordered ring with even number of electrons is given by $$I(\phi)=2\sum_{n} I_n(\phi)
\label{fermioniceven}$$ where $-\lfloor N_e/4 \rfloor \leq n < \lfloor N_e/4 \rfloor$, and here we restrict $\phi$ in the domain $0 \leq \phi < 1$ if $N_e/2$ is even while in the domain $-0.5 \leq \phi < 0.5$ if $N_e/2$ is odd. If the system contains odd number of electrons, we have $$I(\phi)=2\sum_{n} I_n(\phi) + I_{n^{\prime}}(\phi)
\label{fermionicodd}$$ where $-\lfloor \left(N_e-1\right)/4 \rfloor \leq n < \lfloor
\left(N_e-1\right)/4 \rfloor$. The quantum number $n^{\prime}$ has to be determined in the following way.
Any odd value of $N_e$ can be expressed into the form $(4p\pm 1)$ where $p=1,2,3,\ldots$, and the quantum number $n^{\prime}$ becomes equal to $\pm p$ corresponding to these two forms of $N_e$. In the above expression we restrict $\phi$ in the range $0 \leq \phi < 1$ when $N_e$ is of the form $(4p-1)$, whereas $\phi$ is to be bounded between $-0.5 \leq \phi < 0.5$ when $N_e$ has the form $(4p+1)$. We do not display the $I-\phi$ characteristics for spin fermionic systems as in the absence of electron-electron interaction, the electron spin cannot alter the characteristic features of the persistent currents from those presented in Fig. \[spinless\].
Now we address the problem of persistent current in SID mesoscopic rings using the Hamiltonian Eq. (\[hamil\]), and, in this study we do not consider the spin of the electrons as it will not change the qualitative behavior of the currents within the one-electron picture. We present exact calculation of the currents in the presence of all LRH integrals in the tight-binding Hamiltonian with diagonal disorder, and it involves exact numerical diagonalization of the Hamiltonian matrices. The results for some representative examples are given in Fig. \[random\] and Fig. \[binary\]. In Fig. \[random\], we plot $I(\phi)$ versus $\phi$ curves for the systems with $v=-1,~ \mu=1.4, ~N=50$, and, $N_e=20$ and $23$, where disorder is introduced by random choice of $\epsilon_i$’s from the “box" distribution Eq. (\[site1\]) setting $W=1$. The solid lines correspond to three microscopic configurations of disorder, while the dotted lines are for the ordered cases obtained by setting all $\epsilon_i$’s equal to zero. Fig. \[binary\] is the $I-\phi$ characteristics of the systems with the same set of parameters as those in Fig. \[random\], where site energies are chosen randomly from the “binary-alloy" distribution given by Eq. (\[site2\]) with $c=0.5$ and $\delta=|\epsilon_A-\epsilon_B|=1$. We consider three typical disordered configurations of the ring compatible with the “binary-alloy" distribution, and the $I-\phi$ curves for these configurations are represented by solid lines in Fig. \[binary\]. The dotted lines in this figure are identical to those in Fig. \[random\].
Let us now analyze the results presented in Fig. \[random\] and Fig. \[binary\]. We see that for the present model of SID mesoscopic ring, the persistent currents are always periodic in $\phi$ with periodicity $\phi_0$. Fig. \[random\] and Fig. \[binary\] clearly show that in the presence of all LRH integrals in the Hamiltonian, the $I-\phi$ characteristics of a given SID mesoscopic ring are almost insensitive to the microscopic configuration of disorder of the ring, and, we have checked considering $100$ distinct configurations of the given system that all the $I-\phi$ curves nearly collapse to a single curve. These figures are self-explanatory to revel the fact that as we vary the microscopic configurations of the ring, the persistent currents do not fluctuate in sign and the fluctuation in magnitude becomes exceedingly small. The most interesting result is that in the present model persistent current is not reduced by disorder, and it is apparent from Fig. \[random\] and Fig. \[binary\] that the currents in the disordered rings are of the same order of magnitude as the current in the ordered ring. We have also seen that the decrease in amplitude of the current is quite small even if we increase the strength of disorder. These results dramatically differ from the previous predictions according to which the persistent currents exhibit strong fluctuations both in sign and magnitude depending on the realization of disorder, and the currents are reduced by several orders of magnitude due to disorder. The orders of magnitude reduction of the persistent currents suggested earlier are basically due to the tendency of localization of the electrons in the Anderson-like models. On the other hand, the present tight-binding model with all LRH integrals supports extended electronic eigenstates even in the presence of disorder [@bala], and for this reason persistent currents are not reduced by the impurities. Our results are in good agreement with the experimental observations [@chand; @maily].
We have also noticed certain interesting features of the $I-\phi$ curves that are characteristics of any disordered ring. The discontinuity in $I(\phi)$ as a function of $\phi$ are characteristics of the ordered systems which disappears due to disorder, and the current in the disordered rings are always zero at these points of discontinuity. This result can be easily understood on a very general ground. We can treat disorder as a perturbation over the ordered situation that lifts the degeneracy at the crossing points of the unperturbed energy levels. So gaps open up at the ground level in the presence of disorder making $I(\phi)$ a continuous function of $\phi$, and, also $I(\phi)$ becomes exactly equal to zero at these points of discontinuity.
In conclusion, we have investigated the behavior of persistent currents in SID mesoscopic rings by a simple model including all LRH integrals in the usual Anderson model. Our exact calculations show that both the sign and magnitude of the experimentally observed currents can be explained from the present model. In this work we have convincingly established that the essential physical mechanisms are the LRH integrals that accounts for the observed behavior of persistent currents in SID mesoscopic rings.
[99]{}
M. Büttiker, Y. Imry, and R. Landauer, Phys. Lett. **96A**, 365 (1983). H. F. Cheung, Y. Gefen, E. K. Riedel, and W. H. Shih, Phys. Rev. B **37**, 6050 (1988). H. F. Cheung and E. K. Riedel, Phys. Rev. Lett. **62**, 587 (1989). G. Montambaux, H. Bouchiat, D. Sigeti, and R. Friesner, Phys. Rev. B **42**, 7647 (1990). H. Bouchiat and G. Montambaux, J. Phys. (Paris) **50**, 2695 (1989). B. L. Altshuler, Y. Gefen, and Y. Imry, Phys. Rev. Lett. **66**, 88 (1991). F. von Oppen and E. K. Riedel, Phys. Rev. Lett. **66**, 84 (1991). A. Schmid, Phys. Rev. Lett. **66**, 80 (1991). V. Ambegaokar and U. Eckern, Phys. Rev. Lett. **65**, 381 (1990). M. Abraham and R. Berkovits, Phys. Rev. Lett. **70**, 1509 (1993). G. Bouzerar, D. Poilblanc, and G. Montambaux, Phys. Rev. B **49**, 8258 (1994). T. Giamarchi and B. S. Shastry, Phys. Rev. B **51**, 10915 (1995). G. Burmeister and K. Maschke, Phys. Rev. B **65**, 155333 (2002). V. Chandrasekhar, R. A. Webb, M. J. Brady, M. B. Ketchen, W. J. Gallagher, and A. Kleinsasser, Phys. Rev. Lett. **67**, 3578 (1991). D. Mailly, C. Chapelier, and A. Benoit, Phys. Rev. Lett. **70**, 2020 (1993). L. P. Levy, G. Dolan, J. Dunsmuir, and H. Bouchiat, Phys. Rev. Lett. **64**, 2074 (1990). R. Deblock, R. Bel, B. Reulet, H. Bouchiat, and D. Mailly, Phys. Rev. Lett. **89**, 206803 (2002). B. Reulet, M. Ramin, H. Bouchiat, and D. Mailly, Phys. Rev. Lett. **75**, 124 (1995). E. M. Q. Jariwala, P. Mohanty, M. B. Ketchen, and R. A. Webb, Phys. Rev. Lett. **86**, 1594 (2001). B. Balagurov, D. B. Malyshev, V. A. Adame, and F. Dominguez, Phys. Rev. B **69**, 104204 (2004).
[^1]: Corresponding author: Santanu K. Maiti\
E-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Various integrable geodesic flows on Lie groups are shown to arise by taking moments of a geodesic Vlasov equation on the group of canonical transformations. This was already known for both the one- and two-component Camassa-Holm systems [@GiHoTr2005; @GiHoTr2007]. The present paper extends our earlier work to recover another integrable system of ODE’s that was recently introduced by Bloch and Iserles [@BlIs2006]. Solutions of the Bloch-Iserles system are found to arise from the Klimontovich solution of the geodesic Vlasov equation. These solutions are shown to form one of the legs of a dual pair of momentum maps. The Lie-Poisson structures for the dynamics of truncated moment hierarchies are also presented in this context.'
author:
- |
Darryl D. Holm$^{1,\,2}$, Cesare Tronci$^{3}\!$\
[$^1$ *Department of Mathematics, Imperial College London, London SW7 2AZ, UK*]{}\
[$^2$ *Institute for Mathematical Sciences, Imperial College London, 53 Prince’s Gate, London SW7 2PG, UK*]{}\
[$^3$*Section de Mathématiques, École Polytechnique Fédérale de Lausanne, CH-1015 Lausanne, Switzerland*]{}\
\
title: |
The geodesic Vlasov equation\
and its integrable moment closures
---
Introduction
============
Kinetic equations govern the evolution of probability distributions in the phase space of many-particle systems in non-equilibrium statistical mechanics. For example, the phase-space probability distribution of a many-particle system whose correlations are negligible is governed by the [*collisionless Boltzmann*]{} equation, also known as [*Vlasov equation*]{} [@Vl1961]. This equation encodes evolution of the Vlasov single-particle probability distribution $f({\bf q},{\bf p},t)$ as conservation along phase space trajectories, written as $$\label{VlasovChar}
\frac{df}{dt}=0 \qquad\text{along}\qquad \dot{\bf q}=\frac{{{\partial}}h}{{{\partial}}\bf p}\,, \quad \dot{\bf p}=-\,\frac{{{\partial}}h}{{{\partial}}\bf q}\,,$$ where $h$ is the single-particle Hamiltonian, often expressed as the sum of kinetic and potential energies $$h=\frac{|{\bf p}|^2}{2}+V({\bf q})$$ in physical applications. Applying the canonical Poisson bracket $\{\cdot,\,\cdot\}$ in the phase space variables ($\bf q,p$) expresses the Vlasov equation in its familiar form $$\label{Vlasov}
{\frac{\partial f}{\partial t}} +\big\{f,h\big\}=0 \,.$$ The Hamiltonian structure of this system is well known [@MaWe1981]. Namely, the Vlasov equation possesses a Lie-Poisson bracket defined on the Lie algebra of canonical transformations, such that the Liouville theorem for preservation of the volume on phase space entirely characterizes Vlasov dynamics.
The moment method, widely used in probability theory, provides approximate descriptions of the Vlasov solutions. Moments are functionals of the distribution function $f$ obtained by projections onto the space of phase space polynomials (symmetric tensors). Since the Vlasov distribution depends on both position [**q**]{} and momentum [**p**]{}, one may define two different types of moments. These are the [*kinetic moments*]{} and the [*statistical moments*]{}. Kinetic moments are given by projection of the Vlasov distribution $f({\bf q,p})$ onto the $n$-th power of the single-particle momentum, $\bf p$. In contrast, statistical moments are integrals of $f({\bf q,p})$ against the $n$-th power of the full phase space vector, ${\bf z}=({\bf q,p})$. The remarkable property of these two hierarchies of moment projections is that they each define equivariant momentum maps [@HoLySc1990; @ScWe1994; @GiHoTr2008]. Consequently, the resulting moment dynamics is again Lie-Poisson. Moment equations possess interesting closures, which are given by the particular Lie algebra structure determining their Lie-Poisson bracket. For example, kinetic moments of the Vlasov equation at zero-th and first order yield the familiar closure known as *ideal fluid dynamics*.
Remarkably, these kinetic moment equations are associated to a family of integrable dynamical systems, whose most famous example is probably the Benney system for shallow water dynamics [@Be1973; @Gi1981]. This convergence of different areas of mathematical physics also occurs for several other integrable equations. For example, as shown in [@GiHoTr2005], a specific form of the first-order kinetic moment equation yields the Camassa-Holm equation [@CaHo1993]. Extending the system to include the zero-th order moment yields another integrable system; the two-component Camassa-Holm system [@GiHoTr2007]. Interestingly enough, these Camassa-Holm systems are geodesic flows on *different* Lie groups, arising as moment closures of the *same* kinetic equation, called here the [[******]{}geodesic Vlasov equation]{} or EP${Can}$. The latter acronym refers to the Euler-Poincaré (EP) equation on the group of canonical transformations $Can$ acting on phase space $T^*Q$.[^1] A special case of EP${Can}$ for canonical transformations whose generating functions are *linear* in the canonical momentum has recently appeared in the theory of *metamorphoses* in imaging science [@HoTrYo2007]. This paper reviews the theory of geodesic equations on the statistical moments and shows how such equations possess an additional interesting closure, which is related to the space of purely quadratic Hamiltonian functions. We find that such a closure yields a particular case of yet another integrable system, recently discovered by Bloch and Iserles [@BlIsMaRa2005; @BlIs2006]. Moreover, extending to inhomogeneous quadratic Hamiltonians yields complete equivalence between moment equations and the Bloch-Iserles (BI) system.
#### Plan
The rest of this section adds a few more remarks about the Lie-Poisson bracket for the Vlasov equation. Section 2 is devoted to the Hamiltonian structure of the Vlasov moments and their truncations. Section 3 formulates the geodesic Vlasov equation, presents its dual pair and illustrates the geometric footing of kinetic theory. Section 4 shows how both of the Camassa-Holm systems are obtained as geodesic equations on kinetic moments. The last section derives the BI system from the statistical moment equations and presents the corresponding Klimontovich solutions.
The Vlasov kinetic equation
---------------------------
The Vlasov equation is a Lie-Poisson Hamiltonian system on the group of canonical transformations of the phase space $T^*Q$ for a configuration manifold $Q$ [@MaWe1981]. The dynamics of Lie-Poisson systems takes place on the dual $\mathfrak{g}^*$ of the Lie algebra $\mathfrak{g}$ of the symmetry group $G$. In this case $G={Can}(T^*Q)$ and $\mathfrak{g}=\mathfrak{X}_{\rm can}(T^*Q)$. That is, the Lie algebra is the infinite-dimensional space of Hamiltonian vector fields. Given the Lie algebra isomorphism between Hamiltonian vector fields and phase-space functions ($\mathfrak{X}_{\rm can}\simeq\mathcal{F}$), the dynamical variable is a phase-space distribution $f({\bf q,p})$, i.e., a density on phase space ($f\in\mathcal{F}^*\!\simeq{\rm
Den}$). Upon using the definition of canonical Poisson bracket $\left\{\cdot,\cdot\right\}$, the Vlasov Lie-Poisson structure is found to be $$\{F,H\}[f]=\iint f ({\bf q,p})\left\{\frac{\delta F}{\delta
f},\frac{\delta H}{\delta f}\right\}\,{\rm d}^{K} {\bf q}\ {\rm
d}^{K} {\bf p}$$ where $K={\rm dim}(Q)$. The Vlasov equation (\[Vlasov\]) is recovered upon choosing $F=f$ and $h=\delta H/\delta f$.
In many physical applications, the Vlasov Hamiltonian is the sum of kinetic and potential energy. For example, electrostatic or gravitational interactions are governed in the absence of collisions by the Poisson-Vlasov system whose Hamiltonian is given by $$\label{VlasovPoisson}
H[f]=\iint\! f({\bf q,p})\left(\frac12\,|{\bf
p}|^2+\Delta^{-1\!}\!\int\! f({\bf q,p'})\ {\rm d}^{K\,}{\bf
p}'\right) {\rm d}^K{\bf q}\ {\rm d}^K{\bf p} \,,$$ where $\Delta^{-1}$ denotes convolution with the Green’s function of the Laplace operator.
Hamiltonian structure of Vlasov moments
=======================================
The moment method is a popular approach in kinetic systems theory. This approach is justified geometrically because taking moments of the Vlasov distribution is a momentum map [@HoLySc1990; @ScWe1994; @GiHoTr2008]. This momentum map arises via the dual of a Lie algebra homomorphism arising from the well-known isomorphism between symmetric tensors and polynomials. The main point is that this momentum map endows the space of symmetric tensors with a Lie bracket, thereby generating a well defined Lie algebra. In what follows, we shall analyze the cases of kinetic and statistical moments separately and then discuss their similarities.
Kinetic moments and the Schouten concomitant
--------------------------------------------
[[******]{}Kinetic moments]{} are constructed from the following fiber integral [@QiTa2004] $$\begin{gathered}
\label{FibIntMom}
A_n({\bf q}, t):=\int_{T^*_{\bf q}Q\!}\left({\bf p}\cdot{\rm d}{\bf
q}\right)^n f({\bf q,p},t)\ {\rm d}^K{\bf q}\wedge{\rm d}^K{\bf p}
\\
= \sum_{i_1\dots i_n=1}^K
\int_{T^*_{\bf q}Q} p_{\,i_1}\dots p_{\,i_n} \
{\rm d}q^{i_1}\otimes\dots\otimes
{\rm d}q^{i_n}\,f({\bf q, p}, t)\ {\rm d}^K{\bf
q}\wedge{\rm d}^K{\bf p}
\\
=\sum_{i_1\dots i_n=1}^K\!
\big(A_n({\bf q}, t)\big)_{i_1\dots i_n\,} \
{\rm d}q^{i_1}\otimes\dots\otimes{\rm d}q^{i_n}
\otimes{\rm d}^K{\bf q} \,,\end{gathered}$$ where ${\bf p}\cdot{\rm d}{\bf q}$ denotes the canonical one form (canonical momentum) and ${\rm d}^K{\bf q}$ is the volume element on the configuration space $Q$. This construction projects the Vlasov distribution onto the space of symmetric tensors. In particular, kinetic moments are defined as symmetric covariant tensor fields carrying the volume element. That is, they are [*symmetric covariant tensor densities*]{}.
The moments are functionals of the Vlasov density $f$. Hence, their variational derivative may be computed by applying the chain rule as $$\begin{aligned}
\frac{\delta F}{\delta f}
= \sum_{n=0}^\infty\,\frac{\delta F}{\delta
A_n}
\contract
\frac{\delta
A_n}{\delta f}
&
:=\sum_{n=0}^\infty\,\sum_{\,i_1\dots i_n=1}^K\!\frac{\delta
(A_n)_{i_1\dots i_n}}{\delta f}\,\frac{\delta F}{\delta
(A_n)_{i_1\dots i_n}}
\\&
=\sum_{n=0}^\infty\,\sum_{\,i_1\dots i_n=1}^K
p_{\,i_1}\dots p_{\,i_n}
\,\frac{\delta F}{\delta
(A_n)_{i_1\dots i_n}}
\\&
=: \sum_{n=0}^\infty\, \frac{\delta F}{\delta A_n} \contract\, {\bf
p}^{n\!}
\,,\end{aligned}$$ which explicitly defines the contraction operation $\contract$. This chain rule formula expresses the Lie algebra homomorphism (isomorphism) from symmetric tensors to polynomials, whose dual is the momentum map associated to the moments [@GiHoTr2008]. Inserting the chain rule formula into the Vlasov bracket yields the Lie-Poisson bracket for moments, $$\left\{F,G\right\}[A]=- \!\sum_{n,m=0\,}^\infty
\int\! A_{m+n-1}({\bf q})\,\contract\!\left[ \frac{\delta F}{\delta
A_n},\frac{\delta G}{\delta A_m}\right]
\,{\rm d}^3{\bf q}
\label{moment-LPB}$$ in which the bracket[^2] $$\label{MomSchouten}
\Bigg[ \frac{\delta F}{\delta A_n}\,,\,\frac{\delta G}{\delta
A_m}\Bigg]=\, \text{\Large$\mathcal{S}$}\!\left(n
\,\Bigg(\frac{\delta F}{\delta
A_n}\cdot\nabla\Bigg)\otimes\frac{\delta G}{\delta A_m} \, -\,
m\, \Bigg(\frac{\delta G}{\delta
A_m}\cdot\nabla\Bigg)\otimes\frac{\delta F}{\delta A_n}
\right)$$ is inherited from the canonical Poisson bracket. Here the notation $A\cdot B$ for one-index contraction between covariant and contravariant tensors is written as $(A\cdot B)_{ij...}^{hl...} = A_{ij...k}B^{khl...}$ and analogously for $B\cdot A=(B\cdot A)^{km...}_{jl...} =
B^{km...i}A_{ijl...}$. This bracket is well known in differential geometry as an invariant differential operator of first order [@Ni1955]. In fact, this operation is a Lie bracket, which is known as the [*Schouten concomitant*]{} or [*symmetric Schouten bracket*]{}. See, e.g., [@GiHoTr2008] for more discussions and references.
$\quad$\
In one dimension, the moment Lie-Poisson structure (\[moment-LPB\]) is the Kupershmidt-Manin bracket [@KuMa1978] which was found in the context of the integrable Benney system in shallow water theory. Lebedev was the first to establish its relation with the Lie algebra of Hamiltonian vector fields in [@Le1979] and Gibbons recognized later [@Gi1981] its direct relation to the Vlasov flow. In higher dimensions, Kupershmidt introduced a multi-index notation [@Ku1987], corresponding to the occupation number representation of the symmetric Schouten bracket. This observation suggested the quantum-like framework for kinetic moments in [@GiHoTr2008], where the moment space is described in terms of a bosonic Fock space.
The moment algebra comprises symmetric contravariant tensor fields and these may be characterized as the Fock space represented by a direct sum of symmetric powers of vector fields given by $$\mathfrak{g}:=\bigoplus_{n=0}^\infty\left(\,\bigvee_{i=0}^{n}\,\mathfrak{X}(Q)\,\right)
\qquad\text{with}\quad
\bigvee_{i=0}^{n}\mathfrak{X}
:=\,\text{\large$\mathcal{S}$}\!\left(\bigotimes_{i=0}^n\,\mathfrak{X}\right)
=:\mathfrak{g}_n
\,.
\label{V-devil-def}$$ This is reminiscent of the [*universal enveloping algebra of the diffeomorphism group*]{} Diff($Q$), which is the enveloping algebra $\mathcal{U}(\mathfrak{X})$ of vector fields $\mathfrak{X}(Q)$ on the configuration space $Q$. It is a standard result that the graded structure of an enveloping algebra possesses a Poisson bracket structure [@DaSWe1999].
$\,$\
Interestingly enough, the Schouten concomitant identifies the kinetic moment algebra with the Lie algebra of symbols of differential operators. This identification is quite suggestive, since differential symbols are known to be a subalgebra of the Lie algebra of pseudo-differential symbols [@KhZa1995]. While the differential symbols (the moment algebra) are supposed to have no underlying Lie group structure, the group ${\Psi}D$ of pseudo-differential operators is a well defined Poisson-Lie group [@KhZa1995]. This suggests that the characterization of coadjoint orbits for moment dynamics requires the complete Poisson-Lie group structure of pseudo-differential operators. A similar direction involving vector fields was followed by Ovsienko and Roger [@OvRo1999]. Also, the appearance of the Wick-ordered product from quantum theory in this Poisson-Lie group context [@KhZa1995; @OvRo1999] implies a further relation to moment dynamics, whose quantum-like creation and annihilation operators were presented in [@GiHoTr2008].
The moment algebra carries a graded structure ($\mathfrak{g}=\text{\large$\oplus$}_i\,\mathfrak{g}_i$) with filtration $$\label{KinMomFiltr}
\big[\,\mathfrak{g}_{n\,},\mathfrak{g}_m\big]\subseteq
\mathfrak{g}_{n+m-1}$$ thereby recovering the space of vector fields $\mathfrak{g}_1=\mathfrak{X}$ as a particular subalgebra. The *largest* subalgebra is however $\mathfrak{g}_{0\,}\text{\large$\oplus$}\,\mathfrak{g}_1\simeq\mathfrak{X}\,\circledS\,\mathcal{F}$, i.e. the semidirect product of vector fields with scalar functions. This space appears in the description of ideal compressible fluids, where the $\mathfrak{X}$-variable is the fluid velocity and the $\mathcal{F}$-variable is associated to the fluid density. Geodesic motion on this space has also recently appeared in the *metamorphosis* process in imaging science [@HoTrYo2007].
The moment equations may be written in Lie-Poisson form as $${\frac{\partial A_m}{\partial t}}=\,-\!\sum_{m,n=0}^\infty
{\textrm{\large ad}}^*_\text{\!\small${\frac{\delta H}{\delta A_n}}$}\, A_{n+m-1}$$ where ad$^*$ is the Lie algebraic coadjoint operation defined using the pairing $$\sum_{k,n=0}^\infty\big\langle\,
{\textrm{\large ad}}^*_{\beta_n} \,A_{k\,},\,\alpha_{k-n+1}\big\rangle:=
\sum_{k,n=0}^\infty\big\langle\,
A_{k\,},\,\left[\beta_n,\,\alpha_{k-n+1}\right]\big\rangle \,.$$ The explicit expression for ad$^*$ is given in [@GiHoTr2008] in any number of dimensions. Here we present the one-dimensional case which will be needed in the following sections. The Schouten concomitant in equation (\[MomSchouten\]) assumes a particularly simple form in 1D $$\left[\alpha_m,\beta_n\right]\,=\,m\,\alpha_n\,\partial_q\beta_n\,-\,n\,\beta_n\,\partial_q\alpha_m$$ where $q$ is the spatial variable. Simple use of integration by parts yields the following covariant tensor density of rank $k-n+1$: $${\textrm{\large ad}}_{\beta_{n}}^{\ast}A_{k}=\left( k+1\right) \,
A_{k}\,\partial_q \beta_{n} + n\,\beta_{n}\,\partial_q A_{k}
\,.
\label{Kirillov-ad-action}$$ This operation was introduced by Kirillov [@Ki1982], who first envisioned the possibility of a Lie-Poisson bracket on the symmetric Schouten algebra. Familiar versions of this operator with $n=1$ or $k=n$ arise in the theory of ideal fluid dynamics, soliton dynamics and image matching, while very little is known for other values on $n,k$. Some features of this intriguing open question are investigated further below, in dealing with truncations of moment hierarchies.
Statistical moments and their Lie-Poisson structure
---------------------------------------------------
As we have seen, the fiber integral defining the [*kinetic moment*]{} hierarchy in (\[FibIntMom\]) requires a kinetic equation on a cotangent bundle. In contrast, the notion of [*statistical moments*]{} is given on a symplectic vector space. Upon denoting $\bf z=(q,p)$, the definition of the $n$-th statistical moment is given by $$X^n(t):=\int {\bf z}^n f({\bf z},t)\ {\rm d}^N{\bf z}$$ where the upper index $n$ in the integrand denotes tensor power (${\bf z}^n=\otimes^n{\bf z}$), while for the time-dependent tensors $X^n(t)$ it denotes the tensor rank. This definition places the statistical moments and kinetic moments into the same mathematical framework. The first observation is that statistical moments are symmetric contravariant tensors on phase space, which is now a symplectic vector space $V$ of even dimension $N=2K$ (eventually $V=\mathbb{R}^N$) with elements ${{\bf z}}=z^i {\bf
e}_i\in V$.
The moment Poisson bracket for statistical moments may be obtained by following exactly the same steps as in the previous discussion for kinetic moments. That is, one inserts the chain rule formula $$\begin{aligned}
\frac{\delta F}{\delta f}
&=&
\sum_{n=0}^\infty\,
\left({\frac{\partial F}{\partial X^n}}\right)_{i_1\dots i_n}\left({\frac{\delta X^n}{\delta f}}\right)^{i_1\dots
i_n}
\nonumber\\
&:=& \sum_{n=0}^\infty\ {\frac{\partial F}{\partial X^n}}\contract{\frac{\delta X^n}{\delta f}}
=
\,\sum_{n=0}^\infty\, {\frac{\partial F}{\partial X^n}}\contract \,{{\bf z}}^{n\!\!}
\label{contract-def}\end{aligned}$$ (and definition of $\contract$) into the Vlasov structure, which may then be written as $$\begin{aligned}
\{F,G\}[f]=\iint\! f({{\bf z}})\left\{\frac{\delta F}{\delta
f},\frac{\delta G}{\delta f}\right\}\, {\rm d}^N{{\bf z}}=\iint\!
f({{\bf z}})\left[\, \mathbb{J}\contract \left( \frac{\partial}{\partial
{{\bf z}}}\frac{\delta F}{\delta f}\otimes\frac{\partial}{\partial
{{\bf z}}}\frac{\delta G}{\delta f}\right)\right] {\rm d}^N{{\bf z}}\end{aligned}$$ where $\mathbb{J}$ is a non-degenerate two form. That is, $\mathbb{J}$ is a $N\times N$ antisymmetric matrix of maximal rank. Since equation (\[contract-def\]) implies $$\frac{\partial}{\partial {{\bf z}}}\frac{\delta F}{\delta f} = \sum_n\,
n\,\frac{\partial F}{\partial X^n}\contract\,{{\bf z}}^{n-1}\,,$$ it follows that the moment Poisson structure is $$\begin{aligned}
\nonumber
\{F,G\}(X)\,&=\!\sum_{n,m=0}^\infty
X^{n+m-2}\contract\left[\frac{\partial F}{\partial
X^n},\frac{\partial G}{\partial X^m}\right]
\\
&=:\!\sum_{n,m=0}^\infty \left\langle
X^{n+m-2\,},\left[\frac{\partial F}{\partial X^n},\frac{\partial
G}{\partial X^m}\right]\right\rangle\end{aligned}$$ where $$\label{mombrkt}
\left[\frac{\partial F}{\partial X^n},\frac{\partial G}{\partial
X^m}\right]:=n\,m\
\text{\large$\mathcal{S}$}\!\left(\frac{\partial F}{\partial
X^n}\cdot \mathbb{J}\cdot \frac{\partial G}{\partial
X^m}\right)$$ is the moment Lie bracket, in which again $\mathcal{S}$ operates to take the symmetric part of its argument. Recall that $\mathbb{J}$ is considered as a contravariant antisymmetric matrix, i.e. it possesses upper indexes $\mathbb{J}^{ij}=-\mathbb{J}^{ji}$.
Thus, again, the isomorphism between symmetric tensors and polynomials produces the momentum map associated with the moments [@ScWe1994]. In turn, this means that the Lie-Poisson bracket for statistical moments is inherited from the Vlasov Lie-Poisson structure. In contrast to the Schouten concomitant (\[MomSchouten\]) for kinetic moments, the Lie bracket for statistical moments in (\[mombrkt\]) still involves the symplectic matrix $\mathbb{J}$ (Poisson tensor). Thus, the dynamics of the statistical moments depends explicitly on the original symplectic structure. This allows, for example, the direct construction of moment invariants (Casimirs) as presented in [@HoLySc1990]. Also, the moment algebra (\[mombrkt\]) involves symmetric tensors that are [*covariant*]{}, rather than contravariant as happens for the Schouten concomitant.
$\quad$\
Statistical moments are important, for example, in the study of beam dynamics in particle accelerators. In this framework, they are defined as [@Ch1983] $$\begin{gathered}
\mathcal{M}_n^{\widehat{n}}(t):=\iint {\bf p}^n\, {\bf
q}^{\widehat{n}}\,f({\bf q,p},t)\ {\rm d}^p{\bf q}\ {\rm d}^p{\bf q}
\\
=\left(\mathcal{M}_n^{\widehat{n}}(t)\right)_{j_1...j_n}^{i_1...i_{\widehat{n}}}
{\bf e}_{i_1}\otimes...\otimes{\bf e}_{i_{\widehat{n}}}\otimes {\bf
e}^{j_1}\otimes...\otimes{\bf e}^{j_n}\end{gathered}$$ where ${\bf e}_k$ is a basis element of the configuration vector space $Q$, while ${\bf e}^k$ is its dual (so that $V=Q\times Q^*$). In 1D, the [*beam emittance*]{} $$\epsilon:=\left(\mathcal{M}_0^2\,\mathcal{M}_2^0-(\mathcal{M}_1^1)^2\right)^{1/2}\,,$$ known as the Courant-Snyder invariant [@CoSn1958], was recognized as a moment invariant (Casimir). This observation led to the study of more general moment invariants [@Dr1990; @HoLySc1990]. A geometric investigation of statistical moments was carried out in [@ScWe1994], where the moment algebra was related to the Heisenberg algebra on phase space. For particle accelerator design, moments are often used in computational efforts to account for space charge effects and other beam-related problems.
As for the kinematic moments, one characterizes the Lie algebra of statistical moments by using the grading, $$\mathfrak{g}:=\bigoplus_{n=0}^\infty\left(\,\bigvee_{i=0}^{n}\,V^*\!\right)=:\bigoplus_{i=0}^\infty\,\mathfrak{g}_i$$ with the filtration $$\label{StatMomFiltr}
\big[\,\mathfrak{g}_{n\,},\mathfrak{g}_m\big]\subseteq
\mathfrak{g}_{n+m-2}$$ which shows how symmetric matrices $\mathfrak{g}_2=V^*\vee V^*={\rm
Sym}^*(N)$ form a particular subalgebra (here we denote by ${\rm
Sym}^*(N)$ covariant symmetric matrices). The largest subalgebra is given by $\mathfrak{g}_0\,\text{\large$\oplus$}\,\mathfrak{g}_1\,\text{\large$\oplus$}\,\mathfrak{g}_2=\mathbb{R}\,\text{\large$\oplus$}\,
V^*\text{\large$\oplus$}\, {\rm Sym}^*(N)$ and it will play a central role in the remainder of this paper.
$\quad$\
Analogously to the Lie algebra of kinetic moments, the statistical moments also carry a bosonic Fock space structure, where appropriate occupation numbers may be defined through the introduction of the multi-index notation ${{\bf z}}^\sigma:=(z^1)^{\sigma_1}\dots(z^N)^{\sigma_N}$. Indeed, taking moments by $X^\sigma(t)=\int {{\bf z}}^\sigma f({{\bf z}},t)\,{\rm d}^N{{\bf z}}$ yields the occupation number representation for statistical Vlasov moments.
The moment equations are written in terms of the Lie algebra coadjoint operator $\rm ad^*$ as $$\begin{aligned}
\nonumber
\frac{dX^m}{dt} &= -\sum_{n=0}^\infty \textrm{\large
ad}^*_\text{\!\small${\frac{\partial H}{\partial X^n}}$}\, X^{n+m-2}
\\
&= -\,m\sum_{n=0}^\infty n\ \text{\large$\mathcal{S}$}\left(
\left({\frac{\partial H}{\partial X^n}}\cdot\mathbb{J}\right)\!\contract\,
X^{m+n-2}\right) \label{Vlasov-moment-eqn}\end{aligned}$$ so that $$\Big(\dot{X}^m\Big)^{\!i_1...i_m}= -\,m\sum_{n=0}^\infty n\
\text{\large$\mathcal{S}$}\!\left(
\left({\frac{\partial H}{\partial X^n}}\right)_{\!\!j_1...j_{n-1}k}\mathbb{J}^{k\,i_m}\Big(X^{m+n-2}\Big)^{\!i_1...i_{m-1}\,j_1...j_{n-1}}\!\right)$$ in which repeated tensorial indexes are summed. Having established the notation for the various operations among moments, we turn next to their application in geodesic Vlasov flows.
Truncation of moment hierarchies
--------------------------------
The truncation of Vlasov moment hierarchies is a typical problem in kinetic theory and, for statistical moments, this question was addressed by Channell [@Ch1995] by using Levi’s decomposition theorem. Channell’s result may be summarized by saying that, for a moment Hamiltonian not depending on the first-order moment (${{\partial}}H/{{\partial}}X^1=0$), the moment hierarchy can be truncated at any order, thereby yielding a truncated Lie-Poisson system. In order to see how this works in practice, let us write the truncated equations at the order $K$ for the hierarchy of statistical moments $\dot{X}^n={\rm ad}^*_{h_m}\,X^{m+n-2}$, where $h_m={{\partial}}H/{{\partial}}X^m$. We have $$\begin{aligned}
\nonumber
\dot{X}^1&={\textrm{\large ad}}^*_{h_2}\,X^{1}+{\textrm{\large ad}}^*_{h_3}\,X^{2}+\,\dots+{\textrm{\large ad}}^*_{h_K}\,X^{K-1}
\\\nonumber
\dot{X}^2&={\textrm{\large ad}}^*_{h_2}\,X^{2}+{\textrm{\large ad}}^*_{h_3}\,X^{3}+\,\dots+{\textrm{\large ad}}^*_{h_K}\,X^{K}
\\
\dot{X}^3&={\textrm{\large ad}}^*_{h_2}\,X^{3}+\,\dots+{\textrm{\large ad}}^*_{h_{K-1}}\,X^{K}
\\\nonumber
&\vdots
\\\nonumber
\dot{X}^{K-1}&={\textrm{\large ad}}^*_{h_2}\,X^{K-1}+{\textrm{\large ad}}^*_{h_3}\,X^{K}
\\\nonumber
\dot{X}^{K}&={\textrm{\large ad}}^*_{h_2}\,X^{K}\end{aligned}$$ and we recognize that the equation for $X^1$ is decoupled, so we restrict to the equations for $X^2,\dots,X^K$. At this point, one verifies that the Lie Poisson bracket for the truncated moment system is given by $$\{F,G\}(X)\,=\!\sum_{n=2}^K\, \sum_{m=2}^{K-n+2}\left\langle
X^{n+m-2\,},\left[\frac{\partial F}{\partial X^n},\frac{\partial
G}{\partial X^m}\right]\right\rangle$$ with the same notation as in (\[mombrkt\]). We recognize that the truncated structure is completely determined by the Lie algebra filtration (\[StatMomFiltr\]) and does not depend on the particular expression of the Lie bracket, which was not used in deriving the truncated system above.
Following similar arguments, one can write the truncated system for kinetic moments as $$\begin{aligned}
\nonumber
{{\partial}}_t A_0&=-{\textrm{\large ad}}^*_{h_1}\,A_{0}-{\textrm{\large ad}}^*_{h_2}\,A_{1}-\,\dots-{\textrm{\large ad}}^*_{h_K}\,A_{K-1}
\\\nonumber
{{\partial}}_t A_1&=-{\textrm{\large ad}}^*_{h_1}\,A_{1}-{\textrm{\large ad}}^*_{h_2}\,A_{2}-\,\dots-{\textrm{\large ad}}^*_{h_K}\,A_{K}
\\
{{\partial}}_t A_2&=-{\textrm{\large ad}}^*_{h_1}\,A_{2}-\,\dots-{\textrm{\large ad}}^*_{h_{K-1}}\,A_{K}
\\\nonumber
&\vdots
\\\nonumber
{{\partial}}_t A_{K-1}&=-{\textrm{\large ad}}^*_{h_1}\,A_{K-1}-{\textrm{\large ad}}^*_{h_2}\,A_{K}
\\\nonumber
{{\partial}}_t A_{K}&=-{\textrm{\large ad}}^*_{h_1}\,A_{K}\end{aligned}$$ where $h_m=\delta H/\delta A_m$ and we have assumed $h_0\equiv0$. The zero-th moment equation decouples and one is left with the remaining equations for $A_1,\dots,A_K$. These equations possess the following bracket structure: $$\{F,G\}(A)\,=-\,\sum_{n=1}^K\, \sum_{m=1}^{K-n+1}\left\langle
A_{n+m-1\,},\left[\frac{\delta F}{\delta A_n},\frac{\delta
G}{\delta A_m}\right]\right\rangle$$ with the same notation as in (\[MomSchouten\]). One recognizes again that the truncated structure is uniquely determined by the Lie algebra filtration (\[KinMomFiltr\]). This fact suggests that a similar approach would also apply to the BBGKY moments of the Liouville equation (reduced probability distributions), whose corresponding Lie algebra is known to possess a similar filtration [@MaMoWe84]. However, we leave the investigation of the BBGKY moment hierarchy for another time.
The geodesic Vlasov equation
============================
We have seen that moment hierarchies are equivalent descriptions of the Vlasov equation, which allow for geometric closures of the kinetic system (e.g. the ideal fluid closure for kinetic moments). The Vlasov equation is a Lie-Poisson equation on the Lie algebra of the group $Can(T^*Q)$ of canonical transformations, and this property is reflected in the Lie-Poisson structure of moment dynamics. It is well known that physical systems with interesting geometric behavior are often geodesic flows on Lie groups with respect to a metric provided by the system’s kinetic energy. The most familiar example is probably rigid body motion, which is governed by geodesic motion on SO(3). Likewise, Euler’s equations for ideal fluids may be interpreted as geodesic motion on the volume-preserving diffeomorphisms $\rm Diff_{\rm vol}(\mathbb{R}^3)$ of the 3D flow domain $\mathbb{R}^3$ [@Ar1966]. Another interesting example of geodesic motion is provided by the EPDiff equation [@HoMaRa1998], which governs geodesic motion on the full diffeomorphism group Diff($\mathbb{R}^n$). In many cases, geodesic flows on Lie groups turn out to be completely integrable Hamiltonian systems. For example, in a one-dimensional flow domain $\mathbb{R}$, EPDiff recovers the Camassa-Holm equation for shallow water waves [@CaHo1993].
The problem of geodesic flow on the symplectic group ([*symplecto-hydrody-namics*]{}) was introduced by Arnold and Khesin in [@ArKe1998]. The present paper pursues this idea by considering geodesic flow on the canonical transformations within the context of Vlasov dynamics. In particular, we develop a [[******]{}geodesic Vlasov equation]{} called EP$Can$ ([*Euler-Poincaré equation on the canonical transformations*]{}) as an extension of previous work in [@GiHoTr2005].
$\quad$\
The name [EP]{.nodecor}$Can$ refers to the Euler-Poincaré equation for geodesic motion on the subgroup of the symplectic transformations arising from Hamiltonian vector fields. The subgroup $Can$ (which could equally well be called [Diff]{.nodecor}$_{\rm can}$) may be identified as the group of smooth invertible *canonical* transformations with smooth inverses. These transformations coincide with symplectic transformations in simple domains such as $\mathbb{R}^{2K}$. In those simpler cases, the geodesic Vlasov equation is also known as ${\rm EP}$Symp [@GiHoTr2005; @GiHoTr2007]
As mentioned in the introduction, the idea to investigate [EP]{.nodecor}$Can$ was motivated by the observation that geodesic equations for kinetic moments were found to include the Camassa-Holm equation [@GiHoTr2005]. In fact, geodesic moment equations arise from EP$Can$ whenever the norm may be expanded as a Taylor series. Later, it was recognized [@GiHoTr2007] that the geodesic moment equations also recover the [*two-component Camassa-Holm equation*]{} [@ChLiZh2005]. The latter is a geodesic flow on a semidirect-product Lie group [@Kuzmin2007]. In order to explain these issues, we shall introduce the EPCan equation and show how it specializes to each integrable case.
Given a symplectic manifold $\mathcal{P}$ of even dimension $N=2K$, the EP$Can$ Vlasov Hamiltonian is defined by $$H[f]=\frac12 \iint f(z)\, \mathcal{G}(z,z')\, f(z') \
{{\rm d}}^{N\!}z\ {{\rm d}}^{N\!}z'
=\frac12\,\big\|f\big\|_\text{\scriptsize$\mathcal{G}$}^{\,2}
\label{EPsympHam}$$ where $z\in \mathcal{P}$ and the kernel $\mathcal{G}$ is chosen so that it defines an appropriate norm on ${\rm Den}(\mathcal{P})$. When dealing with moments, we shall restrict to the special cases $\mathcal{P}=T^*Q$, with $Q$ a general configuration manifold, and $\mathcal{P}=V$, with $V$ a symplectic vector space. The geodesic Vlasov equation (aka EP$Can$) is written simply as $$\label{epcan}
{\frac{\partial f}{\partial t}}=-\left\{ f, {\frac{\delta H}{\delta f}}\right\}=-\Big\{ f,\,
\mathcal{G}*f\Big\}$$ which coincides with Euler’s vorticity equation in 2D when $\mathcal{G}=(-\Delta)^{-1}$.
Euler-Poincaré equations on Hamiltonian vector fields
-----------------------------------------------------
In order to understand how the geodesic Vlasov equation arises from an Euler-Poincaré approach, one starts with an invariant Lagrangian defined on the tangent space of the canonical transformations $\mathcal{L}:TCan\to\mathbb{R}$, which is purely quadratic. By the invariance property, we can write the associated variational principle on the Lie algebra of Hamiltonian vector fields as follows $$\delta \int _{t_0}^{t_1}L[{\bf X}_h]\ {\rm d}t=0$$ where ${\bf X}_h=\mathbb{J}\nabla h$ and $$\begin{aligned}
L[{\bf X}_h]
&:=&
\frac12
\Big\langle\widehat{Q}{\bf X}_\textit{\small h},\,{\bf X}_\textit{\small h}\Big\rangle
=\frac12
\Big\langle\widehat{Q}\,\mathbb{J}\, \nabla h_{\,},\,\mathbb{J} \nabla h\Big\rangle
\nonumber\\
&=&
\frac12
\Big\langle{\rm div}\big(\mathbb{J}\,\widehat{Q}\,\mathbb{J}\, \nabla h\big),\, h\Big\rangle
=:L[h]
\,,\end{aligned}$$ in which $\widehat{Q}$ is taken to be a positive-definite symmetric operator so that $L[{\bf X}_h]$ defines a nondegenerate norm. The Legendre transform $$f=\frac{\delta L}{\delta h}={\rm div}\big(\mathbb{J}\,\widehat{Q}\,\mathbb{J}\, \nabla h\big)
\,\Rightarrow\,
h=\left({\rm div}\,
\mathbb{J}\,\widehat{Q}\,\mathbb{J}\, \nabla
\right)^{\!-1}\!f$$ yields the EP$Can$ Hamiltonian in the Vlasov form $$H[f]=\frac12\,\big\langle f,\widehat{O}^{-1}f\big\rangle$$ with $$\widehat{O}:={\rm div}\,\mathbb{J}\,\widehat{Q}\,\mathbb{J}\, \nabla
\,.$$ This formula specifies the relation between the geodesic Vlasov equation and the geodesic motion on the Hamiltonian vector fields. An interesting case occurs when $\widehat{Q}$ is the ‘flat’ operation $\widehat{Q}\,{\bf X}_\textit{\small h}={\bf X}_{\it h}^{\,\flat}$ which takes contravariant vectors to covariant vectors, so that $${\rm div}\,
\mathbb{J}\left(\mathbb{J}\,\nabla h\right)^{\flat}=
-\,\Delta h
\,.$$ Then the operator $\widehat{O}$ reduces to minus the Laplacian $$\widehat{O}=-\,\Delta$$ and in two dimensions one obtains the Euler Hamiltonian for vorticity dynamics, $$H[\omega]=1/2\,\big\langle\omega,(-\Delta)^{-1\,}\omega\big\rangle
\,,$$ with $\omega=f$. In the more general case when $\widehat{Q}$ is a purely differential operator, one finds that $\widehat{Q}$ and $\mathbb{J}$ commute and thus $\widehat{O}=-\,{\rm div}\,\widehat{Q}\, \nabla$. Also if $\widehat{Q}$ commutes with the divergence, then, one has $\widehat{O}=-\,\widehat{Q}\, \Delta$. However in general, $\widehat{Q}$ is a matrix differential operator that does not commute with $\mathbb{J}$.
The ideal fluid dual pair for the Vlasov equation
-------------------------------------------------
Both the Poisson-Vlasov system (\[VlasovPoisson\]) in plasma physics and the geodesic Vlasov equation (\[EPsympHam\]) possess the single-particle solution (Klimontovich solution) expressed as a delta function in phase space, $$\label{klimsol}
f({{\bf z}},t)=\sum_{a=1}^\mathcal{N}\, w_a\,\delta({{\bf z}}-\boldsymbol{\zeta}_a(t))
\,,$$ where the index $a$ is summed over $a=1,\dots,\mathcal{N}$ and the number $w_a$ is a constant weight associated to each particle. This apparently trivial solution is of fundamental importance in physics and leads to the Klimontovich theory of kinetic equations [@Kl1982]. In the remainder of this paper, we shall see how such a solution emerges within the analysis of integrable moment closures of EP$Can$.
In particular, an interesting situation occurs when one allows for more general solutions of the form $$f({{\bf z}},t)=\sum_{a=1}^\mathcal{N}\int\!
w_a(s)\,\delta({{\bf z}}-\boldsymbol{\zeta}_a(s,t))\ {\rm d}^k s
\label{sing-fsoln-momap}$$ where $s$ is a coordinate on the immersed submanifold $S_a\subset\mathbb{R}^{2K}$ with $k={\rm dim}(S)$ (e.g. $k=1$ for a curve, or $k=2$ for a surface). For simplicity of notation, we have already suppressed the implied subscript $a$ in the arclength $s$ for each $w_a$ and $\boldsymbol{\zeta}_a$. For further simplicity and without loss of generality, we take $\mathcal{N} = 1$ and so suppress the index $a$ in what follows. This is equivalent to treating an isolated singular solution. As one might expect, this is only a notational simplification; not a real restriction.
The ${\boldsymbol{\zeta}}$’s in (\[sing-fsoln-momap\]) belong to the space of embeddings ${\rm Emb}(S,\mathbb{R}^{2K})$. That is, $\boldsymbol{\zeta}:S\hookrightarrow \mathbb{R}^{2K}$. Remarkably, these solutions define a momentum map[^3] $${\bf J}_{\rm Sing}:\, {\rm
Emb}(S,\mathbb{R}^{2K})\to \mathfrak{X}_{\rm can}^*(\mathbb{R}^{2K})$$ where $\mathfrak{X}_{\rm can}^*(\mathbb{R}^{2K})\simeq{\rm
Den}(\mathbb{R}^{2K})$. This momentum map is produced by the [*left*]{} action of canonical transformations on $\mathcal{P}$ by composition of functions; that is, $$\eta\cdot{\boldsymbol{\zeta}}=\eta\circ{\boldsymbol{\zeta}}\,.$$ Importantly, the same kind of momentum map arises in the motion of ideal fluids (e.g., for point vortices in 2D). See [@MaWe1983], where these momentum maps are shown to possess a dual pair structure [@We1983]. Namely, if one considers $S$ as a manifold with volume form $$\omega_{\rm vol}=w(s)\,{\rm d}^k s
\,,$$ then the *right* action of $\rm Diff_{\rm vol}$ on ${\rm
Emb}(S,\mathbb{R}^{2K})$ $${\boldsymbol{\zeta}}\cdot\eta={\boldsymbol{\zeta}}\circ\eta
\,,$$ yields another momentum map $${\bf J}_S:\,{\rm
Emb}(S,\mathbb{R}^{2K})\to\mathfrak{X}^*_{\rm vol}(S)
\,.$$ In more generality, if $(S,w)$ is a volume manifold and $(\mathcal{P},\omega)$ is a symplectic manifold, then the right action momentum map is (cf. [@MaWe1983]) $${\bf J}_S:{\boldsymbol{\zeta}}\mapsto{\boldsymbol{\zeta}}^*\omega:{\rm
Emb}(S,\mathcal{P})\to\mathfrak{X}^*_{\rm vol}(S)
\,.$$ To summarize, we have the following dual pair structure\
(150,100)(-60,0)(110,75)[${\rm Emb}(S,\mathcal{P})$]{} (78,50)[$\mathbf{J}_{\rm Sing}$]{} (160,50)[$\mathbf{J}_S$]{} (72,15)[$\mathfrak{X}^{\ast}_{\rm can} (\mathcal{P})$]{} (170,15)[$\mathfrak{X}^{\ast}_{\rm vol}(S)$]{} (130,70)[(-1, -1)[40]{}]{} (135,70)[(1,-1)[40]{}]{}
\
which is formally equivalent to the dual pair structure for ideal fluids [@MaWe1983]. Moreover, the left leg yields a solution of the Vlasov equation regardless the number of dimensions and this makes the above dual pair a natural object in kinetic theory.
In order to write explicit formulas, we specialize to the case $\mathcal{P}=\mathbb{R}^{2K}$. Upon denoting ${\boldsymbol{\zeta}}(s)=({\bf
Q}(s),{\bf P}(s))$, one defines the following Poisson structure on ${\rm Emb}(S,\mathcal{P})$ $$\big\{F,G\big\}_{\rm Emb}=\sum_{i=1}^K\int \! \frac1{w(s)}\left(
{\frac{\delta F}{\delta Q^i}}{\frac{\delta G}{\delta P_i}}- {\frac{\delta G}{\delta Q^i}}{\frac{\delta F}{\delta P_i}}\right) {\rm
d}^k s$$ where we see that the factor $1/w(s)$ is needed for functionals of the form [$G({{\boldsymbol{\zeta}}})=\int\omega_{\rm vol}\,g({\boldsymbol{\zeta}})=\int
w(s)\,g({\boldsymbol{\zeta}}(s))\, {\rm d}^k s$]{}, whose functional derivative $\delta G/\delta \zeta=w(s)\,{\rm d}{g}/{\rm d}{\zeta}$ takes values in ${\rm Den}(S)$.
$\quad$\
The factor $1/w(s)$ is reminiscent of the vortex strength factors in the Poisson bracket for point vortices [@MaWe1983]. Indeed, the bracket above appears as the higher dimensional version of the vortex bracket, so that the one dimensional vortex strengths are replaced by appropriate densities (the weights $w(s)$) on the embedded space $S$.
Finally one checks that, for any Hamiltonian function $h\in
\mathcal{F}(T^*\mathbb{R}^{2K})$, $$\big\{F,\langle{\bf J}_{\rm Sing},h\rangle\big\}_{\rm Emb}={\bf
X}_h[F]$$ where ${\bf X}_h[F]$ is the infinitesimal generator of the action of canonical transformations $Can$($\mathbb{R}^{2K}$) on ${\rm
Emb}(S,\mathbb{R}^{2K})$. Thus, ${\bf J}_{\rm Sing}$ satisfies the classical definition of a momentum map.
The singular solution momentum map ${\bf J}_{\rm Sing}$ produces the collective Vlasov Hamiltonian $H\circ{\bf J}_{\rm Sing}$. In particular, substituting the singular solution momentum map (\[sing-fsoln-momap\]) into the EP$Can$ Hamiltonian (\[EPsympHam\]) yields the collective Hamiltonian $$\label{EPCan-collective}
H_\mathcal{N}=\frac12\sum_{a,b=1}^\mathcal{N}\iint
w_a(s)\,w_b(s')\,\mathcal{G}\big({\bf Q}_a(s),{\bf P}_a(s),{\bf
Q}_b(s'),{\bf P}_b(s')\big)\ {\rm d}^ks\,{\rm d}^ks'$$ thereby producing the following collective equations of motion $$\begin{aligned}
\frac{\partial {\bf Q}_a(s,t)}{\partial t} &= \frac{\delta
H_\mathcal{N}}{\delta {\bf P}_a} =
w_a(s)\sum_{b=1}^\mathcal{N}\int\! w_b(s')\,\frac{\partial
}{\partial {\bf P}_a}\,\mathcal{G}\big({\bf Q}_a(s),{\bf
P}_a(s),{\bf Q}_b(s'),{\bf P}_b(s')\big)\, {\rm d}^ks'
\\
\frac{\partial {\bf P}_a(s,t)}{\partial t} &=- \frac{\delta
H_\mathcal{N}}{\delta {\bf Q}_a} =-\,
w_a(s)\sum_{b=1}^\mathcal{N}\int\! w_b(s')\, \frac{\partial
}{\partial {\bf Q}_a}\,\mathcal{G}\big({\bf Q}_a(s),{\bf
P}_a(s),{\bf Q}_b(s'),{\bf P}_b(s')\big)\, {\rm d}^ks' .\end{aligned}$$
$\,$\
The collective dynamics of singular solutions deserves some care, depending on the form of the Vlasov Hamiltonian. The existence of such solutions does not guarantee the existence of a well defined collective Vlasov Hamiltonian $H\circ{\bf J}_{\rm Sing}$. This is because the singular solution momentum map ${\bf J}_{\rm Sing}$ may produce divergent terms in the collective Hamiltonian [@We1983]. For example, this is the case of the Vlasov Poisson system (\[VlasovPoisson\]), where the divergence is generated by potential terms such as $1/2\sum w_a w_b|{\bf Q}_a-{\bf Q}_b|^{-1}$, when $a=b$. The same situation occurs for point vortex solutions of the planar Euler’s vorticity equation; these solutions correspond to the 2D phase space Hamiltonian $H_\mathcal{N}=1/2\,\sum
w_a\,w_b\,\log|(Q_a-Q_b,P_a-P_b)|$. On the other hand, these problems are absent, for instance, in the Vlasov-Helmholtz system (see [@GiHoTr2008] and references therein), since the potential terms there are given by $1/2\sum w_a w_b \,e^{|{\bf Q}_a-{\bf
Q}_b|}$.
As for the right-action momentum map, the expression $${\bf
J}_S({\boldsymbol{\zeta}})={\boldsymbol{\zeta}}^*\omega={\boldsymbol{\zeta}}^*({\rm d}{\bf q}\wedge{\rm
d}{\bf p})={\boldsymbol{\zeta}}^*{\rm d}{\bf q}\wedge{\boldsymbol{\zeta}}^*{\rm d}{\bf p}={\rm d}\!\left({\boldsymbol{\zeta}}^*{\bf
q}\right)\wedge{\rm d}\!\left({\boldsymbol{\zeta}}^*{\bf p}\right)$$ yields the following simple expression $$\label{right-momap}
{\bf J}_S({\bf Q,P})={\rm d}{\bf Q}(s)\wedge {\rm
d}{\bf P}(s)=\sum_{n,m=1}^k{\frac{\partial \bf Q}{\partial s^n}}{\frac{\partial \bf P}{\partial s^m}}\,{\rm
d}s^n\!\wedge{\rm d}s^m$$ The conservation law ${\rm d}{\bf J}_S/{\rm d}t=0$ is recovered by Noether’s theorem, due to the Diff$(S)$-invariance of the collective Vlasov Hamiltonian $H_\mathcal{N}=H\circ{\bf
J}_{\rm Sing}$ in (\[EPCan-collective\]).
When $S$ is a Lagrangian submanifold (this requires ${\rm
dim}(S)=1/2\,{\rm dim}(\mathcal{P})$), the momentum map ${\bf J}_S$ restricts to ${\bf J}_S({\boldsymbol{\zeta}})=0$. Likewise, the case ${\rm
dim}(S)=0$ recovers the usual Klimontovich solution (\[klimsol\]) of particle motion used in kinetic theory. This fact, together with the geometric results on moment hierarchies of kinetic equations, illustrates the geometric basis of kinetic theory, in analogy to Arnold’s formulation of the ideal fluid [@Ar1966].
Klimontovich solution and the Lagrange-to-Euler map
---------------------------------------------------
This section discusses the two limiting cases of the singular solution momentum map ${\bf J}_{\rm Sing}$, that is ${\rm dim}(S)=0$ and $S=\mathcal{P}$. As mentioned above, the first case yields the Klimontovich solution (\[klimsol\]), which is then a momentum map . In this case, the solution identifies the particle trajectories, subject to initial conditions ${\boldsymbol{\zeta}}_a(0)={{\bf z}}_a^0$, so that the particles are transported in the phase space $\mathcal{P}$ by canonical transformations as $\left\{{\boldsymbol{\zeta}}_a(t)\right\}=\psi_t\circ\left\{{{\bf z}}_a^0\right\}$, where $\psi_t\in{Can}(\text{\large$\times$}_{\!a\,}\mathcal{P})$ is generated by the collective Hamiltonian $H\circ{\bf J}_{\rm Sing}$. In two phase-space dimensions, the Klimontovich solution is the usual point vortex solution of the Euler’s vorticity equation.
Another suggestive case of the above treatment is given by $S=\mathcal{P}$, so we may denote $s={{\bf z}}^0$. Then, one has ${\boldsymbol{\zeta}}_a(\cdot,t)=\eta^{(a)}_t\in{Can}(\mathcal{P})$ and the momentum map ${\bf J}_{\rm
Sing}:\text{\large$\times$}_{\!a\,}{Can}(\mathcal{P})\to\mathfrak{X}_{\rm can}^*(\mathcal{P})$ is written as $$\begin{aligned}
f({{\bf z}},t)=\sum_a\int\!w_a({{\bf z}}^0)\,
\text{\large$\delta$}\!\left({{\bf z}}-\eta_t^{(a)\!\!}\cdot{{\bf z}}^{0}\right)\,
{\rm d}^{2K} {{\bf z}}^{0}
\,.
{} $$ This expression coincides with the well known Lagrange-to-Euler map for fluids, whose importance is well established in continuum dynamics. The Lagrange-to-Euler map is equivalent to the characteristic form of the Vlasov equation (\[VlasovChar\]). Notice that both the Klimontovich and the Lagrange-to-Euler maps are produced by the [*same*]{} Lie group ${Can}(\mathcal{P})$ acting on $\text{\large$\times$}_{\!a\,}\mathcal{P}$ and $\text{\large$\times$}_{\!a\,}{Can}(\mathcal{P})$ respectively, with the same left action by composition of functions, that is $\eta\cdot\left\{{\boldsymbol{\zeta}}_a\right\}=\left\{\eta\circ{\boldsymbol{\zeta}}_a\right\}$ in the first case and $\eta\cdot\left\{\eta^{(a)}\right\}=\left\{\eta\circ\eta^{(a)}\right\}$ in the second. On the other hand, for the Klimontovich case, the collective dynamics generated by the Hamiltonian $H\circ{\bf J}_{\rm
Sing}:\text{\large$\times$}_{\!a\,}\mathcal{P}\to\mathbb{R}$ produces the canonical transformations $\psi\in{Can}(\text{\large$\times$}_{\!a\,}\mathcal{P})\neq\text{\large$\times$}_{\!a\,}{Can}(\mathcal{P})$. This point is of fundamental importance because the Lie group ${Can}(\text{\large$\times$}_{\!a\,}\mathcal{P})$ is the symmetry group of the Liouville equation [@MaMoWe84] and the difference between ${Can}(\text{\large$\times$}_{\!a\,}\mathcal{P})$ and $\text{\large$\times$}_{\!a\,}{Can}(\mathcal{P})$ is related to the particle correlations, which are neglected in the second situation. (The latter is the Vlasov mean field approximation.)
The fact that these two fundamental maps each arise from the left leg of a dual pair of momentum maps again illuminates the geometric footing of kinetic theory. The above arguments also provide mathematical support for the wide success of Klimontovich method in kinetic equations [@Kl1982].
Geometric kinetic theory
------------------------
The presence of dual pairs in kinetic theory illuminates the Liouville and Vlasov equations in the light of their Lie symmetry properties. (Something similar happens for Euler’s vorticity equation.) Namely, the presence of momentum maps is not accidental in kinetic approaches. Indeed, a reasonable summary of the results in [@MaMoWe84; @HoLySc1990; @GiHoTr2008] could be made by saying that the process $$\hspace{-.3cm} \fbox{Liouville equation}
\,\to\fbox{Vlasov equation}
\begin{matrix}
\,\
\nearrow\fbox{\,\ ideal fluid\,\ }\quad
\\\,\\
\!\!\!
\searrow\fbox{beam optics}
\end{matrix}$$ is given by a composition of momentum maps. In other words, taking the moments (BBGKY, kinetic or statistical) of a Lie-Poisson kinetic equation is always a momentum map [@MaMoWe84; @HoLySc1990; @GiHoTr2008]. Moreover, the closures adopted to obtain Vlasov from BBGKY, fluid theory from kinetic moments and beam optics from statistical moments are also momentum maps arising from particular subgroups of the symmetry group of the starting system. More explicitly, passing from Liouville to Vlasov requires the subgroup ${Can}(\mathcal{P})\subset{Can}(\text{\large$\times$}_{\!i\,}\mathcal{P})$. Likewise, passing from Vlasov to fluid requires the fiber preserving subgroup $Can_\pi(T^*Q)\subset{Can}(T^*Q)$. Finally, passing from Vlasov to beam optics requires the subgroup ${\rm
Sp}(2K,\mathbb{R})\subset{Can}(\mathbb{R}^{2K})$. We can summarize the situation in the following statement
*All these moment approximations in kinetic theory are momentum maps.*
That the BBGKY distributions are momentum maps is a remarkable fact. One may ask whether the Klimontovich averages in plasma theory also share this property. In this case, the autocorrelations considered in the latter approach would again be naturally included in the geometry of the theory. We leave this promising question open, as a direction for future research.
Moment closures of EP$Can$: integrable cases
============================================
As explained in [@GiHoTr2005; @GiHoTr2007], the geodesic Vlasov equation may be represented in terms of the moments. Indeed, upon supposing that the metric $\mathcal{G}$ in (\[EPsympHam\]) and (\[epcan\]) is sufficiently smooth, may can expand $\mathcal{G}({{\bf z}},{{\bf z}}')$ in a Taylor series.
Integrable closures of kinetic moments
--------------------------------------
In this section, we present the kinetic moment hierarchy for EP$Can$. Upon denoting ${{\bf z}}=(\bf q,p)$, one may expand $\mathcal{G}$ in a Taylor series, as follows, $$\begin{aligned}
\mathcal{G}({{\bf z}},{{\bf z}}')\,=&\sum_{n,m=0}^\infty{\bf p}^n\otimes{\bf
p}^{\prime\,m\!} \contract \,G_{nm}({\bf q,q'})
\\
=&\ \sum_{n,m}\ \sum_{i_1...i_n}\sum_{j_1...j_m}\big({\bf
p}^n\big)_{i_1,...,i_n}\left({\bf
p'}^m\right)_{j_1,...,j_m}\big(G_{nm}({\bf
q,q'})\big)^{i_1,...,i_n,j_1,...,j_m}\end{aligned}$$ where $G_{nm}$ is now a contravariant tensor field of rank $n+m$. Inserting this Taylor expansion in the EPCan Hamiltonian yields (with the notation above) $$H=\frac12\sum_{n,m=0}^\infty\int A_n({\bf q}) \otimes A_m({\bf
q}')\,\contract\, G_{nm}({\bf q,q'})\ {\rm d}^K{\bf q}\, {\rm
d}^K{\bf q}' = \frac12\,\big\|\{A_n\}\big\|_G$$ Thus, upon denoting $$G_{nm}*A_m:=\int G_{nm}({\bf q,q'})\,\contract \,
A_m({\bf q}')\ {\rm d}^K{\bf q}'$$ the geodesic moment equations become $${\frac{\partial A_n}{\partial t}}=-\sum_{m,k=0}^\infty \text{\large
ad}^*_{\,G_{m\!k\,}*A_k}\,A_{n+m-1}$$ where ad$^*$ is the coadjoint Lie-Schouten operator. The singular solutions of the Vlasov moment hierarchy may be expressed in the following form: $$\label{KlimKinMom}
A_n({\bf q},t)=\int w(s)\ {\bf P}^n(s,t)\,\delta({\bf q-Q}(s,t))\ {\rm
d}^k s$$ which is a momentum map ${\bf J}:{\rm Emb}(S,T^*Q)\to
\mathfrak{g}^*$, where $\mathfrak{g}$ is the kinetic moment algebra.
As we have seen in the preceding discussions, the moment algebra possesses the important subalgebra $\mathfrak{g}_1=\mathfrak{X}(Q)$ of vector fields on the configuration manifold. In terms of canonical transformations, this corresponds to Hamiltonian generating functions that are linear in the momentum coordinate, i.e. point transformations. These are cotangent lifts $T^*$Diff($Q$) of diffeomorphisms on the configuration manifold $Q$ [@HoMa2004; @GiHoTr2007]. Remarkably, when the moment hierarchy of EPCan is closed such that $G_{11}=:G_1$ is the only non-vanishing term of $G_{nm}$, we obtain the Hamiltonian on the one-form density $A_1=:{\bf m}({\bf q})\cdot{\rm d}{\bf q}\otimes{\rm d}^K{\bf q}\in
\mathfrak{X}^*$ $$H=\frac12\iint {\bf m}({\bf q})\cdot G_1({\bf q,q'})\, {\bf m}({\bf
q}')\ {\rm d}^K{\bf q}\ {\rm d}^K{\bf q}'$$ By using the property of the Schouten bracket $[\beta_1,\,\alpha_n]=\pounds_{\beta_1}\,\alpha_n$ (where $\pounds$ denotes Lie derivative), one finds the EPDiff equation, $${\frac{\partial \bf m}{\partial t}}+\textit{\large\pounds}_{G_1*\bf m}\, {\bf m}=0
\,.$$ EPDiff is the [*Euler-Poincaré equation on the diffeomorphisms*]{} [@HoMa2004]. This equation has the important property of exhibiting emergent singular $\delta$-like solutions from any confined smooth initial configuration. In 1D, the particular case $G_1=(1-\alpha_2\partial^2)^{-1}$ yields the integrable [*Camassa-Holm equation*]{}, which is well known in the community of integrable systems.
Remarkably, if we also allow for $G_{00}=:G_0\neq0$, we obtain the Hamiltonian $$\begin{gathered}
H[A_0,A_1]=\frac12\iint A_1({\bf q})\cdot G_1({\bf q,q'})\, A_1({\bf
q}')\ {\rm d}^K{\bf q}\ {\rm d}^K{\bf q}' \\
+ \frac12\iint A_0({\bf q})\ G_0({\bf q,q'})\, A_0({\bf q}')\ {\rm
d}^K{\bf q}\ {\rm d}^K{\bf q}'\end{gathered}$$ which yields a geodesic flow on the semidirect-product Lie group ${\rm Diff}\circledS\,\mathcal{F}$, introduced in [@HoTrYo2007] in the context of image matching, discussed further in [@HoTr2009] and shown in numerical simulations to exhibit emergent singularities in both of its variables [@HoOnTr2009]. Interestingly enough, the EP(${\rm Diff}\circledS\,\mathcal{F}$) equations (with notation $\beta_0\diamond
A_0=-\,{\textrm{\large ad}}^*_{\beta_0}\,A_0$) $$\begin{aligned}
A_{0,t}+\textit{\large\pounds}_{G_1*A_1}\,A_0 &= 0
\,,\\
A_{1,t}+\textit{\large\pounds}_{G_1*A_1}\,{A_1} &= A_{0\,}
\text{\large$\diamond$}\ (G_0
* A_0) \,.
$$ are a geodesic flow on the [*extended point transformations*]{}, i.e. compositions of cotangent lifts and fiber translations [@MaWeRaScSp1983].
$\quad$\
The semidirect-product Lie group ${\rm
Diff}_{\,}\circledS_{\,}\mathcal{F}$ is identified with the compositions of cotangent lifts with fiber translations on the phase space $T^*Q$ with coordinates $({\bf q,p})$. This identification is of fundamental importance in plasma physics [@MaWeRaScSp1983]. It also yields the interpretation of $EP\!\left({\rm
Diff}_{\,}\circledS_{\,}\mathcal{F}\right)$ as a geodesic motion on the Lie group ${Can}_{\pi}(T^*Q)$ of [[******]{}fiber-preserving canonical transformations]{} on the cotangent bundle $T^*Q$. In fact, any transformation by ${Can}_{\pi}(T^*Q)$ can be realized as the composition of a fiber translation and a cotangent lift (or viceversa) [@BaWe1997]. On the other hand, such a transformation is always a canonical transformation characterized by a generating function that is *linear* (and inhomogeneous) in the canonical momentum. Therefore, since ${\rm
Diff}(Q)_{\,}\circledS_{\,}\mathcal{F}(Q)\simeq {Can}_{\pi}(T^*Q)$ then $EP({\rm Diff}\circledS_{\,}\mathcal{F})\simeq EP{Can}_{\pi}$.
In 1D, the special case $G_0=\delta$-function yields yet another integrable system, $$\begin{aligned}
\lambda_t&=-\left(u\lambda\right)_q
\\
u_t-u_{qqt}&=-3uu_q+2u_q u_{qq}+u u_{qqq} -\lambda\lambda_q\end{aligned}$$ known as [*two-component Camassa-Holm equation*]{} [@ChLiZh2005] (here $u=(1-\partial^2)^{-1}A_1$ and $\lambda=A_0$). This system first appeared in [@OlRo1996]. Its singular solutions were studied in [@CoIv2008] who pointed out their relation between this system and shallow water equations. Upon slightly modifying the Hamiltonian by $G_0=(1-\partial^2)^{-1}$, one has the spontaneous emergence of the Klimontovich solution (\[KlimKinMom\]) in $(A_0,A_1)$, as shown in [@HoOnTr2009]. The two-component Camassa-Holm system and its Klimontovich solutions have also found applications in the *metamorphosis* approach to image matching, e.g., for magnetic resonance images [@HoTrYo2007].
Geodesic flow on statistical moments
------------------------------------
Section 5 shows that representing the geodesic Vlasov equation in terms of statistical moments in a special case recovers the well-known Bloch-Iserles integrable system [@BlIsMaRa2005; @BlIs2006]. If $\mathcal{G}$ in (\[EPsympHam\]) and (\[epcan\]) is assumed to be analytic in [*both*]{} position and momentum, its Taylor expansion may be written as: $$\mathcal{G}({{\bf z}},{{\bf z}}')=\sum_{n,m=0}^\infty\Gamma_{nm}\contract\,{{\bf z}}^{n}\otimes{{\bf z}}^{\prime\,m\!\!}$$ where $\Gamma_{nm}$ is a covariant tensor of rank $n+m$. Then, the EP$Can$ Hamiltonian (\[EPsympHam\]) is written in terms of the moments $X^n(t)=\int{\bf z}^n f({\bf z},t)\ {\rm d}^N{\bf z}$ as $$H=\frac12 \sum_{n,m=0}^\infty \Gamma_{nm}\contract\, X^n\otimes X^m$$ so that the moment equations for $(n,m)\in \mathbb{Z}_+$ turn out to be $$\dot{X}^{m}\,= -\,m\sum_{n=0}^\infty n\
\text{\large$\mathcal{S}$}\bigg( \Big(\left(\Gamma_{nk}\contract\,
X^k\right)\cdot\mathbb{J}\Big)\contract \,X^{m+n-2}\bigg)
\label{mom-eqn-nm}$$ where, as before, the symbol $\contract$ denotes contraction between upper and lower indices of the various tensors. In explicit index notation this reads as $$\begin{gathered}
\Big(\dot{X}^m\Big)^{\!i_1...i_m}=
\\
-\,m\sum_{n=0}^\infty n\ \text{\large$\mathcal{S}$}\!\left(
\big(\Gamma_{n k}\big)_{\!j_1...j_n\,l_1...l_k}
\left(X^{k}\right)^{l_1...l_k}
\mathbb{J}^{j_n\,i_m}\Big(X^{m+n-2}\Big)^{\!i_1...i_{m-1}\,j_1...j_{n-1}}\!\right)
\,.\end{gathered}$$ An example of this equation is discussed in Section \[BlIs-eqn-sec\] for the case $m=n=2$.
Klimontovich dynamics for statistical moments
---------------------------------------------
As in the case of kinetic moments, the single-particle Klimontovich solution $$f({{\bf z}},t)=\sum_a w_a\,\delta({{\bf z}}-{\boldsymbol{\zeta}}_a(t))
\label{Klim-soln}$$ of the geodesic Vlasov equation offers an interesting opportunity for presenting solutions of the moment hierarchy. These solutions may be represented as $$X^n(t)=\sum_a w_a\,{\boldsymbol{\zeta}}_a^n(t)$$ where ${\boldsymbol{\zeta}}_a^n:={\boldsymbol{\zeta}}_a\otimes\dots\otimes{\boldsymbol{\zeta}}_a$ ($n$ times) and ${\boldsymbol{\zeta}}_a$ satisfies Hamilton’s canonical equations with Hamiltonian $$H_\mathcal{N}
=
\frac12
\sum_{a,b}\sum_{n,m} w_a\,w_b\,\Gamma_{nm}\contract\ {\boldsymbol{\zeta}}_a^n\otimes{\boldsymbol{\zeta}}_b^m
\,.$$ This means that ${\boldsymbol{\zeta}}_a$ satisfies $$\dot{{\boldsymbol{\zeta}}}_a=\mathbb{J}\,\nabla_{\!{\boldsymbol{\zeta}}_a}H_{\!\mathcal{N}}
=
\mathbb{J}\sum_{nm}n\,w_a\,\sum_{b}w_b\,\Gamma_{nm}\contract\ {\boldsymbol{\zeta}}_a^{n-1}\otimes{\boldsymbol{\zeta}}_b^m
\,,$$ where the moment order $m$ or $n$ ranges from zero to infinity, and $a,b$ range over the number of particles $a,b=1,\dots,\mathcal{N}$. Each ${\boldsymbol{\zeta}}_a$ undergoes Hamiltonian dynamics because the Vlasov single-particle solution is an equivariant momentum map. Equivariance of the corresponding momentum map for dynamics of point vortices in an Euler fluid was proven in [@MaWe1983]. The same argument holds for the singular solutions of kinetic moment equations, including the EPDiff equation [@HoMa2004].
$\,$\
It is important to notice that higher-order moment truncations do not possess the Klimontovich solution. The latter exists only for genuine moment *closures*, such as the fluid closure for the kinetic moments (consisting of $A_0,A_1$), or the $2^{nd}$-order closure for statistical moments used in linear beam optics (consisting of $X^1,X^2$).
Bloch-Iserles system as a moment equation {#BlIs-eqn-sec}
=========================================
The case $(n,m)=(2,2)$ of the moment equation (\[mom-eqn-nm\]) yields an important moment subalgebra, given by [[******]{}homogeneous quadratic polynomials on phase space]{}, i.e. quadratic forms on $V$. In this case, formula (\[mombrkt\]) implies the following Lie-Poisson moment bracket $$\{F,G\}(X^2)=4 \left\langle X^{2},\,
\mathcal{S} \left(\frac{\partial F}{\partial
X^2}\ \mathbb{J}\,\frac{\partial G}{\partial
X^2}\right)\right\rangle$$ where the moment $X^2$ is now a $N\times N$ symmetric matrix. Because of the antisymmetry of $\mathbb{J}$, the bracket above may be rewritten as $$\{F,G\}(X^2)=\left\langle X^{2\,},\left[\frac{\partial F}{\partial
X^2},\frac{\partial G}{\partial
X^2}\right]_{2\mathbb{J}}\right\rangle$$ where $$\left[\frac{\partial F}{\partial X^2},\frac{\partial G}{\partial
X^2}\right]_{2\mathbb{J}\!} := \frac{\partial F}{\partial
X^2}\,2\mathbb{J}\,\frac{\partial G}{\partial X^2} - \frac{\partial
G}{\partial X^2}\,2\mathbb{J}\,\frac{\partial F}{\partial X^2}$$ which is the Lie bracket for the integrable Bloch-Iserles (BI) system of equations introduced in [@BlIsMaRa2005; @BlIs2006]. In that case, an antisymmetric matrix $\mathbb{N}$ of any dimension defines the following Lie bracket on the space of symmetric matrices of the same dimension. In formulas one has $$\label{BILB}
[X,Y]_\mathbb{N}:=X\mathbb{N}Y-Y\mathbb{N}X$$ where $X$ and $Y$ are symmetric matrices. The Bloch-Iserles system $$\label{BIeq}
\dot{X}=\left[(X)^2,\mathbb{N}\right]$$ is Lie-Poisson on this Lie algebra, with Hamiltonian $H_{BI}=\frac12{\rm Tr}(^{t\!}XX)$. Here the notation $(X)^2$ denotes standard matrix multiplication of $X$ by itself, in order to distinguish from second-order moments $X^2$. In addition, $^{t\!}X$ denotes the transpose of the matrix $X$, so that $^{t\!}X=X$ when $X$ is symmetric. One concludes the following.
In the even-dimensional case, the integrable Bloch-Iserles system is the Vlasov moment equation (\[Vlasov-moment-eqn\]) obtained from the quadratic Hamiltonian $$H_{BI}=\frac12\langle X^2,X^2\rangle$$ associated with the antisymmetric matrix $\mathbb{N}:=2\mathbb{J}$.
Let us now look at the moment bracket for functions of $$(X^0,X^1,X^2)\in
\mathfrak{g}_0^*\oplus\mathfrak{g}_1^*\oplus\mathfrak{g}_2^* \simeq
\mathbb{R}\oplus V \!\oplus \left(V\vee V\right)$$ where $\vee$ is again the symmetric tensor product defined in (\[V-devil-def\]) and $X^0=const$ is naturally taken as the probability normalization. The moment bracket (\[mombrkt\]) becomes $$\begin{gathered}
\label{mombrkt-symtens}
\{F,G\}(X)
=\, X^0 \, \frac{\partial F}{\partial
{{\boldsymbol{X}}}^1}\cdot\mathbb{J}\frac{\partial G}{\partial {{\boldsymbol{X}}}^1}
\\
+
{{\boldsymbol{X}}}^{1}\cdot\left(\frac{\partial F}{\partial X^2}\,2\mathbb{J}\frac{\partial G}{\partial {{\boldsymbol{X}}}^1}
-\,\frac{\partial G}{\partial X^2}\,2\mathbb{J}\frac{\partial
F}{\partial {{\boldsymbol{X}}}^1}\right)
\\
+
\left\langle X^{2\,},\left[\frac{\partial F}{\partial X^2},\frac{\partial G}{\partial X^2}\right]_{2\mathbb{J}}\right\rangle
$$ which is given by the direct sum of the canonical Poisson bracket on $V$ in first term, plus the semidirect-product Lie-Poisson bracket on $\mathfrak{g}_{2}^{*\,}\circledS\ \mathfrak{g}_{1}^*\simeq\,{\rm
Sym}\,\circledS\ V$ in the second and third terms. Thus, the specialization of the moment bracket (\[mombrkt\]) to (\[mombrkt-symtens\]) in this case defines a Lie-Poisson bracket on $\left({\rm
Sym}\,\circledS \,V\right)\oplus \mathbb{R}$.
We now turn our attention to the odd-dimensional BI system. In this system, one has a degenerate antisymmetric matrix $\mathbb{N}$ of odd dimension $n$ and rank $2K$. Upon defining $\bar{\mathbb{N}}$ as the non-degenerate minor of maximal dimension ($2K$), the degenerate matrix $\mathbb{N}$ $$\mathbb{N} = \begin{bmatrix}
\bar{\mathbb{N}} & 0 \\
0 & 0
\end{bmatrix}$$ produces the Lie bracket (\[BILB\]) associated to the BI equation (\[BIeq\]).
The odd-dimensional BI system is known to be a geodesic flow on the space endowed with the Lie bracket (cf. equation (2.14) in [@BlIsMaRa2005]) $$\begin{gathered}
\label{ExtendedSDBracket}
\left[(S, A, B), (S', A', B')\right] : =
\\
\left(S\,\bar{\mathbb{N}} S' - S'\,
\bar{\mathbb{N}} S\,,\, S \,\bar{\mathbb{N}} A' - S'\,
\bar{\mathbb{N}} A\,,\, ^{t\!}A \,\bar{\mathbb{N}} A' -\,^{t\!}A'\,
\bar{\mathbb{N}} A\right)\end{gathered}$$ where one denotes $d=n-2K$, for any $S, S' \in \operatorname{Sym}
(2K)$, $A, A' \in \mathcal{M}_{2K\times d}$, and $B, B' \in
\operatorname{Sym}(d)$.
We will show that the bracket (\[mombrkt\]) for the Vlasov moment system is Lie-Poisson on the dual to the Lie algebra . That is, we choose $d=1$, which is the case of interest here. Proposition 2.5 in [@BlIsMaRa2005] shows that the geodesic equations on ${\rm Sym}(2K+1)$ are equivalent to the geodesic equations on , where $V$ is a $2K$-dimensional symplectic space carrying a non-degenerate symplectic structure $\mathbb{\bar{N}}/2$. This leads us to the identification of the Bloch-Iserles system with geodesic moment dynamics.
When $d=1$, the odd-dimensional BI system is a Vlasov moment equation of the form (\[Vlasov-moment-eqn\]) on the statistical moments. This equation is generated by the quadratic Hamiltonian $$H(X)=\frac12 \left\langle
X^{2\,},X^2\right\rangle+\frac12\,{{\boldsymbol{X}}}^1\cdot{{\boldsymbol{X}}}^1$$ yielding a geodesic moment flow. The corresponding BI Hamiltonian $H_{BI}(X)=1/2\,{\rm Tr}(^{t\!}X X)$ is written in terms of the symmetric $(2K+1)$-dimensional matrices of the form $$X = \begin{bmatrix}
X^2 & {{\boldsymbol{X}}}^1 \\
^{t\!}{{\boldsymbol{X}}}^{1} & 4X^0
\end{bmatrix}$$ These matrices are endowed with the Bloch-Iserles Lie bracket (\[BILB\]), where the (degenerate) antisymmetric matrix $\mathbb{N}$ takes the form $$\mathbb{N} = \begin{bmatrix}
2\mathbb{J} & 0 \\
0 & 0
\end{bmatrix}$$ and $\mathbb{J}$ is the canonical symplectic matrix.
Upon denoting $$\bar{\mathbb{N}}=2\mathbb{J}\,,\qquad S=\frac{\partial F}{\partial
X^2}\,, \qquad S'= \frac{\partial G}{\partial X^2}\,, \qquad
A=\frac{\partial F}{\partial {{\boldsymbol{X}}}^1}\,, \qquad A'=\frac{\partial
G}{\partial {{\boldsymbol{X}}}^1}\,,$$ one sees that the only difference between the Lie bracket in (\[ExtendedSDBracket\]) and the bracket in (\[mombrkt-symtens\]) resides in a constant factor in the first term of (\[mombrkt-symtens\]) $$\left(\frac{\partial F}{\partial
{{\boldsymbol{X}}}^1}\right)\cdot\mathbb{J}\frac{\partial G}{\partial {{\boldsymbol{X}}}^1} =
\frac14\left[\left(\frac{\partial F}{\partial
{{\boldsymbol{X}}}^1}\right)\cdot2\mathbb{J}\frac{\partial G}{\partial {{\boldsymbol{X}}}^1} -
\left(\frac{\partial G}{\partial
{{\boldsymbol{X}}}^1}\right)\cdot2\mathbb{J}\frac{\partial F}{\partial {{\boldsymbol{X}}}^1}
\right]$$ where the square bracket in the right hand side is identical to the last component of (\[ExtendedSDBracket\]). This difference however can be easily overcome. Indeed, one can always re-define the Lie bracket on as $$\begin{gathered}
\left[(S, A, B), (S', A', B')\right] : =
\\
\left(S\,\bar{\mathbb{N}} S' - S'\,
\bar{\mathbb{N}} S\,,\, S \,\bar{\mathbb{N}} A' - S'\,
\bar{\mathbb{N}} A\,,\,\frac14\left( ^{t\!}A \,\bar{\mathbb{N}} A' -
\,^{t\!}A'\, \bar{\mathbb{N}} A\right)\right)\end{gathered}$$ and verify that the map $$\begin{aligned}
\Psi:\makebox{$\left({\rm Sym}(2K)\,\circledS \,V\right)\oplus
\mathbb{R}$}&\ \to {\rm Sym}(2K+1)
\\
(S,A,B)&\ \mapsto
\begin{bmatrix}
S & A \\
^{t\!} A & 4B
\end{bmatrix}\end{aligned}$$ is a Lie algebra isomorphism for ${\rm Sym}(2K+1)$, which is endowed with the BI Lie bracket $[X,Y]_{\mathbb{N}}=X\mathbb{N}Y-Y\mathbb{N}X$. The isomorphism property is a direct verification identical to Proposition 2.5 in [@BlIsMaRa2005]. In particular, let $(S, A, B), (S', A', B') \in
\left({\rm Sym}(2K)\,\circledS \,V\right)\oplus \mathbb{R}$ and compute directly that $$\begin{aligned}
\Psi\big([(S, A, B),\, &(S', A', B')]\big) =
\\
&= \Psi\big(S\,\bar{\mathbb{N}} S' - S' \,\bar{\mathbb{N}} S,\, S
\,\bar{\mathbb{N}} A' - S' \,\bar{\mathbb{N}} A,\, 1/4\, (\,^{t\!}A
\,\bar{\mathbb{N}} A' -\, ^{t\!}A'
\,\bar{\mathbb{N}} A)\big) \\
& =
\begin{bmatrix}
S\,\bar{\mathbb{N}} S' - S' \,\bar{\mathbb{N}} S & S \,\bar{\mathbb{N}} A' - S' \,\bar{\mathbb{N}} A \\
^{t\!}(S \,\bar{\mathbb{N}} A' - S' \,\bar{\mathbb{N}} A) & ^{t\!}A
\,\bar{\mathbb{N}} A' -\, ^{t\!}A' \,\bar{\mathbb{N}} A
\end{bmatrix} \\
& =
\begin{bmatrix}
S & A \\
^{t\!}A & 4B
\end{bmatrix}
\begin{bmatrix}
\bar{\mathbb{N}} & 0 \\
0 & 0
\end{bmatrix}
\begin{bmatrix}
S' & A' \\
^{t\!}A' & 4B'
\end{bmatrix}
- \begin{bmatrix}
S' & A' \\
^{t\!}A' & 4B'
\end{bmatrix}
\begin{bmatrix}
\bar{\mathbb{N}} & 0 \\
0 & 0
\end{bmatrix}
\begin{bmatrix}
S & A \\
^{t\!}A & 4B
\end{bmatrix} \\
& = \big[\Psi(S,A,B), \Psi(S',A',B')\big]_{\!\mathbb{N}}\end{aligned}$$ as required.
Thus, we conclude that the BI system and the geodesic moment equations are equivalent.
Klimontovich solutions of the Bloch-Iserles system {#Ksoln-BI-subsec}
--------------------------------------------------
The Klimontovich map for the geodesic Vlasov equation provides simple solutions of the Bloch-Iserles system in any dimension. For example, in even dimensions, the dynamics of the BI solution $$\label{BIKlim-even}
X(t)=\sum_a w_a\,{\boldsymbol{\zeta}}_a^2(t)$$ is given by the system $$\dot{{\boldsymbol{\zeta}}}_a= w_a\sum_{b}w_b\, \mathbb{N}\,{\boldsymbol{\zeta}}_b^2\, {\boldsymbol{\zeta}}_a$$ where $\mathbb{N}=2\,\mathbb{J}$ and ${\boldsymbol{\zeta}}_b^2\, {\boldsymbol{\zeta}}_a$ is a covector such that $({\boldsymbol{\zeta}}_b^2\, {\boldsymbol{\zeta}}_a)_i=({\boldsymbol{\zeta}}_b^2)_{ij}\,
({\boldsymbol{\zeta}}_a)^j$. In explicit index notation, one has $$\big(\dot{{\boldsymbol{\zeta}}}_a\big)^i= 2\,w_a\sum_{b}w_b\,
\mathbb{J}^{ij}\left({\boldsymbol{\zeta}}_b^2\right)_{jk} \left({\boldsymbol{\zeta}}_a\right)^k
\,.$$ This system is a Hamiltonian system with the homogeneous quartic Hamiltonian $$H_\mathcal{N}=\frac12\sum_{a,b=1}^\mathcal{N}w_a\,w_b\,{\rm Tr}\left(
\,^{t\!}({\boldsymbol{\zeta}}_a^2)\,{\boldsymbol{\zeta}}_b^2\right)$$ Notice that, by writing the equation for ${\boldsymbol{\zeta}}_a$ as $$\dot{{\boldsymbol{\zeta}}}_a= w_a\sum_{b\neq a}w_b\, \mathbb{N}\,{\boldsymbol{\zeta}}_b^2\,
{\boldsymbol{\zeta}}_a+w_a^2\,\|{\boldsymbol{\zeta}}_a\|^2\,\mathbb{N}\, {\boldsymbol{\zeta}}_a$$ we can specialize the above to the simple case when $w_a=1$ for a fixed $a$ and $w_b=0\ \forall b\neq a$, so that $$X(t)={\boldsymbol{\zeta}}^2(t) \qquad\text{ with }\qquad \dot{{\boldsymbol{\zeta}}}=
\left\|{\boldsymbol{\zeta}}\right\|^2 \mathbb{N} \,{\boldsymbol{\zeta}}$$ This case leads however to trivially linear dynamics, since the norm $\left\|{\boldsymbol{\zeta}}\right\|$ is evidently conserved. This does not happen for different norms in the quadratic moment Hamiltonian, such as ${H}=1/2\left(\Gamma_{22}\contract\, X^2\otimes X^2\right)$.
The above arguments also provide solutions to the Bloch-Iserles system in $2K+1$ dimensions. Indeed, the particle solution of EP$Can$ (\[Klim-soln\]) becomes a solution of the BI system in the following form $$\label{BIKlim-odd}
X(t) = \sum_a w_a\! \begin{bmatrix}
\ \boldsymbol{\zeta}_a^2(t) & \boldsymbol{\zeta}_a(t) \ \\
^{t\!}\boldsymbol{\zeta}_a(t) & 4 \
\end{bmatrix}
\in{\rm Sym}(2K+1)
\,.$$ Here ${\boldsymbol{\zeta}}_a$ undergoes Hamiltonian dynamics with $$H_\mathcal{N}=\frac12\sum_{a,b}w_a\,w_b\,\Big({\boldsymbol{\zeta}}_a\cdot{\boldsymbol{\zeta}}_b+
{\rm Tr}\left( \,^{t\!}({\boldsymbol{\zeta}}_a^2)\,{\boldsymbol{\zeta}}_b^2\right)\Big)
\,,$$ whose collective EP$Can$ equations are $$\dot{{\boldsymbol{\zeta}}}_a= w_a\sum_{b}w_b\, \mathbb{N}\,{\boldsymbol{\zeta}}_b^2\,
{\boldsymbol{\zeta}}_a+\frac12\,w_a\sum_{b}w_b\,\mathbb{N}\,{\boldsymbol{\zeta}}_b
\,.$$ As we shall see, these solutions are also momentum maps in both the even and odd-dimensional cases, since the Klimontovich solution (\[Klim-soln\]) is a momentum map. In particular, upon fixing $\mathcal{N}=1$, these are solution momentum maps $V\to{\rm Sym}(n)$ (${\boldsymbol{\zeta}}\mapsto X$), where $V$ is a symplectic space and $X\in{\rm
Sym}(n)$ is the Bloch-Iserles dynamical variable. This construction arises from a special case of the momentum map in (\[sing-fsoln-momap\]), with dim($S$)=0 (Klimontovich case), $\mathcal{P}=V$ (symplectic vector space) and where the Lie algebra $\mathfrak{X}_{\rm can}(V)$ is restricted to the Lie subalgebra of linear Hamiltonian vector fields. In the more general case when dim$(S)\geq1$ one has also the conserved quantity in (\[right-momap\]). That is, the operation of taking moments preserves the dual pair structure of the Vlasov equation.
Upon fixing $\mathcal{N}=1$, the solution (\[BIKlim-odd\]) of the odd-dimensional Bloch-Iserles system is a momentum map $${\bf J}_{2K+1}:(V,w\mathbb{J})\to{\rm Sym}(2K+1)$$ where $w\mathbb{J}$ is the symplectic form on the vector space $V$. Moreover, the solution (\[BIKlim-even\]) in the even-dimensional case is also a momentum map $${\bf J}_{2K}:(V,w\mathbb{J})\to{\rm Sym}(2K)
\,.$$
In what follows we shall use the isomorphisms $${\rm Sym}(2K+1)\simeq \left({\rm Sym}(2K)\,\circledS\,
V\right)\oplus\mathbb{R}$$ and $${\rm Sym}(2K)\simeq\mathfrak{sp}^*(2K,\mathbb{R}) \,,$$ where $\mathfrak{sp}(2K,\mathbb{R})$ denotes the Lie algebra of Hamiltonian matrices. We prove the first statement, which comprises the second as a particular case. Let $V$ be endowed with the Poisson structure $$\left\{F,G\right\}=\frac1w\ \Big.^{t}\!\!\left(\frac{\partial
F}{\partial {\boldsymbol{\zeta}}}\right)\cdot\mathbb{J} \frac{\partial G}{\partial
{\boldsymbol{\zeta}}}$$ then, the definition of momentum map can be verified by inserting $G=\left\langle{\bf J},\beta\right\rangle$, with ${\bf
J}=w({\boldsymbol{\zeta}}^2,{\boldsymbol{\zeta}},1)\in\left({\rm Sym}(2K)\,\circledS\,
V\right)\oplus\mathbb{R}$ and $\beta=(\beta_2,\boldsymbol{\beta}_1,\beta_0)$ its dual. Then, we have $$\left\{F,\left\langle{\bf J},\beta\right\rangle\right\}=
\Big.^{t}\!\Big(\mathbb{J}\cdot\boldsymbol{\beta}_1+2\mathbb{J}\beta_{2\,}{\boldsymbol{\zeta}}\Big)
\cdot\frac{\partial F}{\partial {\boldsymbol{\zeta}}}$$ which identifies the infinitesimal action of linear (inhomogeneous) Hamiltonian vector fields on the phase space functions $F\in\mathcal{F}(V)$.
Restricting to even $2K$ dimensions requires setting $\beta_0=0=\boldsymbol{\beta}_1$, thereby producing the action of (homogeneous) Hamiltonian vector fields, i.e. Hamiltonian matrices in $\mathfrak{sp}(2K,\mathbb{R})$.
The integrability properties of these solutions will be discussed elsewhere.
Conclusions and outlook
=======================
After reviewing the geometric basis of Vlasov moment dynamics, this paper showed how moment closures of the geodesic Vlasov equation (\[epcan\]) produce interesting known integrable systems, which include CH, CH2 and Bloch-Iserles (BI) equations in both odd and even dimensions. While the CH and CH2 cases were already known to arise in 1D [@GiHoTr2005; @GiHoTr2007], the higher dimensional moment bracket showed that these moment closures also recover the EPDiff equation and extend CH2 to higher dimensions. The paper also recovered the BI system from (\[epcan\]) by a finite-dimensional moment closure, corresponding to inhomogeneous quadratic phase-space functions. Thus, a kinetic theory approach led to special solutions of BI.
The moment closures preserve the two equivariant momentum maps in the Vlasov dual pair. Preservation of this structure guaranteed that the resulting closed moment systems discussed here were still Poisson. This preservation also enabled reduction to finite-dimensional systems by using the Klimontovich particle solutions from plasma theory. For example, the peakon solutions of the CH equation arose from a Klimontovich approach in [@GiHoTr2005]. Singular solutions also arose upon allowing extra smoothing in the CH2 moment Hamiltonian [@GiHoTr2007; @HoOnTr2009]. In addition, this paper showed in Section \[Ksoln-BI-subsec\] that the same approach also produced solutions of the finite-dimensional BI system. However, Klimontovich solutions are not admitted by arbitrary approximations. They are prevented, for example, when moment hierarchies are simply truncated at a certain weight. That is, moment [*closures*]{} preserve the Vlasov dual pair, while moment [*truncations*]{} do not, even though they may be shown to still be Lie-Poisson. Open questions concern both the construction of a Lax pair for the Klimontovich dynamics of the BI system and potential integrability properties of the truncated equations (e.g. the ($A_1,A_2$) truncation for kinetic moments).
The geometric setting showed how the left momentum map in the Vlasov dual pair recovers both the Klimontovich solution and the Lagrange-to-Euler map. That this geometry applies also to the Liouville equation illuminates the geometric footing of kinetic theory. Indeed, this paper explained how all the standard moment approximations in kinetic theory are momentum maps preserving the same dual pair. This construction would certainly be destroyed by introducing the collision integral, whose celebrated Boltzmann version implies irreversibility and produces a preferred direction of time via the $H$-theorem. This irreversibility prevents the Klimontovich solutions, which are solutions of a time-reversal invariant system.
The kinetic-theory interpretation of geodesic Vlasov moment dynamics also provided insight into the physical description of the integrable cases. For example, the CH2 case was interpreted in this light as a charged fluid in the context of the one-component CH equation, thereby extending the CH model to include space-charge effects. On the other hand, CH2 has also been related to shallow water dynamics by applying a series of approximations to the Green-Naghdi equations [@CoIv2008]. Remarkably, a modified version of CH2 dynamics has also been applied in image matching [@HoTrYo2007]. (In image matching, the Hamiltonian is the norm in which one applies optimal control.) Emergent peakon solutions were found to result from applying $H^1$ smoothing to the Hamiltonian in [@HoOnTr2009].
The paper also identified several other potentially interesting open problems. One of these is the problem of making physical applications of the Kirillov ad$^*$-action (\[Kirillov-ad-action\]) for the Lie-Poisson bracket on the symmetric Schouten algebra for arbitrary values of $(n,k)$. Another is to determine whether the Klimontovich average in plasma kinetic theory is a momentum map. One may also ask how the family of symplectically conserved quantities corresponding to statistical-moment versions of the Poincaré invariants found in [@HoLySc1990] may fit into the theory of kinetic moments.
The kinetic approach used here may also provide physical interpretations of use in applying the BI system. For example, particle beams in linear accelerator lattices are described in terms of symplectic transfer matrices. (The same holds for linear ray optics.) In formulas, one has the relation ${{\bf z}}(t)=\mathcal{M}(t)\,{{\bf z}}(0)$, where $\mathcal{M}(t)$ is a one parameter subgroup of ${\rm Sp}(6,\mathbb{R})$ determining the beam evolution ${{\bf z}}(t)\subset \mathbb{R}^{6}$. In this sense, geodesics in ${\rm Sp}(6,\mathbb{R})$ would correspond to [*optimal transfer maps*]{} for particle or optical beams. It is interesting that similar approaches have recently emerged in quantum computation, where the Hamiltonian of the system is constrained to an optimal trajectory (i.e., a geodesic) by a cost function from optimal control theory [@BrElHo08; @NiDoGuDo08]. These additional open problems bode well for the potential success in future applications of using the geometric approach to Vlasov dynamics discussed here. The application of ideas from optimal control to the Vlasov moments may be especially fruitful.
Acknowledgments {#acknowledgments .unnumbered}
---------------
We are indebted with François Gay-Balmaz, Simon Hochgerner, Boris Khesin, David Levermore, Cornelia Vizman, Tudor Ratiu and Harvey Segur for useful and stimulating discussions. The work of DDH was also partially supported by the Royal Society of London Wolfson Research Merit Award.
[99]{}
V.I. Arnold, [*Sur la géométrie différentielle des groupes de Lie de dimension infinie et ses applications à l’hydrodynamique des fluides parfaits*]{}. Ann. Inst. Fourier (Grenoble) 16 1966 fasc. 1, 319–361
V. I. Arnold, B. A. Khesin, Topological methods in hydrodynamics. Applied Mathematical Sciences, 125, Springer-Verlag, New York, 1998.
S. Bates, A. Weinstein, Lectures on the geometry of quantization. Berkeley Math. Lect. Notes 8 (1997).
D. J. Benney, *Properties of long nonlinear waves*. Stud. App. Math. 52 (1973) 45–50.
A. M. Bloch, A. Iserles, [*On an isospectral Lie-Poisson system and its Lie algebra.*]{} Found. Comput. Math. 6 (2006), no. 1, 121–144.
A. M. Bloch, A. Iserles, J. E. Marsden, T. S. Ratiu, *A class of integrable geodesic flows on the symplectic group and the symmetric matrices.* DAMTP Tech. Rep. NA2006/02, University of Cambrige UK, 2006 (arXiv:math-ph/0512093)
D. C. Brody, D. C. P. Ellis, D. D. Holm, [*Hamiltonian statistical mechanics*]{} J. Phys. A: Math. Theor. 41 (2008), 502002
R. Camassa, D. D. Holm, *An integrable shallow water equation with peaked solitons.* Phys. Rev. Lett. 71 (1993), no. 11, 1661–1664.
A. Cannas da Silva, A. Weinstein, Geometric models for noncommutative algebras. Berkeley Math. Lect. Notes 10 (1999).
P. J. Channell, [*The moment approach to charged particle beam dynamics.*]{} IEEE Trans. Nucl. Sci. 30 (1983), no. 4, 2607–2609
P. J. Channell, [*Canonical integration of the collisionless Boltzmann equation.*]{} Ann. N Y Acad. Sci. 751 (1995), 152–161
M. Chen, S. Liu, Y. Zhang, [*A two-component generalization of the Camassa-Holm equation and its solutions.*]{} Lett. Math. Phys. 75 (2006), no.1, 1–15
A. Constantin, R. Ivanov, [*On an integrable two-component Camassa–Holm shallow water system*]{}. Phys. Lett. A 372 (2008), no. 48, 7129–7132
E. D. Courant, H. D. Snyder, Ann. Phys., NY 3 (1) (1958).
A. J. Dragt, F. Neri, G. Rangarajan, D. R. Douglas, L. M. Healy, R. D. Ryne, *Lie algebraic treatment of linear and nonlinear beam dynamics.* Ann. Rev. Nucl. Part. Sci. 38 (1990), no. 38, 455–496
J. Gibbons, *Collisionless Boltzmann equations and integrable moment equations.* Phys. D 3 (1981), no. 3, 503–511.
J. Gibbons, D. D. Holm, B. A. Kupershmidt, *The Hamiltonian structure of classical chromohydrodynamics.* Phys. D 6 (1982/83), no. 2, 179–194.
J. Gibbons, D. D. Holm, C. Tronci, [*Singular solutions for geodesic flows of Vlasov moments*]{}, Math. Sci. Res. Inst. Publ. 55 (2008), 199–220.
J. Gibbons, D. D. Holm, C. Tronci, [*Vlasov moments, integrable systems and singular solutions*]{}, Phys. Lett. A 372 (2008) 1024–1033
J. Gibbons, D.D. Holm, C. Tronci, [*Geometry of Vlasov kinetic moments*]{}, Phys. Lett. A 408 (2008) 4184-4196
D. D. Holm, W. P. Lysenko, J. C. Scovel, *Moment invariants for the Vlasov equation.* J. Math. Phys. 31 (1990), no. 7, 1610–1615.
D. D. Holm, J. E. Marsden, *Momentum maps and measure valued solutions (peakons, filaments, and sheets) of the Euler-Poincaré equations for the diffeomorphism group.* [Progr. Math.]{}, [232]{} (2004) 203-235.
D. D. Holm, J. E. Marsden, T. S. Ratiu, *Euler-Poincaré equations and semidirect products with applications to continuum theories* Adv. in Math., 137 (1998) 1-81.
D. D. Holm, L. Ó Náraigh, C. Tronci, [*Singular solutions of a modified two-component Camassa-Holm equation*]{}. Phys. Rev. E 79, (2009) 016601.
D. D. Holm, C. Tronci, [*Geodesic flows on semidirect-product Lie groups: geometry of singular measure-valued solutions*]{}. Proc. Roy. Soc. A, 465 (2009) 457-476.
D. D. Holm, A. Trouv[é]{}, L. Younes, *The Euler-Poincaré theory of metamorphosis*. To appear, [*Quart. Appl. Math.*]{}
A. A. Kirillov, [*Invariant operators on geometric quantities*]{}, J. Math. Sci. 18 (1982), no. 1, pp. 1–21
B. Khesin, I. Zakharevich, [*Poisson-Lie group of pseudodifferential symbols*]{}. Comm. Math. Phys. 171 (1995), no. 3, 475–530
Y. L. Klimontovich, The Statistical Theory of Non-equilibrium Processes in a Plasma, M.I.T. Press, Cambridge, Massachusetts, 1967
B. A. Kupershmidt [*Hydrodynamical Poisson brackets and local Lie algebras.*]{} Phys. Lett. A 121 (1987), no. 4, 167–174.
B. A. Kupershmidt, Ju. I. Manin, *Long wave equations with a free surface. II. The Hamiltonian structure and the higher equations.* Funktsional. Anal. i Prilozhen. 12 (1978), no. 1, 25–37
P. A. Kuzmin, [*Two-component generalizations of the Camassa–Holm equation*]{}. Math. Notes 81 (2007), 130–134.
D. R. Lebedev, [*Benney’s long waves equations: Hamiltonian formalism.*]{} Lett. Math. Phys. 3 (1979), no. 6, 481–488.
J. E. Marsden, P. J. Morrison, and A.Weinstein, [*The Hamiltonian structure of the BBGKY hierarchy equations*]{}. Contemp. Math. 28 (1984), 115–124.
J. E. Marsden, A. Weinstein [*Coadjoint orbits, vortices, and Clebsch variables for incompressible fluids.*]{} Physica D 7 (1983), pp. 305-323.
J. E. Marsden, A. Weinstein, *The Hamiltonian structure of the Maxwell-Vlasov equations.* Physica D 4 (1981/82), no. 3, 394–406.
J. E. Marsden, A. Weinstein, T. S. Ratiu, R. Schmid, R. G. Spencer, *Hamiltonian systems with symmetry, coadjoint orbits and plasma physics.* Atti Accad. Sci. Torino, Cl. Sci. Fis. Mat. Natur. 117 (1983), 289–340.
M. A. Nielsen, M. R. Dowling, M. Gu, A. C. Doherty, [*Quantum computation as geometry*]{} Science 311 (2008), 1133–1135
A. Nijenhuis [*Jacobi-type identities for bilinear differential concomitants of certain tensor fields. I, II.*]{} Indag. Math. 17 (1955), 390–403
P. Olver, P. Rosenau, [*Tri-Hamiltonian duality between solitons and solitary-wave solutions having compact support*]{}. Phys. Rev. E 53 (1996) 1900–1906.
V. Ovsienko, C. Roger [*Deforming the Lie algebra of vector fields on $S^1$ inside the Lie algebra of pseudodifferential symbols on $S^1$.*]{} Am. Math. Soc. Transl., Ser. 2, 194 (1999), 211–226
H. Qin, W. M. Tang, [*Pullback transformations in gyrokinetic theory.*]{} Phys. of Plasmas 11 (2004), no. 3, 1052-1063.
C. Scovel, A. Weinstein, *Finite-dimensional Lie-Poisson approximations to Vlasov-Poisson equations.* Comm. Pure Appl. Math. 47 (1994), no. 5, 683–709.
A. A. Vlasov, *On the kinetic theory of an assembly of particles with collective interaction.* J. Phys. (USSR) 9 (1945) 25-40; A. A. Vlasov, *Many-particle theory and its application to plasma.* Gordon and Breach, New York, 1961
Weinstein, A. \[1983\], *The local structure of Poisson manifolds.* J. Differ. Geom. 18 (1983), 523–557.
[^1]: The group of canonical transformations $Can$ is also known as the Hamiltonian diffeomorphisms $Diff_{\rm Ham}$.
[^2]: The operator $\mathcal{S}$ here takes the symmetric part of its argument.
[^3]: Here $\mathfrak{X}_{\rm can}$ denotes the Lie algebra of Hamiltonian vector fields, which should not be confused with the larger Lie algebra corresponding to the tangent space at the identity of the symplectomorphism group $Symp$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Existing and planned optical telescopes and surveys can detect artificially-illuminated objects comparable in total brightness to a major terrestrial city out to the outskirts of the Solar System. Orbital parameters of Kuiper belt objects (KBOs) are routinely measured to exquisite precisions of $< 10^{-3}$. Here we propose to measure the variation of the observed flux $F$ from such objects as a function of their changing orbital distances $D$. Sunlight-illuminated objects will show a logarithmic slope $\alpha \equiv (d\log F/d\log D)= -4$ whereas artificially-illuminated objects should exhibit $\alpha= -2$. Planned surveys using the proposed LSST will provide superb data that would allow measurement of $\alpha$ for thousands of KBOs. If objects with $\alpha=-2$ are found, follow-up observations can measure their spectra to determine if they are illuminated by artificial lighting. The search can be extended beyond the Solar System with future generations of telescopes on the ground and in space, which would be capable of detecting phase modulation due to very strong artificial illumination on the night-side of planets as they orbit their parent stars.'
---
0.8in
**Detection Technique for Artificially-Illuminated**
**Objects in the Outer Solar System and Beyond**
0.7cm
Abraham Loeb$^{1,2}$ and Edwin L. Turner$^{3,4}$
$^1$ [*Astronomy Department, Harvard University, 60 Garden Street, Cambridge, MA 02138, USA*]{}\
$^2$ [*Institute for Theory and Computation, Harvard-Smithsonian Center for Astrophysics, 60 Garden Street, Cambridge, MA 02138, USA*]{}\
$^3$ [*Department of Astrophysical Sciences, Princeton University, Princeton, NJ 08544, USA*]{}\
$^4$ [*Institute for the Physics and Mathematics of the Universe, The University of Tokyo, Kashiwa 227-8568, Japan*]{}\
0.2in
[**Kewords:**]{} astrobiology, SETI, Kuiper belt objects, artificial illumination
Introduction
============
The search for extraterrestrial intelligence (SETI) has been conducted mainly in the radio band [@Wilson; @Tarter; @Shostak], with peripheral attention to exotic signals in the optical [@Howard; @Horowitz; @Ribak; @Dyson_Plant; @Elvis] and thermal infrared [@Dyson]. Possible “beacon” signals broadcasted intentionally by another civilization to announce its presence as well as the ”leakage” of radiation, produced for communication or other purposes ([*e.g.*]{}, radar), have been the usual targets of radio SETI observations.
As technology evolves on Earth, expectations for plausible extraterrestrial signals change. For example, the radio power emission of the Earth has been declining dramatically in recent decades due to the use of cables, optical fibers and other advances in communication technology, indicating that eavesdropping on distant advanced civilizations might be more difficult than previously thought [@Forgan].
Here we are guided instead by the notion that biological creatures are likely to take advantage of the natural illumination provided by the star around which their home planet orbits. As soon as such creatures develop the necessary technology, it would be natural for them to artificially illuminate the object they inhabit during its dark diurnal phases.
Our civilization uses two basic classes of illumination: thermal (incandescent light bulbs) and quantum (light emitting diodes \[LEDs\] and fluorescent lamps). Such artificial light sources have different spectral properties than sunlight. The spectra of artificial lights on distant objects would likely distinguish them from natural illumination sources, since such emission would be exceptionally rare in the natural thermodynamic conditions present on the surface of relatively cold objects. Therefore, [*artificial illumination may serve as a lamppost which signals the existence of extraterrestrial technologies and thus civilizations*]{}. Are there realistic techniques to search for the leakage of artificial illumination in the optical band?[^1]
It is convenient to normalize any artificial illumination in flux units of 1% of the solar daylight illumination of Earth, $f_\oplus\equiv 0.01(L_\odot/4\pi D_\oplus^2)= 1.4\times 10^4~{\rm
erg~s^{-1}~cm^{-2}}$, where $D_\oplus=1.5\times 10^{13}~{\rm cm}\equiv
1~{\rm AU}$ is the Earth-Sun distance. Crudely speaking, this unit corresponds to the illumination in a brightly-lit office or to that provided by the Sun just as it rises or sets in a clear sky on Earth.[^2]
Artificially Illuminated Kuiper Belt Objects
============================================
We first examine the feasibility of this new SETI technique within the Solar System, which offers the best prospects for detecting intrinsically faint sources of light.
The flux reaching an observer from any self-luminous source varies according to the familiar inverse square law, but the flux from scattered sunlight off an object at a distance $D\gg 1~{\rm AU}$ scales as $D^{-4}$ due to the combination of the inverse square dependence of the solar flux which illuminates it combined with the inverse square dependence of the scattered component of that incident flux which reaches an observer on Earth. Thus, the observed flux from an object that is artificially illuminated at a level of $f_{\oplus}$ would be larger than the flux due to its reflected sunlight by a factor of $(A/1\%)^{-1}(D/1~{\rm AU})^2$, where $A$ is the albedo (reflection coefficient) of the object to sunlight. The $A$ values of objects in the outer solar system vary widely [@Albedos] and their colors range from neutral to very red [@Colors]. This implies that the ratio of artificial illumination, with an unknown spectrum, to scattered sunlight could be a strong function of wavelength.
More than $\sim 10^3$ small bodies have already been discovered in the distance range of $30$–$50~{\rm AU}$, known as the Kuiper belt of the Solar System [@Petit]. The number of known Kuiper belt objects (KBOs) will increase by 1-2 orders of magnitude over the next decade through wide-field surveys such as Pan-STARRS[^3] and LSST.[^4] The sizes[^5] of known KBOs ($\sim 1$–$10^3~{\rm km}$) are usually inferred by assuming a typical albedo [@Grundy] of $A\sim 4$–$10\%$. (The albedo of a KBO can sometimes be calibrated more reliably based on measurements of its thermal infrared emission.[^6]) For $A=7\%$ and a distance $D=50~{\rm AU}$, an artificially $f_\oplus$-illuminated object would be brighter by a factor $\sim
3.6\times 10^2$ than if it were sunlight-illuminated. This implies that an $f_\oplus$-illuminated surface would provide the same observed flux $F$ as a sunlight-illuminated object at that distance, if it is $\sim {\sqrt{3.6\times 10^2}}=19$ times smaller in size. In other words, an $f_\oplus$-illuminated surface of size 53 km (comparable to the scale of a major city) would appear as bright as a $10^3~{\rm km}$ object which reflects sunlight with $A=7\%$. Since $\sim 10^3~{\rm
km}$ objects were already found at distances beyond $\sim 50~{\rm
AU}$, [*we conclude that existing telescopes and surveys could detect the artificial light from a reasonably brightly illuminated region, roughly the size of a terrestrial city, located on a KBO.*]{}
Weaker artificial illumination by some factor $\epsilon<1$ relative to the “1% of daylight on Earth” standard represented by $f_\oplus$, would lower the observed flux by the same factor, since the observed flux scales as $F\propto \epsilon$. Correspondingly, the equivalent object size needed for artificial illumination to produce the same observed flux as due to sunlight illumination, would increase by $\epsilon^{-1/2}$. Nevertheless, [*existing telescopes could detect dimly illuminated regions ($\epsilon\sim 1\%$) hundreds of km in size on the surface of large KBOs.*]{}
The current artificial illumination on the night-side of the Earth has an absolute $r$-band magnitude of roughly 44 (corresponding to $1.7\times 10^{13}~{\rm lumens}$ produced from $\sim 2\times
10^{12}~{\rm Watts}$ of electric power).[^7] [^8] Existing telescopes could see the artificially-illuminated side of the Earth out to a distance of $\sim
10^3~{\rm AU}$, where its brightness in scattered sunlight and in artificial lighting (at current levels) would coincidentally be roughly equal. A present-day major terrestrial city, Tokyo for example,[^9] has an absolute $r$-band magnitude of very roughly 48 with apparent $r$-magnitudes of approximately 16 at a distance of 1 AU, 24 at 30 AU, 26 at 100 AU and 31 (about as faint as the faintest detected objects in the Hubble Ultra-Deep Field) at $10^3~{\rm AU}$.
Although precise numbers depend on many detailed properties of the telescope, instrument, observing conditions (sky brightness, image quality [*etc.*]{}), representative exposure times to reach the aforementioned $r$-band apparent magnitudes at high (50-to-1) signal-to-noise ratio are 1, 500 and 1800 seconds, respectively, for the first three cases with an 8-meter class telescope in good observing conditions and using modern CCD detectors. Reaching $r\sim
31$ is not feasible from the ground and took over $3\times 10^5$ seconds with the 2.4-meter Hubble Space Telescope.
Thus, [*existing optical astronomy facilities are capable of detecting artificial illumination at the levels currently employed on Earth for putative extraterrestrial constructs on the scale of a large terrestrial city or greater out to the edge of the Solar System.*]{}
A Flux-Distance Signature of Artificial Illumination
====================================================
Orbital parameters of Kuiper belt objects (KBOs) are routinely measured[^10] to a precision of $< 10^{-3}$ via astrometric observations [@Petit]. [*A simple but powerful and robust method for identifying artificially-illuminated objects is to measure the variation of the observed flux $F$ as a function of its changing distance $D$ along its orbit.*]{} Sunlight-illuminated objects will show a logarithmic slope of $\alpha \equiv (d\log F/d\log D)= -4$ whereas artificially-illuminated objects should exhibit $\alpha= -2$. The required photometric precision of better than a percent for such measurements (over timescales of years) can be easily achieved with modern telescopes.
If objects with $\alpha=-2$ are discovered, follow-up observations with long exposures on $8-10$ meter and space telescopes could determine their spectra and test whether they are illuminated by artificial thermal (incandescent) or quantum (LED/fluorescent) light sources.[^11] The exposure time requirements to achieve moderate signal-to-noise spectra would be extreme, running to millions of seconds or more, at the faint end of the magnitude range under consideration. However, the motivation to determine the nature and properties of an object showing convincing $\alpha=-2$ behavior would be even more extreme. A complementary follow-up search for artificial radio signals could be conducted with sensitive radio observatories [@Loeb], such as VLA,[^12] ATA,[^13] GMRT,[^14] LOFAR,[^15] MWA,[^16] and PAPER,[^17] which would be able to detect extraordinarily low levels of radio emission by current terrestrial standards. In general, follow-up using all available observational resources would be well justified.
KBOs vary in brightness for reasons other than their changing distance from the Earth and the Sun [@Rab; @Sch; @Lightcurves]. Specific causes include a changing phase angle (due largely to the Earth’s orbital motion) leading to changes in the contributions from coherent backscattering and surface shadowing, outgassing ([*i.e.,*]{} cometary activity), rotation of objects with non-spherical shapes or surface albedo variations, and for some objects occultation by a binary companion. Although the brightness changes associated with these effects are typically tenths of a magnitude and can be larger for some objects, their time scales are short (hours to days in most cases) and, with the exception of outgassing, the resulting variations are periodic. For these reasons it will be necessary to monitor KBO brightnesses frequently and for a period of years in order to model or, at worst, average out other contributions to variability on an object-by-object basis and allow the secular trend with changing distance ([*i.e.,*]{} the $\alpha$ value) to emerge. Fortunately, LSST [@Ivezic] will obtain extensive and very high quality data of precisely this nature for unrelated and conventional purposes. Thus, [*the survey we propose can identify KBO (or asteroid) candidates for intensive follow-up with no investment of additional observational resources.*]{}
We note that artificial lights might also vary on short time scales, either due to their being turned on and off, due to beaming, or due to bright spots appearing and disappearing over the limb as the object rotates.
Night Lights Beyond the Solar System
====================================
The next generation of ground-based telescopes (EELT,[^18] GMT,[^19] and TMT[^20]) as well as space telescopes (JWST,[^21] Darwin,[^22] and TPF[^23]) will be able [@Ria] to search for artificial illumination of extra-solar planets [@Sch1; @Sch2]. Although the $\alpha$ test proposed above for objects in the outer Solar System is not relevant for exoplanets, a search for the orbital phase (time) modulation of the observed flux from the artificial illumination of the night-side on Earth-like planets as they orbit their primary could be used in its place. The observer would see stronger artificial illumination when the dark side of the planet is more in view, exactly the opposite of the case with natural day side illumination from the star. Cloud cover would mask some of the artificial illumination of an Earth-like planet in a stochastic time dependent manner, which might significantly complicate the interpretation of such phase curves.
A preliminary broad-band photometric detection could be improved through the use of narrow-band filters which are tuned to the spectral features of artificial light sources (such as LEDs). For this signature to be detectable, the night side needs to have an artificial brightness comparable to the natural illumination of the day side. Clearly, the corresponding extraterrestrial civilization would need to employ much brighter and more extensive artificial lighting than we do currently since the global contrast between the day and night sides is a factor $\sim 6\times 10^5$ for the present-day Earth. In favorable scenarios, some proposed versions of NASA’s TPF mission would have reasonable prospects of detecting the artificial illumination of an exoplanet if it were at levels a few times greater than $f_\oplus$ or more.
City lights would be easier to detect on a planet which was left in the dark of a formerly-habitable zone after its host star turned into a faint white dwarf. The related civilization would need to survive the intermediate red giant phase of its star. If it does, separating its artificial light from the natural light of a white dwarf, would be much easier than for the original star, both in contrast and in absolute brightness.
Concluding Remarks
==================
In addition to the low prior probability that should probably be assigned to the idea of an alien civilization occupying KBOs, the search proposed in this paper could fail for a host of other plausible reasons. The artificially illuminated spaces might be underground or otherwise shielded for a variety of reasons, such as to avoid wasting of energy or to maintain a stealthy presence. Advanced technology, including biological alteration of sensory organs, might be employed to render very low natural illumination levels useable. Moreover, the most easily detectable signatures might well be in very different bands, such as radio emissions. Thus, as for all other known SETI techniques, a null result would have no clear meaning. However, this is not a sufficient reason to refrain from searching since it is clearly impossible to predict the behaviors or capabilities of unknown alien civilizations with any confidence and because a positive result would carry such immense implications.
Artificially-lit KBOs might have originated from civilizations near other stars. In particular, some small bodies may have traveled to the Kuiper belt through interstellar space after being ejected dynamically from other planetary systems [@Moro]. These objects can be recognized by their hyperbolic orbits. A more hypothetical origin for artificially-lit KBOs involves objects composed of rock and water/ice (asteroids or low-mass planets) that were originally in the habitable zone of the Sun, developed intelligent life, and were later ejected through gravitational scattering with other planets (such as the Earth or Jupiter) into highly eccentric orbits. Such orbits spend most of their time at their farthest (turnaround) distance, $D_{\rm max}$. If this distance is in the Kuiper belt, then the last time these objects came close to Earth was more than $\sim 500~(D_{\rm max}/10^2~{\rm
AU})^{3/2}$ years ago, before the modern age of science and technology began on Earth.
0.45in [**ACKNOWLEDGEMENTS.**]{} We thank F. Dyson, M. Holman and A. Parker for helpful comments. AL was supported in part by NSF grant AST-0907890 and NASA grants NNX08AL43G and NNA09DB30A. ELT gratefully acknowledges support from a Princeton University Global Collaborative Research Fund grant and the World Premier International Research Center Initiative (WPI Initiative), MEXT, Japan.
[51]{}
Doressoundiram, A., Boehnhardt, H., Tegler, S., & Trujillo, C. (2008) Color Properties and Trends of the Transneptunian Objects. In: Barucci, M., Boehnhardt, H., Cruikshank, D., & Morbidelli, A. (Eds), The Solar System Beyond Neptune. Univ. of Arizona Press, Tucson, AZ. pp. 91-104.
, F. J. (1960) Search for Artificial Stellar Sources of Infrared Radiation. [*Science*]{} 131: 1667-1668.
, F. J. (2003) Looking for Life in Unlikely Places: Reasons Why Planets May Not Be the Best Places to Look for Life. [*Int. J. of Astrobiology*]{} 2: 103-110.
, D. H., & [Elvis]{}, M. (2011) Extrasolar Asteroid Mining as Forensic Evidence for Extraterrestrial Intelligence. [*International Journal of Astrobiology*]{} 10: 307-313.
, D. H., & [Nichol]{}, R.C. (2011) A Failure of Serendipity: the Square Kilometre Array Will Struggle to Eavesdrop on Human-Like Extraterrestrial Intelligence. [*International Journal of Astrobiology*]{} 10: 77-81.
, W.M., Noll, K.S., & Stephens, D.C. (2005) Diverse Albedos of Small Trans-Neptunian Objects. [*Icarus*]{} 176: 184-191.
, P., et al. (2001) Targeted and all-sky search for nanosecond optical pulses at Harvard-Smithsonian. (2001) [*Proc. SPIE*]{} 4273: 119-127.
, A., et al. (2007) Initial results from Harvard all-sky optical SETI. [*Acta Astronautica*]{} 61:78-87.
, Z., et al. (2008) LSST: from Science Drivers to Reference Design and Anticipated Data Products. [*preprint*]{} arXiv:0805.2366; see also http://www.lsst.org/lsst/.
, A., & [Zaldarriaga]{}, M. (2007) Eavesdropping on Radio Broadcasts from Galactic Civilizations with Upcoming Observatories for Redshifted 21 cm Radiation. [*Journal of Cosmology & Astroparticle Phys.*]{} 1: 20-32.
, A., Turner, E.L., & Loeb, A. (2009) Will the Large Synoptic Survey Telescope Detect Extra-Solar Planetesimals Entering the Solar System? [*Astrophys. J.*]{} 704: 733-742.
, J.-M. et al. (2011) The Canada-France Ecliptic Plane Survey-Full Data Release: The Orbital Structure of the Kuiper Belt. [*Astron. J.*]{} 142: 131-155.
, D. L., Schaefer, B. E., & Tourtellotte, S. W. (2007) The Diverse Solar Phase Curves of Distant Icy Bodies. I. Photometric Observations of 18 Trans-Neptunian Objects, 7 Centaurs, and Nereid. [*Astron. J.*]{} 133: 26-43.
, P., & Schneider, J. (2007) Improving Earth-Like Planets’ Detection with an ELT: the Differential Radial Velocity Experiment. [*Astron. & Astrophys.*]{} 469: 355-361.
, E., (2006) Search for Temporal Coherence in the Sky. [*SPIE Proc.*]{} 6268: 62683G.
, B. E., Rabinowitz, D. L., & Tourtellotte, S. W. (2009) The Diverse Solar Phase Curves of Distant Icy Bodies II. The Cause of the Opposition Surges and Their Correlations. [*Astron. J.*]{} 137: 129-144.
, J. et al. (2010) The Far Future of Exoplanet Direct Characterization. [*Astrobiology*]{} 10: 121-126.
, J. et al. (2010) Reply to “A Comment on ‘The Far Future of Exoplanet Direct Characterization’-The Case for Interstellar Space Probes” by I.A. Crawford. [*Astrobiology*]{} 10: 857-858.
, S., Lacerda, P., & Ortiz, J. (2008) Photometric Lightcurves of Transneptunian Objects and Centaurs: Rotations, Shapes, and Densities. In: Barucci, M., Boehnhardt, H., Cruikshank, D., & Morbidelli, A. (Eds), The Solar System Beyond Neptune. Univ. of Arizona Press, Tucson, AZ. pp. 129-142.
, S. et al. (2011) Are We Any Closer to Finding Intelligent Life Elsewhere? [*Astrobiology*]{} 11: 487-492; see also http://www.seti.org/.
, J., Grundy, W., Brown, M., Cruikshank, D., Spencer, J., Trilling, D. & Margot, J.-M. (2008) Physical Properties of Kuiper Belt and Centaur Objects: Constraints from the Spitzer Space Telescope. In: Barucci, M., Boehnhardt, H., Cruikshank, D., & Morbidelli, A. (Eds), The Solar System Beyond Neptune. Univ. of Arizona Press, Tucson, AZ. pp. 161-179.
, J. (2001) The Search for Extraterrestrial Intelligence (SETI). [*Ann. Rev. Astron. Astrophys.*]{} 39: 511-548.
, T. L. (2001) The search for Extraterrestrial Intelligence. [*Nature*]{} 409: 1110-1114.
[^1]: Here we focus on illumination in the optical band but identical considerations apply to creatures that evolved to sense radiation in the UV and IR bands, in which stars are also highly luminous.
[^2]: http://www.brillianz.co.uk/data/documents/Lumen.pdf
[^3]: http://pan-starrs.ifa.hawaii.edu/public/home.html
[^4]: http://www.lsst.org/lsst/
[^5]: These sizes correspond to diameters for the larger objects, which are spherical in shape, but are merely characteristic linear scales for the smaller objects which have irregular shapes.
[^6]: http://www.minorplanetcenter.org/iau/lists/Sizes.html
[^7]: http://www.lightinglab.fi/IEAAnnex45/guidebook/11$\_$technical$\%$20potential.pdf
[^8]: This value assumes a Sun-like spectrum in the optical band and an illumination efficiency (lumens/watt) similar to that of the Sun, which is in the range of modern fluorescent and LED lights as well. The choice of the $r$-band is obviously somewhat arbitrary and is meant only for illustrative purposes. The artificial illumination employed by an alien civilization might have a wide range of possible spectra, perhaps correlated with that of the primary star hosting the object on which they evolved.
[^9]: http://www.tepco.co.jp/en/forecast/html/kaisetsu$-$e.html
[^10]: Long-term monitoring of KBOs may also serve to limit or detect deviations from Keplerian orbits due to artificial propulsion.
[^11]: One should also examine images of the dark side of Solar System moons, suspected of hosting liquid water. For example, city lights can be searched for in images taken by the Cassini spacecraft of the dark side of Saturn’s moon, Enceladus.
[^12]: http://www.vla.nrao.edu/
[^13]: http://www.seti.org/ata
[^14]: http://gmrt.ncra.tifr.res.in/
[^15]: http://www.lofar.org/
[^16]: http://www.mwatelescope.org/
[^17]: http://astro.berkeley.edu/$\sim$dbacker/eor/
[^18]: http://www.eso.org/public/teles-instr/e-elt.html
[^19]: http://www.gmto.org/
[^20]: http://www.tmt.org/
[^21]: http://www.jwst.nasa.gov/
[^22]: http://www.esa.int/export/esaSC/120382$\_$index$\_$0$\_$m.html
[^23]: http://planetquest.jpl.nasa.gov/TPF/tpf$\_$index.cfm
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $F$ be a field finitely generated and of transcendence degree one over a $p$-adic field, and let $\ell\neq p$ be a prime. Results of Merkurjev and Saltman show that ${{\text{\rm H}}}^2(F,\mu_\ell)$ is generated by ${{\mathbb Z}}/\ell$-cyclic classes. We prove the “${{\mathbb Z}}/\ell$-length” in ${{\text{\rm H}}}^2(F,\mu_\ell)$ equals the $\ell$-Brauer dimension, which Saltman showed to be two. It follows that all $F$-division algebras of period $\ell$ are crossed products, either cyclic (by Saltman’s cyclicity result) or tensor products of two cyclic division algebras. Our result was originally proved by Suresh assuming $F$ contains $\mu_\ell$.'
address:
- |
Department of Mathematics & Computer Science\
Emory University\
Atlanta, GA 30322\
USA
- |
Instituto de Ciências Matemáticas e de Computação\
Universidade de São Paulo\
São Carlos, São Paulo\
Brazil
author:
- Eric Brussel and Eduardo Tengan
bibliography:
- 'hnx.bib'
title: 'Division algebras of prime period $\ell\neq p$ over function fields of $p$-adic curves'
---
Introduction
============
It is widely believed that for a field $F$ and a prime-to-${{\text{\rm char}}}(F)$ number $n$, the $n$-torsion ${{\text{\rm H}}}^2(F,\mu_n)={}_n{{\text{\rm Br}}}(F)$ of the Brauer group of $F$ is generated by ${{\mathbb Z}}/n$-cyclic classes, i.e., the cup product map ${{\text{\rm H}}}^1(F,\mu_n)\otimes_{{\mathbb Z}}{{\text{\rm H}}}^1(F,{{\mathbb Z}}/n)\to{{\text{\rm H}}}^2(F,\mu_n)$ is surjective. This is true, for example, when $F$ contains the $n$-th roots of unity by Merkurjev-Suslin’s theorem ([@MS83 Theorem 16.1]), when $n=3$ by [@MS83 Corollary 16.4], and when $n=5$ by a recent result of Matzri ([@Mat08 Theorem 4.5]). When ${{\text{\rm H}}}^2(F,\mu_n)$ is generated by ${{\mathbb Z}}/n$-cyclic classes we define the [*${{\mathbb Z}}/n$-length*]{} ${}_n{{\text{\rm L}}}(F)$ in ${{\text{\rm H}}}^2(F,\mu_n)$ to be the smallest number of ${{\mathbb Z}}/n$-cyclic classes needed to write any class, or $\infty$ if no such number exists. If ${}_n{{\text{\rm L}}}(F)=c$ then any class in ${{\text{\rm H}}}^2(F,\mu_n)$ has index at most $n^c$, so that the [*$n$-Brauer dimension*]{} ${}_n{{\text{\rm Br.dim}}}(F)$ is (finite and) bounded by $c$. It is not known whether finite $n$-Brauer dimension implies finite ${{\mathbb Z}}/n$-length, or even whether finite $n$-Brauer dimension implies ${{\text{\rm H}}}^2(F,\mu_n)$ is generated by cyclic classes.
We study this problem when $F$ is a field that is finitely generated and of transcendence degree one over the $p$-adic field ${{\mathbb Q}}_p$. In [@Sa97 Theorem 3.4], Saltman showed that then ${}_n{{\text{\rm Br.dim}}}(F)$ equals two, for any (prime-to-$p$) $n$. Recently Suresh showed that when $n=\ell$ is prime and $F$ contains $\mu_\ell$, the ${{\mathbb Z}}/\ell$-length in ${{\text{\rm H}}}^2(F,\mu_\ell)$ is also two ([@Sur10 Theorem 2.4]). The assumption on roots of unity excludes important cases such as the rational function field $F={{\mathbb Q}}_p(T)$ (if $\ell\neq 2,3$ and $p\neq 1\pmod\ell$). But in this case the cup product map is surjective by Merkurjev’s theorem [@Mer83 Theorem 2] and Saltman’s cyclicity result for classes of prime index [@Sa07 Theorem 5.1], so in any case ${{\text{\rm H}}}^2(F,\mu_\ell)$ is known to be generated by ${{\mathbb Z}}/\ell$-cyclic classes.
We show that for a prime $\ell\neq p$ the ${{\mathbb Z}}/\ell$-length in ${{\text{\rm H}}}^2(F,\mu_\ell)$ is two, hence that all $F$-division algebras of period $\ell$ and index $\ell^2$ decompose into two cyclic $F$-division algebras of index $\ell$. It follows immediately that all $F$-division algebras of period $\ell$ are (abelian) crossed products. Noncrossed products of larger $\ell$-power period exist by [@BMT] and [@BT11]. Our results rely heavily on Saltman’s degree-$\ell$ cyclicity result and his hot point criterion [@Sa07 Corollary 5.2], and our lifting results from [@BT11], which use the machinery of Grothendieck’s existence theorem. We show an $F$-division algebra $\Delta$ of period $\ell$ and index $\ell^2$ is decomposable by explicitly constructing a tensor factor of degree $\ell$, lifting a class constructed over the generic points of the closed fiber of a 2-dimensional model $X/{{\mathbb Z}}_p$, as developed in [@BT11]. Since the function fields of the closed fiber are global fields, we can use class field theory (esp. Grunwald-Wang’s theorem) to manipulate the lifted class so that it cancels the hot points of $[\Delta]$, which implies it is part of a decomposition of $\Delta$ by the hot point criterion. The cyclicity result then shows the remaining factor is cyclic. Suresh’s approach in [@Sur10] similarly cancels $\Delta$’s hot points using a tensor factor, but his tensor factor is constructed as a symbol algebra, which requires $\mu_\ell\subset F$. There is no obvious way to get to the general case from that construction.
Background and Conventions
==========================
\[brauer\] In this paper an [*$F$-division algebra*]{} is a division ring that is central and finite-dimensional over $F$. If $D$ is an $F$-division algebra we write $[D]$ for the class of $D$ in the Brauer group ${{\text{\rm Br}}}(F)$, ${{\text{\rm ind}}}(D)$ for the [*index*]{} or [*degree*]{} of $D$, and ${{\text{\rm per}}}(D)$ for the [*period*]{} of $D$. We say $D$ is a [*crossed product*]{} if it contain a maximal subfield that is Galois over $F$. See [@ABGV] for a discussion of crossed product and noncrossed product division algebras.
We write ${{\text{\rm H}}}^2(F,\mu_n)={}_n{{\text{\rm Br}}}(F)$ for the $n$-torsion subgroup, where $n$ is prime-to-${{\text{\rm char}}}(F)$ and $\mu_n$ is the group of $n$-th roots of unity. In the terminology of [@ABGV Section 4], the [*$n$-Brauer dimension*]{} ${}_n{{\text{\rm Br.dim}}}(F)$ of $F$ is the smallest number $d$ such that every class in ${{\text{\rm H}}}^2(F,\mu_n)$ has index dividing $n^d$, or $\infty$ if no such number exists. We say ${{\text{\rm H}}}^2(F,\mu_n)$ is [*generated by ${{\mathbb Z}}/n$-cyclic classes*]{} if the cup product map ${{\text{\rm H}}}^1(F,\mu_n)\otimes_{{\mathbb Z}}{{\text{\rm H}}}^1(F,{{\mathbb Z}}/n)\to{{\text{\rm H}}}^2(F,\mu_n)$ is surjective, and a class is [*${{\mathbb Z}}/n$-cyclic*]{} if it has the form $(f)\cdot\theta$ for some $(f)\in{{\text{\rm H}}}^1(F,\mu_n)$ and $\theta\in{{\text{\rm H}}}^1(F,{{\mathbb Z}}/n)$. If ${{\text{\rm H}}}^2(F,\mu_n)$ is generated by cyclic classes, the [*${{\mathbb Z}}/n$-length*]{} ${}_n{{\text{\rm L}}}(F)$ is the smallest number of ${{\mathbb Z}}/n$-cyclic classes needed to express an arbitrary class, or $\infty$ if no such number exists. See [@ABGV Section 3] for a discussion of known results regarding ${{\mathbb Z}}/n$-length, usually called “symbol length” when $F$ contains an $n$-th root of unity.
It is clear that if ${{\text{\rm H}}}^2(F,\mu_n)$ is generated by cyclic classes and ${}_n{{\text{\rm L}}}(F)$ is finite then ${}_n{{\text{\rm Br.dim}}}(F)\leq {}_n{{\text{\rm L}}}(F)$. Conversely, as mentioned above, it is not known whether a finite Brauer dimension implies a finite ${{\mathbb Z}}/n$-length, or even that ${{\text{\rm H}}}^2(F,\mu_n)$ is generated by ${{\mathbb Z}}/n$-cyclic classes. However, when $n=\ell$ Merkurjev proved that ${{\text{\rm H}}}^2(F,\mu_\ell)$ is generated by classes of index $\ell$ ([@Mer83 Theorem 2]), hence for the fields considered in this paper ${{\text{\rm H}}}^2(F,\mu_\ell)$ is generated by ${{\mathbb Z}}/\ell$-cyclic classes by Saltman’s cyclicity result.
Let $S$ be an excellent scheme and suppose $n$ is invertible on $S$. We write ${{\mathbb Z}}/n(r)$ for the étale sheaf ${{\mathbb Z}}/n$ twisted by an integer $r$, and ${{\text{\rm H}}}^q(S,r)={{\text{\rm H}}}^q(S,{{\mathbb Z}}/n(r))$ for the étale cohomology group. If $S={{\text{\rm Spec}\,}}A$ for a ring $A$, we write ${{\text{\rm H}}}^q(A,r)$. If $T$ is a subscheme of $S$ we write $\kappa(T)$ for its ring of meromorphic functions, which is the localization of ${{\text{\rm O}}}_T$ at all associated points. If $T\to S$ is a morphism of schemes then the restriction ${{\text{\rm res}}}_{T|S}:{{\text{\rm H}}}^q(T,r)\to{{\text{\rm H}}}^q(S,r)$ is defined, and we write $\beta_S={{\text{\rm res}}}_{T|S}(\beta)$, and if $S={{\text{\rm Spec}\,}}A$ we write $\beta_A$. If $Z\subset S$ is a subscheme we write $Z_T$ for the preimage $Z\times_S T$.
If $v$ is a valuation on a field $F$, we write $\kappa(v)$ for the residue field of the valuation ring ${{\text{\rm O}}}_v$, and $F_v$ for the completion of $F$ with respect to $v$. If $v$ arises from a prime divisor $D$ on $S$, we write $v=v_D$, $\kappa(D)$, and $F_D$. If a set $\{v_i\}$ arises from a divisor $D=\sum_i D_i$, we write $F_D=\prod_i F_{D_i}$. Recall that if $F=(F,v)$ is a discretely valued field and $\alpha\in{{\text{\rm H}}}^q(F,r)$, then $\alpha$ has a [*residue*]{} $\partial_v(\alpha)$ in ${{\text{\rm H}}}^{q-1}(\kappa(v),r-1)$. More generally if $\xi$ is a generic point of a scheme $S$, $F=\kappa(\xi)$, and $\alpha\in{{\text{\rm H}}}^q(S,r)$, then for each discrete valuation $v$ on $F$ we define $$\partial_v(\alpha){{\,\overset{\text{\rm df}}{=}\,}}\partial_v(\alpha_F)\in{{\text{\rm H}}}^{q-1}(\kappa(v),r-1)$$ We say $\alpha$ is [*unramified*]{} with respect to $v$ if $\partial_v(\alpha)=0$, and in that case the [*value*]{} of $\alpha$ at $v$ is the element $\alpha(v)={{\text{\rm res}}}_{F|F_{v}}(\alpha)\in{{\text{\rm H}}}^q(\kappa(v),r)\leq{{\text{\rm H}}}^q(F_v,r)$ ([@GMS 7.13, p.19]). If $v$ arises from a prime divisor $D$ on a scheme, we will substitute the notations $\partial_D$ and $\alpha(D)$. If $S$ is noetherian we write $D_\alpha$ for the [*ramification divisor of $\alpha$ on $S$*]{}, which is the sum of (finitely many) prime divisors on $S$ at which $\alpha$ ramifies.
\[setup\] In the following, $F$ will always be a finitely generated field extension of ${{\mathbb Q}}_p$ of transcendence degree one, $n$ will be a prime-to-$p$ number, and $X/{{\mathbb Z}}_p$ will be a connected regular (projective, flat) relative curve over ${{\mathbb Z}}_p$ with function field $F=K(X)$. Such a surface exists for any $F$ by a theorem of Lipman (see [@Liu Theorem 8.3.44]). We write $X_0=X\otimes_{{{\mathbb Z}}_p}{{\mathbb F}}_p$ for the closed fiber, $C=X_{0,{{\text{\rm red}}}}$ for the reduced scheme underlying the closed fiber, $C_1,\dots,C_m$ for the irreducible components of $C$, and ${{\mathcal S}}$ for the set of singular points of $C$. We assume that $X_0$ has normal crossings, hence that each $C_i$ is regular, and at most two of them meet (transversally) at each singular point of $C$. This is permitted by embedded resolution of curves in surfaces (see [@Liu Theorem 9.2.26]).
We say an effective divisor $D$ on $X$ is [*horizontal*]{} if each of its irreducible components maps surjectively to ${{\text{\rm Spec}\,}}{{\mathbb Z}}_p$. By [@BT11 Proposition 2.4] there exists for each closed point $z\in X\backslash{{\mathcal S}}$ a regular irreducible horizontal divisor $D\subset X$ that intersects $C$ transversally. Let $\mathscr D_{{\mathcal S}}$ denote the support of these lifts. We say a divisor $D$ is [*distinguished*]{} and write $D\in\mathscr D_{{\mathcal S}}$ if it is reduced and supported in $\mathscr D_{{\mathcal S}}$. Each $D\in\mathscr D_{{\mathcal S}}$ is a disjoint union of its irreducible components, each of which has a single closed point and meets $C$ transversally.
By weak approximation ([@Sa98 Lemma]), we may choose an element $\pi\in F$ such that ${{\text{\rm div}}}(\pi)=C+E\subset X$, where $E$ is horizontal and avoids all closed points of any finite set containing ${{\mathcal S}}$.
Suppose $\ell\neq p$ is prime, $\alpha\in{{\text{\rm H}}}^2(F,\mu_\ell)$, and $D_\alpha\subset X$ has normal crossings. Following Saltman’s terminology in [@Sa07] we say $\alpha$ has a [*hot point*]{} $z$ on $X$ if (and only if) $z$ is a nodal point of $D_\alpha$, and if $D,D'\subset D_\alpha$ are the two irreducible components meeting transversally at $z$, then $\partial_D(\alpha)$ and $\partial_{D'}(\alpha)$ are unramified at $z$, and ${{\left<{\partial_D(\alpha)(z)}\right>}}\neq{{\left<{\partial_{D'}(\alpha)(z)}\right>}}$. By [@Sa07 Corollary 5.2], $\alpha$ has index $\ell$ if and only if $D_\alpha$ has no hot points (hot point criterion), and by [@Sa07 Theorem 5.1], if $\alpha$ has index $\ell$ then it is cyclic.
\[lambda\] Assume the setup of .
[a)]{} There is a decomposition ${{\text{\rm H}}}^1({{\text{\rm O}}}_{C,{{\mathcal S}}},{{\mathbb Z}}/n){{\;\simeq\;}}({{\mathbb Z}}/n)^{\beta_C}\oplus\Gamma$, where $\beta_C$ is the Betti number of the dual graph of $C$, and $\Gamma\leq{{\text{\rm H}}}^1(\kappa(C),{{\mathbb Z}}/n)$ is the set of tuples $\theta_C=(\theta_1,\dots,\theta_m)\in{{\text{\rm H}}}^1(\kappa(C),{{\mathbb Z}}/n)$ such that each $\theta_i$ is unramified at each $z\in{{\mathcal S}}\cap C_i$, and $\theta_C(z)=\theta_i(z)=\theta_j(z)\in{{\text{\rm H}}}^1(\kappa(z),{{\mathbb Z}}/n)$ whenever $z\in C_i\cap C_j$.
[b)]{} For $q\geq 0$ and any integer $r$ there is a map $$\lambda:{{\text{\rm H}}}^q({{\text{\rm O}}}_{C,{{\mathcal S}}},r)\to{{\text{\rm H}}}^q(F,r)$$ and a commutative diagram $$\xymatrix{
{{\text{\rm H}}}^q({{\text{\rm O}}}_{C,{{\mathcal S}}},r)\ar[r]^-\lambda\ar[d]_{\oplus{{\text{\rm res}}}_i}&{{\text{\rm H}}}^q(F,r)\ar[d]^{\oplus_i{{\text{\rm res}}}_i}\\
{{\text{\rm H}}}^q(\kappa(C),r)\ar[r]^-{{\text{\rm inf}\,}}&\bigoplus_i{{\text{\rm H}}}^q(F_C,r)
}$$ such that if $\alpha_C\in{{\text{\rm H}}}^q({{\text{\rm O}}}_{C,{{\mathcal S}}},r)$ and $\alpha=\lambda(\alpha_C)$ then:
1. $\alpha$ is defined at the generic points of $C_i$, and $\alpha(C_i)={{\text{\rm res}}}_i(\alpha_C)$.
2. The ramification locus of $\alpha$ (on $X$) is contained in $\mathscr D_{{\mathcal S}}$.
3. If $D\in\mathscr D_{{\mathcal S}}$ is prime and $z=D\cap C$, then $\partial_D\cdot\lambda={{\text{\rm inf}\,}}_{\kappa(z)|\kappa(D)}\cdot\partial_z$.
4. If $\alpha_C$ is unramified at a closed point $z$, and $D$ is any (horizontal) prime lying over $z$, then $\alpha$ is unramified at $D$, and has value $\alpha(D)={{\text{\rm inf}\,}}_{\kappa(z)|\kappa(D)}(\alpha_C(z))$.
Computations
============
We first construct the cyclic class $\gamma\in{{\text{\rm H}}}^2(F,\mu_n)$ using a lift from ${{\text{\rm H}}}^1(\kappa(C),{{\mathbb Z}}/n)$.
\[lemma\] Assume the setup of . Suppose $\theta_C\in\Gamma\leq{{\text{\rm H}}}^1({{\text{\rm O}}}_{C,{{\mathcal S}}},{{\mathbb Z}}/n)$ lifts $(\theta_1,\dots,\theta_m)\in{{\text{\rm H}}}^1(\kappa(C),{{\mathbb Z}}/n)$ as in Theorem \[lambda\](a), such that $\theta_C$ is unramified at all $z\in E\cap C$, with value $\theta_C(z)=0$ (in addition to being unramified at ${{\mathcal S}}$). Let $\gamma=(\pi)\cdot\lambda(\theta_C)\in{{\text{\rm H}}}^2(F,\mu_n)$. Then for any prime divisor $D$ on $X$ $$\partial_D(\gamma)=\begin{cases}
\theta_i &\text{ if $D=C_i$}\\
-{{\text{\rm inf}\,}}_{\kappa(z)|\kappa(D)}(\partial_{z}(\theta_C))\cdot(\pi)
&\text{ if $D\in\mathscr D_{{\mathcal S}}$ and $z=D\cap C$}\\
0&\text{ otherwise}\\
\end{cases}$$ The ramification divisor $D_\gamma$ has normal crossings, and consists of each $C_i$ at which $\theta_i$ is nonzero, together with all $D\in\mathscr D_{{\mathcal S}}$ lifting $z:\partial_z(\theta_C)\neq 0$.
Set $\theta=\lambda(\theta_C)$. For any prime divisor $D$ on $X$ we compute $$\partial_D(\gamma)=\big[v_D(\pi)\theta-\partial_D(\theta)\cdot(\pi)
+v_D(\pi)\partial_D(\theta)\cdot(-1)\big]_{F_D}$$ This element is in the subgroup ${{\text{\rm H}}}^1(\kappa(D),{{\mathbb Z}}/n)\leq{{\text{\rm H}}}^1(F_D,{{\mathbb Z}}/n)$. Let $z=D\cap C$. There are several cases to consider.
If $D=C_i$ is an irreducible component of $C$ then since $E$ contains no components of $C$ (by ) we have $v_D(\pi)=1$, and since $\theta_C\in{{\text{\rm H}}}^1({{\text{\rm O}}}_{C,{{\mathcal S}}},{{\mathbb Z}}/n)$ we have $\partial_D(\theta)=0$ by Theorem \[lambda\](b)(ii). Therefore $\partial_D(\gamma)={{\text{\rm res}}}_{F|F_{C_i}}(\theta)=\theta_i\in{{\text{\rm H}}}^1(\kappa(C_i),{{\mathbb Z}}/n)$.
If $D$ is horizontal and runs through a point of ${{\mathcal S}}$, then $v_D(\pi)=0$ since $E$ avoids ${{\mathcal S}}$, and $\partial_D(\theta)=0$ by Theorem \[lambda\](b)(ii), hence $\partial_D(\gamma)=0$.
If $D$ is horizontal, avoids ${{\mathcal S}}$, and $v_D(\pi)\neq 0$, then $D$ is a component of $E$, so by assumption, $\partial_z(\theta_C)=0$ for $z\in E\cap C$ and $\theta_C(z)=0$. Thus $\partial_D(\theta)=0$ and $\theta(D)=0$ by Theorem \[lambda\](b)(iv). Therefore $\partial_D(\gamma)=0$.
If $D$ is horizontal, avoids ${{\mathcal S}}$, and $v_D(\pi)=0$, then $\partial_D(\gamma)=-\partial_D(\theta)\cdot(\pi)$. If $D\not\in\mathscr D_{{\mathcal S}}$ this is zero by Theorem \[lambda\](b)(ii). If $D\in\mathscr D_{{\mathcal S}}$ then $\partial_D(\gamma)=-{{\text{\rm inf}\,}}_{\kappa(z)|\kappa(D)}(\partial_{z}(\theta_C))\cdot(\pi)$ by Theorem \[lambda\](b)(iii), and since $\pi$ is a local equation for $C$ at $z$ and $D$ intersects $C$ transversally at $z$, the image of $\pi$ in the local field $\kappa(D)$ is a uniformizer, hence $(\pi)$ has order $n$ in ${{\text{\rm H}}}^1(\kappa(D),\mu_n)$. Thus if $D$ is horizontal then $\partial_D(\gamma)$ is nonzero if and only if $D\in\mathscr D_{{\mathcal S}}$ and $\partial_z(\theta_C)\neq 0$, and then $\partial_D(\gamma)=-{{\text{\rm inf}\,}}_{\kappa(z)|\kappa(D)}(\partial_{z}(\theta_C))\cdot(\pi)$.
We conclude the ramification divisor $D_\gamma$ of $\gamma$ consists of the components $C_i$ of $C$ for which $\theta_i$ is nonzero, together with the distinguished prime divisors $D\in\mathscr D_{{\mathcal S}}$ lying over points $z$ at which $\theta_C$ is ramified. Since all such $D$ are regular and intersect $C$ transversally, $D_\gamma$ has normal crossings.
Next we show an $F$-division algebra of prime period $\ell\neq p$ and index $\ell^2$ is decomposable by constructing a cyclic factor using Lemma \[lemma\], designed to cancel the division algebra’s hot points.
\[theorem\] Let $F$ be a field finitely generated of transcendence degree one over ${{\mathbb Q}}_p$, and suppose $\Delta$ is an $F$-division algebra of prime period $\ell\neq p$ and index $\ell^2$. Then $\Delta$ is decomposable.
We may assume $\ell$ is odd, since if $\ell=2$ the result is a classical theorem of Albert. Assume the setup , let $\alpha=[\Delta]\in{{\text{\rm H}}}^2(F,\mu_\ell)$, and let $D_\alpha$ be the ramification divisor of $\alpha$ on $X$. We may assume $D_\alpha\cup C$ has normal crossings and horizontal components contained in $\mathscr D_{{\mathcal S}}$, and that we have an element $\pi\in F$ as in with ${{\text{\rm div}}}(\pi)=C+E$, where $E$ is horizontal and avoids the nodal points of $D_\alpha\cup C$.
By Grunwald-Wang’s theorem applied over the global fields $\kappa(C_i)$ there exist elements $\theta_i\in{{\text{\rm H}}}^1(\kappa(C_i),{{\mathbb Z}}/\ell)$, $i=1,\dots,m$, with the following properties:
1. $\partial_z(\theta_i)=0$ when $z\in C_i$ is a singular point of $C\cup D_\alpha\cup E$.
2. $\theta_i(z)=\theta_j(z)$ whenever $z\in C_i\cap C_j$.
3. $\theta_i(z)=0$ at all $z\in E\cap C_i$.
4. If $z\in D_\alpha\cap{{\mathcal S}}$ then (i) ${{\left<{\partial_{C_i}(\alpha)(z)-\theta_i(z)}\right>}}={{\left<{-\theta_i(z)}\right>}}$ if $z\in C_i$ is a curve point of $D_\alpha$; (ii) $\theta_i(z)=0$ if $z\in C_i$ is a not-hot nodal point of $D_\alpha$; (iii) ${{\left<{\partial_{C_i}(\alpha)(z)-\theta_i(z)}\right>}}={{\left<{\partial_{C_j}(\alpha)(z)-\theta_i(z)}\right>}}$ if $z\in C_i\cap C_j$ is a hot point of $\alpha$.
5. If $z\in D_\alpha\backslash{{\mathcal S}}$ then (i) ${{\left<{-\theta_i(z)}\right>}}={{\left<{\partial_D(\alpha)(z)}\right>}}$ if $z\in C_i$ is a curve point of $D_\alpha$; (ii) $\theta_i(z)=0$ if $z\in C_i$ is a not-hot nodal point of $D_\alpha$; (iii) ${{\left<{\partial_{C_i}(\alpha)(z)-\theta_i(z)}\right>}}={{\left<{\partial_D(\alpha)(z)}\right>}}$ if $z\in C_i\cap D$ is a hot point of $\alpha$.
Note that (d)(i,iii) and (e)(i,iii) makes sense since the given residues of $\alpha$ are unramified at the given $z$; (d)(i,iii) and (e)(iii) are possible since $\ell$ is odd; (c) does not conflict with (d)(i,iii) and (e)(i,iii) since $E$ avoids the nodal points of $D_\alpha\cup C$; and (b) does not conflict with (d)(ii,iii) and (e)(ii,iii) by symmetry.
The $\theta_i$ are unramified with equal values at all nodal points $z\in{{\mathcal S}}$ by (a,b), so they glue together to produce an element $\theta_C\in\Gamma\leq{{\text{\rm H}}}^1({{\text{\rm O}}}_{C,{{\mathcal S}}},{{\mathbb Z}}/\ell)$ by Theorem \[lambda\](a). Note that $\theta_C$ is nonzero by (d)(iii) (or (e)(iii)) since $\alpha$ has at least one hot point by the hot point criterion, and then $\theta_C(z)=\theta_i(z)$ is necessarily nonzero.
Let $\gamma_1=(\pi)\cdot\lambda(\theta_C)$. Then $\gamma_1$ and $E$ satisfy the hypotheses of Lemma \[lemma\] by (c) and the assumptions on $E$, hence $D_{\gamma_1}\cup C$ has normal crossings and distinguished horizontal components, and since $\theta_C$ is ramified at all nodal points of $D_{\gamma_1}$, $\gamma_1$ has no hot points, hence it has index $\ell$ by [@Sa07 Corollary 5.2]. Write $$\begin{aligned}
D_\alpha&=C'+H\\
D_{\gamma_1}&=C''+H'\\\end{aligned}$$ where $C',C''\subset C$, and $H,H'\subset\mathscr D_{{\mathcal S}}$ are distinguished horizontal divisors. Set $$\gamma_2=\alpha-\gamma_1$$ We intend to show that $\gamma_2$ has index $\ell$. Since $\theta_C$ is unramified at all singular points of $D_\alpha\cup C$ by (a), $H'$ avoids all of these points by Lemma \[lemma\], hence $H\cap H'=\varnothing$. Evidently then $D_{\gamma_2}\subset C+H+H'.$ Now $D_{\gamma_2}$ has normal crossings on $X$ since $D_\alpha\cup C$ and $D_{\gamma_1}\cup C$ both have normal crossings. Since $D_{\gamma_2}$ has normal crossings, $\gamma_2$ has index $\ell$ if and only if $\gamma_2$ has no hot points on $X$. For the following analysis, note the nodal points ${{\mathcal S}}_{\gamma_2}$ of $D_{\gamma_2}$ are in $H'$, ${{\mathcal S}}$, and $H$, and in the latter two cases $\theta_C$ is unramified at $z$, hence has a value $\theta_C(z)$.
Suppose $z\in{{\mathcal S}}_{\gamma_2}$ and $z\not\in D_\alpha$. Then $z$’s status as a point of $D_{\gamma_2}$ (hot, not hot) is the same as its status as a point of $D_{\gamma_1}$, hence it is not hot by Lemma \[lemma\] (it is a “cold” point in the terminology of [@Sa07]).
Suppose $z\in {{\mathcal S}}_{\gamma_2}\cap D_\alpha\cap H'$. Then $z\in C_i\cap D$ for some $C_i$ and some prime divisor $D\subset H'$, and $\theta_C$ is ramified at $z$ by Lemma \[lemma\]. We have $\partial_D(\alpha)=0$ since $D\not\subset D_\alpha$, and by Lemma \[lemma\] $$\partial_{z}(\partial_D(\gamma_2))=\partial_{z}(-\partial_D(\gamma_1))
=\partial_{z}(\partial_{z}(\theta_C)\cdot(\pi))
=v_{z}(\bar\pi)\partial_{z}(\theta_C)$$ where $\bar\pi$ is the image of $\pi$ in $\kappa(D)$. Since ${{\text{\rm div}}}(\pi)$ has normal crossings at $z$, $v_{z}(\bar\pi)=1$, hence $\partial_{z}(\partial_D(\gamma_2))=\partial_{z}(\theta_C)\neq 0$. Since $\partial_z(\partial_D(\gamma_2))\neq 0$, $z$ is not a hot point of $\gamma_2$.
Suppose $z\in {{\mathcal S}}_{\gamma_2}\cap D_\alpha\cap{{\mathcal S}}$. If $z$ is a curve point of $D_\alpha$ on $C_i$, i.e., $\partial_{C_j}(\alpha)=0$ where $C_j$ is the other component of $C$ at $z$, then $\partial_{C_i}(\gamma_2)(z)=\partial_{C_i}(\alpha)(z)-\theta_C(z)$ and $\partial_{C_j}(\gamma_2)(z)=-\theta_C(z)$ by Lemma \[lemma\], and so $z$ is not a hot point of $\gamma_2$ by (d)(i). If $z$ is a nodal point of $D_\alpha$ on $C_i\cap C_j$, then $\theta_C(z)=0$ if $z$ is not a hot point of $\alpha$ by (d)(ii), so that the status of $z$ for $\gamma_2$ is the same as for $\alpha$ (not hot); otherwise ${{\left<{\partial_{C_i}(\gamma_2)(z)}\right>}}={{\left<{\partial_{C_j}(\gamma_2)(z)}\right>}}$ by (d)(iii), hence $z$ is not a hot point for $\gamma_2$ in any case.
Suppose $z\in{{\mathcal S}}_{\gamma_2}\cap D_\alpha\cap H$. Assume $z\in C_i\cap D$ for a prime divisor $D\subset H$. Then $\partial_D(\gamma_1)=0$ since $H\cap H'=\varnothing$. If $z$ is a curve point of $D_\alpha$, i.e., $\partial_{C_i}(\alpha)=0$, then $\partial_{C_i}(\gamma_2)(z)=-\theta_C(z)$ and $\partial_D(\gamma_2)=\partial_D(\alpha)$, hence ${{\left<{\partial_{C_i}(\gamma_2)(z)}\right>}}={{\left<{\partial_D(\gamma_2)(z)}\right>}}$ by (e)(i), so $z$ is not a hot point for $\gamma_2$. If $z$ is a not-hot nodal point of $D_\alpha$ then $\theta_C(z)=0$ by (e)(ii), so the status of $z$ is unchanged (not hot) for $\gamma_2$. If $z$ is a hot point of $\alpha$ then ${{\left<{\partial_{C_i}(\alpha)(z)-\theta_C(z)}\right>}}={{\left<{\partial_D(\alpha)(z)}\right>}}$ by (e)(iii), hence $z$ is not a hot point for $\gamma_2$. This completes the analysis. We conclude $\gamma_2$ has no hot points on $X$, hence $\gamma_1$ and $\gamma_2$ both have index $\ell$.
Let $\Delta_1$ and $\Delta_2$ be the $F$-division algebras underlying $\gamma_1$ and $\gamma_2$, respectively, so that $[\Delta]=[\Delta_1\otimes_F\Delta_2]$. Since ${{\text{\rm ind}}}(\Delta_1\otimes_F\Delta_2)={{\text{\rm ind}}}(\Delta)=\ell^2$ and ${{\text{\rm ind}}}(\Delta_i)=\ell$, it follows that $\Delta_1\otimes_F\Delta_2$ is a division algebra, hence we have a decomposition $\Delta{{\;\simeq\;}}\Delta_1\otimes_F\Delta_2$.
As mentioned in , it is known that ${}_\ell{{\text{\rm Br.dim}}}(F)=2$, but not known in general whether the ${{\mathbb Z}}/\ell$-length in ${{\text{\rm H}}}^2(F,\mu_\ell)$ is finite. We now have the following.
Let $F$ be a field finitely generated and of transcendence degree one over ${{\mathbb Q}}_p$. Then every element of ${{\text{\rm H}}}^2(F,\mu_\ell)$ is a sum of two ${{\mathbb Z}}/\ell$-cyclic classes.
If $\alpha\in{{\text{\rm H}}}^2(F,\mu_\ell)$ then the index of $\alpha$ is either $\ell$ or $\ell^2$ by [@Sa97 Theorem 3.4]. If it is $\ell$, then $\alpha$ is already ${{\mathbb Z}}/\ell$-cyclic by [@Sa07 Theorem 5.1]. If it is $\ell^2$ then $\alpha=\gamma_1+\gamma_2$ for classes $\gamma_i$ in ${{\text{\rm H}}}^2(F,\mu_\ell)$ of index $\ell$ by Theorem \[theorem\]. These classes are again ${{\mathbb Z}}/\ell$-cyclic by Saltman’s theorem, and the result follows.
Saltman proved that all $F$-division algebras of prime degree $\ell$ are cyclic crossed products in [@Sa07], and Suresh proved the prime period case when $F$ contains the $\ell$-th roots of unity in [@Sur10]. We now have the prime period case in general:
Let $F$ be a field finitely generated and of transcendence degree one over ${{\mathbb Q}}_p$, and let $\Delta$ be a division algebra of prime period $\ell\neq p$. Then $\Delta$ is a crossed product.
The index of $\Delta$ is either $\ell$ or $\ell^2$ by [@Sa97 Theorem 3.4]. If it is $\ell$, then $\Delta$ is a cyclic crossed product by [@Sa07 Theorem 5.1]. If it is $\ell^2$ then $\Delta=\Delta_1\otimes_F\Delta_2$ by Theorem \[theorem\], and each $\Delta_i$ is cyclic by Saltman’s theorem. Let $L_i/F$ be a cyclic Galois maximal subfield of $\Delta_i$. Then $L=L_1\otimes_F L_2$ is a commutative Galois subalgebra of $\Delta$ of degree $\ell^2$, hence it is a Galois field extension of $F$, since $\Delta$ is a division algebra. Since $L$ obviously splits $\Delta$, it is a Galois maximal subfield of $\Delta$, hence $\Delta$ is a crossed product.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We develop an approximate analytical technique for evaluating the performance of multi-hop networks based on beaconless , a popular standard for wireless sensor networks. The network comprises sensor nodes, which generate measurement packets, relay nodes which only forward packets, and a data sink (base station). We consider a detailed stochastic process at each node, and analyse this process taking into account the interaction with neighbouring nodes via certain time averaged unknown variables (e.g., channel sensing rates, collision probabilities, etc.). By coupling the analyses at various nodes, we obtain fixed point equations that can be solved numerically to obtain the unknown variables, thereby yielding approximations of time average performance measures, such as packet discard probabilities and average queueing delays. The model incorporates packet generation at the sensor nodes and queues at the sensor nodes and relay nodes. We demonstrate the accuracy of our model by an extensive comparison with simulations. As an additional assessment of the accuracy of the model, we utilize it in an algorithm for sensor network design with quality-of-service (QoS) objectives, and show that designs obtained using our model actually satisfy the QoS constraints (as validated by simulating the networks), and the predictions are accurate to well within 10% as compared to the simulation results.'
author:
- |
Rachit Srivastava, Sanjay Motilal Ladwa, Abhijit Bhattacharya and Anurag Kumar\
Dept. of Electrical Communication Engineering\
Indian Institute of Science, Bangalore, 560012, India\
email: {rachitsri, sanjofpesit}@gmail.com, {abhijit, anurag}@ece.iisc.ernet.in
bibliography:
- 'journal\_paper.bib'
title: 'A Fast and Accurate Performance Analysis of Beaconless Multi-Hop Networks'
---
Introduction {#sec:introduction}
============
Wireless sensor networks (WSNs), a concept that originated in the mid-1990s, have now reached a stage in their evolution that we are beginning to see their actual deployment. It is not unreasonable to expect that in 10-15 years the world will be covered with wireless sensor networks with access to them via the Internet[@wireless-sensor-network]. [@IEEE802-15-4-06std] is a popular standard for the physical layer and medium access control for low-power wireless sensor networks. With the growing importance of wireless sensor networks in industrial applications [@gungor-hancke09industrial-WSN-survey], we need analysis and design techniques for multi-hop IEEE 802.15.4 networks. The standard uses CSMA/CA for medium access control, and defines two types of CSMA/CA algorithms - beaconed (or slotted) and beaconless (or unslotted), respectively (see [@IEEE802-15-4-06std] for details). In this paper we develop a new approximate analytical technique for multi-hop beaconless networks, and demonstrate the usefulness of such an analytical tool in designing multi-hop networks with Quality of Service (QoS) objectives.
![A WSN consisting of nodes arranged in a tree topology. Source node $i$ has packet generation rate $\lambda_i$. Relay nodes $6$, $7$ and $8$ forward the data of their predecessors to the base-station, BS.[]{data-label="fig:network"}](figures/tree_with_queue.pdf)
Figure \[fig:network\] depicts the queueing schematic of a network of the type that we are concerned with in this paper. There are sensor nodes, namely, 1, 2, 3, 4, 5 and 7, that generate measurement packets according to independent point processes at rate $\lambda_i$ that need to be transported to the base station (BS). The sensor nodes can also serve as *relays* for the traffic of other sensor nodes, e.g., Node 7 is a relay for nodes 2 and 3. Two additional nodes, 6 and 8, serve only as relay nodes, due to the limited range of the radios associated with the sensor nodes. The nodes use unslotted CSMA/CA to contend for the wireless medium and transmit their packets. Even the stability analysis of such networks is a difficult problem (even for networks with single hop traffic; see [@bordenave] for a detailed discussion). In principle, the entire network can be modeled via a large coupled Markov chain, with the state at each node being the number of packets in its queue, and the state of the contention process for the head-of-the-line (HOL) packet. Such an approach is well known to be intractable even for a network with a single contention domain (no hidden network) and saturated queues (see [@bianchi00performance; @kumar-etal04new-insights]). Thus all researchers have taken recourse to developing approximate analyses. Table \[tbl:related-literature\] summarizes some of the recent work on modeling approaches for beaconless IEEE 802.15.4 networks where we have listed the major limitations of the proposed models as well. Owing to the need for low power operation and large coverage, in general, a sensor network is multi-hop with the presence of hidden nodes. Further, for networks that carry measurement traffic, the queue occupancies vary with time. Unlike other models, we consider all of these attributes in our modeling approach.
**Authors** **Network Scenario & Limitations** **Modeling Approach**
------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Kim et al. (2006)[@kim-etal06unslotted-CSMACA] beaconless 802.15.4; **rare packet arrivals, star network; no hidden nodes; no queueing at nodes**; with ACKs 2-dim discrete time Markov chain for each node with states: (backoff stage, backoff counter) while in backoff, and separate states when transmitting and idling; with unknown transition probabilities; coupling the per node models to get the unknowns via fixed point approach
Qiu et al. (2007)[@winet.qiu-etal07general-model-wireless-interference] 802.11 DCF; (un)saturated, **star network**; with hidden nodes and ACKs; infinite buffer space discrete time Markov chain with states as the set of transmitting nodes having unknown transition probabilities, together with RSS measurements between each pair of node to obtain CCA and packet failure probabilities; iterative procedure for solving unknowns
Kim et al. (2008)[@kim-etal08modeling-802.15.4-single-hop] beaconless 802.15.4; **star topology; no hidden nodes; no collisions**; infinite buffer space unsaturated analysis of single hop network by simple mathematical model of M/G/1 queue; no collision is considered; computes QoS as packet delay and packet loss probability; computes battery life time
Singh et al. (2008)[@winet.singh-etal08slotted-zigbee-star] beaconed 802.15.4; (un)saturated, **star network; no hidden nodes**; with ACKs; infinite buffer; **equal traffic arrival rate at all nodes** Markov renewal process at each node with cycle length dependent on the number of nodes available to attempt in the first backoff period of the cycle; transition probabilities and conditional expectations using node decoupling; coupling via stationary probabilites of remaining nodes; saturation analysis results used for the unsaturated case
Buratti-Verdone (2009)[@buratti-verdone09modeling-802.15.4-beaconless] beaconless 802.15.4; **star topology; no hidden nodes; no ACKs and retransmissions; no queuing** bidimensional process with backoff counter and backoff stage at time $t$; evaluation of sensing, transmission and success probability; computes mean energy spent by each node for a transmission
He et al. (2009)[@winet.he-etal09slotted-CSMACA] **beaconed** 802.15.4; **saturated, star network; no hidden nodes**; with ACKs; two Markov chains for each node: one embedded at the end of transmissions in the common channel, and the other during backoffs embedded at each slot; both having states as (backoff stage, backoff counter); chains coupled to get a fixed point approach
Martalo et al. (2009)[@winet.martalo-etal09slotted-CSMACA] **beaconed** 802.15.4; unsaturated, **star network; no hidden nodes; no ACKs; finite buffer space** two semi-Markov processes: one for tagged node having unknown transition probabilities, and the other for the shared radio channel that couples all the per node models to obtain the unknown quantities; MGFs used for finding queueing delays
Goyal et al. (2009)[@goyal-etal09beaconless-zigbee] beaconless 802.15.4; **unsaturated star network; no hidden nodes; no queueing at nodes**; with ACKs tracked the number of nodes with non-empty queues assuming that the probability of $m$ nodes having non-empty queues at any given time is same as the probability of $m-1$ empty nodes getting a new packet to send while a non-empty node is sending its current packet (for coupling individual node models); assumed that the CCA attempt processes of other nodes with non-empty queues are Poisson processes (decoupling)
Lauwens et al. (2009)[@lauwens-etal09unslotted-CSMACA-analysis] beaconless 802.15.4; **saturated, star network; no hidden nodes; no ACKs** semi-Markov model for each node with unknown and non-homogeneous transition probabilities based on the backoff stage of the node; also found the distribution of backoff intervals and iterated over these together with the stationary probabilities of the semi-Markov processes
Jindal-Psounis (2009)[@jindal09] 802.11 DCF; unsaturated, multi-hop network; with hidden nodes and ACKs; infinite buffer space; **modeling involves details that are specific to 802.11 such as the RTS/CTS mechanism**; For each of four possible classes of two-edge topologies, derived collision and idle probabilities for each edge in terms of expected service times of the edges; for a general topology, decomposed the local network topology around each edge into a number of two-edge topologies, and derived the collision and idle probabilities for the edge in terms of those of the two-edge topologies in the decomposition; using these probabilities, set up and solved a Markov chain model for each edge (with states as the current backoff window, backoff counter, and time since the last successful/unsuccessful RTS/CTS exchange) to express the expected service times of the edges via a set of fixed point equations, which were solved iteratively to find the achievable rate region under 802.11 in the given topology
DiMarco et al. (2010)[@dimarco-etal10modeling-ieee-802-15-4-multihop] beaconless 802.15.4; unsaturated, multi-hop network; **no queueing at nodes**; with hidden nodes and ACKs 3-dim discrete time Markov chain for each node having state: (backoff stage, backoff counter, retransmission counter) with CCA and collision probabilities (unknown) as state transition probabilities (decoupling); assumed independence of node processes to find expressions for the CCA and collision probabilities using stationary probabilities of different Markov chains (coupling)
Sen-De (2010)[@sen-de10modelling-802.15.4-beaconless] beaconless 802.15.4; unsaturated, multi-hop network; with hidden nodes; **no queuing; analysis does not match simulation well** 1-dim Markov model for transmitting node with states being in CCA, backoffs states and data transmission; steady state transmission probability and probability of success (throughput) are obtained through fixed point iterations
Shyam-Kumar (2010)[@shyam-kumar10me-thesis] beaconless 802.15.4; (un)saturated, multi-hop network; with hidden nodes and ACKs; infinite buffer; **analysis does not match simulation well** saturated analysis using continuous time Markov chain with states as the set of transmitting nodes having unknown transition probabilities; found CCA and failure probabilities using steady state probability of the Markov chain; unsaturated network modeled as a Markov renewal process constituted by the nodes having non-empty queues, saturation analysis results used in each cycle with the set of non-empty nodes
Marbach et al. (2011)[@marbach11] **non-adaptive CSMA characterized by a *fixed* attempt probability on each link**, and a non-zero sensing period on each link; asynchronous, unsaturated, multihop network; **over-simplified model - does not conform to any existing standard** intuitively formulated fixed point equations involving node idling probabilities, and transmission attempt rates; showed uniqueness of the fixed point, and asymptotic accuracy for large networks with small sensing period and appropriately decreasing link attempt probabilities; for a given sensing period, characterized the achievable rate region of the simplified CSMA policy by defining a set of arrival rate vectors, and showing that for every member in that set, there exists a policy (a vector of attempt probabilities) such that the link service rates obtained from the fixed point equations exceed the link arrival rates
\[tbl:related-literature\]
Our Contributions and Comparison with Related Work {#our-contributions-and-comparison-with-related-work .unnumbered}
--------------------------------------------------
We consider a multi-hop WSN consisting of static sources and relays arranged in a tree topology operating in the beacon-less mode with acknowledgements (ACKs). Each node has an infinite buffer space and may operate in the saturated or unsaturated regime with fixed packet length. Different analysis techniques are developed for networks with hidden nodes and networks without hidden nodes due to the difference in activity lengths perceived by a node in the presence and absence of hidden nodes. Under certain approximations, we model the stochastic process evolving at a node by incorporating the influence of the other nodes in the network by their (unknown) time averaged statistics, and then couple these individual node processes via a system of fixed point equations, which is solved using an iterative scheme to obtain the unknown variables. Although this *decoupling* (or *mean-field*) approximation is popular in such situations, our more detailed model incorporating several issues not considered together hitherto requires a careful handling of the analysis. We identify and calculate two QoS measures for each source node, viz., the packet delivery probability and end-to-end packet delay, in terms of these variables. We observe that in a multi-hop IEEE 802.15.4 network, the packet delivery probability falls sharply before the end-to-end packet delay becomes substantial.
The rest of the paper is organized as follows: In Section \[sec:overview-of-standard\], we give an overview of the unslotted CSMA/CA mechanism. Section \[sec:analytical-model\] explains the node behaviour and elaborates upon the different analysis techniques used for networks without hidden nodes and networks with hidden nodes. Section \[sec:numerical-and-simulation-results\] compares our analysis with simulations. Section \[sec:network-design\] demonstrates how the analysis can be used for designing multi-hop networks with QoS objectives. Finally, we conclude the paper in Section \[sec:conclusion\].
Unslotted CSMA/CA for Beacon-less IEEE 802.15.4 Networks {#sec:overview-of-standard}
========================================================
A typical node behaviour under unslotted CSMA/CA is shown in Figure \[fig:evolution\]. A node with an empty queue remains idle until it generates a packet or receives one from its predecessor nodes. When a node has data to send (i.e., has a non-empty queue), it initiates a random back-off with the first back-off period being sampled uniformly from $0$ to *${2^{macminBE}-1}$*, where *macminBE* is a parameter fixed by the standard. For each node, the back-off period is specified in terms of slots where a slot equals 20 symbol times ($T_{\mathsf{s}}$), and a symbol time equals 16 $\mu s$.[^1] The node then performs a CCA (*Clear Channel Assessment*) to determine whether the channel is idle. If the CCA succeeds, the node does a *Rx-to-Tx turnaround*, which is $12$ symbol times, and starts transmitting on the channel. The failure of the CCA starts a new back-off process with the back-off exponent raised by one, i.e., to *macminBE+1*, provided that this is less than its maximum value, *macmaxBE*. The maximum number of successive CCA failures for the same packet is governed by *macMaxCSMABackoffs*, exceeding which the packet is discarded at the MAC layer. The standard allows the inclusion of acknowledgements (ACKs) which are sent by the intended receivers on a successful packet reception. Once the packet is received, the receiver performs a *Rx-to-Tx turnaround*, which is again $12$ symbol times, and sends a $22$ symbol fixed size ACK packet. A successful transmission is followed by an *InterFrame Spacing*(IFS) before sending another packet.
When a transmitted packet collides or is corrupted by the PHY layer noise, the ACK packet is not generated, which is interpreted by the transmitter as failure in delivery. The node retransmits the same packet for a maximum of *aMaxFrameRetries* times before discarding it at the MAC layer. After transmitting a packet, the node turns to the Rx-mode and waits for the ACK. The *macAckWaitDuration* determines the maximum amount of time a node must wait for in order to receive the ACK before concluding that the packet (or the ACK) has collided. The default values of *macminBE*, *macmaxBE*, *macMaxCSMABackoffs*, and *aMaxFrameRetries* are 3, 5, 4, and 3 respectively.
Description of the Modeling Approach {#sec:analytical-model}
====================================
As shown in Figure \[fig:network\], the network consists of sensors, relays and a base-station (BS). We consider time invariant links and static (immobile) nodes. The links are characterized by a predefined target packet error rate (PER) (e.g., a PER of 1% on each link). We assume that the sensors generate traffic according to independent point processes. These constitute the aggregate external arrival process for the sensor network. Each node transmits its data to the next-hop node according to the topology; throughout this paper, *we shall work with a tree topology so that each transmitter has exactly one receiver node*. The intermediate nodes along a route may be relays in which case they simply forward the incoming traffic, or they may be sensors which transmit their own packets as well as the received packets. Based on the network congestion, the nodes may discard packets due to consecutive failed CCAs or frame retries. Figure \[fig:evolution\] shows a typical sample path of the process evolution at a node.
![Process evolving at a tagged node as it goes through various states.[]{data-label="fig:evolution"}](figures/evolution.pdf){height="4cm" width="9cm"}
Modeling Simplifications {#subsec:model-simplify}
------------------------
The process evolutions at various nodes in the system are coupled, and each node (source or relay) can be present in any state; further, different nodes can interact with different subsets of nodes based on their Carrier Sense (CS) range and positions.[^2] Hence, the random process representing the system is multi-dimensional and non-homogeneous. The exact analysis of such a process is intractable. We, therefore, make several simplifications:
(S1)
: All packets are of the same fixed length. *Hence, the DATA transmission duration is fixed.*
(S2)
: The internode propagation delays are of the order of nanoseconds, and hence can be taken to be zero.
(S3)
: Collision model (zero packet capture): If a receiver is in the communication range of two or more nodes that are transmitting simultaneously, it does not receive any of those transmitted packets, i.e, all the packets involved in a collision get corrupted unlike an SINR model, where some of the interfering packets may still be received successfully, depending on the SINR threshold.
(S4)
: The Inter Frame Spacing (IFS) is neglected. *It means that the receiver does not discard the packet received during the IFS.*
(S5)
: By (S3), a packet is necessarily bad if any other packet is being heard by the same receiver. Even if this is not the case, a packet can be in error due to noise. The packet error probability on each link is fixed and known. *As already mentioned, the links are not time varying and we can empirically obtain the error characteristics of a link.*
(S6)
: ACK packets are short and, therefore, not corrupted by PHY layer noise.
(S7)
: A node’s CCA succeeds when there is no transmission by any node in its Carrier Sense (CS) range at the time of initiation of its CCA. *Recall that in the standard, the channel state is averaged over the $8$ symbol duration.*
(S8)
: The time taken by a transmitting node for the activities of successful transmission and collision are the same. We denote this time by $T_{tx}$. *If the transmitted data collides at the receiver, the macAckWaitDuration for the transmitting node is equal to the sum of the turnaround time of $12~T_s$ and the ACK duration of $22~T_s$, a total of $34~T_s$ (see Figure \[fig:transmission\]).*
(S9)
: We assume symmetry in carrier sensing, and signal reception, i.e., *if a node $i$ can detect (respectively decode) the packet transmissions of another node $j$, then node $j$ can also detect (resp. decode) the packet transmissions of node $i$.*
![The transmission period $T_{\mathsf{tx}}$ which includes the DATA transmission time $T_x$, the turnaround time of $12~T_s$, and the MAC ACK duration of $22~T_s$ (or the *macAckWaitDuration* of $34~T_s$, by **(S8)**).[]{data-label="fig:transmission"}](figures/transmission.pdf)
*Decoupling approximation:* We employ an approximation whereby we model each node separately, incorporating the influence of the other nodes in the network by their average statistics, and as if these nodes were independent of the tagged node. This commonly used approximation is also called a “mean field approximation”, and has been widely used in the wireless networking literature [@bianchi00performance; @kumar-etal04new-insights; @jindal09].
Modeling Node Activity {#subsec:node-activity-modeling}
----------------------
The channel activity perceived by any node $i$ is only due to the nodes in its CS set (i.e., the set of nodes within its CS range) denoted by $\Omega_i$. Note that $\Omega_i$ does not include node $i$. Let us consider the durations during which the queue at node $i$ is non-empty. During these times, node $i$ alternates between performing CCAs and transmitting the packets from its queue. We model the CCA attempt process at node $i$ as a Poisson process of rate $\beta_i$ *conditioned on being in backoff periods*[@goyal-etal09beaconless-zigbee]. For each node $j\in \Omega_i$, consider only those times at which the node is not transmitting (i.e., its queue is empty, or the node is in backoff). We model the CCA attempt process of each node $j\in \Omega_i$ conditioned on these times by an independent Poisson process with rate $\overline{\tau}_j^{(i)}$, $j\in\Omega_i$. By modeling simplification **(S1)**, we assume that all packets entering node $i$ have the same fixed length, and hence take the same amount of time when being transmitted over the medium (denoted by $T_{\mathsf{tx}}$).
Let us now remove the time intervals at node $i$ during which the queue at this node is empty, thereby concatenating all the busy periods at the node. In this conditional time, as a result of the Poisson point process assumption for the attempt process at node $i$, and also for the attempt processes of the nodes in $\Omega_i$, we observe that instants at which packet attempts complete (equivalently, new backoff intervals for the attempts start) are renewal instants. These are denoted by $X^{(k)}_i$ in Figure \[fig:ARP\_calc\_of\_rates\]. The corresponding renewal lifetimes are denoted by $W_i^{(k)}$.
Figure \[fig:system\] shows details of a renewal cycle of node $i$. The dashed transmission durations belong to nodes other than $i$. The cycle always ends with a transmission from node $i$. This last transmission in a cycle always corresponds to a non-collided packet, which, however, could be received in error.
![The renewal process obtained by observing the process at a node $i$ after removing all the idle time periods from the original process shown in Figure \[fig:evolution\]. The renewal epochs are denoted by $\{X_i^{(k)}\}$ and the cycle lengths by $\{W_i^{(k)}\}$.[]{data-label="fig:ARP_calc_of_rates"}](figures/ARP_calc_of_rates.pdf){width="9cm"}
![Renewal cycles at a tagged node, say $i$: (a) a renewal cycle when there are no hidden nodes in the system (all activity periods on the channel are of length $T_{\mathsf{tx}}$); (b) a renewal cycle when there are hidden nodes (due to hidden node collisions, the activity periods on the channel can be longer than $T_{\mathsf{tx}}$.)[]{data-label="fig:system"}](figures/system.pdf)
Before discussing the difference in activity lengths in the two network types (as shown in Figure \[fig:system\]), we list the various packet collision scenarios. Packet collisions in a wireless network with a CSMA MAC can occur due to the presence of hidden nodes. In addition, *there are scenarios where transmissions from two transmitters placed within the CS range of each other can overlap, known as **Simultaneous Channel Sensing** and **Vulnerable Window*** (also known as First and Second Collision window respectively, in [@goyal-etal09beaconless-zigbee]). These two collision scenarios are explained in Figures \[fig:simul\_channel\] and \[fig:CCAfail\], respectively. Note that we are making use of **(S2)** to avoid including insignificant timing details here.
![Node $j\in \Omega_i$ finishes its backoff, performs a CCA, finds the channel idle and starts transmitting the DATA packet. Node $i$ finishes its backoff anywhere in the shown $12~T_s$ duration and there is no other ongoing transmission in $\Omega_i$, its CCA succeeds and it enters the transmission duration. As a result, the DATA packets may collide at $r(i)\in \Omega_j$ and/or $r(j) \in \Omega_i$.[]{data-label="fig:simul_channel"}](figures/simul_channel_sensing.pdf)
![Vulnerable window of $4~T_s$ in the transmission period of node $j\in \Omega_i$ during which node $i$’s CCA attempt would be successful.[]{data-label="fig:CCAfail"}](figures/CCAfail.pdf)
Since the vulnerable window size is small compared to the data transmission duration of a node (see Figure \[fig:CCAfail\]), *we shall assume that the probability of a node’s CCA initiating in the vulnerable window of another node is negligible*, and therefore, neglect the resulting packet collisions. Further, since the ACK packet size is just a small fraction of DATA packet size (e.g. compare an ACK packet ($22$ symbols) with a DATA packet of length $260$ symbols at PHY layer), *we shall assume that the probability of packet collision involving ACK packets is negligible*.
We now return to elucidate the difference in length of activity periods in a renewal cycle.
(a) Absence of Hidden Nodes:
: When a node transmits, all the other nodes in the network can hear it, resulting in CCA failures for other nodes that try to assess the channel in the transmission period. Also if two nodes are involved in simultaneous channel sensing, the activity period may extend from $T_{\mathsf{tx}}$ to a maximum of $(T_{\mathsf{tx}}+12~T_s)$. Since $12~T_s<<T_{\mathsf{tx}}$, we can assume that the activity period is only a single transmission period of duration $T_{\mathsf{tx}}$.
(b) Presence of Hidden Nodes:
: Since a node’s hearing capacity is limited, it may not perceive the activities of all the nodes in the network which may cause dilation of activity period.
![*Dilation of transmission period as perceived by node $i$*.[]{data-label="fig:t_eff"}](figures/t_eff.pdf)
0
![*Typical hidden node case and Evolution processes of Node $i$ and Node $j$.*[]{data-label="fig:hidden_node_explained"}](figures/hidden_node_explaned.pdf)
As shown in the Figure \[fig:t\_eff\], node $j$ and $k$ are hidden from each other with respect to the receiver node $i$. Suppose node $j$ starts transmitting while nodes $i$ and $k$ are silent. Since node $k$ cannot hear node $j$, it can start its own transmission while the transmission from node $j$ is still going on. For node $i$, this looks like a dilated transmission period, whose length is denoted by $T_{i}^{(\mathsf{eff})}$.
Derivation of Node Interaction Equations {#sec:fixed-pt-eqns-derivation}
========================================
In the following subsections, we consider the node processes described in Section \[sec:analytical-model\], and identify certain useful steady state quantities at each node, e.g., CCA attempt rate. Then we employ the decoupling approximation (described in Section \[sec:analytical-model\]), and detailed stochastic analysis, to write down certain fixed point equations relating these quantities. In doing this, *we assume the network to be stable so that these steady state quantities exist*. Then in Section \[sec:iterative-scheme\], we use an iterative scheme to solve for the unknown variables, and finally use the quantities thus obtained to evaluate the network performance, namely, end-to-end delay and packet delivery probability from each source node. The analysis will be presented for the more general case where hidden nodes may be present in the system; wherever necessary, we shall mention the changes required in the expressions for the special case where no hidden nodes are present in the system.
Derivation of Fixed Point Equations {#subsec:fixed-pt-derivation}
-----------------------------------
We begin by analyzing the process at node $i$ shown in Figures \[fig:ARP\_calc\_of\_rates\] and \[fig:system\]. Consider an epoch at which node $i$ starts its backoffs in order to send the first packet in its queue. At this instant, each node $j \in \Omega_i$ is either empty or performing backoffs for its HOL packet or transmitting its HOL packet. Recall that the rate at which node $i$ completes backoffs during its backoff periods is denoted by $\beta_i$. The activity of node $i$ is affected by the residual rate of CCA attempts by $j \in \Omega_i$ after removing those CCA attempts that fail due to nodes that are hidden from node $i$; this rate is denoted by $\overline{\tau}^{(i)}_j$. Modeling these conditional attempt processes as independent Poisson processes, we define: $$\begin{aligned}
\eta_i = \frac{\beta_i}{\beta_i + \displaystyle{\sum_{j\in\Omega_i}\overline{\tau}_j^{(i)}}} \quad &;& \quad g_i = \frac{1}{\beta_i + \displaystyle{\sum_{j\in\Omega_i}\overline{\tau}_j^{(i)}}} \\
\text{and} \qquad c_i &=& \Bigg(1-e^{-12~T_\mathsf{s}\beta_i}\Bigg) \end{aligned}$$ Thus, $\eta_i$ is the probability that node $i$ makes a CCA attempt before any node in $\Omega_i$, $g_i$ is the mean time until the first CCA attempt, and $c_i$ is the probability that after an attempt by a node in $\Omega_i$, node $i$ attempts within the 12 symbol vulnerable period, thus causing a “simultaneous” channel sensing collision (see Figure \[fig:simul\_channel\]).
0 Since ZigBee is designed to operate at fairly low packet generation rates at which packet discard probability (due to successive CCA failures or frame retries) is very low, all the recursive equations written below upto \[subsubsec:packet-failure\] do not take packet discards into account.
### Perceived CCA attempt rate, $\overline{\tau}_j^{(i)}$
In a network with hidden nodes, in general, $\Omega_j \setminus \Omega_i$ is nonempty. Hence, CCA attempts of node $j$ could be blocked by the activities of nodes in $\Omega_j \setminus \Omega_i$, thus leading to the node $i$ perceiving a CCA attempt rate from node $j$ that is less than the CCA attempt rate when $\Omega_j \setminus \Omega_i = \phi$ (absence of hidden nodes). We have denoted this effective CCA attempt rate of node $j$ as perceived by node $i$ by $\overline{\tau}_j^{(i)}$ for all $j \in \Omega_i$. Define
$N_j^{(\mathsf{cca})}(t)$:
: Total number of CCAs attempts of Node $j$ in time $(0, t]$
$N_j^{(\mathsf{ccaf_{-i}})}(t)$:
: Number of failed CCA attempts of Node $j$ due to the nodes in $\Omega_j \setminus \Omega_i$ in time $(0, t]$
$I_j(t)$:
: The time during which node $j$ is not transmitting in time $(0, t]$
$T_{j}^{\mathsf{(bo)}}(t)$:
: Time for which node $j$ is in backoff in the interval $(0, t]$
$T_{j}^{\mathsf{(ne)}}(t)$:
: Queue non-empty period of node $j$ in the interval $(0, t]$
Note that with these definitions, we have $$\begin{aligned}
\beta_j &= \lim_{t\rightarrow \infty}\frac{N_{j}^{(\mathsf{cca})}(t)}{T_{j}^{\mathsf{(bo)}}(t)}\text{ a.s}\end{aligned}$$
We can further define,
$$\begin{aligned}
b_j &= \lim_{t\rightarrow \infty}\frac{T_{j}^{\mathsf{(bo)}}(t)}{T_{j}^{\mathsf{(ne)}}(t)}\\
q_j &= \lim_{t\rightarrow \infty}\frac{T_{j}^{\mathsf{(ne)}}(t)}{t}\end{aligned}$$
where $b_j$ can be interpreted as the long-term fraction of time node $j$ is in backoff provided it is non-empty, and $q_j$ is the long-term fraction of time that node $j$ is non-empty, which is also the queue non-empty probability of node $j$, assuming the system is ergodic.
Then, we can write $$\begin{aligned}
\overline{\tau}_j^{(i)} &= \lim_{t\rightarrow \infty}\frac{N_j^{(\mathsf{cca})}(t) - N_j^{(\mathsf{ccaf_{-i}})}(t)}{I_j(t)} \text{ a.s}\\
&= \lim_{t\rightarrow \infty} \frac{\frac{N_{j}^{(\mathsf{cca})}(t)}{T_{j}^{\mathsf{(bo)}}(t)}\times \frac{T_{j}^{\mathsf{(bo)}}(t)}{T_{j}^{\mathsf{(ne)}}(t)}\times \frac{T_{j}^{\mathsf{(ne)}}(t)}{t}\times \left(1 - \frac{N_j^{(\mathsf{ccaf_{-i}})}(t)}{N_j^{(\mathsf{cca})}(t)}\right) }{\frac{I_{j}(t)}{t}} \end{aligned}$$ Using the earlier definitions, we have $$\begin{aligned}
\overline{\tau}_j^{(i)} &= \frac{\beta_j \times b_j \times q_j \times (1 - \alpha_j^{(-i)})}{1 - q_j + q_j \times b_j}\label{eqn:1}\end{aligned}$$ where $\alpha_j^{(-i)}$ is the probability of CCA failure of node $j$ only due to the nodes in $\Omega_j \setminus \Omega_i$, and is derived below.
### Computation of $\alpha_j^{(-i)}$
Using the notation introduced earlier, we can write $$\begin{aligned}
\alpha_j^{(-i)} &= \lim_{t\rightarrow \infty} \frac{N_j^{(\mathsf{ccaf_{-i}})}(t)}{N_j^{(\mathsf{cca})}(t)} \text{ a.s}\\
&= \lim_{t\rightarrow \infty} \frac{\frac{N_j^{(\mathsf{ccaf_{-i}})}(t)}{T_j^{(\mathsf{ne})}(t)}}{\frac{N_j^{(\mathsf{cca})}(t)}{T_j^{(\mathsf{ne})}(t)}} \\\end{aligned}$$ Applying the Renewal-Reward Theorem (RRT) [@kulkarni95modeling-stochastic-systems], we have $$\begin{aligned}
\alpha_j^{(-i)} &= \frac{\frac{\mathbf{N}_j^{(\mathsf{ccaf_{-i}})}}{W_{j}}}{\frac{\mathbf{N}_j^{(\mathsf{cca})}}{W_{j}}} \\
&= \frac{\mathbf{N}_j^{(\mathsf{ccaf_{-i}})}}{\mathbf{N}_j^{(\mathsf{cca})}}\label{eqn:alpha_j_i_exp}\end{aligned}$$ where $W_j$ is the mean time between the ends of transmission during a busy period of node $j$ (i.e., the mean renewal cycle length). Now, by taking the reward to be the mean number of total CCAs by node $j$ in the renewal cycle, we can write $$\begin{aligned}
\mathbf{N}_j^{(\mathsf{cca})} &= \eta_j + (1 - \eta_j)c_j + (1 - \eta_j)(1 - c_j)(\beta_j T_j^{(\mathsf{eff})} + \mathbf{N}_j^{(\mathsf{cca})}) \end{aligned}$$
0 Here, the first term says that with probability $\eta_j$ , only one CCA is performed by node $j$ after which it starts transmitting, and the cycle ends. The second term corresponds to an event of simultaneous channel sensing in which node $j$ is involved with some other node that had already started its transmission, as a result of which there is only one CCA performed by node $j$ followed by its data transmission and the cycle ends. In the final term, node $j$ neither starts transmitting first nor it is involved in a simultaneous channel sensing event. Note that node $j$ performs CCAs at the rate of $\beta_j$ for the entire dilated transmission period $T_j^{(\mathsf{eff})}$ i.e., a total of $\beta_j T_j^{(\mathsf{eff})}$ CCAs all of which fail, and the cycle continues after this period. Upon rearranging terms, we have $$\begin{aligned}
\label{eqn:N_cca_eqn}
\mathbf{N}_j^{(\mathsf{cca})} &= \frac{\eta_j + (1 - \eta_j)c_j + (1 - \eta_j)(1 - c_j)\beta_j T_j^{(\mathsf{eff})} }{1 - (1 - \eta_j)(1 - c_j)}\end{aligned}$$ Again, let the reward be the mean number of failed CCA attempts by node $j$ in a renewal cycle due to transmissions by nodes in $\Omega_j \setminus \Omega_i$, say $\mathbf{N}_j^{(\mathsf{ccaf_{-i}})}$, which we can write, using **(S7)** as $$\label{eqn:ccafail_node_j}\begin{split}
\mathbf{N}_j^{(\mathsf{ccaf_{-i}})}&=\left(\frac{\displaystyle{\sum_{k\in \Omega_j \cap \Omega_i} \overline{\tau}_k^{(j)}}}{\beta_j + \displaystyle{\sum_{l\in \Omega_j} \overline{\tau}_l^{(j)}}} \right)(1 - c_j)\left(\mathbf{N}_j^{(\mathsf{ccaf_{-i}})}\right) + \\ &\left(\frac{\displaystyle{\sum_{k\in \Omega_j \setminus \Omega_i} \overline{\tau}_k^{(j)}}}{\beta_j + \displaystyle{\sum_{l\in \Omega_j} \overline{\tau}_l^{(j)}}} \right)
(1 - c_j)\left(\beta_j T_{\mathsf{tx}} + \mathbf{N}_j^{(\mathsf{ccaf_{-i}})}\right)
\end{split}$$ Clearly, in case of a simultaneous channel sensing event involving node $j$, the cycle ends with zero reward. Therefore, the terms in Equation only consider the case where node $j$ is not involved in simultaneous channel sensing. The first term in corresponds to transmission attempts by nodes in $\Omega_j \cap \Omega_i$; this results in zero reward, and the cycle continues thereafter. The second term accounts for transmission attempts by nodes in $\Omega_j \setminus \Omega_i$; this results in a total reward of $\beta_j T_{\mathsf{tx}}$ and the cycle continues thereafter. Upon rearranging terms, we have $$\begin{aligned}
\mathbf{N}_j^{(\mathsf{ccaf_{-i}})} &= \frac{\left(\frac{\displaystyle{\sum_{k\in \Omega_j \setminus \Omega_i} \overline{\tau}_k^{(j)}}}{\beta_j + \displaystyle{\sum_{l\in \Omega_j} \overline{\tau}_l^{(j)}}} \right)(1 - c_j)\left(\beta_j T_{\mathsf{tx}}\right)}{1 - (1 - \eta_j)(1 - c_j)}\end{aligned}$$ Hence from \[eqn:alpha\_j\_i\_exp\], the CCA failure probability of node $j$ only due to the nodes in $\Omega_j \setminus \Omega_i$ is given by $$\begin{aligned}
\alpha_j^{(-i)} &= \frac{\left(\frac{\displaystyle{\sum_{k\in \Omega_j \setminus \Omega_i} \overline{\tau}_k^{(j)}}}{\beta_j + \displaystyle{\sum_{l\in \Omega_j} \overline{\tau}_l^{(j)}}} \right)(1 - c_j)\left(\beta_j T_{\mathsf{tx}}\right)}{\eta_j + (1 - \eta_j)c_j + (1 - \eta_j)(1 - c_j)\beta_j T_j^{(\mathsf{eff})}}\label{eqn:6}\end{aligned}$$
For networks with no hidden nodes, $\alpha_j^{(-i)} = 0$.
### The dilated activity period, $T_i^{(\mathsf{eff})}$ {#subsubsec:Ti-eff}
As mentioned earlier, the length of the dilated activity period as perceived by a node depends upon the set of nodes in its CS range. We propose two different models for calculating $T_i^{(\mathsf{eff})}$.
- **$M/D/\infty$ Model :** Since each node has a different set of neighbours, we can make the following approximation to simplify our analysis:
(A1)
: The mean length of Dilated Activity Period as perceived by node $i$ is equal to the length of mean busy period of an $M/D/\infty$ queue where the deterministic service time is equal to a single transmission period $T_{\mathsf{tx}}$, and the arrival process is approximated as a Poisson process having rate equal to the aggregate transmission initiation rate of *all* the nodes in the CS range of node $i$.
Figure \[fig:MDinf\_approx\] represents the $M/D/\infty$ model pictorially. Note that the above assumption is equivalent to saying that all nodes in $\Omega_i$ are hidden from each other, and hence this approximation results in a larger mean for the dilated activity period.
Note that the transmission initiation rate of any node $j\in \Omega_i$, as perceived by node $i$, is $\overline{\tau}_j^{(i)}$. For any node $i$, let the aggregate transmission initiation rate for nodes in $\Omega_i$ be
![The $M/D/\infty$ approximation for the calculation of $T_i^{(\mathsf{eff})}$. The nodes in $\Omega_i$ are assumed to be hidden from each other and their transmission initiations are approximated by a Poisson process with a rate of $\zeta_i$ given by Equation . The service time for each of these transmissions is $T_{\mathsf{tx}}$.[]{data-label="fig:MDinf_approx"}](figures/MDinf_approx.pdf)
$$\begin{aligned}
\zeta_i = \sum_{j\in \Omega_i}\overline{\tau}_j^{(i)}
\label{eqn:zeta}\end{aligned}$$
Assuming node $j\in \Omega_i$ has started transmission and another node $k \in \Omega_i$ starts at time $u < T_{\mathsf{tx}}$, the expression for $T_i^{(\mathsf{eff})}$ can be written recursively as $$\begin{aligned}
\textstyle{T_i^{(\mathsf{eff})}} &\textstyle{=}& \textstyle{T_{\mathsf{tx}}\exp\{-\zeta_i T_{\mathsf{tx}}\} + \int_0^{T_{\mathsf{tx}}}(u+T_i^{(\mathsf{eff})})\zeta_i\exp\{-\zeta_i u \}du} \nonumber \\
&\textstyle{=}& \frac{1}{\zeta_i}(\exp\{\zeta_i T_{\mathsf{tx}}\} - 1)\end{aligned}$$
- **Boorstyn et al. [@boorstyn-etal87CSMA-throughput-analysis] Model:** For node $i$ in Figure \[fig:t\_eff\], Figure \[fig:boorstyn\_model\] shows its evolution process in conditional time. We see that when nodes in $\Omega_i$ attempt packet transmission, then we have a dilated transmission period, $T_i^{(\mathsf{eff})}$ as perceived by Node $i$.
![State transition diagram for computing $T_i^{(\mathsf{eff})}$ for the nodes in Figure \[fig:t\_eff\][]{data-label="fig:state_diagram"}](figures/state_transition_diag.pdf)
Here we will use a CTMC model as suggested by Boorstyn et al. [@boorstyn-etal87CSMA-throughput-analysis] for determining $T_i^{(\mathsf{eff})}$ during times when node $i$ is non empty. Let $\phi_{i}$ be the state when all the nodes in $\Omega_i$ as well as node $i$ are in backoff. Then the state transition diagram for computing $T_i^{(\mathsf{eff})}$ for the nodes in Figure \[fig:t\_eff\] is shown in Figure \[fig:state\_diagram\]. Note that return to $\phi_{i}$ are renewal instants since the transition out of $\phi_{i}$ are all exponentially distributed. By the *insensitivity property* explained in [@boorstyn-etal87CSMA-throughput-analysis], the steady state probability of being in state $\phi_{i}$, say $\pi_{\phi_{i}}$, does not depend on the distribution of packet transmission time except by its mean. So we can take the packet service times to be exponentially distributed. Hence $\pi_{\phi_{i}}$ can be obtained as in [@boorstyn-etal87CSMA-throughput-analysis], i.e.,
$$\begin{aligned}
\label{eqn:phi_eqn_1}
\pi_{\phi_{i}} &= \left[\sum_{D_{i}\in \mathscr{D}_i} \prod_{j \in D_{i}} \overline{\tau}_j^{(i)} T_{\mathsf{tx}}\right]^{-1} \end{aligned}$$
where $D_{i}$ is a set of nodes in $\Omega_i$ that are actively transmitting, and $\mathscr{D}_i$ is the collection of all such sets. Again, from the RRT we obtain $\pi_{\phi_i}$ as the ratio of the time the system is in state $\phi_{i}$ and the mean time between the ends of transmission during a busy period of Node $i$, i.e., $$\resizebox{1.0\hsize}{!}{$\pi_{\phi_{i}} = \frac{\frac{1}{\beta_i + \displaystyle{\sum_{j \in \Omega_i} \overline{\tau}_j^{(i)}}}}{\frac{1}{\beta_i + \displaystyle{\sum_{j \in \Omega_i} \overline{\tau}_j^{(i)}}} + \frac{\beta_i}{\beta_i + \displaystyle{\sum_{j \in \Omega_i} \overline{\tau}_j^{(i)}}} \times T_{\mathsf{tx}} + \frac{\displaystyle{\sum_{j \in \Omega_i} \overline{\tau}_j^{(i)}}}{\beta_i + \displaystyle{\sum_{j \in \Omega_i} \overline{\tau}_j^{(i)}}} \times T_{i}^{(\mathsf{eff})}}$}$$ which yields $$\label{eqn:phi_eqn_2}
\pi_{\phi_{i}} = \frac{1}{1 + \beta_i \times T_{\mathsf{tx}} + \left(\displaystyle{\sum_{j \in \Omega_i} \overline{\tau}_j^{(i)}}\right)\times T_{i}^{(\mathsf{eff})}}$$ Hence, from Equation and , we can obtain $T_{i}^{(\mathsf{eff})}$ for each node $i$. Observe that the dilated activity period computed using the Boorstyn et al. model would be less conservative than the $M/D/\infty$ model. Also note that identification of $\mathscr{D}_{i}$ is an NP-complete combinatorial problem [@boorstyn-etal87CSMA-throughput-analysis]; however, efficient algorithms exist that handle networks of arbitrary topology and moderate size (50-100 nodes).
### Probability of CCA failure, $\alpha_i$
The probability of CCA failure is the probability of the occurrence of at least one transmitting node in the CS range of node $i$, given that node $i$ performs a CCA. Defining $N_i^{(\mathsf{cca})}(t)$ and $N_i^{(\mathsf{f})}(t)$ as the total number of CCAs and number of failed CCAs in the interval $(0,t]$, respectively, we have: $$\begin{aligned}
\alpha_i = \lim_{t \to \infty}\frac{N_i^{(\mathsf{f})}(t)}{N_i^{(\mathsf{cca})}(t)} \text{ a.s }= \lim_{t \to \infty}\frac{\frac{N_i^{(\mathsf{f})}(t)}{T_{i}^{\mathsf{(ne)}}(t)}}{\frac{N_i^{(\mathsf{cca})}(t)}{T_{i}^{\mathsf{(ne)}}(t)}}\end{aligned}$$ Applying Renewal-Reward Theorem (RRT) [@kulkarni95modeling-stochastic-systems], we get $$\begin{aligned}
\alpha_i &=&\frac{\frac{\mathbf{N}_i^{(\mathsf{f})}}{W_i}}{\frac{\mathbf{N}_i^{(\mathsf{cca})}}{W_i}}\nonumber\\
&=&\frac{\mathbf{N}_i^{(\mathsf{f})}}{\mathbf{N}_i^{(\mathsf{cca})}}
\label{eqn:alpha_rates}\end{aligned}$$ where $\mathbf{N}_i^{(\mathsf{f})}$ is the mean number of failed CCAs and $\mathbf{N}_i^{(\mathsf{cca})}$ is the mean number of total CCAs in a cycle. $W_i$ is the mean renewal cycle length.
0 It is given by $$\begin{aligned}
W_i &=& g_i+\eta_i T_{\mathsf{tx}}+(1-\eta_i)c_i T_{\mathsf{tx}}+(1-\eta_i)(1-c_i)(T^{(\mathsf{eff})}_i+W_i)\end{aligned}$$ The first term is the mean time until any node in $\Omega_i \cup \{i\}$ starts transmission as seen by node $i$. In the second term, with probability $\eta_i$, node $i$ finishes its backoff first and attempts a CCA. Since no other node is transmitting, this CCA succeeds and node $i$ transmits a packet and the cycle ends after the transmission period $T_{\mathsf{tx}}$. Note that this transmission period cannot be dilated because it blocks all the nodes in $\Omega_i$ (ignoring the maximum jitter of $12~T_\mathsf{s}$ due to simultaneous channel sensing). The third terms says that some other node finishes its backoff first *and* starts transmitting further getting involved in a simultaneous channel sensing event with node $i$ (which happen with a probability of $c_i$) due to which node $i$ starts transmitting and the cycle ends after the transmission period $T_{\mathsf{tx}}$ (again note that this transmission period cannot be dilated either). The final term depicts the event in which another node starts transmitting first and node $i$ does not perform a CCA in the vulnerable period. This transmission period is susceptible to dilation and hence taken to be $T_i^{(\mathsf{eff})}$. The final expression for $W_i$ turns out to be $$\begin{aligned}
W_i &=& \frac{g_i+\eta_i T_{\mathsf{tx}}+(1-\eta_i)c_i T_{\mathsf{tx}}+(1-\eta_i)(1-c_i)T_i^{(\mathsf{eff})}}{1-(1-\eta_i)(1-c_i)}\end{aligned}$$ Computation of $\mathbf{N}_i^{(\mathsf{cca})}$ is as in Equation . Now the mean number of failed CCA attempts by node $i$ in a renewal cycle, i.e., $\mathbf{N}_i^{(\mathsf{f})}$, can be written using **(S7)** as $$\begin{aligned}
\mathbf{N}_i^{(\mathsf{f})} &=& (1-\eta_i)(1-c_i)\Big(\beta_i T^{(\mathsf{eff})}_i+\mathbf{N}_i^{(\mathsf{f})}\Big)\end{aligned}$$ Rearranging and using Equation , the CCA failure probability is given by $$\begin{aligned}
\alpha_i &=& \frac{(1-\eta_i)(1-c_i)\beta_i T^{(\mathsf{eff})}_i}{\eta_i + (1-\eta_i)c_i + (1-\eta_i)(1-c_i)\beta_i T^{(\mathsf{eff})}_i}
\label{eqn:alpha}\end{aligned}$$
Note that for no hidden node network, $T^{(\mathsf{eff})}_i = T_{\mathsf{tx}}$.
### Packet failure probability, $\gamma_i$ {#subsubsec:packet-failure}
A transmitted packet can fail to be decoded by its intended receiver due to a collision, or due to noise. We do not take packet capture into account here (see **(S3)**). Define $M_i(t)$ as the total number of transmissions in interval $(0,t]$, $M_i^{(\mathsf{c})}(t)$ as total number of collisions in interval $(0,t]$ and $l_i$ as the probability of *data* packet error (known by **(S5)**) on the link between node $i$ and its receiver $r(i)$ due to noise. Recall that ACKs are not corrupted by PHY noise (see **(S6)**). The probability of packet failure, $\gamma_i$, is defined as: $$\begin{aligned}
\gamma_i&=&\Bigg(\lim_{t \to \infty}\frac{M_i^{(\mathsf{c})}(t)}{M_i(t)}\Bigg)+\Bigg(1-\lim_{t \to \infty}\frac{M_i^{(\mathsf{c})}(t)}{M_i(t)}\Bigg)l_i
\text{ a.s }\end{aligned}$$ Dividing each term in limit by $T_{i}^{\mathsf{(ne)}}(t)$ and applying RRT, we get $$\begin{aligned}
\gamma_i &=& \Bigg(\frac{\frac{\mathbf{M}_i^{(\mathsf{c})}}{W_i}}{\frac{\mathbf{M}_i}{W_i}}\Bigg) + \Bigg(1 - \frac{\frac{\mathbf{M}_i^{(\mathsf{c})}}{W_i}}{\frac{\mathbf{M}_i}{W_i}}\Bigg)l_i\end{aligned}$$ where $\mathbf{M}_i^{(\mathsf{c})}$ is the mean number of collided packets, and $\mathbf{M}_i$ is the mean number of transmitted packets in a renewal cycle. If we ignore packet discards, there is a single transmission by node $i$ in each renewal cycle so that $\mathbf{M}_i$ becomes unity. We can rewrite $$\begin{aligned}
\gamma_i &=& p_i +(1-p_i )l_i
\label{eqn:gamma}\end{aligned}$$ where $p_i $ is the probability of packet collision and equals $\mathbf{M}_i^{(\mathsf{c})}$. Since $l_i$ is given for each node $i$, we need to compute probability of packet collision, $p_i$, to compute packet failure probability, $\gamma_i$.
We denote the set of nodes that can cause interference in successful reception at $r(i)$ by $I_{r(i)}$, which is composed of two sets:
$C^{(1)}_{r(i)}=\{j\in\mathscr{N}: j\in\Omega_i\text{ and }j\in I_{r(i)}\}$\
$C^{(2)}_{r(i)}=\{j\in\mathscr{N}: j \notin\Omega_i\text{ and }j \in I_{r(i)}\}$
such that $$\begin{aligned}
C^{(1)}_{r(i)}\cap C^{(2)}_{r(i)}&=&\emptyset\\
C^{(1)}_{r(i)}\cup C^{(2)}_{r(i)}&=&I_{r(i)} \end{aligned}$$ Note that the receiver $r(i)$ is in $C^{(1)}_{r(i)}$ (see Figure \[fig:interference\]).
![Interference region around the receiver. $R^{(CS)}$ is the CS range which is assumed to be equal for all nodes by **(S9)**.[]{data-label="fig:interference"}](figures/interference.pdf)
Let the reward be the mean number of collisions of node $i'$s packets in a renewal cycle, say, $\mathbf{M}_i^{(\mathsf{c})}$ , which can be written as $$\begin{aligned}
\mathbf{M}_i^{(\mathsf{c})} &=& (1-\eta_i)(1-c_i)\mathbf{M}_i^{(\mathsf{c})} + \overline{\mathbf{M}}_i^{(\mathsf{c})}\end{aligned}$$ where $\overline{\mathbf{M}}_i^{(\mathsf{c})}$ is mean number of collisions of node $i'$s packets in a renewal cycle when node $i$ attempts a packet transmission or does a simultaneous channel sensing. On rearranging, the packet collision probability at node $i$ is given by $$\begin{aligned}
p_i &=& \frac{\overline{\mathbf{M}}_i^{(\mathsf{c})}}{1-(1-\eta_i)(1-c_i)}
\label{eqn:p}\end{aligned}$$ To compute $\overline{\mathbf{M}}_i^{(\mathsf{c})}$, we first need to find the probability that nodes in a given set are not transmitting using a product approximation due to the unavailability of joint distribution of processes. Let $\overline{h}_i$ be the fraction of time node $i$ is not transmitting (unconditional) and is equal to $$\begin{aligned}
\overline{h}_i &=& \lim_{t\rightarrow \infty}\frac{I_i(t)}{t} \quad \text{ a.s}\\
&=& (1-q_i) + q_ib_i \end{aligned}$$ The expression for $\overline{\mathbf{M}}_i^{(\mathsf{c})}$ can be found as the sum of several components as follows:
1.
: The first term accounts for the fact that Node $i$ started transmitting in the presence of at least one transmission by the hidden nodes set, i.e., $C^{(2)}_{r(i)}$. $$\begin{aligned}
R_i^{(1)} &=& \eta_i\Bigg(1-\prod_{j\in C^{(2)}_{r(i)}}\overline{h}_j\Bigg)\end{aligned}$$
2.
: The second term accounts for the scenario when Node $i$ started its transmission as a simultaneous channel sensing with a node in $\Omega_i$ in the presence of at least one transmission by the hidden nodes set, i.e., $C^{(2)}_{r(i)}$.. $$\begin{aligned}
R_i^{(2)} &=& (1-\eta_i)c_i\Bigg(1-\prod_{j\in C^{(2)}_{r(i)}}\overline{h}_j\Bigg)\end{aligned}$$
3.
: The third term considers the case where Node $i$ starts transmitting in the absence of any ongoing transmission by a hidden node but it encounters a simultaneous transmission by a node in $C^{(1)}_{r(i)}$ anywhere in the corresponding $12~T_s$ period and/or a transmission by a node in $C^{(2)}_{r(i)}$, i.e., a hidden node anywhere in its activity period $T_{\mathsf{tx}}$. $$\begin{aligned}
\scriptstyle{R_i^{(3)}} &\scriptstyle{=}& \scriptstyle{\eta_i\Bigg(\prod_{j\in C^{(2)}_{r(i)}}\overline{h}_j\Bigg)\Bigg(1-\exp \Bigg\{-12~T_s\Bigg( \displaystyle{\sum_{j\in C^{(1)}_{r(i)}}}\overline{\tau}_j^{(i)} \Bigg) \Bigg\}.} \\
&& \scriptstyle{\exp \Bigg\{-T_{\mathsf{tx}}\Bigg(\displaystyle{\sum_{j\in C^{(2)}_{r(i)}}}\overline{\tau}_j\Bigg)\Bigg\}\Bigg) }\end{aligned}$$ where $\overline{\tau}_j$ is the rate of successful CCA attempts of node $j$ over non-transmitting period, $I_j(t)$ and is equal to $$\begin{aligned}
\overline{\tau}_j &= \lim_{t \rightarrow \infty }\frac{N^{(\mathsf{cca})}_j(t) - N^{(\mathsf{f})}_j(t)}{I_j(t)} \quad \text{ a.s}\\
&= \frac{\beta_j b_j q_j (1 - \alpha_j)}{1 - q_j + q_jb_j}\end{aligned}$$
4.
: The fourth term says that Node $i$ started its transmission as a simultaneous channel sensing with a node in $C^{(1)}_{r(i)}$ in the absence of any hidden node. This event surely ends up in a collision at the receiver $r(i)$. $$\begin{aligned}
R_i^{(4)} &=& \Bigg(\frac{\displaystyle{\sum_{j\in C^{(1)}_{r(i)}}\overline{\tau}_j^{(i)}}}{\beta_i+\displaystyle{\sum_{j\in \Omega_i}\overline{\tau}_j^{(i)}}}\Bigg)c_i\Bigg(\prod_{j\in C^{(2)}_{r(i)}}\overline{h}_j\Bigg) \end{aligned}$$
5.
: The final term says that Node $i$ started its transmission as a simultaneous channel sensing with a node in $\Omega_i \setminus C^{(1)}_{r(i)}$ in the absence of any hidden node but it encounters a simultaneous transmission by a node in $C^{(1)}_{r(i)}$ and/or by a node in $C^{(2)}_{r(i)}$, i.e., a hidden node. $$\begin{aligned}
R_i^{(5)} &= \Bigg(\frac{\displaystyle{\sum_{j\in \Omega_i\setminus C^{(1)}_{r(i)}}\overline{\tau}_j^{(i)}}}{\beta_i+\displaystyle{\sum_{j\in \Omega_i}\overline{\tau}_j^{(i)}}}\Bigg)c_i\Bigg(\prod_{j\in C^{(2)}_{r(i)}}\overline{h}_j\Bigg) \nonumber\\
& \bigg(1-\exp \Bigg\{-12~T_\mathsf{s}\Bigg( \displaystyle{\sum_{j\in C^{(1)}_{r(i)}}}\overline{\tau}_j^{(i)} \Bigg) \Bigg\} \nonumber\\
&\exp \Bigg\{-T_{\mathsf{tx}}\Bigg(\displaystyle{\sum_{j\in C^{(2)}_{r(i)}}}\overline{\tau}_j\Bigg)\Bigg\}\Bigg) \end{aligned}$$
The overall packet collision probability is given by $$\begin{aligned}
p_i &=& \frac{R_i^{(1)}+R_i^{(2)}+R_i^{(3)}+R_i^{(4)}+R_i^{(5)}}{\eta_i+(1-\eta_i)c_i}
\label{eqn:p_hidden}\end{aligned}$$ The packet failure probability, $\gamma_i$, can now be calculated using Equations and .
Note that for no hidden node network $C^{(2)}_{r(i)} = \phi$ and $\prod_{j\in C^{(2)}_{r(i)}}\overline{h}_j = 1$
### Average Service Rate ($\sigma_i$)
Each packet that reaches the HOL position in the queue at a node can have multiple backoffs and transmissions before it is successfully transmitted or discarded. We define $\overline{Z}_i$ as the average time spent in backoff, and $\overline{Y}_i$ as the average transmission time until the packet is successfully transmitted or discarded at the MAC layer. Then the average service rate, $\sigma_i$, at node $i$ can be calculated as: $$\begin{aligned}
\frac{1}{\sigma_i}=\overline{Z}_i + \overline{Y}_i
\label{eqn:sigma}\end{aligned}$$ Using the default values from the standard, the mean backoff durations can be calculated (shown in Table \[tab:mean\_backoff\_durations\]).
-------------- ----------------------------------------- -------------------------------------
**Backoff** **Mean Backoff Duration** **Mean Backoff Duration**
**Exponent** **with successful CCA($T_\mathsf{s}$)** **with failed CCA($T_\mathsf{s}$)**
$3$ $70+20$ $70+8$
$4$ $150+20$ $150+8$
$5$ $310+20$ $310+8$
-------------- ----------------------------------------- -------------------------------------
: Mean backoff durations in symbol times, $T_\mathsf{s}$[]{data-label="tab:mean_backoff_durations"}
Define the following quantities: $$\begin{aligned}
\scriptstyle{\overline{B}_i} &\scriptstyle{=} \scriptstyle{\left(70+8+158\alpha_i + 318\alpha_i^2 + 318\alpha_i^3 + 318\alpha_i^4\right)}\\
\scriptstyle{T_i^{(1)}}&\scriptstyle{=}\scriptstyle{\Bigg(\frac{(70+8)(1-\alpha_i)}{(1-\alpha_i^5)}+\frac{236\alpha_i(1-\alpha_i)}{(1-\alpha_i^5)}+\frac{554\alpha_i^2(1-\alpha_i)}{(1-\alpha_i^5)}}\\
&\scriptstyle{+\frac{872\alpha_i^3(1-\alpha_i)}{(1-\alpha_i^5)}+\frac{1190\alpha_i^4(1-\alpha_i)}{(1-\alpha_i^5)}\Bigg) }\\
\scriptstyle{T_i^{(2)}}&\scriptstyle{=}\scriptstyle{(78+158+318+318+318)}\end{aligned}$$ where $\overline{B}_i$ refers to the mean backoff duration until the packet is transmitted or discarded due to successive CCA failures, $T_i^{(1)}$ has the interpretation of the mean backoff duration given that the packet transmission was successful, and $T_i^{(2)}$ is the mean time spent in backoff given that the packet was discarded due to successive CCA failures. Then the quantities $\overline{Z}_i$ and $\overline{Y}_i$ can be calculated as: $$\begin{aligned}
\scriptstyle{\overline{Z}_i}&\scriptstyle{=}&\scriptstyle{\alpha_i^5T_i^{(2)}+(1-\alpha_i^5)[T_i^{(1)}+\gamma_i(\alpha_i^5T_i^{(2)}+(1-\alpha^5)} \\
&&\scriptstyle{[T_i^{(1)}+\gamma_i(\alpha_i^5T_i^{(2)}+(1-\alpha_i^5)[T_i^{(1)}+\gamma_i(\alpha_i^5T_i^{(2)}+(1-\alpha_i^5)T_i^{(1)})])])]}\\
\scriptstyle{\overline{Y}_i}&\scriptstyle{=}&\scriptstyle{(1-\alpha_i^5)[T_{\mathsf{tx}}+\gamma_i(1-\alpha_i^5)[T_{\mathsf{tx}}+\gamma_i(1-\alpha_i^5)[T_{\mathsf{tx}}+\gamma_i(1-\alpha_i^5)T_{\mathsf{tx}}]]]}\end{aligned}$$
### Aggregate Arrival Rate ($\nu_i$), Goodput ($\theta_i$) and Discard Probability ($\delta_i$) for a Node
The arrival process at each node consists of packets which are generated at the same node (if it is a sensor) and the packets to be forwarded. The aggregate arrival rate at a node $i$ can be written as: $$\begin{aligned}
\nu_i=\lambda_i+\sum_{k\in\mathcal{P}_i}\theta_k
\label{eqn:nu}\end{aligned}$$ where $\lambda_i$ is the packet generation rate at sensor node $i$, $\mathcal{P}_i$ is the set of all its in-neighbours, and $\theta_i$ is the rate of packet transmission by the node which are *successfully received* at $r(i)$ (known as goodput). An enqueued packet at a node is successfully received by the intended receiver in a manner shown in Figure \[fig:goodput\].
![The goodput of a node is defined as the rate of successfully received packets by its receiver.[]{data-label="fig:goodput"}](figures/goodput.pdf)
Assuming that the queueing system is stable and has a steady-state solution, the goodput $\theta_i$ is given by $$\begin{aligned}
\theta_i=\nu_i(1-\delta_i)
\label{eqn:theta_arrival}\end{aligned}$$ where $\delta_i$ is the probability of discarding a packet due to consecutive CCA failures, or successive failed retransmission attempts, and is given by $$\begin{aligned}
\scriptstyle{\delta_i=\alpha_i^5+(1-\alpha_i^5)\gamma_i\Big[\alpha_i^5+(1-\alpha_i^5)\gamma_i\Big[\alpha_i^5+(1-\alpha_i^5)\gamma_i\Big[\alpha_i^5+(1-\alpha_i^5)\gamma_i\Big]\Big]\Big]}
\label{eqn:delta}\end{aligned}$$ Note that if the queue at node $i$ is saturated, then the goodput $\theta_i$ is equal to $\sigma_i$. Note that while calculating $\sigma_i$, we have taken the packet discards into account.
### The node non-empty probability, $q_i$ {#non_empty_prob}
To find the expression for $q_i$, assuming that all the arriving packets reach the HOL position (i.e., no tail drops) and applying Little’s Theorem, we get $$\begin{aligned}
q_i=\frac{\nu_i}{\sigma_i}
\label{eqn:q_arrival}\end{aligned}$$ Further, for a saturated node, the quantity $q_i$ is equal to $1$.
### Obtaining $b_i$ and $\beta_i$
In order to find $b_i$, the fraction of time a node is in backoff provided it is non-empty, we embed a renewal process in conditional time where the renewal epochs are those instants at which the node enters the random backoff period after a packet transmission or packet discard. We use the RRT to find the expression for $b_i$ as $$\begin{aligned}
b_i=\frac{\overline{B}_i}{\overline{B}_i + (1-\alpha_i^5)T_{\mathsf{tx}}}
\label{eqn:calc_of_b}\end{aligned}$$
The quantity $\beta_i$, the rate of CCA attempts in backoff times, is based on the backoff completion times irrespective of the transmission attempt of the packet since a packet retransmission is considered the same as a new packet transmission. Thus, for the calculation of $\beta_i$, it suffices to observe the process at node $i$ only in the backoff times. Figure \[fig:backoff\_times\] shows the residual backoff process where after completion of the backoff duration, the node performs a CCA.
![Evolution of backoff periods in conditional (backoff) time. The $y$-axis is in units of a backoff slot ($20 ~T_\mathsf{s}$). The range indicates that the random backoff duration is uniformly distributed within it. The CCA duration, and Rx-to-Tx turnaround duration are 8 and 12 symbol times respectively, as indicated in the boxes.[]{data-label="fig:backoff_times"}](figures/backoff_times.pdf){height="30mm" width="93mm"}
Using the result of Kumar et al. [@kumar-etal04new-insights] with their collision probability replaced by our CCA failure probability, $\alpha_i$, the expression for $\beta_i$ can be written as $$\begin{aligned}
\beta_i=\frac{1+\alpha_i+\alpha_i^2+\alpha_i^3+\alpha_i^4}{\overline{B}_i}
\label{eqn:beta}\end{aligned}$$
Iterative Solution and Calculation of Performance Measures {#sec:iterative-scheme}
==========================================================
![The global iteration scheme. The square boxes indicate the derived quantities and the round boxes indicate the known quantities.[]{data-label="fig:iterations"}](figures/iterations.pdf)
Global Iteration Scheme {#sec:iterations}
-----------------------
The global iteration scheme to solve the fixed point equations derived above is shown in Figure \[fig:iterations\]. We start with a vector $\{0,10,\lambda_i,l_i\}$ corresponding to $\{\alpha_i,\overline{\tau}_j^{(i)},\nu_i,l_i\}$ for each node $i$ and node $j \in \Omega_i$, and repeat the procedure until the quantities converge.
Existence of a fixed point
--------------------------
A careful look at the derivation of the fixed point equations in Section \[subsec:fixed-pt-derivation\] reveals that the fixed point variables are $\{(\alpha_i,q_i)\}_{i=1}^N$, where $\alpha_i$ is the CCA failure probability at node $i$, and $q_i$ is the queue non-empty probability at node $i$. These variables are related via the set of equations , , , , , , , and -. We observe that all the functions involved in the fixed point equations are continuous (since compositions, sums, and products of continuous functions are continuous, and also minimum of continuous functions is continuous). Hence, the fixed point equations can be represented as a *continous* map from $[0,1]^{2N}$ to $[0,1]^{2N}$. Hence, the existence of a solution to the fixed point equations follows from Brouwer’s fixed point theorem.
Proving the uniqueness of the fixed point, and convergence of the iterative procedure to the fixed point is out of the scope of this work. For a detailed discussion of the complexity associated with these proofs for general multi-hop topologies, see Section IV.D in [@jindal09]. To the best of our knowledge, the only work to have attempted a formal proof of the uniqueness of the fixed point in the context of multi-hop networks is [@marbach11], but that too for a rather simplified version of CSMA policies. However, *in our numerical experiments, the fixed point iterations always converged to a solution*.
End-to-End Delay Calculation
----------------------------
We use the two parameter Whitt’s QNA [@whitt83queueing-network-analyzer] to calculate the mean sojourn time at each node based on an approximate Moment Generating Function (MGF) of service time. We can write an expression for service time in a recursive manner by allowing infinite CCAs and retransmissions. To account for the discarded packets, we allow only a fraction of transmitted packets to join the next-hop neighbour’s queue based on the discard probability at that node. We denote the service time at node $i$ by $S_i$, and let $B_i$ denote the length of random backoff duration, which is assumed to have an exponential distribution with rate $\beta_i$ for node $i$. Then, $S_i$ can be written as $$\begin{aligned}
S_i &=& \begin{cases}
B_i + \tilde{S}_i & \text{w.p. } \alpha_i \\
B_i + T_{\mathsf{tx}} & \text{w.p. } (1-\alpha_i)(1-\gamma_i) \\
B_i + T_{\mathsf{tx}} + \tilde{S}_i & \text{w.p. } (1-\alpha_i)\gamma_i
\end{cases}\end{aligned}$$ where $\tilde{S}_i$ is a random variable with the same distribution as $S_i$. The MGF of $B_i$, denoted by $M_{B_i}(z)$, is equal to $\frac{\beta_i}{z+\beta_i}$. Therefore, we can express the MGF $M_{S_i}(z)$ of service time $S_i$ as $$\begin{aligned}
M_{S_i}(z) &=& \frac{\beta_i(1-\alpha_i)(1-\gamma_i)e^{-zT_{\mathsf{tx}}}}{z+\beta_i(1-\alpha_i)(1-\gamma_ie^{-zT_{\mathsf{tx}}})}.\end{aligned}$$ The first two moments of the service time, $\mathbb{E}(S_i)$ and $\mathbb{E}(S_i^2)$ can be calculated by differentiating the MGF $M_{S_i}(z)$: $$\begin{aligned}
\mathbb{E}(S_i) &=& -\frac{d(M_{S_i}(z))}{dz}\Bigg \vert_{z=0} \\
\mathbb{E}(S_i^2) &=& \frac{d^2(M_{S_i}(z))}{dz^2}\Bigg \vert_{z=0} \end{aligned}$$ The QNA procedure commences from the leaf nodes and converges to the base-station in a sequential manner. We first calculate the squared coefficient of variance of service time, denoted by $c_{S_i}^2$, at each node, i.e., $$\begin{aligned}
c_{S_i}^2 &=& \frac{\mathbb{E}(S_i^2)}{(\mathbb{E}(S_i))^2} - 1\end{aligned}$$ Ignoring the packet discards at the MAC layer, the net arrival rate $\Lambda_i$ at a node $i$ is given by: $$\begin{aligned}
\Lambda_i &=& \begin{cases}
\lambda_i + \displaystyle{\sum_{k\in \mathcal{P}_i}}\theta_k & \text{for source-cum-relay nodes} \\
\displaystyle{\sum_{k\in \mathcal{P}_i}}\theta_k & \text{for relay nodes}
\end{cases}\end{aligned}$$ For any node $i$, let $\rho_i$ be defined as $$\begin{aligned}
\rho_i &=& \begin{cases}
\lambda_i\mathbb{E}(S_i) & \text{for leaf nodes} \\
\Lambda_i\mathbb{E}(S_i) & \text{for internal nodes}
\end{cases}\end{aligned}$$ We now calculate the squared coefficient of variance for the departure process ($c_{D_i}^2$) for a node which uses the squared coefficient of variance for the interarrival times at that node ($c_{A_i}^2$). Here we include the probability of a packet discard which is equal to $\delta_i$. $$\begin{aligned}
c_{D_i}^2 &=& (1 - \delta_i)(1 + \rho_i^2(c_{S_i}^2 - 1) + (1-\rho_i^2)(c_{A_i}^2 - 1))\end{aligned}$$ Further, $c_{A_i}^2$ is calculated as $$\begin{aligned}
c_{A_i}^2 &=& \frac{1}{\Lambda_i}\Bigg(\lambda_i + \displaystyle{\sum_{j\in \mathcal{P}_i}}\Lambda_jc_{D_j}^2\Bigg)\end{aligned}$$ Finally, the mean sojourn time at a node $i$ is given by $$\begin{aligned}
\overline{\Delta}_i &=& \frac{\rho_i\mathbb{E}(S_i)(c_{A_i}^2 + c_{S_i}^2)}{2(1-\rho_i)} + \mathbb{E}(S_i)\end{aligned}$$ The end-to-end mean packet delay for a source node $j$, provided that the set of nodes along the path from this node to the BS is $L_j$, is given by $$\begin{aligned}
\Delta_j &=& \displaystyle{\sum_{i\in L_j}\overline{\Delta}_i}\end{aligned}$$
Packet Delivery Probability for each Source Node
------------------------------------------------
$p_i^{(del)}$ for each source node $i$ is defined as the fraction of generated packets at source node $i$ that reach the base station without any time bound. Let the set of nodes constituting the path from a source node $i$ to the BS be $L_i$. Assuming that the drop events are independent from node to node, the expression for $p_i^{(del)}$ is $$\begin{aligned}
p_i^{(del)} &=& \prod_{j\in L_i}(1-\delta_j)\end{aligned}$$ where $\delta_j$ is the packet discard probability of node $j$.
Discussion on Validity of the Fixed Point Approach {#sec:dtmc-stability}
==================================================
We recall that we developed the analysis under the premise that the system of queues is stable, and, hence, all the steady state quantities involved in the fixed point equations exist. The next question we ask is whether, having performed the above analysis, we can use the results to conclude that the system of queues is indeed stable, which would provide a consistency in the overall approach.
We proceed by modeling the CSMA/CA multihop network analyzed in Section \[sec:analytical-model\] as a Discrete Time Markov Chain (DTMC), and therefrom, deriving a sufficient condition for the stability of the network.
A DTMC Model {#subsec:dtmc}
------------
The system evolves *synchronously* over slotted time (this is an idealization only for the analysis in this section), with the duration of each slot being 1 symbol time [^3]. We assume that the *start instants of all backoffs are aligned with the slot boundaries*. Further, the external arrival process to each source node is assumed to be Poisson; hence the number of arrivals in successive slots are independent (independent increment property). The evolution of the queueing system can then be modeled as a DTMC embedded at these slot boundaries, with the state of each queue comprising of the following components:
1. Queue length at each node
2. Residual backoff at each node
3. Residual CCA time at each node
4. Backoff stage (i.e., the number of CCA attempts for the current packet) at each node
5. Transmission state (whether the node is transmitting a packet or not) at each node
6. Retransmission stage (the number of retransmissions so far for the current packet) at each node
7. Residual packet length when a node is transmitting (in symbol times)
Note that except the queue lengths, all the other components of the state space are *finite*, while the state space of the queue lengths is countable. Hence, the state space is countable. We denote the state at time step $n$ by $({\mathbf{X}}(n),{\mathbf{Y}}(n))$, where ${\mathbf{X}}(n)=(X_1(n),\ldots,X_N(n))$ denotes the queue length process, and ${\mathbf{Y}}(n)$ denotes the rest of the components of the state taken together. Thus, ${\mathbf{X}}(n)\in\mathbb{Z}^n_+$, and ${\mathbf{Y}}(n)\in K_y$, where $K_y$ is a finite set of finite valued vectors.
We adopt the following convention from [@panda] (see Figure 4 in Section 3.1 in [@panda]) for counting the queue lengths at each time step: all arrivals that occur during a slot are counted immediately before the end boundary of the slot, all packets that leave the queues during a slot are counted at the end boundary of the slot, and the queue lengths are computed immediately after the end boundary of the slot, so as to account for all the arrivals and departures during the slot.
With this setup, the queueing system evolves as a DTMC over the state space $\mathbb{Z}^n_+\times K_y$. The transition probabilities of the DTMC are governed by the distribution of the arrival process, and the backoff distribution. Note that a similar DTMC model was proposed in [@kumar-etal04new-insights] for IEEE 802.11 CSMA/CA under saturation assumption.
Now note that the operation of the system starts with all the queues empty, i.e., in the state $({\mathbf{X}}(0)={\mathbf{0}},{\mathbf{Y}}(0)={\mathbf{0}})$. *We are, therefore, only interested in the communicating class of the DTMC containing the all-zero state*, $({\mathbf{X}}(0)={\mathbf{0}},{\mathbf{Y}}(0)={\mathbf{0}})$. Let us denote this class by $\mathcal{C}_{0}$.
\[prop:irreducible\] The class $\mathcal{C}_{0}$ is closed, and aperiodic.
See the Appendix.
It follows from Proposition \[prop:irreducible\] that the DTMC evolving from the all-zero state is confined to the class $\mathcal{C}_{0}$. *With abuse of notation, from now on, we denote by $({\mathbf{X}}(n),{\mathbf{Y}}(n))$, the DTMC evolving from the all-zero state.*
A Sufficient Condition for Network Stability {#subsec:stability-conditions}
--------------------------------------------
Since we have a DTMC, by “stability”, we mean that the DTMC $({\mathbf{X}}(n),{\mathbf{Y}}(n))$ is *positive recurrent*. Then, since $({\mathbf{X}}(n),{\mathbf{Y}}(n))$ is also aperiodic, it can be verified that the steady state rates introduced in Section \[sec:analytical-model\] *exist*. Therefore, we shall derive conditions for positive recurrence of the DTMC.
Note that the queue non-empty probability of queue $i$, $1\leq i\leq N$, in the steady state is[^4] $$q_i = \lim_{n\to \infty}{\mathsf{Pr}}[X_i(n) > 0|{\mathbf{X}}(0)=\mathbf{0},{\mathbf{Y}}(0)=\mathbf{0}]$$
We have the following sufficient condition on $q_i$, $1\leq i\leq N$, for positive recurrence of the DTMC $({\mathbf{X}}(n),{\mathbf{Y}}(n))$.
\[thm:stability\] If $\sum_{i=1}^N q_i\:<\:1$, then the DTMC $({\mathbf{X}}(n),{\mathbf{Y}}(n))$ (evolving from the all-zero state) is positive recurrent.
See the Appendix.
**Discussion:** We began this section by asking the question whether the results from our analytical model can be used to verify the initial assumption of stability. Since the analysis involves many approximations, evidently no definite answer can be given to this question. However, we have found from extensive simulations (see Section \[sec:numerical-and-simulation-results\] for details of the simulation procedure) that the solution to the fixed point analysis models $\sum_{i=1}^N q_i$ accurately (*within an error of up to 10%*) for external arrival rates of up to about 6 packets/sec (see Figure \[fig:measure\_sum\_qi\] for example).
![Plot of measure $\sum_{i=1}^N q_i$ for the topology in Figure \[fig:hidden\_10nodes\][]{data-label="fig:measure_sum_qi"}](plots/10_Nodes_q_sum_plot_upto_rate_20.pdf)
This observation along with Theorem \[thm:stability\] can support the claim that if $\sum_{i=1}^N q_i < 0.9$ in the solution to the fixed point analysis, then we can safely assume the network to be stable.
Numerical and Simulation Results {#sec:numerical-and-simulation-results}
================================
For the verification of our analytical model, we use QualNet (v4.5) simulator [@qualnet] with the default parameter values and a fixed payload size of $70$ bytes. However, the QualNet implementation is devoid of ACKs. Therefore, we compare results only for the ACK*-less* scenarios (although the analysis permits the modeling of ACKs). We use the following simulation models in Qualnet to declare that a packet is in error:
1. Collision and Link Error Model: If a receiver is receiving a packet from a node, and there is another transmission in the carrier sense range of the receiver, then the receiver concludes collision of all the packets that are intended for it. In addition, if a receiver $r(i)$ of node $i$ receives a packet successfully from node $i$ since there is no other packet transmission in the receiver’s carrier sense range, then the received packet is concluded to be in error with probability $l_i$, the probability of data packet error on the link between node $i$ and the receiver $r(i)$ due to noise.
2. Capture Model: If a receiver is receiving a packet from a node, and there is another transmission in the carrier sense range of the receiver, then the receiver computes the PER (from the SINR) of the packet that is meant for it, and rejects the packet if the computed PER is greater than a random number generated between 0 and 1.
To increase the accuracy of simulation for each arrival rate, we generate the packets for 1500 seconds, and average the results over 25 repetitions with different random number seeds. 1500 seconds was chosen so that even at lower arrival rates ($\leq$ 2pkt/sec), sufficient number of packets would be generated from each source, thereby allowing the simulation to reach steady state.
Simulation of an Example Network Topology
-----------------------------------------
The network shown in Figure \[fig:hidden\_10nodes\] has hidden nodes (the dependency graph is also shown). Note that all the nodes are sources (some of which also serve as relays) with identical packet generation rates that are simultaneously increased. Figures \[fig:measure\_alpha\_for\_tree-n10-CS3-PER0.01\], \[fig:measure\_gamma\_for\_tree-n10-CS3-PER0.01\], \[fig:measure\_delta\_for\_tree-n10-CS3-PER0.01\], and \[fig:measure\_pdel\_for\_tree-n10-CS3-PER0.01\] are plots of measures $\alpha_i$, $\gamma_i$, $\Delta_i$ and $p^{(\mathsf{del})}_i$ for the nodes in Figure \[fig:hidden\_10nodes\].
![A 10 nodes TREE topology. Also shown is the dependency graph where the dotted lines connecting two nodes indicate that the nodes are in CS range of each other.[]{data-label="fig:hidden_10nodes"}](figures/hidden_10nodes.pdf)
![Plots of CCA failure rate, $\alpha$, for the nodes in Figure \[fig:hidden\_10nodes\]; blue solid line indicates analytical results, green solid line indicates simulation with Capture Model, and Red dotted line indicates simulation with Collision + PER Model[]{data-label="fig:measure_alpha_for_tree-n10-CS3-PER0.01"}](plots/treen10CS3PER001_alpha_plot.jpg){height="7cm" width="9.5cm"}
![Plots of the packet failure probability, $\gamma$, for the nodes in Figure \[fig:hidden\_10nodes\]; blue solid line indicates analytical results, green solid line indicates simulation with Capture Model, and Red dotted line indicates simulation with Collision + PER Model[]{data-label="fig:measure_gamma_for_tree-n10-CS3-PER0.01"}](plots/treen10CS3PER001_gamma_plot.jpg){height="7cm" width="9.5cm"}
![Plots of the mean delivery delay, $\Delta$, for the nodes in Figure \[fig:hidden\_10nodes\]; blue solid line indicates analytical results, green solid line indicates simulation with Capture Model, and Red dotted line indicates simulation with Collision + PER Model[]{data-label="fig:measure_delta_for_tree-n10-CS3-PER0.01"}](plots/treen10CS3PER001_end-to-end-delay.jpg){height="7cm" width="9.5cm"}
![Plots of the average probability of delivery, $P^{(\mathsf{del})}$, for the nodes in Figure \[fig:hidden\_10nodes\]; blue solid line indicates analytical results, green solid line indicates simulation with Capture Model, and Red dotted line indicates simulation with Collision + PER Model[]{data-label="fig:measure_pdel_for_tree-n10-CS3-PER0.01"}](plots/treen10CS3PER001_pdel.jpg){height="7cm" width="9.5cm"}
0 We observe that error in measure $\alpha_i$ decreases at higher arrival rates. Despite significant error in measure $\alpha_i$ at lower arrival rates ($\leq 0.5$ pkts/sec), other measures are within $\pm10\%$ error (though not shown here). This can be attributed to the following reasons: (1) values of measure $\beta_i$ and $b_i$ are more governed by mean time of first and second backoff, and not so much by $\alpha_i$, especially for small values of $\alpha_i$. Same is true for measure $q_i$; (2) since the other measures are obtained from measures $\beta_i$, $b_i$ and $q_i$, they are also not affected much by inaccuracies in measure $\alpha_i$; most importantly, (3) measure $\delta_i$ for hidden node networks is more affected by measure $\gamma_i$ but not by measure $\alpha_i$, since a packet is discarded at the MAC layer only after 5 consecutive CCA failures, the probability of which is very small at such low rates. Error in measure $\gamma_i$ is upto $25\%$ even at higher arrival rates. Error in measure $\Delta_i$ is within $\pm10\%$ for all the arrival rates plotted, and so is true for measure $p^{(\mathsf{del})}_i$. Also at higher arrival rates, the probability of delivery, $p^{(\mathsf{del})}_i$, for relay nodes has more error compared to leaf nodes since the error accumulates due to the error in the discard probability, $\delta_i$, at each node along a path.
**Observations:**
1. We observe from the plots that the errors in the measures increase at higher arrival rates, i.e., the accuracy of the analysis decreases at higher arrival rates. For example, whereas for Nodes 1 to 4, and Nodes 6 to 10, the errors in the packet failure probability, $\gamma_i$, were between 1.5% and 20%, for Node 5, the error was up to 22.2% with arrival rate of 20 packets per second. We will see, however, that the practical operating range of these networks is closer to 1 measurement per second from each node, for which value the approximation is excellent.
2. The differences in the simulation measures under the Capture Model and the Collision + Link Error Model were well within 10% for all nodes at lower arrival rates, and within 20% even at an arrival rate of 20 packets per second. Hence, we conclude that the Collision + PER Model does not yield performance significantly different from the Capture Model. For other examples illustrating this fact, see [@sanjaymeth].
Extensive Simulation Results
----------------------------
We now present a summary of an extensive simulation study, where we only use Collision + Link Error as the model. We have performed simulations with a variety of scenarios where one or more of the following features were varied: the topology, number of nodes in the network, average number of nodes in the CS range of a node, and the probability of data packet error on the link due to noise. For compact representation of the various cases, we use the following notation:
- tree-n$N$-CS$m$-PER$l$: These are the cases in which the network is a tree with $N$ nodes (excluding the base station), all of which are sources. On an average, $m$ nodes are in the CS range of a node, i.e., $m=\frac{\sum_{i=1}^N \text{number of nodes in the CS range of node $i$}}{N}$. Moreover, the PER takes the same value, $l$, on all the links.
- tree$\mathsf{R}$-n$N$-CS$m$-PER$l$: These are the cases in which the network is a tree with $N$ nodes (excluding the base station) such that 10 nodes among them are source nodes, and the rest of the nodes are relay nodes (the “R” in “treeR” signifies that there are relays as well as sources). On an average, $m$ nodes are in the carrier sense range of a node, and the PER takes the same value, $l$, on all the links.
- star-n$N$-CS$m$-PER$l$: These are the cases in which the network is a star with $N$ nodes (excluding the base station), all of which are sources, positioned symmetrically on a circle centred at the base station. The $m$ nodes nearest to a node are in the carrier sense range of it, and the PER on all the links has the same value $l$.
- line-n$N$-CS$m$-PER$l$: These are the cases in which the network is a line with $N$ nodes (excluding the base station) all of which are sources. $m$ nodes on either side of a node are in the carrier sense range of it. The PER on all the links is the same, and is equal to $l$.
For example, the case in Figure \[fig:hidden\_10nodes\] is tree-n10-CS3-PER0.01.
We compute the fractional errors in performance measures as $\frac{\text{Simulation} - \text{Analysis}}{\text{Simulation}}$. The Boorstyn et al. [@boorstyn-etal87CSMA-throughput-analysis] model based analysis (see \[subsubsec:Ti-eff\]) is employed for computing $T_i^{(\mathsf{eff})}$.
Table \[tab:error for various topologies\] tabulates the accuracy of the analysis compared to simulation, and also the computation times of both simulation and analysis. We use the following notation to indicate the range of errors: (i) indicates that the error is within $\pm$ $10\%$; (ii) $+$ indicates overestimate by analysis with error from $10\%$ to $25\%$; (iii) $++$ indicates overestimate by analysis with error more than $25\%$; (iv) $-$ indicates underestimate by analysis with error from $10\%$ to $25\%$; and (v) $--$ indicates underestimate by analysis with error more than $25\%$. The entry in every row is $a, b$, where $a, b \in \{\checkmark, +, ++, -, -- \}$ with $a$ denoting the error summary for a range of $\lambda$ where, the discard probability, $\delta_i \leq 0.01$, and $b$ denoting the error summary for $\lambda > 2$ pkts/sec.
-------- ---------------------------------- --------------------------------- --------------------- ----------------------------------- -----------------------------------
Sl. No Topology $\overline{P}^{(\mathsf{del})}$ $\overline{\Delta}$ Simulation Time Analysis Time
(Min, Mean, Max) in seconds in seconds
per arrival rate until 10pkts/sec per arrival rate until 10pkts/sec
1 tree$\mathsf{R}$-n23-CS4-PER0.01 , $-$ , $3120, 9720, 17580$ $16$
2 tree$\mathsf{R}$-n20-CS3-PER0.01 , $-$ , $2220, 7740, 15060$ $12$
3 tree$\mathsf{R}$-n20-CS3-PER0.01 , $-$ , $2580, 8640, 16800$ $14$
4 tree$\mathsf{R}$-n19-CS3-PER0.01 , $-$ , $2460, 7680, 16020$ $10$
5 tree$\mathsf{R}$-n19-CS2-PER0.01 , , $1800, 7440, 14340$ $10$
6 tree$\mathsf{R}$-n20-CS3-PER0.01 , $-$ , $2880, 9480, 17040$ $16$
7 tree$\mathsf{R}$-n20-CS2-PER0.01 , $--$ , $2280, 8340, 15180$ $12$
8 tree$\mathsf{R}$-n22-CS4-PER0.01 , $--$ , $3420, 11820, 19680$ $36$
9 tree$\mathsf{R}$-n19-CS2-PER0.01 , $-$ , $2040, 7440, 14340$ $14$
10 tree$\mathsf{R}$-n19-CS3-PER0.01 , $-$ , $2460, 8760, 17040$ $17$
11 tree$\mathsf{R}$-n20-CS3-PER0.01 , $-$ , $2640, 9180, 17460$ $17$
12 tree$\mathsf{R}$-n19-CS3-PER0.01 , $-$ , $2340, 7920, 16740$ $15$
13 tree$\mathsf{R}$-n19-CS3-PER0.01 , $-$ , $2400, 8280, 16980$ $12$
14 tree$\mathsf{R}$-n20-CS3-PER0.01 , $-$ , $2340, 8340, 16440$ $11$
15 tree$\mathsf{R}$-n19-CS3-PER0.01 , $-$ , $2280, 7860, 16260$ $11$
16 tree$\mathsf{R}$-n19-CS2-PER0.01 , $--$ , $1740, 7260, 14220$ $9$
17 tree$\mathsf{R}$-n20-CS3-PER0.01 , $-$ , $2520, 8520, 16020$ $11$
18 tree$\mathsf{R}$-n19-CS2-PER0.01 , $-$ , $2040, 7560, 14760$ $10$
19 tree$\mathsf{R}$-n19-CS2-PER0.01 , $-$ , $1920, 7140, 14460$ $9$
20 line-n10-CS2-PER0.01 , , 780, 6240, 12060 5
21 line-n10-CS3-PER0.01 , , 1080, 7860, 13980 8
22 line-n10-CS4-PER0.01 , , $+$ 1260, 8700, 16080 12
23 star-n20-CS9-PER0.01 , , 900, 6480 ,12660 36
24 star-n20-CS11-PER0.01 , , 1140, 7620, 14880 48
25 tree-n13-CS2-PER0.01 , , 960, 7320, 13260 5
-------- ---------------------------------- --------------------------------- --------------------- ----------------------------------- -----------------------------------
**Observations and Discussion:**
1. As can be observed from Table \[tab:error for various topologies\], for small arrival rates at which the discard probability, $\delta_i \leq 0.01$, the errors in both $\overline{P}^{(\mathsf{del})}$, and $\overline{\Delta}$ are less than 10% in all the scenarios tested, whereas at higher arrival rates, the error in $\overline{P}^{(\mathsf{del})}$ sometimes exceeds 25% (scenarios 7,8, and 16), which is in agreement with our earlier observation from Figure \[fig:measure\_pdel\_for\_tree-n10-CS3-PER0.01\] as well. However, the error in $\overline{\Delta}$ even at higher arrival rates was within 10% for all but one scenario (scenario 22), where the analysis overestimated the delay within an error of 25%.
One possible explanation for the degradation in the accuracy of the analysis at higher arrival rates can come from the discussion in Section \[sec:dtmc-stability\]. Note that at higher arrival rates, the queue non-empty probability of the nodes are typically higher, and hence the condition $\sum_{i=1}^N q_i < 1$ (see Theorem \[thm:stability\]) is more likely to be violated, in which case, the system may not be stable, and the validity of the fixed point analysis is questionable. For example, note from Figure \[fig:measure\_sum\_qi\] that for the example topology in Figure \[fig:hidden\_10nodes\], $\sum_{i=1}^N q_i$ approaches 1 at around $\lambda = 10$ pkts/sec, and it can be seen from Figures \[fig:measure\_alpha\_for\_tree-n10-CS3-PER0.01\], \[fig:measure\_delta\_for\_tree-n10-CS3-PER0.01\], \[fig:measure\_gamma\_for\_tree-n10-CS3-PER0.01\], and \[fig:measure\_pdel\_for\_tree-n10-CS3-PER0.01\] that the errors in the measures are more pronounced for $\lambda \geq 10$ pkts/sec.
2. We also observe from Table \[tab:error for various topologies\] that the average time to analytically compute the performance measures for each arrival rate (using the Boorstyn et.al [@boorstyn-etal87CSMA-throughput-analysis] model for computing $T^{(\mathsf{eff})}_i$) was of the order of seconds, whereas the average simulation time ran into several hours.
Further, it was observed that for hidden-nodes networks with 20 nodes, and an average of 3 nodes in the CS range of a node, the analysis with Boorstyn et.al [@boorstyn-etal87CSMA-throughput-analysis] model for computing $T^{(\mathsf{eff})}_i$ takes around $13$ secs per arrival rate, and if $T^{(\mathsf{eff})}_i$ is computed using $M/D/\infty$, then the analysis takes around $9$ secs per arrival rate. Hence we can trade-off between the accuracy of analysis and computation time since analysis with $M/D/\infty$ model usually incurs more error ($> 25\%$) at higher arrival rates, but less than $10\%$ for $\lambda < 0.5$ pkts/sec.
3. Finally, it was observed that the packet discard probability at a node increases to an impractical value much before the average delay at the node becomes substantial, e.g., tree-n10-CS3-PER0.01-Node-1 discards $2$-$5$ packets for every $1000$ packets when the average node delay is $5$-$6$ msec. Hence, the overall packet delivery probabilities for the sources act as performance bottlenecks for these networks.
Network Design {#sec:network-design}
==============
In this section, we shall consider a problem of QoS constrained network design for a given positive traffic arrival rate to demonstrate the usefulness of the analytical model in an iterative network design process.
Consider the following setting: *a set of sensor nodes and a base station (BS) are deployed over an area; each sensor node is a data source, and can also act as a relay. The transmit power level of each node can be adjusted over a range.* Consider the complete graph over these nodes. Each directed link in this graph will have a certain packet error rate (PER) that *depends on the transmit power level of the sender node on that link.*
We consider the following network design problem: for a given traffic arrival rate $\lambda > 0$ at the sensor nodes, *minimize the maximum transmit power level used by the sensor nodes, such that the resulting network has the following properties.*
1. Each sensor node has a path to the BS, with the PER on each link in each path being upper bounded by a predefined target threshold $p$.
2. For the given $\lambda$, the end-to-end packet delivery probability (i.e., the probability that a packet is not discarded) on any path is at least $p_{\mathrm{del}}$.
3. For the given $\lambda$, the mean delay (computed over the successfully delivered packets) on any path is upper bounded by a predefined target $d_{\max}$.
Note that for a design (network) to satisfy the QoS objectives for a given arrival rate, it is *necessary* that the network satisfies the QoS objectives under zero/light traffic load, i.e., in the limit as $\lambda \to 0$, in which limit a packet that enters the network departs from the network before another packet arrives, i.e., each packet traverses the network $\emph{alone}$, prompting us to call this limit the $\emph{lone packet}$ traffic model (for a more formal proof[^5] of this fact, see [@fullpaper]). We, therefore, adopt the following two step approach:
1. We first focus on the QoS constrained network design problem under the lone-packet traffic model in Section \[subsec:lone-packet-design\]; we formulate the problem as a network design problem on graphs, and propose an algorithm to solve the problem optimally.
2. Then, in Section \[subsec:cont-traffic-design\], we combine the lone-packet based design algorithm with the analytical tool developed in Section \[sec:analytical-model\] to address the more general (and more complex) problem of QoS constrained network design for a given positive traffic arrival rate.
Note that we can ensure a given target PER on a link by ensuring that the *average* received power across the link (averaged over shadowing and fading) meets a received power target. Such a target average received power would be obtained by deriving a margin above the minimum value of received power, from statistics of shadowing and fading. If we take this worst case approach over the joint pdf of shadowing and fading, then for a given target PER, the required transmit power on a link is a non-decreasing function of the link length. Hence, the problem of minimizing the maximum transmit power is *equivalent to minimizing the maximum link length in the resulting network*.
Network Design under the Lone-Packet Model {#subsec:lone-packet-design}
------------------------------------------
Given the target PER $p$ on any link, the mean delay on a link under the “lone packet” model can be computed using an elementary analysis (see [@abhijitmeth]), taking into account the backoff behavior of 802.15.4 CSMA/CA, and using the backoff parameters given in the standard [@IEEE802-15-4-06std]. Then, to meet the mean delay requirement of $d_{\max}$ on a path with $h$ hops, we require that
$$h\leq \left\lfloor\frac{d_{\max}}{\overline{d}_{single-hop}}\right\rfloor\triangleq h_{\max}^{delay}$$
where, $\overline{d}_{single-hop}$ is the mean link delay computed as explained earlier.
Again, given the target link PER of $p$, and the number of retries $r$ (obtained from the standard) before a packet is discarded on a link, the packet discard probability on a link can be obtained as $q\:=\:p^{r+1}$. Hence, to ensure a packet delivery probability of at least $p_{del}$ on a path with $h$ hops, we require that
$$\begin{aligned}
h &\leq \frac{\ln{p_{del}}}{\ln{(1-q)}}&\triangleq h_{\max}^{delivery}\end{aligned}$$
Hence, to ensure the QoS constraints under the “lone packet” model, we require that the *hop count on each path is upper bounded by $h_{\max}\:=\:\min\{h_{\max}^{delay},h_{\max}^{delivery}\}$*.
Thus, the problem of QoS constrained network design under the “lone packet” model can be reformulated as the following graph design problem:
*Given a set of source nodes $Q$, and a base station (BS), indexed as node 0, consider the graph $G=(V,E)$, where $V=Q\cup \{0\}$, and the edge set $E$ consists of all feasible edges, i.e., edges with PER $\leq\: p$ at maximum possible transmit power. Given a hop count constraint $h_{\max}$, the problem is to extract from $G$, a spanning tree on $Q$ rooted at the BS, such that the hop count on any path is upper bounded by $h_{\max}$, and the maximum edge length in the spanning tree is minimized*.
We call this, the MinMax Spanning Tree with Hop Constraint (MMST-HC) problem.
The following is an algorithm to obtain an optimal solution to the MMST-HC problem. At each iteration, we prune all edges with edge length more than or same as the maximum edge length in the current feasible solution, form a shortest path tree (SPT) with hop count as cost using only the remaining edges, and keep doing this until the resulting SPT violates the hop constraint, at which point we stop and declare the last feasible solution as the final solution.
### **SPTiEP: Shortest Path Tree based iterative Edge Pruning Algorithm**
- **Initialize:** Set $k\leftarrow 0$, $G^{(0)}\leftarrow G$
- **Checking feasibility:** In iteration $k$, find a shortest path tree (SPT) $T^{(k)}$ on the graph $G^{(k)}$. Check if all the paths from the sources to the BS in $T^{(k)}$ satisfy the hop constraint.
- If the hop constraint is not met for some of the sources in $T^{(k)}$, **STOP**.
- If $k=0$, declare the problem *infeasible*.
- If $k>0$, output $T^{(k-1)}$ as the final solution.
- If the hop constraint is met for all the sources in $T^{(k)}$, proceed to the next step.
- **Edge pruning:** Let $\overline{w}^{(k)}$ be the maximum edge length in $T^{(k)}$. Remove from $G^{(k)}$, all edges of length $\geq \overline{w}^{(k)}$ to obtain the graph $G^{(k+1)}$.
- **Iterate:** Set $k\leftarrow k+1$. Go to Step 2.
Since the time complexity of finding an SPT in a graph with $N$ nodes is $O(N\log N)$, it can be easily verified that the time complexity of the SPTiEP algorithm is $O(N^3\log N)$ for a graph with $N$ nodes. Thus, *the SPTiEP algorithm is polynomial time*.
### **Proof of correctness of SPTiEP**
We define
$\mathcal{F}^{(k)}\:=\:\{T\subset G^{(k)}: \text{$T$ satisfies the hop constraint}\}$: set of all feasible solutions contained in the graph $G^{(k)}$
$w_{opt}$: The minmax edge length, i.e., the maximum edge length in an optimal solution
Also recall the definitions of $T^{(k)}$ and $\overline{w}^{(k)}$ from the description of the algorithm.
Clearly, $\mathcal{F}^{(k)}$ is non-empty *if and only if* $T^{(k)}\in \mathcal{F}^{(k)}$.
To prove the correctness of SPTiEP algorithm, it is enough to show the following:
\[prop:sptiep-correctness\] Given that $\mathcal{F}^{(k)}\:\neq \emptyset$, $$\mathcal{F}^{(k+1)}\:=\emptyset\: \Rightarrow w_{opt}=\overline{w}^{(k)}$$
See the Appendix.
Network Design for a Given Positive Arrival Rate {#subsec:cont-traffic-design}
------------------------------------------------
Now we come back to the more general problem of QoS constrained network design for a given positive arrival rate. This problem is rendered much more difficult compared to its lone-packet version due to the complex stochastic interaction between contending nodes, which, unlike the lone-packet model, makes it hard to map the QoS constraints to simple explicit constraints on certain graph properties. Therefore, unlike the lone-packet version where we posed the problem as a pure graph design problem, in the positive traffic case, we shall use the analytical model explicitly in an iterative design process to evaluate for QoS, the designs obtained systematically in every iteration. While traditional network simulation tools can also be used, in principle, to evaluate a given network for QoS, *the time required for such network simulation is significantly more than that required by the analysis* (as we saw in Table \[tab:error for various topologies\], and shall see again in our numerical experiments in Section \[subsec:design-exp\]), and that makes network simulation, an impractical option in an iterative design process.
Also, a naive approach to the design problem would be to consider all possible trees from the given graph $G$, and evaluate each of them for QoS (using either the analysis, or network simulation), and choose one whose maximum edge length is minimum among all those that meet the QoS objectives. While this exhaustive search appears to be the only way that is guaranteed to obtain a feasible solution whenever there exists one, this approach has exponential time complexity (since the number of possible trees is exponential in the number of nodes), and is therefore, not practical.
We shall present below, a polynomial time algorithm for the proposed positive traffic design problem, using the SPTiEP algorithm along with the analytical model; note that because of the stochastic nature of the interaction between contending nodes in different possible networks, in absence of an exhaustive evaluation of all possibilities, the algorithm is not theoretically guaranteed to return a feasible solution whenever there exists one; however, in our numerical experiments, the algorithm was always found to return a feasible solution, and moreover, simulations confirmed that designs proposed with the analytical tool did meet the QoS requirements (see Section \[subsec:design-exp\]).
### **Extended SPTiEP: An algorithm for QoS constrained network design at given positive arrival rate $\lambda > 0$**
As we had mentioned before, to meet the QoS objectives at a positive arrival rate, it is *necessary* (but not sufficient) to meet the objectives under the lone-packet model. Hence, our design still needs to satisfy the hop count constraint $h_{\max}$ derived from the QoS constraints under the lone-packet model. With this in mind, we proceed as follows: if the outcome of the SPTiEP algorithm meets the QoS constraints at the given $\lambda$, then that is an optimal solution (since it is an optimal solution for the lone-packet design problem, and it is also QoS feasible for the given $\lambda>0$). If, however, the SPTiEP outcome does not meet the QoS constraint, we need to change the design by *adding some of the edges of greater length that were pruned in course of the SPTiEP algorithm*; note that adding edges is the only option as pruning any more edge from the SPTiEP solution will cause us to violate the hop constraint, and hence the lone-packet QoS. The detailed steps are presented below.
1. **Lone-packet design:** Run the SPTiEP algorithm on the graph $G$ to obtain a tree $T_0$.
- If $T_0$ does not satisfy the hop constraint, declare the problem infeasible, as we cannot satisfy the QoS objectives even for $\lambda = 0$.
- Else, go to the next step.
2. Set $k\leftarrow 0$. Mark all edge lengths in $G$ as *not examined*.
3. **Checking feasibility for $\lambda>0$:** Evaluate $T_k$ for QoS requirements (i.e., $p_{del}$ and $d_{\max}$) at the given arrival rate $\lambda$, using the analytical model.[^6]
- **Stopping criteria 1:** If QoS is met, output $T_k$ as the final solution.
- **Stopping criteria 2:** If QoS is not met, and all edge lengths in $G$ have been *examined*, declare the problem *possibly* infeasible.
- Else, go to next step.
4. Let $\overline{w}_k$ be the maximum edge length in $T_k$. Identify the least edge length $> \overline{w}_k$ in the graph $G$ that is not yet *examined*; let us denote this as $w_{least,k}$.
5. **Edge augmentation:** Augment $T_k$ with all edges (in $G$) of length $\leq w_{least,k}$, to obtain the graph $G_{k+1}$. Mark all the edge lengths in $G_{k+1}$ as *examined*.
[**Comment:** ]{}Since our objective is to minimize the maximum edge length, in Steps 4 and 5, we add back the edges pruned during SPTiEP *in increasing order of their lengths*.
6. **Redesign:** Find an SPT $T_{k+1}$ in $G_{k+1}$. Observe that if $T_k$ satisfies the hop constraint, $T_{k+1}$ also satisfies the hop constraint.
7. **Iterate:** Set $k\leftarrow k+1$. Go to Step 3.
[**Comment:** ]{}In Steps 6 and 7, we check if the resulting shortest path tree in the augmented graph satisfies the QoS constraints for the given positive arrival rate.
Note that in Step 6, there could be several possible SPTs on the graph $G_{k+1}$; while the total network wide traffic load, as well as the total number of nodes is the same in all of these SPTs, the individual loads on the various nodes may vary from one SPT to another, resulting in potentially different delay and delivery performance. Since it is not possible to evaluate all the possible SPTs in polynomial time, we find and evaluate only one of them, and move on to a higher power/edge length design in case of QoS not being satisfied; this may lead to suboptimal design. Moreover, for the same reason, the algorithm may not always return a feasible solution even when there exists one.
**Remarks:**
1. It is interesting to note that the above heuristic is basically the reverse procedure of the SPTiEP algorithm, with the additional step of evaluating each design using the analytical model. At the end of each iteration, the maximum edge length in the solution changes at most to the immediate higher value not used in previous iterations. We stop as soon as a QoS feasible solution is obtained in some iteration.
2. An alternate approach could be to start with a shortest path tree on the entire graph, and then prune edges, checking for feasibility with positive load in each iteration. But since meeting the lone-packet QoS is necessary to meet the positive-load QoS, we have adopted the above approach of first designing a network satisfying the lone-packet QoS, and then backtracking.
Numerical Experiments {#subsec:design-exp}
---------------------
----------- --------- -------- -------- --------- -------- -------- --------- -------- -------- --------- -------- --------
Scenarios
Average Max Min Average Max Min Average Max Min Average Max Min
30 0.0468 0.2532 0.0300 0.3827 0.5084 0.1609 2412.3 3000.6 2041.2 1765.7 2233.8 1326.6
----------- --------- -------- -------- --------- -------- -------- --------- -------- -------- --------- -------- --------
To demonstrate the strengths and limitations of the analysis in an iterative design process, we performed several test runs of the proposed algorithm on randomly generated network scenarios.
30 random networks were generated in a $50\times 50\:m^2$ area as follows: The entire area was partitioned into square cells of side 10 meters. Consider the lattice created by the corner points of the cells. 10 source nodes were placed at random over these lattice points. The minimum power level of the nodes is assumed to be sufficient so that all nodes are within CS range of one another, i.e., we have a “no hidden node” scenario. We chose the target link PER to be 0.01, $p_{del}=95\%$, and $d_{\max}=25\:msec$. These result in a hop constraint of $h_{\max}=5$ for the “lone packet” model. Also, the traffic arrival rate at each source was chosen to be $\lambda = 1$ packet/s, which is actually quite adequate for many wireless sensing applications.
The design algorithm using the analytical model was run on the 30 random scenarios. *In each case, the algorithm was found to return a feasible solution.*
Also, to validate the solution provided by the algorithm, we performed Qualnet simulation on both the outcome of the SPTiEP algorithm as well as the final solution for each of the 30 cases. The observations are as follows:
**Observations:**
1. In all 30 cases, simulations of the final solution showed QoS performance better than the target. That is, *the solution provided by the algorithm using the analytical model did actually satisfy the QoS constraints as verified by the simulation in all cases*.
2. In 12 out of the 30 cases, simulations suggested that the SPTiEP solution (“lone packet” design) met the QoS constraints even for the given $\lambda = 1$ packet/s (and hence was optimal), but the analytical model suggested otherwise. Thus, *the analytical model was somewhat conservative in its prediction of QoS*, which eventually led to a design that used more power than the optimal.
3. However, while the analytical model was conservative in its prediction of QoS, the final design met the QoS objectives and predicted the performance to within 10%., i.e., the model was *quite accurate* in its prediction.
4. Finally, we compared the time taken to run the design algorithm using the analytical model against the time taken to run the Qualnet simulations on the designed networks in the 30 cases. To save time, *in each case, we performed Qualnet simulations only for the lone packet design, and the final design obtained using E-SPTiEP algorithm*. While the algorithm was run in MATLAB 7.11 on a Windows Vista based Dell Inspiron 1525 laptop with 3 GB RAM, and 2 GHz Intel Core 2 Duo CPU, the simulations were run in Qualnet 4.5 on a Linux based Dell server with 32 GB main memory, and 3 Ghz clock speed. The results are summarized in Table \[tbl:exec\_time\_compare\].
From Table \[tbl:exec\_time\_compare\], we see that while *the execution time of the analysis based algorithm is of the order of milliseconds*, simulation of each of the designed networks takes several minutes, which indicates that *if the QoS evaluation step in the iterative design is performed using simulation instead of the analysis, the execution time of the algorithm can go well beyond an hour*.
From the above discussion, we can summarize the following strengths and limitation of the analytical model for use in a network design process.
1. **Strengths:**
- Design based on analysis is much *faster* compared to that based on network simulation, and the predictions are *accurate* to well within 10% compared to the simulation results. Such speed and accuracy also make the analytical tool a good choice for *online/field-interactive* design process, or even for as-you-go deployment of an impromptu wireless network.
- Designs provided by the analysis based algorithm do actually work, i.e., satisfy the QoS constraints in practice, as validated by simulations of the designed networks.
2. **Limitation:** The analysis, while quite accurate compared to the simulation, is somewhat *conservative in its prediction of QoS performance*. This may sometimes lead to *suboptimal* design, or even *declaration of infeasibility*, when actually there may exist a feasible solution.
Conclusion {#sec:conclusion}
==========
We have developed an approximate stochastic model for the performance analysis of beacon-less multihop wireless networks, with arrivals. The model permits the estimation of several time-average performance measures. Our model is accurate at small arrival rates (at which packet discard probability is small) in terms of the packet discard probability, failure probability and throughput, and delay. We calculated the mean end-to-end delays and packet delivery probabilities for each source in the network for specific packet generation rates at the source nodes. The results suggest that, for the relatively small size tree networks that we have studied, to operate in the low-discard low-delay region, the packet arrival rates at nodes should not be greater than a packet every few seconds (e.g. a packet inter-generation time of 5 to 10 seconds at the sources). Finally, we have formulated and solved a problem of QoS constrained network design for given positive traffic arrival rate to demonstrate the usefulness of the analytical model in an iterative network design process.
Appendix {#appendix .unnumbered}
========
Proof of Proposition \[prop:irreducible\]
-----------------------------------------
First observe that transition from any state to the all-zero state is possible if we consider a series of “only down” transitions, i.e., no further external arrival occurs into any of the queues, and the packets leave the queues either due to successful transmission, or drop due to excessive retries or CCA failures. *No external arrival* would mean that the number of packets in the system can only decrease in each transition, and eventually we will reach the all-zero state. Note that this is possible since the maximum number of retransmission attempts, and maximum allowed CCA failures are *finite*.
Now, suppose the class is open. Then, there exists a state $i\in \mathcal{C}_{0}$, and a state $j\notin \mathcal{C}_{0}$ such that $i\rightarrow j$, but $j\nrightarrow i$. But from our earlier argument, we know, state $j$ can reach the all-zero state. Since both state $i$ and the all-zero state belong to the class $\mathcal{C}_{0}$, the all-zero state can reach state $i$. It follows that $j\rightarrow i$, which is a contradiction. Hence, the class $\mathcal{C}_{0}$ is closed.
Clearly, $\mathcal{C}_{0}$ is aperiodic since starting in the all-zero state, the system can remain in the all-zero state if no external arrivals occur in a slot (since the arrival process is Poisson, this event has a positive probability).
Proof of Theorem \[thm:stability\]
----------------------------------
For any $n>0$, $$\begin{aligned}
{\mathsf{Pr}}[\cap_{i=1}^N\{X_i(n)=0\}|{\mathbf{X}}(0)={\mathbf{0}},{\mathbf{Y}}(0)={\mathbf{0}}]\nonumber\\
= 1\:-\:{\mathsf{Pr}}[\cup_{i=1}^N\{X_i(n) > 0\}|{\mathbf{X}}(0)={\mathbf{0}},{\mathbf{Y}}(0)={\mathbf{0}}]\nonumber\\
\geq 1\:-\:\sum_{i=1}^N {\mathsf{Pr}}[X_i(n) > 0|{\mathbf{X}}(0)={\mathbf{0}},{\mathbf{Y}}(0)={\mathbf{0}}]\label{eqn:union-bnd}\end{aligned}$$ where, in writing , we have used the union bound.
Taking limit as $n\to \infty$ on both sides,
$$\begin{aligned}
\lim_{n\to \infty}{\mathsf{Pr}}[\cap_{i=1}^N\{X_i(n)=0\}|{\mathbf{X}}(0)={\mathbf{0}},{\mathbf{Y}}(0)={\mathbf{0}}]\nonumber\\
\geq 1\:-\:\sum_{i=1}^N q_i\nonumber\\
> 0,\:\text{when $\sum_{i=1}^N q_i\:<\:1$}\nonumber\end{aligned}$$
Thus, if $\sum_{i=1}^N q_i<1$, we have that $\lim_{n\to \infty}{\mathsf{Pr}}[\cap_{i=1}^N\{X_i(n)=0\}|{\mathbf{X}}(0)={\mathbf{0}},{\mathbf{Y}}(0)={\mathbf{0}}]>0$, which, in conjunction with Proposition \[prop:irreducible\], implies that $({\mathbf{X}}(n),{\mathbf{Y}}(n))$ is positive recurrent.
Proof of Proposition \[prop:sptiep-correctness\]
------------------------------------------------
Since $\mathcal{F}^{(k)}\:\neq \emptyset$, $$\label{eqn:leq}
w_{opt}\leq\overline{w}^{(k)}$$ This is because $\overline{w}^{(k)}$ is the maximum edge length of $T^{(k)}$, a feasible solution, while $w_{opt}$ is the maximum edge length of an optimal solution.
Again, $\mathcal{F}^{(k+1)}\:=\emptyset\:\Rightarrow\:\nexists\: T\subset G^{(k+1)}$ such that $T$ satisfies hop constraint.
But by construction (Step 3 of SPTiEP algorithm), $G^{(k+1)}$ contains in it, all trees with maximum edge length $<\:\overline{w}^{(k)}$. Therefore, no tree with maximum edge length $<\:\overline{w}^{(k)}$ is a feasible solution to the MMST-HC problem. Hence, $$\label{eqn:notless}
w_{opt}\:\nless \overline{w}^{(k)}$$ Combining and , it follows that $w_{opt}=\overline{w}^{(k)}$.
Acknowledgements {#acknowledgements .unnumbered}
================
This work was supported by a research grant from the Department of Electronics and Information Technology (DeitY), Government of India, through the Automation Systems Technology (ASTEC) program, and by a DeitY-NSF funded Indo-US project on Wireless Sensor Networks for Protecting Wildlife and Humans.
[^1]: Note, however, that there is no central coordinator that maintains synchrony of the slots *across the nodes*, and hence the CSMA/CA protocol is unslotted.
[^2]: A node $j$ is said to be within the Carrier Sense (CS) range of a node $i$ if any channel activity (transmission) due to node $j$ can be detected (but not necessarily decoded) by node $i$.
[^3]: Note that in a CSMA/CA network, all durations (e.g., backoff, transmission, CCA) are multiples of the symbol time
[^4]: under the condition of aperiodicity, this limit always exists
[^5]: A formal proof is necessary for this seemingly obvious statement since, in CSMA/CA networks, in general, the performance is not monotone with the arrival rates (see, e.g., [@borst]); hence, the statement about the lone-packet traffic model needs to be made with care.
[^6]: for the purposes of this design, we completely trust the outcome of the analytical model, i.e., we assume the analytical model to be 100% accurate.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a comparative study of the (magneto)transport properties, including Hall effect, of bulk, thin film and nanostructured MnSi. In order to set our results in relation to published data we extensively characterize our materials, this way establishing a comparatively good sample quality. Our analysis reveals that in particular for thin film and nanostructured material, there are extrinsic and intrinsic contributions to the electronic transport properties, which by modeling the data we separate out. Finally, we discuss our Hall effect data of nanostructured MnSi under consideration of the extrinsic contributions and with respect to the question of the detection of a topological Hall effect in a skyrmionic phase.'
address:
- '$^{1}$Institut für Physik der Kondensierten Materie, Technische Universität Braunschweig, D-38106 Braunschweig, Germany\'
- '$^{2}$Physikalisch-Technische Bundesanstalt, D-38116 Braunschweig, Germany'
author:
- 'D Schroeter$^{1}$, N Steinki$^{1}$, M Schilling$^{1}$, A Fernández Scarioni$^{2}$, P Krzysteczko$^{2}$, T Dziomba$^{2}$, H W Schumacher$^{2}$, D Menzel$^{1}$ and S Süllow$^{1}$'
title: 'MnSi-nanostructures obtained from thin films: magnetotransport and Hall effect'
---
Introduction
============
In recent years, [*skyrmions*]{}, [*i.e.*]{}, a unique form of complex magnetic spin texture, have evolved as a research topic of prime interest both for basic research as from the perspective of possible applications [@muehlbauer2009; @krause2016]. With respect to the latter issue, heterostructures carrying skyrmionic spin textures hold the prospect of being used for new types of data storage devices [@fert2013; @tomasello2014]. Conversely, the former aspect bears relevance in the context of a multitude of fundamental topics such as topologically protected states, spin-orbit effects in magnetic materials or the collective dynamics of the skyrmionic phase [@neubauer2009; @schuette2014; @schwarze2015].
The basic structural element required to stabilize skyrmions is a lack of inversion symmetry in the magnetic lattice, leading to a spin-orbit controlled type of magnetic exchange, the Dzyaloshinskii-Moriya (DM) interaction. For two spins $S_i$, $S_j$, in contrast to the ordinary Heisenberg exchange, the DM interaction results in an energetic term $\propto S_i \times S_j$. If combined with the Heisenberg exchange, and in a Ginzburg-Landau type approach, it has been demonstrated that a skyrmionic phase can be stabilized under these conditions [@muehlbauer2009]. In turn, with respect to materials exhibiting skyrmionic phases, the occurrence of a DM interaction can be attributed either to crystal structures lacking inversion symmetry (as for instance in the B20 compounds [@pfleiderer2010]), or to the symmetry-breaking at surfaces [@bode2007; @heinze2011]. Correspondingly, in terms of the appearance of skyrmionic phases these are either surface induced thin film or material intrinsic skyrmions that are studied.
Belonging to the latter class, the cubic helimagnetic B20 compound MnSi represents probably the most iconic system exhibiting a skyrmionic phase. The material being known for decades (see Ref. [@pfleiderer2010] for a review) was originally studied in the context of spin fluctuation theory [@moriya1985]. Later, the pressure induced suppression of helical magnetic order (helix length $\sim 19$nm, ordered magnetic moment $\sim 0.4 \mu_B$ at ambient pressure) became the focus of studies in the context of quantum criticality in itinerant $d$-metals [@pfleiderer2010; @thompson1989]. Finally, it was noted that the early-reported field-induced [*A-*]{}phase in MnSi [@sakakibara1982] does represent a skyrmion phase [@muehlbauer2009], this way establishing the material as the model compound for studies of skyrmion physics.
Subsequently, one avenue that was followed in investigations on MnSi was the production of thin film material [@karhu2010; @karhu2011; @karhu2012; @geisler2012; @wilson2012; @engelke2012; @li2013; @suzuki2013; @menzel2013; @wilson2013; @yokouchi2014; @wilson2014; @engelke2014; @meynell2014a; @meynell2014b; @yokouchi2015; @lancaster2016]. Conceptually, it was argued that in thin films the skyrmionic phase ought to be stabilized [@butenko2010; @banerjee2014; @rowland2016], and thus it should be possible to better control and investigate skyrmionic behavior. However, in spite of these extensive studies, as yet, there is no consensus regarding the issue if in thin film material of MnSi a skyrmionic phase exists in a similar form as in bulk material.
Initial studies established that MnSi thin films do show helical magnetic order, although for films thicker than $\sim 5$nm at temperatures significantly higher ($T_C$ up to $\sim 48$K) than in bulk material ($T_C = 29.5$K). This effect is attributed to tensile strain induced in MnSi thin films by the lattice mismatch with the Si substrate, effectively leading to a state of negative pressure [@karhu2010; @engelke2012; @engelke2014]. As well, the suppression of the ordering temperature for films with less than 5nm thickness can be understood in terms of the reduction of spin-spin interactions caused by the interface [@engelke2012].
For films thicker than the helix length of bulk material there is evidence that the magnetic ground state is essentially the same as in the bulk, be it that uniaxial anisotropies induced by thin film strain need to be taken into account [@karhu2010; @karhu2011; @karhu2012; @engelke2012; @li2013; @wilson2013; @engelke2014]. For the in-field behavior, however, it turned out that the film anisotropies affect the material properties quite significantly, leading to a variety of proposed magnetic phase diagrams [@wilson2012; @li2013; @menzel2013; @yokouchi2014; @lancaster2016]. Most notably, while for single crystals MnSi that have been mechanically thinned to a few 10nm thickness, enabling studies by Lorentz microscopy [@tonomura2012; @mochizuki2014; @yu2015], there is direct evidence for a skyrmionic in-field phase, as yet there is no conclusive evidence for such a phase in MnSi thin films.
In this situation, an attempt has been made to identify the skyrmion phase in MnSi thin films by a unique feature of the skyrmionic state, the [*topological Hall effect*]{} [@neubauer2009; @li2013; @yokouchi2014; @meynell2014b; @yokouchi2015]. Experimentally, it requires accurate measurements of the Hall effect of the system studied, as the topological Hall effect in bulk material represents a minor additional Hall contribution aside from normal and anomalous Hall effect. Surprisingly, even though the studies on thin film material in the Refs. [@li2013; @yokouchi2014; @meynell2014b; @yokouchi2015] appear to verify that in terms of the general magnetic behavior the different MnSi thin films behave in a similar way, the measurements of the Hall effect in these studies exhibit widely varying experimental features.
As a reference, the analysis to extract the topological Hall effect in bulk MnSi has been performed on data taken for high quality single crystalline plates of about 100$\mu$m thickness (room temperature resistivity $\rho_{xx} = 180$$\mu \Omega$cm, RRR$\sim$100) [@neubauer2009]. The Hall effect and magnetoresistivity have been measured in a standard 6-point configuration up to 9T, with the signal being symmetric/antisymmetric in a magnetic field corresponding to the magnetoresistivity/Hall effect. Overall, the magnitude of the Hall signal $\rho_{yx}$ varies between -150 and 200n$\Omega$cm, and its field dependence is dominated by the normal contribution and an anomalous contribution reflecting primarily the magnetization of the sample. After correction for the normal Hall contribution (carrier density $\sim 4 \times 10^{22}$cm$^{-3}$) an anomaly in the field dependence of the anomalous Hall contribution is observed. This anomaly, with a magnitude of about 5n$\Omega$cm on the background of $\sim 50$n$\Omega$cm for the ordinary anomalous Hall contribution, has been demonstrated to map the [*A-*]{}phase of MnSi.
In comparison, in Table \[tab:comp\] we summarize basic physical parameters for different MnSi thin films reported in literature. As noted before, for a film thickness above 5nm an enhanced ordering temperature $\sim$45K is observed. However, already the absolute values of the resistivity vary significantly, even though they are in the range of single crystalline material. Clearly, disorder in the films is an issue, as exemplified by residual resistivity ratios significantly smaller than for the best single crystals, and varying from sample to sample. As well, sample dependencies become very apparent if the Hall resistivities are considered. As indicated by the values reported at 20K in 1 and 5T, for the different samples they vary in magnitude and even by sign.
Ref. [@engelke2012] [@li2013] [@li2013] [@yokouchi2014] [@meynell2014b] [@yokouchi2015]
------------------------------------------------------------------------- ---------------- ----------- ----------- ----------------- ----------------- -----------------
thickness (nm) 19 10 50 20 25.4 26
$T_C$ (K) 45 45 45 45 42 48
$\rho_{xx: 80 \textrm{K}} (\mu \Omega \textrm{cm}$) 99 165 – 120 113 –
$\rho_{xx: 80 \textrm{K}} / \rho_{xx: 2 \textrm{K}}$ 11 7 – 12 14 –
$\rho_{yx: 20 \textrm{K}/1 \textrm{T}} (\textrm{n} \Omega \textrm{cm}$) – 5 -10 -10 10 –
$\rho_{yx: 20 \textrm{K}/5 \textrm{T}} (\textrm{n} \Omega \textrm{cm}$) – 65 60 40 50 –
$n (10^{22}$cm$^{-3}$) – 3.5 3.5 8.8 4.3 –
: List of main physical characteristics of MnSi thin film samples reported in the literature: ordering temperature $T_C$, resistivity $\rho_{xx: 80 \textrm{K}}$ at 80K, resistivity ratio $\rho_{xx: 80 \textrm{K}} / \rho_{xx: 2 \textrm{K}}$, Hall resistivity $\rho_{yx: 20 \textrm{K}/x \textrm{T}}$ in 1 and 5T, carrier density $n$ extracted from the normal Hall contribution.[]{data-label="tab:comp"}
To extract the topological contribution to the Hall effect from the experimental data, the conceptual idea is to separate the Hall resistivity into three parts: $$\rho_{yx} = \rho_{yx}^N + \rho_{yx}^A + \rho_{yx}^T \label{Hall}$$ Here, $\rho_{yx}^N$ denotes the normal Hall contribution ([*i.e.*]{}, the carrier density dependent term), $\rho_{yx}^A$ the anomalous part proportional to the magnetization and a resistivity-dependent factor, while $\rho_{yx}^T$ represents the topological Hall effect. Given that $\rho_{yx}^N$ is roughly the same for all samples and varies linearily with field, the difficulties in extracting a topological Hall contribution is directly related to the accurate determination of $\rho_{yx}^A$. From the experimental data, however, it is apparent that $\rho_{yx}^A + \rho_{yx}^T$ is not well-controlled, a point that has been made previously in an attempt to more accurately parametrize the Hall resistivity [@meynell2014b].
In this situation, we have set out to reinvestigate the Hall effect in MnSi thin films. We do so by 1.) carefully comparing bulk and thin film data and 2.) nanostructuring MnSi thin films into Hall bar geometries for optimization of the experimental set-up. We analyze the electronic transport properties of different nanostructures MnSi, this way assessing which contributions to the resistive and Hall behavior of thin film material are intrinsic or extrinsic, respectively.
The paper is organized as follows: First, we report the sample preparation and experimental analysis to characterize our MnSi thin films, this in order to compare our samples to thin film samples from previous reports. Next, we describe our steps to nanostructure the thin films and document our resulting Hall bar structures. Subsequently, we characterize our structures regarding their electronic transport properties, [*i.e.*]{}, the resistivity, magnetoresistivity and Hall effect. Finally, we discuss our findings on nanostructured MnSi, this in particular in comparison to thin film and single crystalline material and with respect to the issue of the existence of a skyrmion phase in these samples.
Sample preparation {#sample}
==================
MnSi thin film samples were prepared by molecular beam epitaxy (MBE) in ultra-high vacuum with a base pressure of below 5$\times$10$^{-11}$ mbar. Essentially, the procedure has been described previously [@engelke2012], but was carried out here with some alterations. $P$-doped Si(111) wafers with a size of $10 \times 10 \times 0.28$mm$^3$ and a resistivity of $1 - 10$$\Omega$cm serve as substrates. After a standard cleaning process the substrates were loaded into the MBE chamber, where they were degassed at 750$^\circ$C for two hours.
To remove the native silicon-oxide layer the temperature was then raised at a rate of 1$^\circ$C$/$s up to 1150$^\circ$C were the system was held for 10 min after which the substrate was cooled back to room temperature at a rate of less than 1$^\circ$C$/$s. After this processing step a $7 \times 7$ reconstruction of the Si surface can be clearly seen in reflection high energy electron diffraction (RHEED), confirming the high crystalline quality of the substrate. Subsequently a seed layer for the MnSi films was formed by evaporating 1nm Mn from a Knudsen cell onto the substrate, which is held at a temperature of 180$^\circ$C. As the next step, the substrate is heated to 300$^\circ$C, at which an epitaxial MnSi seed layer is formed. In our experiments, a series of different Mn layer thicknesses between 0.5 and 4nm were tested, with a layer thickness of 1nm showing the clearest $\sqrt{3} \times \sqrt{3}$ pattern visible in RHEED.
Following the growth of the seed layer, MnSi thin films of various thicknesses were grown by simultaneous deposition of Mn and Si in a stoichiometric ratio, with Si evaporated using an electron beam evaporator. The overall deposition rate was 0.02nm$/$s. After growing the MnSi layer the sample is annealed at 350$^\circ$C for 1 hour, always showing the characteristic $\sqrt{3} \times \sqrt{3}$ RHEED pattern. The single phase nature of the samples annealed at 350$^\circ$C were checked by x-ray diffraction, and revealed no detectable impurity phase. As reported in Ref. [@karhu2010], we observe that annealing at higher temperatures than 400$^\circ$C tends to produce MnSi$_{1.7}$ precipitates. Altogether, following above recipe we produced films of a nominal thickness of 10 and 30nm. Below, we present measurements of the electronic transport properties of the 30nm thick samples.
Surface analysis {#surface}
================
In our earlier studies on the (magneto)transport and magnetization of MnSi thin films [@engelke2012; @menzel2013; @engelke2014] we have noticed that the smoothness of the surface is decisive for the sample quality. So far, no roughness data of the film surface has yet been published. Earlier investigations on film morphology have probed only the MnSi/Si interface via x-ray reflectometry, and which has been reported to be of the order of 1nm [@karhu2010; @karhu2011]. Therefore, in order to assess the quality of our thin film samples, we have carried out atomic force microscopy (AFM) measurements on various of our 30nm thick samples with a SIS Nanostation II non-contact AFM system (by Surface Imaging Systems SIS, Germany; now known as N8 NEOS by Bruker), using a PPP-NCLR cantilever (Nanosensors) with scans ranging from 100$\mu$m$\times$100$\mu$m down to 3$\mu$m$\times$3$\mu$m recorded with 1024$\times$1024 data points each.
In the Figs. \[afmglobal\] and \[afmzoom\] we display the results of AFM measurements on two nominally 30nm thick films. Overall, the AFM images show a flat closed MnSi layer with a very uniform surface, with Sq values (RMS roughness) well below 3nm. A major contribution to the roughness is made by needle-like features of 10 to 20nm height arranged in three discrete orientations (see white needles in Figs. \[afmglobal\] and \[afmzoom\]), and some grains (larger white dots in Figs. \[afmglobal\] and \[afmzoom\]) mainly of some ten nanometers height. Needles and grains occupy less then 1% of the surface, and the remaining 99% of the surface shows Sq values of about 1nm only. Considering that the film surface roughness increases during the growth process, the low roughness measured here indicates a good film quality.
Throughout the AFM measurements, also the phase signal of the AFM was monitored in order to check for contamination with other materials, but the phase images show very small contrasts only. These should be attributed to topographic features rather than secondary phase materials. In particular, even the pronounced white structures (needles and grains in Fig. \[afmglobal\]) show only topological features in the phase signal, indicating that in fact we are dealing with a single phase MnSi film. Conversely, considering the well defined topology of the needles and grains, we assume that these are whisker crystals MnSi growing out of the thin films.
Lithography
===========
One experimental issue while measuring the Hall effect is to place the Hall voltage contacts directly opposite to each other on the sample. In reality, imperfections in this step always lead to a magnetoresistive component in the measured signal, effectively reducing the signal-to-noise ratio for the intrinsic Hall signal. Therefore, in order to minimize the magnetoresistive contribution, and in perspective to allow studies on the controlled movement of skyrmions, we have set out to nanostructure our MnSi thin film material by means of electron beam lithography. In a first step, we have produced Hall bar structures, allowing simultaneously to measure the resistivity, magnetoresistivity and Hall effect on our samples.
Standard electron beam lithography in combination with Ar-ion etching was used to pattern the 30nm thick MnSi films into Hall bar geometry with a measured width (by means of scanning electron microscopy (SEM), see below) of the current path ranging from 175nm to 10$\mu$m. The basic design for most of our samples consists of three crossings (Fig. \[rem\] (b)). The samples have been prepared by using the electron beam resist with an additional precleaning step by sonicating the sample in , which significantly improves the adhesion between the sample and the resist.
In the next lithography step platinum (Pt) contacts to the MnSi structure were fabricated by sputtering deposition. The Pt leads are 100nm thick, including a 5nm Ta adhesion layer. The MnSi surface was cleaned in situ beforehand by low energy (0.7keV) Ar-ions to ensure good electrical contact. This step does not influence the magnetic properties as verified by magnetization measurements (not shown). Moreover, simulations of Ar-ion implanting in our films indicate that the low-energy Ar-ions will penetrate the surface by roughly 1nm. We thus conclude that the Ar-etching does not affect the properties of our structures.
As results of these different processing steps we have obtained various Hall bar structures from MnSi thin films. For final characterization, we have carried out scanning electron microscopy on all nanostructures reported on here, with the example of the 175nm structure depicted in Fig. \[rem\](b). The structure contact pads were bonded with aluminum wires to the sample carrier, which was attached to the sample holder of our cryogenic system for measurements of (magneto)resistivity and Hall effect.
Experimental
============
To characterize the MnSi structures we have measured the temperature and field dependent electronic transport properties in a $^4$He cryostat via standard [*ac*]{}, [*dc*]{} four-point probe and van der Pauw configuration in the temperature range 2–300K in fields up to 5T. The magnetic field was applied perpendicular to the film surface (see Fig. \[rem\](a)).
We have carried out these experiments for different structure sizes with a current path width between 10$\mu$m and 175nm. Effectively, with the [*ac*]{} excitation voltage from a low-power resistance bridge, we use measurement currents in the range of 10nA up to 1$\mu$A, which corresponds to current densities of $1\cdot 10^{6}\,$A/m$^2$ for $\mu$m sized structures and up to $2\cdot10^{7}\,$A/m$^2$ for the 175nm structures. For the [*dc*]{} measurements (performed on the 200nm and 230nm structures) a current of 100$\mu$A was used which corresponds to current densities of roughly $2\cdot10^{10}\,$A/m$^2$. In terms of skyrmionic behavior, if skyrmions are present in our samples, already the [*ac*]{} currents could possibly be large enough to electrically drive such skyrmions through our structures [@jonietz2010].
To directly compare data of MnSi single crystals, thin films and patterned structures, in addition to the nanostructured material, we have measured on a 30nm MnSi thin film without lithographical treatment the temperature dependent (magneto)resistivity (current density of $2\cdot 10^{6}\, $A/m$^2$) and Hall effect (current density of $4\cdot 10^{7}\, $A/m$^2$). Furthermore we have determined the corresponding properties of a 2mm$\times$1mm$\times$0.1mm single crystalline sample which was synthesized by tri-arc Czochralski growth.
Geometrical factors required to obtain absolute resistivity values for our nanostructures were determined by the analysis of the scanning electron microscopy pictures. This way, the length $l$ probed in the resistance measurements of our nanostructured samples can be determined easily to an accuracy of about one percent. As well, the width of the current path is a well-controlled quantity, with an uncertainty (depending on structure size) of a few percent. The largest error margin comes from the film thickness, with an uncertainty of about 10% between nominal and real thickness. Only, this error does not affect our comparison of thin film and nanostructured samples, as the latter ones are produced from the films. Altogether, the geometry induced uncertainty of absolute resistivity values is of the order of 10%, the relative error margin in sample-to-sample comparison in the percentage range.
Results
=======
Resistivity {#elcharacter}
-----------
In Fig. \[substrate\] we summarize the resistivities of different MnSi samples, ranging from bulk material to nanostructured thin films (thickness 30nm). For bulk material, we reproduce the essential findings from literature [@pfleiderer2010; @kadowaki1982], that is a resistivity from an itinerant $d$-metal magnetic system, with a room temperature resistivity of about 120$\mu \Omega$cm (Fig. \[substrate\]). A residual resistivity ratio $\rho_{xx: 300 \textrm{K}} / \rho_{xx: 2 \textrm{K}}$ of 16 signals a decent crystalline quality of our specimen.
For a thin film sample that was contacted by Cu wires with silver epoxy to the surface of the film, again we reproduce the resistivity reported previously [@engelke2012; @li2013; @meynell2014b]. Overall, there is a metallic resistivity with a kink-like feature denoting the magnetic transition at $T_C$. A residual resistivity $\rho_{xx: 2 \textrm{K}} = 20 \mu \Omega$cm and a resistivity ratio $\rho_{xx: 80 \textrm{K}} / \rho_{xx: 2 \textrm{K}} = 8$ fully agrees with literature values.
For thin films nanostructured in Hall bar geometry on $P$-doped Si(111)-substrates, in addition to the resistive behavior typical for MnSi film there is a downturn of $\rho_{xx}$ at high temperatures (Fig. \[substrate\]). If we specifically consider the largest structures, say the structure with a lateral extent of 10$\mu$m structure width and 30$\mu$m structure length, the obvious expectation is that at this structure size there should be no size effect, [*i.e.*]{}, the behavior of the MnSi structure and the film should be identical. Indeed, the direct comparison of the data for a 10$\mu$m structure and the thin film reveals the resistivity of film and structure to be almost the same up to 230K. But then, the downturn at higher $T$ must reflect the influence of the substrate and the contact pads on the measurement, effectively providing a resistive path parallel to the MnSi structure.
To demonstrate this, we have modeled our measured signal, assuming that the total resistance $R_{total}$ of nanostructured MnSi (current path width 10$\mu$m) on Si-substrate can be understood as a set of parallel resistors (Fig. \[modeling\]): $$R_{total} = \frac{R_{MnSi} \cdot R_{sub}}{R_{MnSi} + R_{sub}}.$$ Here, we use the total resistance of the substrate $R_{sub}$ which was measured using the contact pads of two different and separated MnSi nanostructures, displaying an exponential temperature dependence. In addition, we use the temperature dependent part of the experimentally determined resistivity $\rho_{exp}(T)$ of thin film MnSi to calculate $R_{MnSi} = (l/a) \rho_{exp}(T)$ for the nanostructure. Here, we use the MnSi structure length $l$ and the cross-section $a$. In Fig. \[modeling\] we demonstrate that with these simple assumptions we can model the temperature dependence of the resistance of a Si-substrate chip carrying a Hall-bar structure MnSi of 10$\mu$m width by comparing the experimental data with the calculated resistance following the parallel resistor approach. It proves that because of the strong temperature dependence of the resistance of the substrate, at high temperatures of roughly $T>$200K the resistivity of the substrate increasingly affects the overall measurement.
The resistance $R_{sub}$, which exponentially increases with decreasing temperature, does not reflect the intrinsic resistive behavior of a $P$-doped Si(111) wafer. Instead, it appears that the resistive contribution short circuiting the MnSi structure at high temperatures is related to a Schottky barrier at the interface between the contact pads and the Si wafer. This is indicated by the current-voltage characteristic of the contact pads connected to MnSi nanostructures (see inset of Fig.\[modeling\]). For the measurement plotted here (at a temperature of 300K), two pads on one MnSi structure were used as $I$ and $V$ contacts, while for the second $I/V$ contacts the pads on a second MnSi structure on the same wafer were used (distance between structures $\sim 80$$\mu$m). In this configuration we thus probe the resistance through the contacts and substrate. The measured $I-V$ characteristic is typical for a Schottky barrier. Moreover, as we lower the temperature, the non-conducting voltage range and the resistance increases, reflecting an increasingly insulating resistance path through the wafer.
Effectively, our finding implies that for the nanostructured samples MnSi there are Schottky barriers between the contact pads and the Si wafer. At high temperatures, these short-cut the nanostructures, while at lower temperatures the barrier becomes impassable, in result decoupling the MnSi nanostructures from the substrate. The quantitative agreement between experimental data and our parallel resistor model proves that at high temperatures (above $\sim 200$K) the measured resistance of the nanostructures is significantly affected by the substrate/contact pads. Conversely, with the exponentially increasing resistivity of the substrate/contact pads for lower temperatures, in the temperature range considered below, $T < 80$K, an influence of the substrate on $\rho_{xx}(T)$ of MnSi can be neglected.
Having thus identified the different contributions to the resistive behavior, in Fig. \[resistivity\] we summarize the resistivities of different samples MnSi, ranging from bulk material and unpatterned films to nanostructured films, at temperatures $T < 80$K. In addition, in Tab. \[tab:samples\] we list the essential parameters obtained from these measurements, [*i.e.*]{}, transition temperatures, resistive coefficients, residual resistivities and resistivity ratios.
Sample bulk film 10$\mu$m 1$\mu$m 300nm 230nm 175nm
------------------------------------------------------ ------- ------- ---------- --------- ------- ---------- -------
$T_C$ (K) 29.5 48 48 47 45 40 44
$A$ ($\mu \Omega$cm/K$^2)$ 0.026 0.046 0.046 0.053 0.034 0.122 0.070
$\rho_{xx: 2 \textrm{K}} (\mu \Omega \textrm{cm}$) 7.5 20 25 60 82 102 258
$\rho_{xx: 80 \textrm{K}} / \rho_{xx: 2 \textrm{K}}$ 7.9 8.0 6.2 3.0 2.0 $\sim$ 3 1.6
: List of main characteristics of various MnSi samples from resistivity measurements: ordering temperature $T_C$, resistive coefficient $A$, resistivity $\rho_{xx: 2 \textrm{K}}$ at 2K, resistivity ratio $\rho_{xx: 80 \textrm{K}} / \rho_{xx: 2 \textrm{K}}$; for details see text [@note2].[]{data-label="tab:samples"}
For bulk material, a kink in the resistivity reflects a magnetic transition with an ordering temperature $T_C = 29$K, consistent with previous reports [@pfleiderer2010; @kadowaki1982]. For thin film material, with the measurement of the resistivity we again reproduce the behavior reported in literature [@engelke2012; @li2013; @yokouchi2014; @meynell2014b; @yokouchi2015]. The residual resistivity is significantly larger than that of bulk material, be it that the resistivity ratio is also slightly larger, indicating a reasonable crystalline quality of the thin film material. Overall, the values we obtain for these quantities for our samples are comparable to those reported in literature (see Tab.\[tab:comp\]), implying an overall similar film quality. As noted before, the magnetic transition temperature is enhanced as compared to bulk, reflecting the tensile strain/effective negative pressure.
The next processing step, nanostructuring the thin films, does not affect the overall appearance of the resistivity $\rho_{xx}$. All nanostructured samples exhibit a kink in the resistivity at $T_C \sim 44 - 48$K (for the 230nm sample, see Ref. [@note2]), denoting the magnetic transition, as in thin film material. As well, below $T_C$ the resistivity evolves as $\rho_{xx} = \rho_{0} + A T^2$. We have included the values $A$ from fits to the data below $T_C$ in Tab. \[tab:samples\], yielding values slightly larger than for bulk material for all samples, and consistent with the thin film result.
Surprisingly, nanostructuring the thin film material affects a characteristic parameter of the electronic transport, [*i.e.*]{}, the residual resistivity increases with smaller nanostructure size. While down to a structure width of 300nm the resistivities are comparatively large, but still in a metallic range, for our smallest structures we find (residual) resistivities in the range of a few hundred $\mu \Omega \textrm{cm}$. Correspondingly, the resistivity ratio decreases with nanostructure size, reflecting the increase of $\rho_{xx: 2 \textrm{K}}$.
Assuming a homogeneous current path in the sample, resistivities above the Mooji rule [@mooji; @tsuei] ($\sim 200$$\mu \Omega$cm) are typically accounted for in terms of either a semi-metallic system or disorder-induced localization. Clearly, the second scenario of disorder-induced localization is not consistent with observations, as in this case (a tendency towards) a negative resistive coefficient $d \rho/d T$ should occur. As well, a transition of MnSi into an intrinsically semi-metallic state induced by nanostructuring appears unlikely, as the magnetic behavior and the character of the temperature dependent resistivity is not affected by the structuring.
But then, in order to account for the large resistivity values in our smallest structure MnSi, either the concept of a homogeneous current path in our nanostructure is not fulfilled, or we are dealing with interfacial effects of an unknown nature in a system consisting of Si and the correlated electron material MnSi.
Magnetoresistance
-----------------
To further characterize our nanostructure samples MnSi we have studied the magnetotransport properties. As an example, in Fig. \[magnetores\] we plot the transverse magnetoresistivity $(\rho_{xx}(B) - \rho_{xx}(B=0))/\rho_{xx}(B=0)$ of a MnSi structure (10$\mu$m current path width). Above $T_C$, the magnetoresistivity is negative and evolves $\propto B^2$, analogous to the findings in single crystalline material [@sakakibara1982; @kadowaki1982] and in agreement with previous thin film studies. Such behavior has been interpreted in terms of spin fluctuation theory [@ueda1976].
As the temperature is lowered below $T_C$, the shape of the magnetoresistivity qualitatively transforms into an inverted S-shape. Moreover, in the temperature range $\sim 10$K a small kink in $\rho(B)$ appears near 1.2T. Such a feature has been seen in previous measurements on MnSi thin film [@menzel2013] and been associated to the phase transition from the magnetically ordered into the ferromagnetically polarized state at $B_{C}$ [@kadowaki1982]. Weak additional kinks in the magnetoresistivity reported previously below $B_{C}$ and tentatively also associated to skyrmion phase formation [@meynell2014b] were not observed in our studies.
The magnetoresistive behavior reported here for the 10$\mu$m structure is similarily seen for all the other structures. This is illustrated in Fig. \[magnetorescomp\], where we plot the magnetoresistivity of different nanostructured samples MnSi at a reduced temperature $T/T_C \sim 0.43$ as function of the reduced magnetic field $B/B_{C}$, and scaled to a value of the magnetoresistivity MR $:= -1$ at $3 \times B_C$. Clearly, for all samples there is a close resemblance in the field evolution, and the features possibly denoting phase transitions coincide.
Altogether, in terms of the resistivity and magnetoresistivity, with our measurements on nanostructured MnSi samples we qualitatively and quantitatively reproduce the findings previously reported on thin film material. Correspondingly, we can safely assume that the physical behavior observed for our samples may be compared to that reported previously for thin film material. Conversely, we still need to explain the enhancement of the residual resistivity for our nanostructured samples.
Hall effect {#hallsection}
-----------
Finally, we have performed Hall effect measurements on our various nanostructured samples MnSi using an out-of-plane magnetic field $B$, the current $I$ along the structure, and the Hall voltage $U_H$ measured $\perp I$ and $\perp B$ (see Fig. \[rem\](a)). In addition, in order to allow for a direct comparison to bulk material MnSi, we have carried out corresponding measurements using the same experimental set-up on the single crystal, for which we have reported the (magneto)resistivity. As well, we have measured the Hall effect for our thin film sample MnSi.
Surprisingly, for our Hall bar structures, in our initial experiments we find - in addition to a Hall voltage $U_H$ - a significant and sample dependent magnetoresistive contribution $U_R$. Following these observations, we have simplified our experimental geometry by producing a single Hall cross (see SEM-picture in the inset of Fig. \[cross\]; current path width 290nm) and measured the cross voltage $U_{cross}$ in zero magnetic field for this structure (Fig. \[cross\]).
In a Hall effect configuration the experimentally measured zero-field offset and magnetoresistive contribution is caused by the residual geometrical offset of the voltage leads. For our configuration, clearly, this offset is technically minimized to a writing mismatch from the e-beam lithography ([*i.e.*]{}, nanometer range), and which should produce only a very small resistive signal. Still, experimentally we do observe a large temperature dependent cross voltage that clearly resembles the resistance measurements presented above. By comparing the absolute resistance scale of the single cross seen in Fig. \[cross\] with measurements performed on a Hall bar structure with comparable current path width between two neighboring voltage taps (single cross current path width $\sim 290$nm, Hall bar structure width $\sim 300$nm and length of $\sim 2.5$$\mu$m, same unpatterned film origin) we see a resistive signal of the same order, even though we are now measuring the voltage drop on nominally equipotential voltage leads.
To quantify the anomalously large cross voltage/resistance for this particular nanostructure, we can estimate the geometrical shift of the voltage leads with respect to each other that would be required to account for the observations. With the measured cross resistance $R_{cross}$ of the structure with a current path width $w = 290$nm, the resistivity at 5K, $\rho_{300 {\rm nm}} = 82 \mu \Omega$cm, and film thickness $d = 30$nm, we can calculate the geometrical shift $x$ as $$x = \frac{R_{cross}\cdot d \cdot w}{\rho_{300 {\rm nm}}}.$$ This way, we find a value $x = 1066$nm, which is of the order of the structure size. Clearly, such behavior is at odds with our assumption of a homogeneous metallic resistance path that we are probing.
One might argue that the very large apparent resistance path that we report for the simple Hall cross in Fig. \[cross\] represents a device failure, even though the temperature dependence of the resistance nicely follows the behavior expected for thin film MnSi. Therefore, we have evaluated the resistive component for our different Hall voltage structures across the equipotential contacts. From this analysis, we find that for all nanostructure samples there is a detectable sample/device-dependent zero field transverse Hall resistance: For the 290nm cross we find the largest resistance value, about 100$\Omega$, while for the other samples it varies between 0.01$\Omega$ (thin film) and 20$\Omega$ (175nm Hall cross), with seemingly a tendency towards larger resistance values for smaller structures. Even disregarding the more extreme values we still see a variance of the zero field transverse Hall resistance of nominal equipotential voltage taps in the order of 0.01 to 1$\Omega$. These resistance values need to be viewed in relation to the Hall resistance across the contacts, which is of the order of 0.05$\Omega$. In this situation, for a large residual resistance on equipotential contacts the signal-to-noise ratio for the determination of the Hall resistance becomes rather small. Below, we will discuss the consequences of this finding.
Next, to finalize our comparative analysis of nanostructured thin film MnSi we have determined the Hall resistivity $\rho_{yx}$ for our different samples (see Figs. \[hallbulk\] – \[hallstruct\]). Because of the (magneto)resistive signal component discussed before, to extract the Hall resistivity, even for the nanostructured samples we have to determine the field-symmetric and antisymmetric signal contribution, with the latter one representing the Hall signal.
Again, we start our discussion with the data for bulk material. With the approach from Eq. (1) to separate the Hall effect into normal and anomalous contribution, for the normal part we can write $\rho_{yx}^N = R_0 B = (n e)^{-1} B$, with the Hall coefficient $R_0$ and the carrier density $n$ of electron charges $e$. Thus, the observation of a linear-in-field behavior of $\rho_{yx}$ at high temperatures (100K $\gg T_C$), in agreement with the findings of Ref. [@neubauer2009], allows corresponding fits to the data, with a summary of the results listed in Tab. \[tab:table1\]. We find a quantitative difference in the absolute value of the Hall resistivity between our data and that from Ref. [@neubauer2009], which translates into a carrier density for our crystal being the corresponding factor larger. It is not clear, if this difference in the value of $n$ reflects a sample dependence. We note that the room temperature resistivity of our sample is 30 % smaller than that from Neubauer et al. [@neubauer2009]. If this difference arises from an inaccurate determination of the (effective) sample geometries in the one or other case, it would translate into a difference of the Hall coefficient broadly consistent with the present data.
[lcc]{} & )&\
Bulk & 0.8 $\pm$ 0.3 & 7.7 $\pm$ 1.9\
Thin film & 3.4 $\pm$ 0.3 & 1.8 $\pm$ 0.2\
Structure 1 (10 $\mu$m) & 3.7 $\pm$ 0.3 & 1.7 $\pm$ 0.2\
Structure 2 (1 $\mu$m) & 3.3 $\pm$ 0.2 & 1.9 $\pm$ 0.1\
Structure 3 (300 nm) & 6.4 $\pm$ 1.7 & 1.0 $\pm$ 0.3\
Structure 4 (230 nm) & 2.4 $\pm$ 0.8 & 2.6 $\pm$ 0.2\
Structure 5 (200 nm) & 3.6 $\pm$ 0.4 & 1.7 $\pm$ 0.1\
Structure 6 (175 nm) & 11.8 $\pm$ 2.5 & 0.5 $\pm$ 0.1\
\[tab:table1\]
After subtraction of the normal Hall contribution we obtain the anomalous contribution, [*i.e.*]{}, $\rho_{yx}^{A} = \rho_{yx} - \rho_{yx}^N (300 {\rm K})$ for bulk material, which we plot in Fig. \[hallbulk\]. Qualitatively and – as discussed – semiquantitatively our data reproduce the behavior reported in Ref. [@neubauer2009]. Overall, our measured anomalous Hall resistivity is a factor of two smaller than that reported in Ref. [@neubauer2009], but otherwise the temperature and field evolution of $\rho_{yx}^{A}$ is the same. The anomalous Hall contribution is positive, its field dependence essentially reflects that of the magnetization, with a prefactor that increases with temperature up to about $T_C$, in good agreement with Ref. [@neubauer2009].
Next, we have performed Hall effect measurements on our thin film sample. The overall appearance roughly resembles the single crystal data, but with some differences in detail. As before, at high temperatures we find a linear-in-field behavior that we use to extract the carrier density. We find it to be about a factor of four smaller than for our single crystal, and a factor of two smaller than the crystal value reported in Ref. [@neubauer2009] (see Tab. \[tab:table1\]).
Again, the anomalous contribution we have extracted following the same procedure as for the single crystal. The overall evolution of the anomalous contribution is similar to that of the crystal: The field dependence shows basically a magnetization behavior, and the overall signal size increases with increasing temperature up to about 50K, [*i.e.*]{}, $T_C$ (Fig. \[hallfilm\]). Only, compared to the single crystal data, the magnitude of the anomalous contribution again is somewhat larger, with an overall signal change of about 100n$\Omega$cm as compared to 70n$\Omega$cm for the single crystal.
Finally, we have carried out Hall effect measurements on our nanostructured samples MnSi, with the main results summarized in the Figs. \[hall01mu\] and \[hallstruct\]. Again, the structures exhibit a behavior which overall is qualitatively similar to that of the crystal, but with some differences in detail. First, from the high-temperature behavior we extract the carrier densities as before (Tab. \[tab:table1\]). Here, the carrier density determined this way appears to change with structure size, with for the smallest structure being an order of magnitude smaller than for single crystalline material.
After subtraction of the normal contribution we obtain the anomalous contribution. As before, qualitatively it resembles that of the crystal and the thin film, producing a positive signal with a magnetization-like field dependence. As well, for increasing temperature up to $\sim 50$K, the signal amplitude increases. As before, the amplitude is enhanced as compared to the crystal, and also slightly larger than for the thin film (about 120n$\Omega$cm).
We have also attempted to measure the Hall effect for our smallest structure (path width 175nm). Here, however, in the cross voltage $U_{cross}$ the magnetoresistive contribution to the signal was much larger than the Hall part. Therefore, because of the low signal-to-noise ratio mentioned above it was impossible to clearly resolve the field dependence of the Hall voltage. We note that already for the structure with path width 300nm an enhanced noise level is observable in the Hall effect data (see Fig. \[hallstruct\]), and which stems from the same difficulty.
Discussion
==========
Summarizing our observations so far, we have performed a thorough and comparative study on single-crystalline, thin film and nanostructured MnSi. Qualitatively and quantitatively, we reproduce the findings on resistivity and magnetotransport previously reported for crystalline and thin film samples. Moreover, we have produced a consistent set of Hall effect data for our set of samples, which globally exhibit a similar behavior and are qualitatively in agreement with some of the prior reports [@karhu2010; @karhu2011; @karhu2012; @geisler2012; @wilson2012; @engelke2012; @li2013; @suzuki2013; @menzel2013; @wilson2013; @yokouchi2014; @wilson2014; @engelke2014; @meynell2014a; @meynell2014b; @yokouchi2015; @lancaster2016], thus allowing us to discuss our observations in comparison to and in the context of those studies.
The essential findings from our comparative study on thin film and nanostructured MnSi to be discussed in detail are thus:
- the measured (residual) resistivity increases with decreasing structure size;
- for the smallest structures the measured resistivity values reach into the ”non-metallic” range, while the overall metallic behavior ($d \rho /d T > 0$) of and magnetic transition signatures in the resistivity are preserved;
- the magnetoresistivity is not affected in its appearance by the structure size;
- even dedicated Hall bar structures produce a resistive signal in a Hall geometry;
- the carrier density derived from the thin film/nanostructure measurements appears to be much smaller than for single crystalline material, with some indications of a structure size dependence.
Some of these observations can be accounted for in rather simple terms. As pointed out, the observation of a large resistive contribution for the measurement of the Hall cross in Fig. \[cross\] and the measured magnetoresistive component for the Hall structures requires to drop the assumption of a homogeneous current path for all samples. Instead, if we assume that the current follows a (more-or-less) percolative path through the sample, one could easily obtain a significant resistive signal even in Hall cross geometry, [*viz.*]{}, on nominally equipotential points in the structure.
The concept of a percolative current path in thin film/nanostructured MnSi would account for the observed increase of the resistivity with decreasing structure size. Conceptually, a percolative current path in a metallic system reflects that in the material there are spatial regions of high and low conductivity. For a thin film, the regions of high conductivity will dominantly carry the current and can be understood as a network of parallel/series resistors. An extended thin film corresponds to a very large network of resistors, and their collective behavior will be similar to the case of a homogeneously distributed current, [*i.e.*]{}, the film and bulk behavior are similar.
In contrast, when we limit the size of the system by nanostructuring, the extension of the network is gradually reduced. In result, this will typically lead to an effectively increased current path. For sufficiently long current paths it may even lead to measured resistances of the structures, which - if transformed into resistivities by using the nominal geometrical dimensions of the structures - produce $\rho$-values that exceed typical metallic values without loosing the metallic character $d \rho /d T > 0$. As well, with this scenario a pronounced device dependence might be expected, as structural defects such a grain boundaries - if they span the whole device - might have a significant effect on the measured resistance. Conversely, a percolative current path would not show up in the magnetoresistivity, if we use a scaled representation as performed in Fig. \[magnetores\].
The concept of a percolative current path in MnSi films/structures might thus explain certain findings reported here. However, reconciling it with the metallicity of the MnSi films appears not to be easy. If we assume that the current in thin films MnSi follows a percolative path, we imply that we have produced significant spatial conductivity variations in a metallic system. One may argue that the type of island growth that controls the morphology of MnSi is responsible for this, effectively producing many grain boundaries, where local strain etc. will reduce conductivity. Consistent with this concept is the observation that the size of the islands in MnSi is typically of the order of a few 10 to 100nm (see shade variations in the inset of Fig. \[afmzoom\], which track the island growth morphology). With our smallest nanostructures MnSi we would thus probe single island regimes, where we might expect large effects from percolative resistivity. Only, based on experience with common metallic films, the naive expectation would be that local conductivity differences even across different islands would be of the order of $\sim 10$%, and it is not obvious that this would be sufficient to produce a percolative resistance path.
We may speculate that part of this issue relates to the tensile strain in the thin films MnSi. The nanostructuring process might lead to a lateral relaxation of the MnSi thin film. This may increase the strain at the grain boundaries and, thus, enhance the percolative resistive effect. If this simple picture of structural relaxation affecting the electronic transport properties of nanostructured MnSi films carries some truth, it would imply that in order to produce and study nanostructured material it requires improvements to the sample quality by growing unstrained epitaxial films.
Regarding the Hall effect, the idea of thin film MnSi being electronically inhomogeneous might provide explanations for some experimental observations. For instance, the sample-to-sample variations of the anomalous Hall contribution, as summarized in Table \[tab:comp\], might simply reflect that because of an inhomogeneous current path only certain parts of the samples are probed by the Hall effect. As these are different from sample to sample, they produce corresponding differences in the anomalous Hall effect. If this would be true, of course, in the future the search for skyrmions by means of the (topological) Hall effect will need to verify that with the Hall effect the spatial regions containing skyrmions are probed. As well, if working on nanostructure samples, to classify these it will be important to verify that in Hall geometry there is only a small zero-field cross voltage ([*i.e.*]{}, a relatively homogeneous current path can be assumed).
As well, the argument of an electronical inhomogeneity would suggest that it varies on a spatial range of the order of the MnSi islands, [*i.e.*]{}, a few ten nanometers. Most likely, this length scale is also an upper limit for the electronic mean free path, leading to a situation where the mean free path is not significantly larger than the diameter of a skyrmion. In that case, it is not evident if the electrons at all can ”see” a skyrmion, as they might undergo scattering before traversing the skyrmion. Again, this might imply that in thin films MnSi it would be difficult to see skyrmionic phases by means of the (topological) Hall effect.
Finally, the apparent reduction of the carrier density with reduced structure size needs to be discussed. Clearly, our experimental observation can not reflect a true carrier density reduction in our nanostructure samples by an order of magnitude compared to bulk material, as this should significantly affect the magnetic properties. Given that $T_C$ is constant for the thin film/nanostructure samples, the magnetically ordered areas in the samples will have typical metallic carrier densities. But then, the reduction of the carrier density - which corresponds to an apparent increase of the Hall voltage - must be an artifact. Here, we might consider if the concept of a percolative resistance accounts for our observations.
In the most basic approach (neglecting band structure etc.), the Hall voltage over a sample is calculated as $$U_H = \frac{R_H \cdot I \cdot B}{d},$$ with the Hall coefficient $R_H$, the current through the sample $I$, the externally applied magnetic field $B$ and the sample thickness $d$. As a note of caution, we stress that this equation is derived based on the concept of a homogeneous current path, and which possibly is not fulfilled in all our cases to good approximation. Still, using the equation as a starting point, in a comparative study on nanostructures produced from the same or similar thin film samples, an artificial enhancement of $U_H$ can not result from a variation of the film thickness $d$. One might argue that with the Hall effect we are predominantly probing sample areas with low carrier densities (large local $R_H$), but in the metallic environment of the surrounding area a short-cutting of such effects would be expected. It remains the possibility of an artifical enhancement of the local current/current density that might produce an increase of $U_H$ measured across the Hall cross area. Such a scenario might possibly arise from the concept of a percolative resistance path sketched above, as with a more narrow current path local currents might be larger. To definitely answer if this is the case for our nanostructure samples MnSi, however, it would require better knowledge on the local structural and electronic ”morphology” of our samples.
Altogether, for thin film/nanostructured MnSi a picture emerges, where thin film material behaves differently from single crystalline specimens. Our work leads us to conclude that there is a possibility of an electronic inhomogeneity in samples of MnSi. This concept as such would be highly unusual for simple metallic systems, but has been discussed for correlated electron systems related to MnSi. In recent years, it has been demonstrated that disorder effects are very much enhanced for correlated electron materials, and which have been discussed for instance within the context of Griffiths phases, which effectively consider electronic inhomogeneities in correlated metals [@castro1998; @otop2005; @brando2016]. Here, a line of thought might be that the effect of disorder induced strain in the films is enhanced by the correlations, which in in turn might produce an electronic inhomogeneity.
In the context of skyrmionic physics, our work has also some consequences. Notably, the observation of local objects such as skyrmions by rather delocalized measurment techniques such electronic transport/Hall effect might be difficult in the presence of structural inhomogeneities. Conversely, the role of structural inhomogeneities for the creation or destruction of skyrmions has also not been studied in detail. Thus, our work highlights the relevance of understanding the role of structural disorder for skyrmionic systems, especially in terms of testing skyrmions electronically.
We gratefully acknowledge support by the Braunschweig International Graduate School of Metrology B-IGSM and the DFG Research Training Group GrK1952/1 ”Metrology for Complex Nanosystems”.
[00]{} S. Mühlbauer, B. Binz, F. Jonietz, C. Pfleiderer, A. Rosch, A. Neubauer, R. Georgii, and P. Böni, Science [**323**]{}, 5916 (2009). S. Krause, and R. Wiesendanger, Nature Materials [**15**]{}, 493 (2016). A. Fert, V. Cros, and J. Sampaio, Nature Nanotechnology [**8**]{}, 152 (2013) R. Tomasello, E. Martinez, R. Zivieri, L. Torres, M. Carpentieri, and G. Finocchio, Sci. Rep. [**4**]{}, 6784 (2014). A. Neubauer, C. Pfleiderer, B. Binz, A. Rosch, R. Ritz, P. G. Niklowitz, and P. Böni, Phys. Rev. Lett. [**102**]{}, 186602 (2009). C. Schütte, and M. Garst, Phys. Rev. B [**90**]{}, 094423 (2014). T. Schwarze, J. Waizner, M. Garst, A. Bauer, I. Stasinopoulos, H. Berger, C. Pfleiderer, and D. Grundler, Nature Materials [**14**]{}, 478 (2015). C. Pfleiderer, T. Adams, A. Bauer, W. Biberacher, B. Binz, F. Birkelbach, P. Böni, C. Franz, R. Georgii, and M. Janoschek, J. Phys.: Condens. Matter [**22**]{}, 164207 (2010). M. Bode, M. Heide, K. von Bergmann, P. Ferriani, S. Heinze, G. Bihlmayer, A. Kubetzka, O. Pietzsch, S. Blügel, and R. Wiesendanger, Nature [**447**]{}, 190 (2007). S. Heinze, K. von Bergmann, M. Menzel, J. Brede, A. Kubetzka, R. Wiesendanger, G. Bihlmayer, and S. Blügel, Nature Physics [**7**]{}, 713 (2011). T. Moriya, [*Spin Fluctuations in Itinerant Electron Magnetism*]{}, (Solid-State Sciences) (Berlin: Springer) (1985). J. D. Thompson, Z. Fisk, and G. G. Lonzarich, Physica B [**161**]{}, 317 (1989). T. Sakakibara, H. Morimoto, and M.Date: J. Phys. Soc. Jpn. [**51**]{}, 2439 (1982). E. Karhu, S. Kahwaji, T. L. Monchesky, C. Parsons, M. D. Robertson, and C. Maunders, Phys. Rev. B [**82**]{}, 184417 (2010). E. A. Karhu, S. Kahwaji, M. D. Robertson, H. Fritzsche, B. J. Kirby, C. F. Majkrzak, and T. L. Monchesky, Phys. Rev. B [**84**]{}, 060404(R) (2011). E. A. Karhu, U. K. Rößler, A. N. Bogdanov, S. Kahwaji, B. J. Kirby, H. Fritzsche, M. D. Robertson, C. F. Majkrzak, and T. L. Monchesky, Phys. Rev. B [**85**]{}, 094429 (2012). B. Geisler, P. Kratzer, T. Suzuki, T. Lutz, G. Costantini, and K. Kern, Phys. Rev. B [**86**]{}, 115428 (2012). M. N. Wilson, E. A. Karhu, A. S. Quigley, U. K. Rößler, A. B. Butenko, A. N. Bogdanov, M. D. Robertson, and T. L. Monchesky, Phys. Rev. B [**86**]{}, 144420 (2012). J. Engelke, T. Reimann, L. Hoffmann, S. Gass, D. Menzel, and S. Süllow, J. Phys. Soc. Jpn. [**81**]{}, 124709 (2012). Y. Li, N. Kanazawa, X. Z. Yu, A. Tsukazaki, M. Kawasaki, M. Ichikawa, X. F. Jin, F. Kagawa, and Y. Tokura, Phys. Rev. Lett. [**110**]{}, 117202 (2013). T. Suzuki, T. Lutz, B. Geisler, P. Kratzer, K. Kern, G. Costantini, Surface Science [**617**]{}, 106 (2013). D. Menzel, J. Engelke, T. Reimann, and S. Süllow, J. Kor. Phys. Soc. [**62**]{}, 1580 (2013). M. N. Wilson, E. A. Karhu, D. P. Lake, A. S. Quigley, S. Meynell, A. N. Bogdanov, H. Fritzsche, U. K. Rößler, and T. L. Monchesky, Phys. Rev. B [**88**]{}, 214420 (2013). T. Yokouchi, N. Kanazawa, A. Tsukazaki, Y. Kozuka, M. Kawasaki, M. Ichikawa, F. Kagawa, and Y. Tokura, Phys. Rev. B [**89**]{}, 064416 (2014). M. N. Wilson, A. B. Butenko, A. N. Bogdanov, and T. L. Monchesky, Phys. Rev. B [**89**]{}, 094411 (2014). J. Engelke, D. Menzel, H. Hidaka, T. Seguchi, and H. Amitsuka, Phys. Rev. B [**89**]{}, 144413 (2014). S. A. Meynell, M. N. Wilson, H. Fritzsche, A. N. Bogdanov, and T. L. Monchesky, Phys. Rev. B [**90**]{}, 014406 (2014). S. A. Meynell, M. N. Wilson, J. C. Loudon, A. Spitzig, F. N. Rybakov, M. B. Johnson, and T. L. Monchesky, Phys. Rev. B [**90**]{}, 224419 (2014). T. Yokouchi, N. Kanazawa, A. Tsukazaki, Y. Kozuka, A. Kikkawa, Y. Taguchi, M. Kawasaki, M. Ichikawa, F. Kagawa, and Y. Tokura, J. Phys. Soc. Jpn. [**84**]{}, 104708 (2015). T. Lancaster, F. Xiao, Z. Salman, I. O. Thomas, S. J. Blundell, F. L. Pratt, S. J. Clark, T. Prokscha, A. Suter, S. L. Zhang, A. A. Baker, and T. Hesjedal, Phys. Rev. B [**93**]{}, 140412(R) (2016). A. B. Butenko, A. A. Leonov, U. K. Rößler, and A. N. Bogdanov, Phys. Rev. B [**82**]{}, 052403 (2010). S. Banerjee, J. Rowland, O. Erten, and M. Randeria, Phys. Rev. X [**4**]{}, 031045 (2014). J. Rowland, S. Banerjee, and M. Randeria, Phys. Rev. B [**93**]{}, 020404(R) (2016). A. Tonomura, X. Z. Yu, K. Yanagisawa, T. Matsuda, Y. Onose, N. Kanazawa, H. S. Park, Y. Tokura, Nano Lett. [**12**]{}, 1673 (2012). M. Mochizuki, X. Z. Yu, S. Seki, N. Kanazawa, W. Koshibae, J. Zang, M. Mostovoy, Y. Tokura and N. Nagaosa, Nature Materials [**13**]{}, 241 (2014). X. Yu, A. Kikkawa, D. Morikawa, K. Shibata, Y. Tokunaga, Y. Taguchi, and Y. Tokura, Phys. Rev. B [**91**]{}, 054411 (2015). F. Jonietz, S. Mühlbauer, C. Pfleiderer, A. Neubauer, W. Münzer, A. Bauer, T. Adams, R. Georgii, P. Böni, R. A. Duine, K. Everschor, M. Garst, A. Rosch, Science [**330**]{}, 1648 (2010). K. Kadowaki, K. Okuda, and M. Date, J. Phys. Soc. Jpn. [**51**]{}, 2433 (1982). The resistivity of the 230nm sample was measured with a current density of $2\cdot10^{10}\,$A/m$^2$ which is 10$^{3}$ $-$ 10$^{4}$ times higher then the current density used for the other measurements. Consequently the $\sim$ 5K shift of the transition temperature is most likely explained via a heating effect in the structure. J. H. Mooij, Phys. Status Solidi A [**17**]{}, 521 (1973). C. C. Tsuei, Phys. Rev. Lett. [**57**]{}, 1943 (1986). K.Ueda, Solid State Commun. [**19**]{}, 965 (1976). While the carrier density for the bulk material has been obtained from data taken at 300K like in Ref. [@neubauer2009] this method is not possible for the film and the structures due to the substrate. Instead the carrier density was taken from measurements done at 100K (thin film, 10$\mu$m, 1$\mu$m and 200nm) or 60K (300nm, 230nm and 175nm). The difference in temperature are taken into account with a corresponding uncertainty. A. H. Castro Neto, G. Castilla, and B. A. Jones, Phys. Rev. Lett. [**81**]{}, 3531 (1998). A. Otop, S. Süllow, M. B. Maple, A. Weber, E. W. Scheidt, T. J. Gortenmulder, and J. A. Mydosh, Phys. Rev. B [**72**]{}, 024457 (2005). M. Brando, D. Belitz, F. M. Grosche, and T. R. Kirkpatrick, Rev. Mod. Phys. [**88**]{}, 025006 (2016).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $ T _{P } f (x) = \int e ^{i P (y)} K (y) f (x-y) \; dy $, where $ K (y)$ is a smooth Calderón-Zygmund kernel on $ \mathbb R ^{n}$, and $ P$ be a polynomial. We show that there is a sparse bound for the bilinear form $ \langle T_P f, g \rangle$. This in turn easily implies $ A_p $ inequalities. The method of proof is applied in a random discrete setting, yielding the first weighted inequalities for operators defined on sparse sets of integers.'
address:
- ' School of Mathematics, Georgia Institute of Technology, Atlanta GA 30332, USA'
- ' School of Mathematics, Georgia Institute of Technology, Atlanta GA 30332, USA'
author:
- 'Michael T. Lacey'
- Scott Spencer
title: |
Sparse Bounds for Oscillatory\
and Random Singular Integrals
---
Introduction
============
Singular integral operators can be pointwise dominated by sparse operators, which are positive localized operators, something that singular integrals are not. This paper extends this theme to the settings of (a) oscillatory singular integrals, and (b) discrete random operators. In both cases, we easily derive weighted inequalities. In the latter case, these are the first such weighted inequalities known. We state our results before providing a broader context.
Call a collection of cubes $ \mathcal S $ in $ \mathbb R ^{n}$ a *sparse* collection if there is a set $ E_Q \subset Q$ for each $ Q\in \mathcal S$ so that (a) $ \lvert E_Q\rvert >c \lvert Q\rvert $ for each $ Q\in \mathcal S$, and (b) the collection of sets $\{ E_Q \;:\; Q\in \mathcal S\}$ are pairwise disjoint. Here $ 0< c< 1$ will be a dimensional constant that we do not track. Define a *sparse bilinear form* to be $$\label{e:sparseOp}
\Lambda _{r,s} (f,g) = \sum_{Q\in \mathcal S} \langle f \rangle _{Q,r} \langle g \rangle _{Q,s} \lvert Q\rvert, \qquad 1 \leq r, s < \infty .$$ Above, $ \langle f \rangle _{Q,r} ^{r} := \lvert 3 Q\rvert ^{-1} \int _{3Q} \lvert f\rvert ^{r} \;dx $, and if $ r=s$, then $ \Lambda _{r} = \Lambda _{r,r}$. We frequently suppress the collection of sparse cubes $ \mathcal S$.
We consider Calderón-Zygmund singular integral operators $ T $, defined to be an $ L ^2 (\mathbb R ^{n})$ bounded convolution operator given by $$\langle T f, g \rangle = \int\!\int K (x-y) f (y) g (x) \; dx\,dy .$$ for compactly supported functions $ f, g$ with disjoint supports. Moreover, the kernel $ K (y)$ satisfies $$\begin{aligned}
\label{e:size}
\lvert \nabla ^{t} K (x,y)\rvert \leq C_t \lvert x-y\rvert ^{-n-t} , \qquad x\neq y \in \mathbb R ^{n}, \end{aligned}$$ for $ t \in \{ 0,1\} $. Key examples are $ K (y) = 1/y$ in dimension one, and the Riesz transform kernels $ y/ \lvert y\rvert ^{n+1} $, in dimension $ n$.
Such operators are of course non-local, and involve subtle cancellative effects. It is thus something of a surprise that such operators are dominated by sparse operators, which have none of these features. This is a special case of [@150105818; @14094351; @150805639].
\[t:czo\] For each Calderón-Zygmund singular integral operator $ T$ and bounded compactly supported function $ f$, there is a sparse operator $ \Lambda = \Lambda _{T, f}$ so that $
\lvert T f \rvert \lesssim \Lambda _{1} f
$.
An immediate corollary are weighted inequalities that are sharp in the $ A_p$ characteristic. See [@MR3085756; @150105818; @14094351].
We consider polynomials of a fixed degree $ d$, given by $ P (x,y) = \sum_{\alpha , \beta \;:\; \lvert \alpha \rvert + \lvert \beta \rvert \leq d } \lambda _{\alpha, \beta } x ^{\alpha } y ^{\beta }$, where we use the usual multi-index notation. The polynomial modulated Calderón-Zygmund operators are $$\label{e:TK}
T_P f (x) = \int e ^{i P (x,y)}K (y) f (x-y) \; dy .$$ The $ L ^{p}$ result below is a special case of the results of Ricci and Stein [@MR822187; @MR890662], and the weak-type result is due to Chanillo and Christ [@MR883667].
\[t:fixedPoly\] For $ 1 < p < \infty $, the operator $ T_P$ is bounded on $ L ^{p}$, that is $$\lVert T_P \;:\; L^{p} \mapsto L ^{p}\rVert \lesssim 1,$$ where the implied constant depends on the degree of $ P$, and in particular is independent of $ \lambda $. Moreover, $ T_P$ maps $ L ^{1}$ to weak $ L ^{1}$, with the same bound.
The dependence on the polynomial being felt only through the degree of $ P$ is important to the application of these bounds to the setting of nilpotent groups, like the Heisenberg group, see [@MR890662]. This dependence continues to hold in the Theorems below.
\[t:main1\] For each $ 1< r < 2$ Calderón-Zygmund operator $ T$, polynomial $ P = P(y)$ of degree $ d$ and bounded supported functions $f, g $ there is a bilinear form $ \Lambda _{r} $ so that $$\lvert \langle T _{P} f ,g \rangle\rvert \lesssim \Lambda _{r} (f,g).$$ The implied constant depends only on $ T$, the degree $ d$, and dimension $ n$ and choice of $ r>1$.
The bound above continues to hold for polynomials $ P $ of two variables, but we suppress the details, as the estimate above can most likely be improved. And, as written is quite easy to prove, yet yields a non-trivial corollary.
\[c:wtd\] For $ 1 < p < \infty $, the operator $ T_P$, where $ P = P (y)$ is of degree $ d$, is bounded on $ L ^{p} (w)$, where $ w $ is a Muckenhoupt weight $ w\in A_p $.
Weak-type and weighted estimates for oscillatory singular integrals have been studied in this and more general contexts by various authors, see for instance [@MR2900003; @MR2910762; @MR2949870; @MR2115460; @MR1782909]. Y. Ding and H. Liu [@MR2900003] were interested in $ L ^{p} (w)$ inequalities for more general operators $ T$. The approach of these authors entails many complications.
The method of proof of Theorem \[t:main1\] is very simple. And, so we suspect that stronger results are possible. For instance, this Conjecture would imply nearly sharp $ A_p$ bounds, for all $ 1< p < 2$.
\[j:oscillatory\] For $ 1 < r < \infty $, the operator $ T_P$, where $ P = P (y)$ is of degree $ d$, for each bounded compactly supported function $ f$, there is a sparse operator $ \Lambda _{1,r}$ so that $$\lvert \langle T _{P} f ,g \rangle\rvert \lesssim \Lambda _{1,r} (f,g).$$
It seems likely that the weak type argument of Chanillo and Christ [@MR883667] would establish the Conjecture for $ r=2$. Also see [@160901564].
We turn to weighted inequalities for *discrete random Hilbert transforms* acting on functions on $ \ell ^2 (\mathbb Z )$. Define a sequence of Bernoulli rvs $ \{ X _{n} \;:\; n\neq 0\}$ with $ \mathbb P (X_n =1) = \lvert n\rvert ^{- \alpha } $, where $ 0 \leq \alpha < 1$. Then, the set $ \{n \;:\; X_n=1\}$ is a.s. infinite, by the Borel-Cantelli Lemma. Then, we consider the random Hilbert transform, and maximal function below. $$\begin{aligned}
\label{e:Ha}
H _{\alpha } f (x)& = \sum_{n\neq 0} \frac {X_n} {n ^{1- \alpha }} f (x-n).
\\ \label{e:Ma}
M _{\alpha } f (x)& = \sup_{n > 0} \Bigl\lvert \frac 1 { S_N} \sum_{n=1} ^{N} X_n f (x-n) \Bigr\rvert, \qquad S_N = \sum_{n=1} ^{N} X_n . \end{aligned}$$
Our sparse bound here is more restrictive, with the value of the sparse index $ r$ depending upon random parameter $ \alpha $.
\[t:R\] For any $ 0< \alpha < 1$, $ 1+ \alpha < r < 2$, almost surely, the following holds: For all functions $ f, g$ finitely supported on $ \mathbb Z $, there is a bilinear sparse operator $ \Lambda _{r} $ so that $$\lvert \langle H _{\alpha } f, g \rangle\rvert \lesssim \Lambda _{r} (f,g).$$ The same inequality holds for $ M _{\alpha }$. (The sparse operator can be taken non-random, but the implied constant is random.)
Weighted inequalities are a corollary. They are *the first we know of* holding for operators defined on sets of the integers with zero asymptotic density.
\[c:R\] For any $ 0< \alpha < 1$, almost surely, the following holds: For all $ 1 + \alpha < p < \frac {1+ \alpha } \alpha $, and weights $ w$ so that $$\label{e:WW}
w ^{1+ \alpha } \in A _{ (1+ \alpha ) (p-1)+1 }, \qquad w \in A _{1 + \frac 1 {(1+\alpha) (p'-1)}} ,$$ we have $
\lVert H _{\alpha } \;:\; \ell ^{p} (w) \mapsto \ell ^{p} (w)\rVert < \infty
$. The implied constant only depends upon $ [ w ^{1+ \alpha }] _{A _{ (1+ \alpha ) (p-1)+1 }}$, and $ [w] _{A_{1 + \frac 1 {\alpha (p'-1)}} }$. The same inequality holds for $ M _{\alpha }$.
The study of these questions was initiated by Bourgain [@MR937581], as an elementary example of a sequence of integers for which one could derive $ \ell ^{p}$ inequalities, with the sequence of integers also having asymptotic density zero. Various aspects of these questions have been studied, both in $ \ell ^{p}$, in the weak $ (1,1)$ endpoints [@MR1325697; @MR2680392; @MR3421994; @MR2318621; @MR2576702]. We are not aware of any result in the literature that proves a weighted estimate in this sort of discrete setting. (If the set of integers has full density, it is easy to transfer weighted estimates.)
There is a subtle difference between the Hilbert transform and the maximal function in this random setting. In particular, more should be true for the maximal function. Prompted by the work of LaVictoire [@MR2576702], we pose
For $ 0< \alpha < 1/2$, almost surely, for all $ 1< r < 2$, and finitely supported functions $ f, g$, there is a sparse operator $ \Lambda _{1,r}$ so that $$\langle M _{\alpha } f , g \rangle \lesssim \Lambda _{1,r} (f,g).$$
We turn to the context for our paper. The concept of sparse operators arose from Lerner’s remarkable median inequality [@MR2721744]. It’s application to weighted inequalities was advanced by several authors, with a high point of this development being Lerner’s argument [@MR3085756] showing that the weighted norm of Calderón-Zygmund operators is comparable to that of the norms of sparse operators. This lead to the question of pointwise control, namely Theorem \[t:czo\]. First established by Conde-Alonso and Rey [@14094351], also see Lerner and Nazarov [@150805639], the author [@150105818] established Theorem \[t:czo\] with a stopping time argument. The latter argument was extended by Bernicot, Frey and Petermichl [@MR3531367] to a setting where the operators are generated by semigroups, including examples outside the scope of classical Calderón-Zygmund theory. For closely related developments see [@MR3484688; @151005789]. The sparse bounds for commutators [@160401334; @160405506] are remarkably powerful. Edging beyond the Calderón-Zygmund context, Benau, Bernicot and Frey [@160506401] have supplied sparse bounds for certain Bochner-Riesz multipliers.
Very recently, Culiuc, di Plinio and Ou [@160305317] have established a sparse domination result in a setting far removed from the extensions above: The trilinear form associated to the bilinear Hilbert transform is dominated by a sparse form. This is a surprising result, as the bilinear Hilbert transform has all the difficult features of the Hilbert transform, with additional oscillatory and arithmetic-like aspects. This paper is an initial effort on our part to understand how general a technique ‘domination by sparse’ could be. There are plenty of additional directions that one could think about.
For instance, the interest in the oscillatory singular integrals is driven in part by their application to singular integrals defined on nilpotent groups. Implications of the sparse bound in this setting are unexplored.
There are two approaches to sparse bounds, the bilinear form method [@160305317], and the use of the maximal truncation inequality [@150105818]. We use neither approach. After applying the known sparse bounds for singular integrals, for the remaining parts of the operator, there is a very simple interpolation argument which you can use in the bilinear setting. The notable point about the proofs are that they are quite easy, and yet deliver striking applications.
Proof of Theorem \[t:main1\]
============================
Our conclusion is invariant under dilations of the operator. Hence, we can proceed under the assumption that $ \lVert P\rVert = \sum_{\alpha } \lvert \lambda _{\alpha }\rvert =1$. We can also assume that the polynomial $ P$ has no linear term, as it can be absorbed into the function $ f$. Under these assumptions we prove
\[t:s\] Let $ P$ be a polynomial without linear terms, and $ \lVert P\rVert=1$. Then, for bounded compactly supported functions $ f, g$ and $ 1< r < \infty $, there is a sparse form $ \Lambda_1 $ and a $ \eta >0$ so that $$\label{e:s}
\lvert \langle T_P f, g \rangle\rvert \lesssim \Lambda _1 (f,g)
+ \sum_{ Q \in \mathcal D\;:\; \lvert Q\rvert \geq 1 }
\langle f \rangle _{Q,r} \langle g \rangle _{Q,r} \lvert Q\rvert ^{1- \eta }$$
It is easy to see that this implies Theorem \[t:main1\], since the second term on the right is restricted to dyadic cubes of volume at least one, and there is a gain of $ \lvert Q\rvert ^{- \eta } $. Moreover, we will see that this Theorem implies the weighted result.
Let $ e (\lambda ) = e ^{i \lambda } $ for $ \lambda \in \mathbb R $. If the kernel $ K$ of $ T$ is supported on $ 2B= \{y \;:\; \lvert y\rvert \leq 2\}$, then we have $$\lvert e (P (y)) K (y) - K (y)\rvert \lesssim \mathbf 1_{2B} (y) \lvert y \rvert ^{-n+1},$$ so that $ \lvert T _{P} f - T f \rvert \lesssim M f $. Both $ T$ and $ M$ admit pointwise domination by sparse forms, hence also by bilinear forms. (This is the main result of [@150105818].)
Thus, we can proceed under the assumption that the kernel $ K $ is not supported on $ B$. We can then write $$K = \sum_{j =1} ^{\infty } \varphi _j$$ where $ \varphi _j$ is supported on $ 2 ^{j-1} B \setminus 2 ^{j-2}B$, with $ \lVert \nabla ^{s} \varphi _ j\rVert _{\infty } \lesssim 2 ^{-nj -sj}$, for $ s=0,1$.
We use shifted dyadic grids, $ \mathcal D _{t}$, for $ 1\leq t \leq 3 ^{n}$. These grids have the property that $$\{ \tfrac 13 Q \;:\; Q\in \mathcal D_t,\ \ell Q= 2 ^{k}, 1\leq t \leq 3 ^{n}\}$$ form a partition of $ \mathbb R ^{n}$. Throughout, $ \ell Q = \lvert Q\rvert ^{1/n} $ is the side length of the cube $ Q$. We fix a dyadic grid $ \mathcal D_t$ throughout the remainder of the argument, and set $ \mathcal D_+ = \{Q \;:\; \ell Q > 2 ^{10}\}$. Define $$I _{Q} f = \int e (P (y)) \varphi _{ k} (y) (\mathbf 1_{\tfrac 13Q} f ) (x-y)\; dy, \qquad \ell Q = 2 ^{k+2}.$$ Note that $ I _Q f $ is supported on $ Q$, and that we have suppressed the dependence on $ P $, which we will continue below.
The basic estimate is then this Lemma.
\[l:fixed\] For each cube $ Q$ with $ \lvert Q\rvert \geq 1 $ and $ 1< r < 2$, there holds $$\label{e:fixed}
\lvert \langle I _{Q} f, g \rangle\rvert
\lesssim 2 ^{- \eta k} \langle f \rangle _{Q, r} \langle g \rangle _{Q,r} \lvert Q\rvert ,$$ where $ \eta = \eta (d,n,r) >0 $.
Theorem \[t:s\] follows immediately from this Lemma. The oscillatory nature of the problem exhibits itself in the next Lemma. Write $$\label{e:K}
I ^{\ast} _Q I_Q \phi (x) = \mathbf 1_{\tfrac 13Q} (x) \cdot \int _{\tfrac 13 Q} K _Q (x,y) \phi (y) \; dy .$$
\[l:K\] For each cube $ Q \in \mathcal D_+$, and $ x\in \tfrac 13 Q$, we have $$\label{e:K<}
\lvert K_Q (x,y)\rvert
\lesssim \lvert Q\rvert ^{-1} \mathbf 1_{Z_ {Q}} (x - y) + \lvert Q\rvert ^{-1 - \epsilon } \mathbf 1_{ Q} (x) \mathbf 1_{Q} (y),$$ where $ Z_ {Q} \subset Q $ has measure at most $ (\ell Q) ^{- \epsilon } \lvert Q\rvert $, where $ \epsilon = \epsilon (n,d)>0$.
This Lemma is well known, see for instance [@MR1879821]\*[Lemma 4.1]{}. Here is how we use the Lemma. Using Cauchy-Schwartz, we have $$\begin{aligned}
\label{e:2}
\lVert I_Q f \rVert_2 ^2 &\lesssim \lvert Q\rvert ^{-1} \int _{Q} \int _{Z_Q} \lvert f (x)\rvert \lvert f (x-y)\rvert \; dy dx
+ \lvert Q\rvert ^{ - \epsilon } \langle f \rangle _{Q,1} ^{2} \lvert Q\rvert
\\&
\lesssim
\lvert Q\rvert ^{- \epsilon /n} \lVert f \mathbf 1_{Q}\rVert _{2 } ^2 . \end{aligned}$$ We also have the trivial but rarely used $ \lVert I _{Q} f \rVert_ \infty \lesssim \lvert Q\rvert ^{-1} \lVert f \mathbf 1_{Q}\rVert _1 $. By Riesz Thorin interpolation, there holds with $ \ell Q = 2 ^{k}$, $$\lVert I _{Q} f \rVert _{r'} \lesssim 2 ^{- \eta k} \lvert Q\rvert ^{-1+ 2/ r'} \lVert f \mathbf 1_{Q}\rVert _r, \qquad 1 < r \leq 2,\ r'= \tfrac r {r-1}.$$ Above, $ \eta = \eta (\epsilon , r)$ But, this immediately implies . Namely, $$\begin{aligned}
\lvert \langle I_Q f, g \rangle \rvert & \lesssim
\lVert I _{Q} f \rVert _{r'} \lVert g \mathbf 1_{Q}\rVert _{r}
\\&\lesssim 2 ^{- \eta k} \lvert Q\rvert ^{-1+ 2/ r'} \lVert f \mathbf 1_{Q}\rVert _r \lVert g \mathbf 1_{Q}\rVert _{r}
\\&= 2 ^{- \eta k} \langle f \rangle _{Q,r} \langle g \rangle _{Q,r} \lvert Q\rvert. \end{aligned}$$ (Alternatively, one can just use bilinear interpolation.)
We now give the weighted result.
The qualitative result that $ T_P $ is bounded on $ L ^{p} (w)$ for $ w \in A_p$, $ 1< p < \infty $ is as follows. Given $ w \in A_p$, recall that the dual weight is $ \sigma = w ^{1- p'}$. Then, it is equivalent to show that $$\lvert \langle T _{P} (f \sigma ), g w \rangle\rvert \lesssim
C _{[w] _{A_p}}\lVert f\rVert _{L ^{p} (\sigma) } \lVert g\rVert _{L ^{p'} (w)}.$$ Using the sparse domination from , we see that we need to prove the corresponding bound for the terms on the right in . Now, it is well known [@MR3085756] that $$\Lambda _{1} ( f, g ) \lesssim
[w] ^{\max \{ 1, \frac {1} {p-1} \}} _{A_p} \lVert f\rVert _{L ^{p} (w ) } \lVert g\rVert _{L ^{p'} (w)}.$$ Indeed, this is a key part of the proof of the $ A_2$ Theorem by sparse operators.
So, it remains to consider the second term on the right in . For each $ k \in \mathbb N $, we have by Proposition \[p:scale\], $ k\in \mathbb Z $, $$\sum_{ Q \in \mathcal D\;:\; \lvert Q\rvert = 2 ^{nk} }
\langle f \rangle _{Q,r} \langle g \rangle _{Q,r} \lvert Q\rvert
\lesssim
[w] _{A_p} ^{1/p} [w] _{RH_r} [ \sigma ] _{RH_r}
\lVert f\rVert _{L ^{p} (w ) } \lVert g\rVert _{L ^{p'} (w)} .$$ As we recall in § \[s:wtd\], there is a $ r = r ([w] _{A_p}) >1$ so that $ [w] _{RH_r} [ \sigma ] _{RH_r} < 4 $. And so the proof of the Corollary is complete.
Indeed, it is easy enough to make this step quantitative. For $ 2 < p < \infty $, the choice of $ r$ can be taken to satisfy $ r-1 > c [w] _{A_p} ^{-1} $, which then means that the choice of $ \eta = \eta (r)$ in is at least as big is $ c [w] _{A_p} ^{-1} $. Then, our bound is $$\label{e:sharp?}
\langle T_P ( \sigma f), g w \rangle _{} \lesssim [w] ^{1+\frac 1p } _{A_p}\lVert f\rVert _{L ^{p} (\sigma )} \lVert g\rVert _{L ^{p'} (w)}, \qquad 2 < p < \infty .$$ We have no reason to believe that this estimate is sharp.
Random Hilbert Transforms
=========================
The discrete Hilbert transform $$H f (x) = \sum_{n\neq 0} \frac {f (x-n)} {n}$$ satisfies a sparse bound: For all finitely supported functions $ f$ and $ g$, there is a sparse operator $ \Lambda $ so that $$\label{e:dHS}
\lvert \langle H f, g \rangle\rvert \lesssim \Lambda _{1,1} (f,g).$$ This is a consequence of the main results of Theorem \[t:czo\]. Recalling the definition of $ H _{\alpha } $ in , we see that $ \mathbb E H_ \alpha f = Hf$, so it remains to consider the difference $$\begin{aligned}
H_ \alpha f (x) - H f (x) &:= \sum_{k=1} ^{\infty }
\sum_{ n \;:\; 2 ^{k-1} \leq \lvert n\rvert < 2 ^{k} } \frac { X_n - n ^{- \alpha }} {n ^{1- \alpha } } f (x-n)
\\&
:= \sum_{k=1} ^{\infty } T_k f (x). \end{aligned}$$ Above, we have passed directly to the distinct scales of the operator. We will subsequently write $ Y_n = X_n - n ^{- \alpha }$, which are independent mean zero random variables.
The crux of the matter are these two estimates:
\[l:almost\] Almost surely, for all $ 0< \epsilon < 1$, and for all integers $ k$, and $ f, g$ supported on an interval $ I$ of length $ 2 ^k$, we have $$\label{e:almost}
\lvert \langle T_k f, g \rangle\rvert
\lesssim
\begin{cases}
2 ^{- k \frac {1- \alpha } 2 + \epsilon } \langle f \rangle _{I,2} \langle g \rangle _{I,2} \lvert I\rvert
\\
2 ^{k \alpha } \langle f \rangle _{I,1} \langle g \rangle _{I,1} \lvert I\rvert
\end{cases}.$$ The implied constant is random, but independent of $ k \in \mathbb N $ and the choice of functions $ f,g $.
The second bound follows trivially from $ \lvert Y_n\rvert /n ^{1- \alpha } \mathbf 1_{ 2 ^{k-1} \leq \lvert n\rvert < 2 ^{k} } \lesssim 2 ^{k (\alpha-1)} $. For the first bound, we clearly have $$\lvert \langle T_k f, g \rangle\rvert
\leq \lVert T_k \;:\; \ell ^2 \to \ell ^2 \rVert \cdot \langle f \rangle _{I,2} \langle g \rangle _{I,2} \lvert I\rvert ,$$ so it suffices to estimate the operator norm above. The assertion is that with high probability, the operator norm is small: $$\label{e:small}
\mathbb P \bigl( \lVert T_k \;:\; \ell ^2 \to \ell ^2 \rVert > C \sqrt k 2 ^{- k \frac {1- \alpha } 2 } \bigr) \lesssim 2 ^{-k},$$ provided $ C$ is sufficiently large. Combine this with the Borel Cantelli Lemma to prove the Lemma as stated.
By Plancherel’s Theorem, the operator norm is equal to $ \lVert Z (\theta )\rVert _{L ^{\infty } ( d\theta )}$, where $$Z (\theta ) :=
\sum_{ n \;:\; 2 ^{k} \leq \lvert n\rvert < 2 ^{k+1} } Y_n \frac { e ^{ 2 \pi i \theta }} {n ^{1- \alpha } }.$$ The expression above is a random Fourier series, with frequencies at most $ 2 ^{k+2}$. By Bernstein’s Theorem for trigonometric polynomials, the $ L ^{\infty } (d \theta )$ norm can be estimated by testing the norm on at most $ 2 ^{k+3}$ equally spaced points in $ \mathbb T $, that is, we have $$\begin{aligned}
\mathbb P \bigl( \lVert Z (\theta )\rVert_ \infty > C \sqrt k 2 ^{- k \frac {1- \alpha } 2 } \bigr)
\lesssim 2 ^{k}
\sup _{\theta }
\mathbb P \bigl( \lvert Z (\theta )\rvert > C \sqrt k 2 ^{- k \frac {1- \alpha } 2 } \bigr), \end{aligned}$$ where we have simply used the union bound.
Now, $ Z (\theta )$ is the sum of independent, mean zero random variables, which are bounded by one, and have standard deviation bounded by $ c 2 ^{ - k \frac {1- \alpha } 2}$. So by, for instance, the Bernstein inequality, it follows that $$\mathbb P ( \lvert Z (\theta )\rvert > C \sqrt k 2 ^{- k \frac {1- \alpha } 2 } ) \lesssim 2 ^{-2k},$$ for appropriate $ C$. This completes the proof.
From the previous Lemma, we have the Corollary below. It with the sparse bound for the Hilbert transform completes the proof of Theorem \[t:R\], for the random Hilbert transform. The case for maximal averages is entirely similar.
\[c:almost\] Almost surely, for $ {1+ \alpha } < r < 2$, there is a $ \eta >0$ so that for all integers $ k$, and all functions $ f, g$ supported on an interval $ I$ of length $ 2 ^{k}$, we have $$\label{e:calmost}
\lvert \langle T_k f, g \rangle\rvert \lesssim
2 ^{- \eta k } \langle f \rangle _{I,r} \langle g \rangle _{I,r} \lvert I\rvert .$$
This follows from Lemma \[l:almost\] by interpolation. The relevant interpolation parameter $ \theta _0$ at which we have only an epsilon loss in the interpolated estimate is given by $$\begin{aligned}
(1-\theta_0)\alpha & = \theta_0 \frac {1- \alpha } 2 ,
\\
\textup{and then} \quad
\frac 1 {r_0} &= \frac {1-\theta _0} 1 + \frac {\theta _0} 2 . \end{aligned}$$ We see that $ r_0 = 1+ \alpha $. And so we conclude that for $ r_0 = 1+ \alpha < r < 2$, we have the required gain in the interpolated bound, which proves the Corollary.
We now turn to the weighted inequalities of Corollary \[c:R\].
For the deterministic Hilbert transform, we have the sharp bound of Petermichl [@MR2354322], namely $$\lVert H \;:\; \ell ^{p} (w) \mapsto \ell ^{p} (w)\rVert \lesssim [w] _{A_p} ^{ \max \{1, \frac 1 {p-1}\}}.$$ So, it remains to bound the terms in . By Proposition \[p:scale\], we then need to see that the hypotheses on $ w$, namely , imply that for some choice of $ r > 1+ \alpha $, we have $$\label{e:ww}
w \in A_p, \quad w \in RH _{r}, \quad \sigma = w ^{1- p'} \in RH _{r}.$$
Recall that $ v \in A_q \cap RH _{s} $ if and only if $ v ^{s} \in A _{s(q-1)+1}$. Now, by assumption, $ w ^{1+ \alpha }\in A _{ (1+ \alpha ) (p-1)+1}$. So, there is a $ t>1$ so that $ w ^{t(1+ \alpha )}\in A _{ (1+ \alpha ) (p-1)+1}$, and the $ A_q$ classes increase in $ q$, so we conclude that $ w \in A_p \cap RH _{r} $, for a $ r > 1+ \alpha $.
The second hypothesis is $ w \in A _{1 + \frac 1 {(1+\alpha) (p'-1)}} $. This is equivalent to $$(w ^{(1-p')}) ^{ 1+\alpha } \in A _{ (1+\alpha ) (p'-1) +1 }.$$ Now, $ w ^{1- p'} = \sigma $ is the dual weight. So by the argument in the previous paragraph, $ \sigma \in RH _{r}$, for some $ r > 1+ \alpha $. So the proof is complete.
Sparse Bounds and Weighted Inequalities {#s:wtd}
=======================================
Let us recall the weighted estimates that we need for our corollaries. A function $ w >0$ is a *Muckenhoupt $ A_p$ weight* if $$[w] _{A_p} = \sup _{Q} \Bigl[\frac { w ^{\frac 1 {1-p}} (Q) } {\lvert Q\rvert }\Bigr] ^{p-1} \frac {w (Q)} {\lvert Q\rvert } < \infty .$$ Above, we are conflating $ w$ as a measure and a density, thus $w ^{\frac 1 {1-p}} (Q) =\int _{Q} w (x) ^{\frac 1 {1-p}} \;dx $. We have these estimates, which are sharp in the $ A_p$ characteristic. They are an element of the sparse proof of the $ A_2$ conjecture. (See [@MR3085756] for a proof.)
\[t:wts\] These estimates hold for all $ 1 < p < \infty $. $$\lVert \Lambda _{1,1} \;:\; L ^{p} (w) \mapsto L ^{p} (w) \rVert \lesssim [w] _{A _{p}} ^{\max \{1, \frac 1 {p-1}\}} .$$
For our applications, we have a second class of operators, a simplified form of those introduced by Benau-Bernicot-Petermichl [@160506401]. For our purposes, we need a much simplified version of their result. Define an additional characteristic of a weight, namely the *reverse Hölder* property. $$\label{e:RH}
[w] _{RH_r} = \sup _{ Q} \frac {\langle w \rangle _{Q, r}} {\langle w \rangle _{Q}}.$$
\[p:scale\] Fix an integer $ k$, and $ 1< r < 2$. We have the bound below for all $ w\in A_p$, where $ r \leq p \leq r' = \frac r {r-1}$. $$\label{e:scale}
\sum_{ Q \in \mathcal D \;:\; \lvert Q\rvert ={2 ^{nk}} }
\langle f \rangle _{Q,r} \langle g \rangle _{Q,r} \lvert Q\rvert
\lesssim [w] _{A_p} ^{1/p} [w] _{RH_r} [ \sigma ] _{RH_r} \lVert f\rVert _{L ^{p} (w)} \lVert g\rVert _{L ^{p'} (w)}$$ where $ \sigma = w ^{1 - p'}$ is the ‘dual’ weight to $ w$.
Let us recall these well known facts.
1. We always have $ [w] _{A_p}, [w] _{RH_r} \geq 1$.
2. For $ w\in A_p$ and $ \sigma = w ^{1-p'}$, the weight $ \sigma $ is locally finite, its ‘dual’ weight is $ w$, and $ [\sigma ] _{A _{p'} } = [w ] _{A_p} ^{p'-1}$.
3. For every $ w \in A _{p}$ there is a $ r = r ([w] _{A_p}) >1$ so that $ w \in RH _{r}$. (In particular, we can take $ r$ so that $ r -1 \simeq [w] _{A_p} ^{-1} $, by [@MR3092729]\*[Thm 2.3]{}. )
4. For every $ w \in A_p$, there is a $ r = r ([w] _{A_p}) >1$ so that $ w ^{r} \in A_p$.
5. We have $ w \in A_p \cap RH_r $ if and only if $ w ^{r} \in A _{r (p-1)+1}$, by [@MR1018575].
This inequality is rephrased in the self-dual way, namely setting $ \sigma = w ^{1-p'}$, it is equivalent to show that for $ k\in \mathbb Z $, $$\label{e:ss}
\sum_{ \substack{Q \in \mathcal D \\ \lvert Q\rvert = 2 ^{nk} } }
\langle f \sigma \rangle _{Q,r} \langle g w \rangle _{Q,r} \lvert Q\rvert
\lesssim
[w] ^{ \frac1p } _{A_p} [ \sigma ] _{RH_r} [ w ] _{RH_r} \lVert f\rVert _{L ^{p} (\sigma ) } \lVert g\rVert _{L ^{p'} (w)}.$$
Fix the integer $ k$. We can assume that for $ \lvert Q\rvert= 2 ^{nk} $, if $ f$ is not zero on $ Q$, then $ f \mathbf 1_{3Q \setminus Q} \equiv 0$, and we assume the same for $ g$. Then, set $$f' = \sum_{ Q \in \mathcal D\;:\; \lvert Q\rvert = 2 ^{nk} } \mathbf 1_{Q}
\Bigl[
\frac 1 { \sigma (Q) } \int _{Q} \lvert f\rvert ^{r} \; d \sigma
\Bigr] ^{1/r}$$ and likewise for $ g'$. It is immediate that $ \lVert f'\rVert _{L ^{p} ( \sigma )} \lesssim \lVert f\rVert _{L ^{p} (\sigma )}$, thus in , it suffices to assume that $ f = f'$. Then, we can even assume that $ f $ and $ g$ are supported on a single cube $ Q$, and take the value 1 on that cube.
Then, write $$\begin{aligned}
\langle \sigma \mathbf 1_{Q} \rangle _{Q,r} &\langle w \mathbf 1_{Q}\rangle _{Q,r} \lvert Q\rvert
\leq
[ \sigma ] _{RH_r} [ w ] _{RH_r}
\langle \sigma \mathbf 1_{Q} \rangle _{Q,1} \langle w \mathbf 1_{Q}\rangle _{Q,1} \lvert Q\rvert
\\
& \leq
[ \sigma ] _{RH_r} [ w ] _{RH_r}
\langle \sigma \mathbf 1_{Q} \rangle _{Q,1} ^{1/p'}\langle w \mathbf 1_{Q}\rangle _{Q,1} ^{1/p}
\cdot \sigma (Q) ^{1/p} w (Q) ^{1/p'}
\\
& \leq [ \sigma ] _{RH_r} [ w ] _{RH_r} [w] _{A_p} ^{1/p}\sigma (Q) ^{1/p} w (Q) ^{1/p'} . \end{aligned}$$ This is the inequality claimed.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present deep 3.6 cm radio continuum observations of the H II region NGC 2024 in Orion B obtained using the Very Large Array in its A-configuration, with $0\rlap.{''''}2$ angular resolution. We detect a total of 25 compact radio sources in a region of $4'' \times 4''$. We discuss the nature of these sources and its relation with the infrared and X-ray objects in the region. At least two of the radio sources are obscured proplyds whose morphology can be used to restrict the location of the main ionizing source of the region. This cluster of radio sources is compared with others that have been found in regions of recent star formation.'
author:
- 'Luis F. Rodríguez and Yolanda Gómez'
- Bo Reipurth
title: 'A Cluster of Compact Radio Sources in NGC 2024 (Orion B)'
---
INTRODUCTION
============
NGC 2024 is a prominent H II region located in the Orion B (L 1630) star-forming complex, at a distance of $\sim$415 pc (Anthony-Twarog 1982). The H II region is sharply ionization-bounded to the south (Barnes et al. 1989). With a total flux density of $\sim$60 Jy at cm wavelengths (Rodríguez & Chaisson 1978; Barnes et al. 1989), the H II region requires about $10^{48}$ ionizing photons per second to maintain its ionization, a flux that can be provided by an O9 ZAMS star with a luminosity of $\sim 5.0 \times 10^4~L_\odot$. Until recently, the required exciting star had not been identified, but Bik et al. (2003) have proposed that the infrared source IRS2b is an embedded late O or early B star responsible for the ionization.
A prominent dust lane is seen in optical images to run across the region from north to south. This dust lane is evident also in molecular line observations of the region (i. e. Schulz et al. 1991). Mezger et al. (1988) identified 6 small-scale submm condensations embedded along this dense ridge, suggesting that star formation is presently taking place here.
Clusters of near-infrared (Lada et al. 1991; Beck et al. 2003; Haisch et al. 2000), mid-infrared (Haisch et al. 2001), and X-ray sources (Skinner, Belzer, & Gagner 2003) are known to exist in association with NGC 2024. Furthermore, a number of submillimeter continuum sources were found and studied by Mezger et al. (1988) and Visser et al. (1998). To provide a more complete census of the young stellar population in this region, we have undertaken deep, high angular resolution 3.6 cm observations in an attempt to detect a cluster of compact radio sources similar to those found in Orion A (Garay, Moran, & Reid 1987; Churchwell et al. 1987), NGC 1333 (Rodríguez, Anglada, & Curiel 1999), the Arches region near the galactic center (Lang, Goss, & Rodríguez 2001), and in GGD 14 (Gómez, Rodríguez, & Garay 2000; 2002).
OBSERVATIONS
============
We have used the Very Large Array of the NRAO[^1] in its A-array configuration to observe NGC 2024 at 3.6 cm. The region was observed on 2002 March 2, 3, and 8. Our phase center was at $\alpha(J2000) = 05^h 41^m 44\rlap.{^s}9;~ \delta(J2000) = -01^\circ 55' 54{''}$. The amplitude calibrator was 1331+305, with an adopted flux density of 5.18 Jy, and the phase calibrator was 0541$-$056, with an average flux density of 0.748$\pm$0.004 Jy over the three epochs of observation. The data were analyzed in the standard manner using the package AIPS of NRAO. The data were self-calibrated in phase. Individual images were made at each epoch to search for fast variability (on a timescale of days) between the epochs observed. To diminish the presence of extended emission, in particular that originating from the very sharp ionization front of the H II region that is evident in Fig. 1b of Barnes et al. (1989), we used only visibilities with baselines larger than 100 k$\lambda$, thus suppressing the emission of structures larger than $\sim$2${''}$. The images were restored with a circular beam of $0\rlap.{''}24$, the average value of the angular resolution of the individual maps made with the ROBUST parameter of IMAGR set to 0. The three maps were then averaged to obtain an rms noise of 15 $\mu$Jy.
A total of 25 sources were detected in a region of $4' \times 4'$. The distribution of these sources is shown in Figure 1. Following Fomalont et al. (2002), we estimate that in a field of $4' \times 4'$ the *a priori number of expected 3.6 cm sources above 0.1 mJy is $\sim$0.6. We then conclude that probably one out of the 25 sources could be a background object, but that we are justified in assuming that practically all the members of the radio cluster are associated with NGC 2024. In Figure 2 we show the positions of the radio sources overlapped on the red DSS2 image of the region. From this figure it is evident that the cluster of radio sources is closely associated with the central parts of the dust lane that runs across NGC 2024.*
In Table 1 we list the positions and flux densities of the sources, averaged over the three observations. We also note in column 5 if they were found to be variable or not. The number given in parenthesis for the variable sources is the ratio between the largest and smallest flux density observed. Finally, in the last three columns we list counterparts, when found. A counterpart was taken as such if its position was within 3$''$ of the radio position. The radio positions are estimated to be accurate to $\sim 0\rlap.{''}05$.
OVERALL CHARACTERISTICS OF THE SAMPLE OF RADIO SOURCES
======================================================
Out of the 25 radio sources detected, 13 have a 2MASS counterpart and 15 have a Chandra counterpart (Skinner et al. 2003). Only the brightest source in the radio cluster, VLA 19, had been previously reported at radio wavelengths (Snell & Bally 1986; Gaume, Johnston, & Wilson 1992; Kurtz, Churchwell, & Wood 1994). Only 4 of the 25 radio sources, VLA 8, VLA 12, VLA 17, and VLA 18, lack a previously reported counterpart.
Of the 25 sources detected, 8 were found to be time variable over the observed timescale of a few days (see column 5 in Table 1). We searched for linear and circular polarization in the sample and only one source (VLA 24) showed left circular polarization at the 3% level. We measured the angular size of the sources using the AIPS task IMFIT, with a correction for bandwidth smearing. With the exception of sources VLA 8, VLA 13, and VLA 19, all sources were found to be unresolved, $\theta_s \leq 0\rlap.{''}2$.
We found radio continuum sources closely associated with three of the six submm sources of Mezger et al. (1988), namely FIR 4 (VLA 9), FIR 5 (VLA 10) and FIR 6 (VLA 14). All three sources are relatively weak and do not show time variability. Remarkably, none of these three radio sources have a near-IR or X-ray counterpart, suggesting very large extinction toward them.
Two close ($\leq 2''$) double systems, formed by sources VLA 12 and 13, as well as VLA 15 and 16, respectively, are part of the cluster. Most probably they constitute physical binaries.
COMMENTS ON SELECTED INDIVIDUAL SOURCES
=======================================
VLA 8
-----
This source has no reported counterparts. It is one of the few sources that is clearly resolved, elongated in the north-south direction (see Fig. 3). Its deconvolved dimensions are $0\rlap.{''}59 \pm 0\rlap.{''}03 \times 0\rlap.{''}25 \pm 0\rlap.{''}02;
PA = 177^\circ \pm 3^\circ$. These angular dimensions correspond to 170 AU $\times$ 95 AU. One possible explanation is that we are observing a thermal radio jet (e. g. Rodríguez 1997) that could be powering the unipolar redshifted CO jet in the region (Richer, Hills, & Padman 1992), since both the radio source and the CO jet are well aligned and elongated in the north-south direction. However, the CO jet seems to emanate from a point about 1$'$ south of VLA 8, a position much closer to VLA 10 (=FIR 5). Furthermore, the detailed shape of the source is curved, unlike most thermal jets that are rather straight. This curved morphology is reminiscent of that seen in cometary H II regions or in ionized proplyds. It is known that the ionized Orion proplyds, when observed in the radio continuum (Henney et al. 2002), show an arc-shaped structure. Furthermore, the physical dimensions of VLA 8 are similar to those of the Orion proplyd LV 2 (Henney et al. 2002). Clearly, additional observations are needed to establish if this radio source is a thermal jet, a cometary H II region, or a proplyd. As we will see below, the fact that the arc “points” to the region where the ionizing star of the region is believed to be located favors the identification of this source as a radio proplyd.
VLA 9
-----
This source has no reported near-IR or X-ray counterpart. It is associated with FIR 4, one of the small-scale submm condensations reported by Mezger et al. (1988), and is thus likely to represent a deeply embedded protostar. It is associated with an infrared reflection nebula (Moore & Yamashita 1995) and a unipolar redshifted CO outflow (Chandler & Carlstrom 1996).
VLA 10
------
This source has no reported near-IR or X-ray counterpart, but it is associated with FIR 5 (Mezger et al. 1988). It could be the powering source of the unipolar redshifted CO jet studied by Richer et al. (1992) and Chandler & Carlstrom (1996) since it is located at the position from where this jet seems to originate. Wiesemeyer et al. (1997) present interferometric 3 mm continuum observations, and show that FIR 5 separates into two clumps, FIR 5e and 5w. VLA 10 coincides precisely with FIR 5w. Lai et al. (2002) studied the detailed magnetic field structure around FIR 5.
VLA 12 and VLA 13
-----------------
These two sources (see Fig. 4) form a close binary system separated by $0\rlap.{''}9$, which at 415 pc corresponds to 375 AU in projection. Source 12 has no reported counterpart. The source VLA 13 is angularly resolved, with deconvolved dimensions of $0\rlap.{''}46 \pm 0\rlap.{''}03 \times 0\rlap.{''}12 \pm 0\rlap.{''}04;
PA = 110^\circ \pm 3^\circ$. Its arc-shaped morphology and the fact that it points to the region where the ionizing star of the region is believed to be located favors, as in the case of VLA 8, the identification of VLA 13 as a radio proplyd.
VLA 14
------
This source has no reported near-IR or X-ray counterpart. It is associated with FIR 6 (Mezger et al. 1988) and could be the powering source of a compact bipolar CO outflow found in this region (Richer 1990; Chandler & Carlstrom 1996) since the outflow emanates from a position coincident with VLA 14.
VLA 15 and VLA 16
-----------------
Both these VLA objects are 2MASS and Chandra sources. They form a close binary system (see Fig. 5) separated by $1\rlap.{''}7$. The source VLA 15 coincides positionally with the infrared source IRS2b, that has been proposed to be the ionizing source of the NGC 2024 H II region by Bik et al. (2003). The source IRS2b, first found by Jiang et al. (1984) and Nisini et al. (1994), is located about 5$''$ north-west of IRS2. IRS2 is the brightest near-IR source of this region (detected by us as VLA 19, see below). The source VLA 16 is not detected in the near-IR, but it is a Chandra source (Skinner et al. 2003).
VLA 17
------
This source has no reported counterparts. It is located exactly over the sharp ionization front that runs east-west (Barnes et al. 1989). It could be a bright knot in the ionization front, with no young star directly associated.
VLA 19
------
This is the brightest radio source in the cluster, with near-IR and X-ray counterparts. It was reported previously as a radio source by Snell & Bally (1986), Gaume et al. (1992), and Kurtz et al. (1994). We found no significant variability at the three epochs observed by us. However, our 3.6 cm flux density of 17.4 mJy is about twice the value of 8.9 mJy reported by Kurtz et al. (1994), from VLA observations taken 13 years before. This radio source coincides with the infrared source IRS2 (Grasdalen 1974; Barnes et al. 1989). The radio source (Fig. 6) clearly shows structure. There is a bright, dominant component to the west, with a fainter extension to the east. At present it is not possible to establish if we are observing a binary source, an ultracompact H II region, or a radio proplyd. Multifrequency observations at high angular resolution are needed to favor one of these possibilities.
VLA 21
------
This source is quite bright in the radio and exhibits large variation with a factor of 5.7. It has a Chandra counterpart.
VLA 24
------
This source, with 2MASS and Chandra counterparts, is the only one that showed evidence of polarization, showing left circular polarization at the 3% level at all three epochs observed. It is also time variable with variations of about a factor of 2. The combination of fast time variability with circular polarization is an indicator of gyrosynchrotron emission (Feigelson & Montmerle 1999).
DISCUSSION
==========
There are several mechanisms that can produce compact centimeter radio sources in regions of star formation. In regions of low mass star formation, thermal jets and gyrosynchrotron emitters can be present. In regions of high mass star formation we also have strong ionizing radiation available and, in addition to the two mechanisms present in low mass star-forming regions, we can have ionized stellar winds, ultracompact H II regions, and radio proplyds. It is possible to distinguish between these various possibilities with high angular resolution, multifrequency observations made with the required sensitivity. For NGC 2024 we only have available the 3.6 cm observations presented here. We can, however, argue that because of its time variability and circular polarization, VLA 24 is most probably a young low mass star showing gyrosynchrotron emission. In the case of sources VLA 2, 5, 6, 11, 20, 21, and 23, their fast time variability suggests also a gyrosynchrotron nature. Sources VLA 8 and VLA 13 are probably radio proplyds given their morphology and orientation. If so, they can be used in an attempt to search for the ionizing source of the region. These two sources are also affected by bandwidth smearing, but the deconvolution made correcting for this effect clearly shows that they are truly extended. In Figure 7 we show an image that includes these two sources as well as other nearby sources. In this image we have passed two lines by each of the sources. These lines have position angles of $\pm$3 $\sigma$ with respect to the perpendicular to the major axes of the sources. The set of lines defines a four-sided region that includes VLA 15 (= IRS2b) and VLA 19 (=IRS2), as can be seen in Fig. 7. Given the uncertainties of the method, we cannot favor either of the sources conclusively. We note, however, that the morphology of VLA 19 (=IRS2) is suggestive also of a proplyd nature (see discussion in section 4.8 as well as Fig. 6). If this is the case, then the object VLA 15 (= IRS2b) is clearly favored as the ionizing source of the region, in agreement with Bik et al. (2003). Over the years, a few examples of radio clusters associated with regions of recent star formation have been appearing in the literature (Garay et al. 1987; Churchwell et al. 1987; Becker & White 1988; Stine & O’Neal 1998; Rodríguez et al. 1999; Gómez et al. 2000; Lang et al. 2001). In Table 2 we summarize the parameters of these radio clusters. Although more such clusters should be studied to have a reliable statistical base, some interesting trends are evident. The diameters of the radio clusters are in the 0.2 to 0.7 pc range. The radio luminosity of the most luminous member is correlated with the bolometric luminosity of the cluster. We believe that sensitive, high angular resolution studies similar to that presented here are needed to establish if these clusters are always present in other relatively nearby regions of massive star formation. If properly understood, these radio clusters could be an invaluable observational tool to study the stellar population of heavily obscured regions of star formation.
CONCLUSIONS
===========
We have presented sensitive, high angular resolution ($0\rlap.{''}2$) VLA observations at 3.6 cm toward the NGC 2024 region of recent star formation. Our main conclusions are summarized below.
1\. We detected a total of 25 compact radio sources in a region of $4' \times 4'$. Only four of these sources do not have a previously reported counterpart at any wavelength. However, only one of the radio sources had been reported in the literature.
2\. We found radio continuum sources closely associated with three of the six submm sources in the region. Remarkably, none of these three radio/submm sources has a near-IR or X-ray counterpart, suggesting very large extinction toward them.
3\. Two of the sources (VLA 8 and VLA 13) are proposed to be radio proplyds whose study may help pinpoint the origin of the ionizing radiation in the region. Our attempts to do this suggest a region containing both VLA 15 (= IRS2b) and VLA 19 (=IRS2) as the area where the ionizing source is located.
4\. The source VLA 19, the counterpart of IRS2, is found to exhibit spatial structure, although its precise nature remains undetermined. We also detected a radio counterpart to IRS2b, the source that has been recently proposed to be ionizing NGC 2024.
5\. Eight of the sources detected (VLA 2, 5, 6, 11, 20, 21, 23, and 24) show fast time variability and probably are young low mass stars exhibiting gyrosynchrotron emission.
6\. Two close ($\leq 2''$) binary systems, formed by sources VLA 12 and 13 and VLA 15 and 16, respectively, are part of the cluster.
7\. We have briefly discussed the parameters of other radio clusters found in regions of star formation and suggest that sensitive, high angular resolution studies of other relatively nearby massive star formation regions are likely to detect similar clusters.
We thank Will Henney for his comments on proplyds and Steve Skinner for providing us with his list of X-ray sources before publication. We are also grateful to George Herbig for calling our attention to the radio cluster in NGC 1579. LFR and YG acknowledge the support of DGAPA, UNAM, and of CONACyT (México). BR acknowledges support from NASA grant NAG5-8108. The Second Palomar Observatory Sky Survey (POSS-II) was made by the California Institute of Technology with funds from the National Science Foundation, the National Aeronautics and Space Administration, the National Geographic Society, the Sloan Foundation, the Samuel Oschin Foundation, and the Eastman Kodak Corporation. The Oschin Schmidt Telescope is operated by the California Institute of Technology and Palomar Observatory. The 2MASS project is a collaboration between The University of Massachusetts and the Infrared Processing and Analysis Center (JPL/ Caltech), with funding provided primarily by NASA and the NSF. This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France.
Anthony-Twarog, B. J. 1982, AJ, 87, 1213
Barnes, Peter J., Crutcher, R. M., Bieging, J. H., Storey, J. W. V., & Willner, S. P. 1989, ApJ, 342, 883
Beck, T. L., Simon, M., & Close, L. M. 2003, ApJ, 583, 358
Becker, R. H. & White, R. L. 1988, ApJ, 324, 893
Bik, A., Lenorzer, A., Kaper, L., Comerón, F., Waters, L. B. F. M., de Koter, A., & Hanson, M. M. 2003, A&A, 404, 249
Chandler, C.J., & Carlstrom, J.E. 1996, ApJ, 466, 338
Churchwell, E., Wood, D. O. S., Felli, M., & Massi, M. 1987, ApJ, 321, 516
Feigelson, E. D. & Montmerle, T. 1999, ARAA, 37, 363
Fomalont, E. B., Kellermann, K. I., Partridge, R. B., Windhorst, R. A., & Richards, E. A. 2002, AJ, 123, 2402
Garay, G., Moran, J. M., & Reid, M. J. 1987, ApJ, 314, 535
Gaume, R. A., Johnston, K. J., & Wilson, T. L. 1992, ApJ, 388, 489
Gómez, Y., Rodríguez, L. F., & Garay, G. 2000, ApJ, 531, 861
Gómez, Y., Rodríguez, L. F., & Garay, G. 2002, ApJ, 571, 901
Grasdalen, G.L. 1974, ApJ, 193, 373
Haisch, K. E., Lada, E. A., & Lada, C. J. 2000, AJ, 120, 1396
Haisch, K. E., Lada, E. A., Piña, R. K., Telesco, C. M., & Lada, C. J. 2001, AJ, 121, 1512
Henney, W. J., O’Dell, C. R., Meaburn, J., Garrington, S. T., & Lopez, J. A. 2002, ApJ, 566, 315
Jiang, D.R., Perrier, C., & Léna, P. 1984, A&A, 135, 249
Kurtz, S., Churchwell, E., & Wood, D. O. S. 1994, ApJS, 91, 659
Lada, E. A.; Evans, N. J., Depoy, D. L., & Gatley, I. 1991, ApJ, 371, 171
Lai, S.-P., Crutcher, R. M., Girart, J. M., & Rao, R. 2002, ApJ, 566, 925
Lang, C. C., Goss, W. M., & Rodríguez, L. F. 2001, ApJ, 551, L143
Menten, K. M. & Reid, M. J. 1995, ApJ, 445, L157
Mezger, P. G., Chini, R., Kreysa, E., Wink, J. E., & Salter, C. J. 1988, A&A, 191, 44
Moore, T. J. T., & Chandler, C. J. 1989, MNRAS, 241, 19
Moore, T.J.T., & Yamashita, T. 1995, ApJ, 440, 722
Nisini, B., Smith, H.A., Fischer, J., & Geballe, T.R. 1994, A&A, 290, 463
Richer, J. S., Hills, R. E., & Padman, R. 1992, MNRAS, 254, 525
Richer, J. S. 1990, MNRAS, 245, 24
Rodríguez, L. F. & Chaisson, E. J. 1978, ApJ, 221, 816
Rodríguez, L.F. 1997, in IAU Symp. No. 182 [*Herbig-Haro Flows and the Birth of Low Mass Stars*]{}, Eds. B. Reipurth & C. Bertout, Kluwer, p. 83
Rodríguez, L.F., Anglada, G., & Curiel, S. 1999, ApJS, 125, 427
Schulz, A., Guesten, R., Zylka, R., & Serabyn, E. 1991, A&A, 246, 570
Skinner, S., Belzer, E., & Gagne, M. 2003, ApJ, submitted.
Snell, R. L. & Bally, J. 1986, ApJ, 303, 683
Stine, P. C. & O’Neal, D. 1998, AJ, 116, 890
Visser, A.E., Richer, J.S., Chandler, C.J., & Padman, R. 1998, MNRAS, 301, 585
Wiesemeyer, H., Güsten, R., Wink, J.E., & Yorke, H.W. 1997, A&A, 320, 287
[l c c c c c c c]{} 1 & 05 41 37.745 & -01 54 41.20 & 0.30 & N & & & HLP 2\
2 & 05 41 37.858 & -01 54 31.82 & 0.16 & Y(1.9) & 05413786-0154323 & SGB 78 &\
3 & 05 41 41.346 & -01 53 32.54 & 0.28 & N & 05414134-0153326 & SGB 120 & HLL 52, BSC 69\
4 & 05 41 41.385 & -01 54 44.64 & 0.26 & N & 05414138-0154445 & SGB 123 &\
5 & 05 41 41.487 & -01 54 39.35 & 0.33 & Y(3.7) & 05414148-0154390 & SGB 124 & HLL 87, BSC 65\
6 & 05 41 43.446 & -01 56 42.74 & 0.10 & Y(6.5) & 05414344-0156425 & SGB 147 & BCB IRS24\
7 & 05 41 43.559 & -01 53 56.67 & 0.20 & N & 05414356-0153567 & SGB 152 &\
8 & 05 41 43.912 & -01 54 34.13 & 0.63 & N & & &\
9 & 05 41 44.136 & -01 54 46.04 & 0.33 & N & & & M FIR 4, MC89-4\
10 & 05 41 44.221 & -01 55 41.32 & 0.11 & N & & & M FIR 5, LCGR FIR 5 4\
11 & 05 41 44.827 & -01 54 25.16 & 0.22 & Y(1.6) & 05414482-0154251 & SGB 171 & HLL 77, HLP 22\
12 & 05 41 45.024 & -01 54 53.94 & 0.11 & N & & &\
13 & 05 41 45.056 & -01 54 54.73 & 0.24 & N & 05414504-0154546 & &\
14 & 05 41 45.168 & -01 56 00.56 & 0.29 & N & & & M FIR 6, LCGR FIR 6 n\
15 & 05 41 45.492 & -01 54 28.70 & 0.24 & N & 05414550-0154286 & SGB 182 & B IRS2b\
16 & 05 41 45.522 & -01 54 27.02 & 0.97 & N & & SGB 183 &\
17 & 05 41 45.554 & -01 55 25.86 & 0.27 & N & & &\
18 & 05 41 45.733 & -01 54 57.93 & 0.49 & N & & &\
19 & 05 41 45.809 & -01 54 29.92 & 17.4 & N & 05414580-0154297 & SGB 187 & IRS2, KCW 206.543-16.347\
20 & 05 41 45.905 & -01 54 10.98 & 0.18 & Y(5.0) & & SGB 188 & BSC 93\
21 & 05 41 46.157 & -01 56 22.20 & 4.54 & Y(5.7) & & SGB 193 &\
22 & 05 41 46.236 & -01 56 54.80 & 0.10 & N & & SGB 196 &\
23 & 05 41 46.570 & -01 54 46.88 & 1.17 & Y(6.2) & 05414655-0154469 & SGB 200 &\
24 & 05 41 48.224 & -01 56 02.01 & 8.60 & Y(2.2) & 05414821-0156020 & SGB 210 &\
25 & 05 41 50.983 & -01 55 06.97 & 0.24 & N & 05415096-0155070 & & HLP 29, HLL 98\
[ l c c c c c ]{} Orion & $\sim 2 \times 10^5$ & $\sim$50 & 8.8 & 0.3 & Menten & Reid (1995)\
NGC 1579 & $\sim 2 \times 10^3$ & 16 & 0.6 & 0.3 & Stine & O’Neal (1998)\
GGD 14 & $\sim 10^4$ & 6 & 0.2 & 0.2 & Gómez et al. (2002)\
NGC 1333 & $\sim 120$ & 44 & 0.3 & 0.7 & Rodríguez et al. (1999)\
Arches & $\sim 10^7$ & 8 & 122.8 & 0.7 & Lang et al. (2001)\
NGC 2024 & $\sim 5 \times 10^4$ & 25 & 3.0 & 0.5 & This paper\
[^1]: The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The exclusive charge exchange reaction $pD\to n(pp)$ at intermediate and high energies is studied within the Bethe-Salpeter formalism. The final state interaction in the detected at zero-near excitation energy $pp$-pair is described by the $^1S_0$ component of the Bethe-Salpeter amplitude. Results of numerical calculations of polarization observables and differential cross-section persuade that, as in the non relativistic case, this reaction can be utilized for a relativistic deuteron tensor polarimeter and as a source of information about the elementary nucleon-nucleon charge exchange amplitude.'
---
-1cm
s0[[\^1S\_0\^[++]{}]{}]{}
[ **Final State Interaction within the Bethe-Salpeter Approach in Charge Exchange $pD\to n(pp)$ process.** ]{}\
S.S. Semikh$^{1 \dag}$, L.P. Kaptari$^{1}$, S.M. Dorkin$^{2}$, and B. Kämpfer$^{3}$
[(1) [*BLTP, Joint Institute for Nuclear Research, Dubna, Russia* ]{}\
(2) [*SNPI, MSU, Dubna, Russia* ]{}\
(3) [*Forschungszentrum Rossendorf, Germany* ]{}\
$\dag$ [*E-mail: [email protected]* ]{}]{}
Introduction {#introd}
============
Nowadays large programs of experimental study of processes with polarized particles are in progress. The setups with deuteron targets (beams) occupy the leading place [@alexa; @kox0; @preliminar; @cosy_proposal]. For an investigation of the $NN$ interaction in the deuteron at short distances the three deuteron form factors, magnetic, electric and quadrupole, should be determined. In the elastic $eD$-scattering with unpolarized particles one can measure only two quantities, e.g. the magnetic form factor and the deuteron function $A(Q^2)$, which is a kinematical combination of all three form factors. Even these two quantities provide important information about the quark physics and dynamics at short distances (see, for instance recent measurements [@alexa] at TJNAF). However, for a full determination of the deuteron form factors, one needs measurements with polarized particles. For example, measurements of the tensor analysing power $T_{20}$ of recoil deuterons in elastic $eD$-scattering allow the determination of the charge form factor $G_c$ at high transferred momenta. The hadron-deuteron processes can be considered as complementary tools in investigation of phenomena at short distances and also as a source of unique information unavailable in electromagnetic reactions (study of nucleon resonances, checking the non relativistic effective models, $NN$ potentials etc.). Experimental and theoretical investigation of the proton-deuteron processes at intermediate and high energies has started some decades ago by studying elastic $pD$ scattering [@elsatic], exclusive and inclusive break-up [@inclusive; @cosy]. In $pD$ processes it is possible to completely restore the reaction amplitude by measuring a full set of polarization observables (see, e.g. [@rekalo; @ladygin; @ourprc]). Hence, as in the electromagnetic case one needs to measure different polarizations of the recoil deuteron. Since polarization observables can be studied only by an additional secondary scattering of the reaction products (in polarimeter), it is obvious that second process must possess a high enough cross section to assure a good efficiency of the polarimeter. In ref. [@wil1] Bugg and Wilkin have proposed as an effective deuteron polarimeter the process $p{\vec D}\to (pp)n$ where the final $pp$-pair is detected with extremely low excitation energy (see also ref. [@polder]). Later investigations [@ishida; @wil1; @moto; @wil2; @kox] confirmed the theoretical predictions and the charge exchange processes were suggested also for use in investigation of a number of reactions with deuterons, e.g. $pp\to D\pi^+$ [@bugg], $NN\pi$-systems, inelastic $({\vec D},{\vec D'})$-reactions off heavy nuclei to study isoscalar transitions $\Delta T = 0,\, \Delta S = 1$ [@morlet] etc. The direct consequence of these facts is that nowadays the interest in investigation of charge exchange processes does not abate. In our previous paper [@nash_yaf] we investigated the process $pD\to n(pp)$ within the impulse approximation. The goal of the present work is to consider theoretically the effects of the final state interaction in this reaction at relativistic initial energies (COSY, Dubna) and to confess whether in this case the non relativistic predictions [@wil1] hold and the reaction can be still regarded as a deuteron polarimeter tool. We propose a covariant generalization of the spectator mechanism [@wil1] based on the Bethe-Salpeter formalism and on numerical solution of the Bethe-Salpeter (BS) equation with a realistic one-boson exchange kernel [@solution; @parametrization]. Here we focus our attention in calculations of the cross section and the tensor analysing power $T_{20}$ within the COSY kinematical conditions at, as in non relativistic case, zero vector analyzing powers of the deuteron.
Kinematics and the invariant amplitude {#gl1}
======================================
We select those processes which, in the deuteron center of mass system, correspond to final states with one fast neutron and a slowly moving proton-proton pair, i.e. reactions of the type $$p\,+\,\vec D
\,=\, n + (p_1+p_2). \label{reaction}$$ The transferred momentum from the proton to the neutron is low, hence the main mechanism of the reaction can be described as a charge exchange process of the incoming proton off the internal neutron whereas the proton in the deuteron remains merely as a spectator. If so, then the resulting $pp$-pair will be detected with low total and relative momenta. In Fig. \[pict1\] the diagram of such processes is schematically depicted. The following notations are adopted: $p=(E_p,\bp)$ and $n=(E_n,\bn)$ are the 4-momenta of the incoming proton and outgoing neutron, $P'$ is the total 4-momentum of the $pp$-pair, which is a sum of the corresponding 4-momenta of detected protons, $p_1=(E_1,\bp_1)$, $p_2=(E_2,\bp_2)$: $P'=p_1+p_2$. The notion of the invariant mass of the pair $s_f,~s_f=P'^2=(2m+E_x)^2$, where $m$ stands for the nucleon mass and $E_x$ for the excitation energy of the pair, is also explored in what follows. The excitation energy $E_x$ ranges from zero to few $MeV$, $E_x\sim 0-8~MeV$. At such low values of $E_x$ the main contribution to the final state of the $pp$ pair in the continuum comes from the the $^1S_0$-configuration [@wil2]. The differential cross section for the reaction (\[reaction\]) reads [@nash_yaf]: $$\frac{d^2\sigma}{dt\,ds_f}\,=\,\frac{1}{2}\frac{1}{64\pi\lambda(p,D)}\,
\sqrt{1-\frac{4m^2}{s_f}}\,\frac{1}{(2\pi)^2}\,|M_{fi}|^2.
\label{cross3}$$ The experimental data from Saclay [@kox] are binned into the following intervals of excitation energy: $$\begin{aligned}
I: && \quad 0\leq E_x \leq 1~ MeV, \nonumber\\
II: && \quad 1~MeV\leq E_x \leq 4~ MeV,\nonumber \\
III: && \quad 4~MeV\leq E_x \leq 8~MeV.\nonumber\end{aligned}$$ To compare the cross section (\[cross3\]) with the experimental data it is necessary to integrate over the invariant mass of the pair in the regions $I$, $II$ and $III$ according to $$\begin{aligned}
\label{crint}
\left(\frac{d\sigma}{dt}\right)_{I,II,III}\,=\,
\frac{1}{(8\pi)^3\lambda}\,
\int\limits_{I,II,III}\,ds_p\,\sqrt{1-\frac{4m^2}{s_f}}\,|M_{fi}|^2.\end{aligned}$$ By using the Mandelstam technique [@mandel] the covariant matrix element corresponding to the diagram on Fig. \[pict1\] can be written in the form $$\begin{aligned}
\nonumber
{{\cal T}}_{r'r}^M &=& \sum\limits_{ss'}\frac{1}{(2m)^2}\int d^4k\,
f_{r's',sr}\times\\
&\times&{\bar u}^{s}(p_n) {\Psi}_M(k)\,(\frac{\hat D}{2}-{\hat k}+m)
{\bar\Psi}_{P'}(k-\frac{q}{2})\,u^{s'}(p_p).
\label{fin}\end{aligned}$$ The elementary charge-exchange amplitude ${{\cal A}}^{ce}$ is incorporated in the matrix element (\[fin\]) by the on-shell amplitudes $f_{r's',sr}$ and the Dirac spinors. In doing so the off-shell effects are neglected and the elementary subprocess is considered as real process with on-shell particles. In our numerical calculations we use the helicity amplitudes of pn scattering, resulting from both the Nijmegen partial wave analysis [@swart; @www] and the well-known results of SAID [@said]. For the final $^1S_0$-state within the $\rho$-classification the BS amplitude ${\bar\Psi}_{P'}$ in the center of mass of the $NN$ pair is represented by four partial amplitudes $^1S_0^{++}$, $^1S_0^{--}$, $^3P_0^{+-}$ and $^3P_0^{-+}$ [@nashi], which for the sake of brevity are denoted as $\phi_1,\dots,\phi_4$. The partial amplitudes $\phi_i$ may be found from the BS equation, which, in the simplest case of pseudo scalar exchanges reads as $$\label{neodn}
{\bar\Psi}_{P'}(p)={\bar\Psi}_{P'}^0(p)+ig^2_{\pi
NN}\int\frac{d^4p'}{(2\pi)^4}\,
\Delta(p-p'){\tilde S}(p_2)\gamma_5{\bar\Psi}_{P'}(p')\gamma_5 S(p_1),$$ where $\Delta$ and $S$ are the scalar and spinor propagators, respectively, ${\tilde S}\equiv U_C\,S\,U_C^{-1}$, and ${\bar\Psi}_{P'}^0(p)$ is the free amplitude corresponding to two non interacting nucleons (the relativistic plane wave). The solution of eq. (\[neodn\]) may be presented as a Neuman-like series, the first term of which is the free term from eq. (\[neodn\]): $$\label{rash}
{\bar\Psi}_{P'}(p)={\bar\Psi}_{P'}^0(p)+{\bar\Psi}_{P'}^i(p).$$ The second part in eq. (\[rash\]) is entirely determined by the interaction and may be symbolically referred to as scattered wave. To determine the scattered wave in eq. (\[rash\]) it is necessary to solve the BS equation of the type (\[neodn\]). Solving the BS equation in the continuum is a much more cumbersome procedure than, e.g. for the homogenous equation. Besides difficulties encountered in solving the latter (singularities of amplitudes, poles in propagators, cuts etc.) the former even does not allow the Wick rotation [@wick] to the Euclidian space, and in the Minkowsky space there are no rigourous mathematical methods of finding solutions [^1]. However, an approximate solution of eq. (\[neodn\]) may be obtained by applying the so-called “one-iteration approximation” [@ourprc]. Within that one may obtain a rather good estimate of the interaction term.
The one-iteration approximation
===============================
For a consistent relativistic analysis of the reaction (\[reaction\]) one should solve the BS equation for both bound state and scattering state with the same interaction kernel. We have found a numerical solution for the deuteron bound state with a realistic one-boson exchange potential [@solution]. The Bete-Salpeter equation, after a partial decomposition, has been solved numerically by using an iteration method. We found that the iteration procedure converges rather quickly if the trial function is properly chosen. In such case even after the first iteration the BS solution coincides with the exact one up to relative momentum $p\sim 0.6-0.7~GeV/c$. This circumstance becomes useful if one needs an approximate solution of the BS equation at not too large momenta $p \leq 0.5-0.7~ GeV/c$. This is just our case, since in reaction (\[reaction\]) the relative momentum of the $pp$-pair is expected to be rather small and the scattering part of the amplitude (\[rash\]) can be obtained from the equation (\[neodn\]) by one iteration. To solve eq. (\[neodn\]) we proceed as follows (see also ref. [@ourprc]): i) for simplicity, in the inhomogeneous BS equation we leave only the pseudo scalar isovector exchanges ($\pi$-mesons) ii) by disregarding the dependence upon $p_0$ in the meson propagator in eq. (\[neodn\]) and then using the standard representation of propagators via generalized Legandre polynomials $Q_l$, one obtains for the main partial amplitude $$\begin{aligned}
\nonumber
&&
\phi_\1s0(p_0,|\bp|)=\phi_\1s0^0(p_0,|\bp|)-\frac{g_{\pi NN}^2}{4\pi}
\frac{1}
{\left(\frac{\sqsf}{2}-E_p\right)^2-p_0^2}\times\\
&&\int\limits_0^\infty\frac{d|\bp'|}{2\pi}
\,\frac{|\bp'|}{|\bp|}\,\frac{1}{E_pE_{p'}}\left[(E_pE_{p'}-m^2)
Q_0({\tilde y}_\mu)-|\bp||\bp'|Q_1({\tilde y}_\mu)\right]
u_{^1\!S_0}(s_f,|\bp'|),
\label{before_int}\end{aligned}$$ where ${\tilde y}_\mu=\displaystyle\frac{\bp^2+\bp^{'2}+\mu^2}{2|\bp||\bp'|}$. In obtaining (\[before\_int\]) the integration over $p_0'$ has been carried out in the residium ${\tilde p}_0=\displaystyle\frac{\sqsf}{2}-E_{p'}$ and, by definition, the BS wave function in the continuum is $$\begin{aligned}
u_{^1\!S_0}(s_f,|\bp'|)=\frac{g_\1s0({\tilde p}_0, |\bp'|)}{\sqsf-2E_{p'}},
\label{u_contin}\end{aligned}$$ where $g_\1s0$ is the vertex function (for more details see [@ourprc; @nashi]). Now if we restrict ourself to only one iteration in (\[before\_int\]) and take the trial function (\[u\_contin\]) as a non relativistic solution of the Schrödinger equation, e.g. the Paris wave function $u^{NR}_{^1S_0}(s_f,|\bp'|)$, the BS amplitude is obtained by formula $$\begin{aligned}
\phi_\1s0(p_0,|{\bf p}|)=\phi_\1s0^0(p_0,|{\bf p}|) -
\frac{G^{o.i.}(\tilde p_0,|{\bf p}|)}{\left(\frac{\sqsf}{2}-E_p\right)^2-p_0^2},
\label{foi}\end{aligned}$$ where the “one-iteration” BS vertex $G^{o.i.}(\tilde p_0,|{\bf p}|)$ is defined as $$\begin{aligned}
&&
G^{o.i.}(\tilde p_0,|{\bf p}|)
=\frac{1}{\pi}\,\frac{g_{\pi NN}^2}{4\pi}
\left\{ \left[1-\frac{E_p}{m}\right]
\int\limits_0^\infty dr\, e^{-\mu r}j_0(pr)\,u^{NR}_{^1\!S_0}(r)
+\right.
\nonumber\\
&&
\left .\frac{|\bp|}{m E_p} \int\limits_0^\infty dr\,
\frac{u^{NR}_{^1\!S_0}(r)}{r}\,e^{-\mu r}\,(1-\mu r)\,j_1(pr)
\right\}.
\label{fin_oi}\end{aligned}$$ Now from eqs. (\[foi\]) and (\[fin\_oi\]) one may easily find the non relativistic analogue of the obtained formulae. So, the free term in eq. (\[foi\]) together with the first term in eq. (\[fin\_oi\]) reflect the non relativistic equation for the $^1S_0$ wave function, while the second term in (\[fin\_oi\]) turns out to be corrections of pure relativistic origin.
Numerical results
=================
In Figs. \[pict7\] and \[pict8\] we present the results of numerical calculations of the cross section $\displaystyle\frac{d\sigma}{dt}$ and tensor analysing power $T_{20}$. The elementary charge exchange amplitude has been taken from ref. [@saidpap] and the non relativistic trial function $u^{NR}_{^1S_0}(r)$ is the solution of the Schrödinger equation with the Paris potential [@paris_cont]. The Bethe-Salpeter amplitudes are those from the numerical solution [@solution] obtained with a realistic one-boson exchange interaction. The dashed lines in Figs. \[pict7\] and \[pict8\] correspond to results within the relativistic impulse approximation [@nash_yaf], while the solid lines denote results with taking into account the final state interaction in one-iteration approximation (as it is described above). It is seen that in all three energy bins the agreement with data for the cross section is essentially improved. Especially it concerns the range $1\leq E_x\leq 4$. For the energy bin close to zero there is still a disagreement with data at low transferred momenta which, probably may be addressed to the fact that in our calculations we have not taken into account the Coulomb interaction in the $pp$-pair. For higher excitation energies ($E_x\sim 8 ~MeV$), other partial waves (e.g. triplet state ) in the $pp$-final state contribute and, within the adopted assumptions one may expect only qualitative agreement with data. From the Fig. \[pict7\] one may conclude that at low excitation energies the supposed mechanism for the reaction (\[reaction\]) (charge exchange subprocess with interaction in $^1S_0$ state of the $pp$-pair in the continuum) seems to be correct. Moreover, from a comparison of the left and right panels in Fig. \[pict7\] one may expect that for the higher initial energy there is larger kinematical region where the mechanism holds. Figure \[pict8\] demonstrates that the tensor analysing power is less sensitive to final state interaction effects. As a matter of fact, the tensor analysing power, being a ratio of non diagonal products of partial amplitudes to the diagonal ones, serves as a measure of the quality of parametrization of partial amplitudes and their mutual relative phases. This has been pointed out in a series of publications (see e.g. refs.[@inclusive; @ourprc]), where a good simultaneous description of cross sections and $T_{20}$ in reactions of the deuteron break-up or elastic scattering from protons, is still lucking. Nevertheless, since in the process (\[reaction\]) the behaviour of the partial amplitudes is mostly governed by the elementary charge exchange ones, an experimental investigation of the tensor analysing power $T_{20}$ in reactions of the type (\[reaction\]) can essentially supplement data on the $NN$ charge exchange amplitudes at high energies. In Figs. \[pict9\] and \[pict10\] we present the predicted cross section and tensor analysing power at high energies (COSY, Dubna). It is immediately seen that the cross section is substantially decreasing with the energy increasing, nevertheless remains large enough to be experimentally investigated. Another peculiarity of the studied process at relativistic energies is that the tensor analysing power $T_{20}$ does not change the sign remaining positive in a large kinematical region, in contrast to the lower energies (cf. Fig. \[pict8\]). Note again, that in the above calculations the vector analyzing power of the deuteron is strictly zero.
From the performed analysis one can conclude that there is a kinematical region for the excitation energy, $E_x <5~ MeV$, and transferred momentum, $|{\bf q}| \le\, 0.3\div 0.4\,~ GeV/c$ (the COSY [@cosy_proposal] kinematics), for which the mechanism of the reaction (\[reaction\]) is fairly well described within the spectator approach by an elementary $pn$ charge exchange subprocess, for active nucleons, with detection of the $pp$-pair in the $^1S_0$ final state. Our covariant approach agrees with previous non relativistic calculations and allows predictions of the cross sections and polarization observables at intermediate and relativistic energies, in particular, for kinematical conditions achievable at COSY. The predicted cross sections and tensor analysing power $T_{20}$ are large enough for the process (\[reaction\]) to be used, in a large range of initial energies, for the determination of properties of polarized deuteron, provided experimentally one simultaneously detects vanishing vector polarization of deuterons.
Summary {#gl4}
=======
In summary, the performed covariant analysis of the reactions $\vec D(p,n)pp$ with two protons in the $^1S_0$ final state allows to conclude that, as in the non relativistic limit, such process can be used as an effective deuteron polarimeter also at relativistic energies, achieved at COSY and Dubna. The effects of final state interaction are found to be substantial and essentially improve the agreement with data.
Acknowledgments
===============
This work was performed in parts during the visits of S.S.S and L.P.K. in the Forschungszentrum Rossendorf, Institute of Nuclear and Hadron Physics. We thank for the support by the program “Heisenberg-Landau” of JINR-FRG collaboration and the grants 06DR921, WTZ RUS 98/678 and RFBR 00-15-96737.
[99]{} L.C. Alexa, B.D. Adrson, K.A. Aniol et al., Phys. Rev. Lett. [**82**]{}, 1374 (1999). S. Kox, E.J. Beise (spokespersons), TJNAF experiments 94-018 “Measurement of the Deuteron Polarization at Large Momentum Transvers in $D(e,e')D$ Scattering”; Nucl. Phys. [**A684**]{} 521, (2001). E. Tomasi-Gustafsson, in Proc. of [*XIV International Seminar On High Energy Physics Problems*]{}, Preprint JINR No. E1,2-2000-166 (Dubna, 2000). V.I. Komarov (spokesman) et al., COSY proposal \#20 updated from 1999, “Exclusive deuteron break-up study with polarized protons and deuterons at COSY” http://ikpd15.ikp.kfa-juelich.de:8085/doc/Proposals.html M.P. Rekalo, I.M. Sitnik, Phys. Lett. [**B 356**]{}, 434 (1995);\
L.S. Azhgirei et al., Phys. Lett. [**B 361**]{}, 21 (1995). V.G. Ableev et al., Nucl. Phys. [**A 393**]{}, 491 (1983);\
C.F. Perdrisat, V. Punjabi, Phys. Rev. [**C 42**]{}, 1899 (1990);\
B. Kühn, C.F. Perdrisat, E.A. Strokovsky, Phys. Lett. [**B312**]{}, 298 (1994). A.K. Kacharava et al., Preprint JINR No. E1-96-42 (Dubna, 1996); C.F. Perdrisat (spokesperson) et al., COSY proposal \#68.1 ”Proton-to-proton polarization transfer in backward elastic scattering”. M.P. Rekalo, N.M. Piskunov, I.M. Sitnik, Few Body Syst. [**23**]{}, 187 (1998). V.P. Ladygin, N.B. Ladygina, J. Phys. [**G 23**]{}, 847 (1997). L.P. Kaptari, B. Kämpfer, S.M. Dorkin, S.S. Semikh, Phys. Rev. [**C57**]{}, 1097 (1998). D.V. Bugg, C. Wilkin, Phys. Lett. [**B152**]{}, 37 (1985);\
D.V. Bugg, C. Wilkin, Nucl. Phys. [**A467**]{}, 575 (1987). S. Kox et. al., Nucl. Instrum. Methods [**A346**]{}, 527 (1994). Y. Satou, S. Ishida, H. Sakai, H.Okamura, et al., Phys. Lett. [**B521**]{}, 153, (2001). T. Motobayashi et al., Phys. Lett. [**B233**]{}, 69 (1989). J. Carbonell, M. Barbaro, C. Wilkin, Nucl. Phys. [ **A529**]{}, 653 (1991). S. Kox et. al., Nucl. Phys. [**A556**]{}, 621 (1993). D.V. Bugg, A. Hasan, R.L. Shypit, Nucl. Phys. [ **A477**]{}, 546 (1988);\
C. Furget et. al., Nucl. Phys. [**A631**]{}, 747 (1998). M. Morlet et. al., Phys. Lett. [**B247**]{}, 228 (1990). S.M. Dorkin, L.P. Kaptari, B. Kämpfer, S.S. Semikh, Phys. Atom. Nucl. 65 (2002) 442; Yad. Fiz. 65 (2002) 469; see also nucl-th/0012088. J. Fleischer, J. Tjon, Nucl. Phys. [**B84**]{}, 375 (1975); Phys. Rev. [**D15**]{}, 2537 (1975)\
M. Zuilhof, J. Tjon, Phys. Rev. [**C22**]{}, 2369 (1980). A.Yu. Umnikov, L.P. Kaptari, F.C Khanna, Phys. Rev. [**C56**]{}, 1700 (1997);\
A.Yu. Umnikov, L.P. Kaptari, K.Yu. Kazakov, F.C. Khanna, Phys. Lett. [**B334**]{}, 163 (1994). A.Yu. Umnikov, Z. Phys. [**A357**]{}, 333 (1997). S. Mandelstam, Proc. Roy. Soc. (London) [**A233**]{}, 123 (1955). S.G. Bondarenko, V.V. Burov, M. Beyer, S.M. Dorkin, Phys. Rev. [**C58**]{}, 3143 (1998). G.C. Wick, Phys. Rev. [**96**]{}, 1124 (1954). http://nn-online.sci.kun.nl http://said.phys.vt.edu V.G.J. Stoks, R.A.M. Klomp, M.C.M. Rentmeester, J.J. de Swart, Phys. Rev. [**C48**]{}, 792 (1993). R.A. Arndt, I.I. Strakovsky, R.L. Workman, Phys. Rev. [**C62**]{}:034005, 2000 M. Lacombe, B. Loiseau, J.M. Richard, R. Vinh Mau et al., Phys. Rev. [**C21**]{}, 861 (1980)
[^1]: Actually there is one realistic solution of the inhomogeneous BS equation in the ladder approximation, obtained by Tjon [@Tjonsol].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We propose a novel space-division based network-coding scheme for multiple-input multiple-output (MIMO) two-way relay channels (TWRCs), in which two multi-antenna users exchange information via a multi-antenna relay. In the proposed scheme, the overall signal space at the relay is divided into two subspaces. In one subspace, the spatial streams of the two users have nearly orthogonal directions, and are completely decoded at the relay. In the other subspace, the signal directions of the two users are nearly parallel, and linear functions of the spatial streams are computed at the relay, following the principle of physical-layer network coding (PNC). Based on the recovered messages and message-functions, the relay generates and forwards network-coded messages to the two users. We show that, at high signal-to-noise ratio (SNR), the proposed scheme achieves the asymptotic sum-rate capacity of MIMO TWRCs within $\frac{1}{2}%
\log(5/4)\approx 0.161$ bits per user-antenna for any antenna configuration and channel realization. We perform large-system analysis to derive the average sum-rate of the proposed scheme over Rayleigh-fading MIMO TWRCs. We show that the average asymptotic sum-rate gap to the capacity upper bound is at most 0.053 bits per relay-antenna. It is demonstrated that the proposed scheme significantly outperforms the existing schemes.
author:
- '[^1]'
title: 'MIMO Two-Way Relaying: A Space-Division Approach'
---
Introduction
============
During the past decade, tremendous progress has been made in the field of network coding [@RAhlswededIT00]. In [@ZhangMobicom06]-[@Petar07], the concept of physical-layer network coding (PNC) was introduced and applied to wireless networks. The simplest model for wireless PNC is a two-way relay channel (TWRC), in which two users $A$ and $B$ exchange information via an intermediate relay. Compared with conventional schemes, PNC allows the relay to deliver linear functions of the users’ messages, which can potentially double the network throughput. It has been shown that the PNC scheme can achieve the capacity of a Gaussian TWRC within 1/2 bit per user [@NamIT10][@WilsonIT10], and its gap to the capacity vanishes at high signal-to-noise ratio (SNR).
Recently, efficient communications over MIMO TWRCs have attracted much research interest, where the two users and the relay are equipped with multiple antennas. Most work on MIMO TWRCs focuses on classical relaying strategies borrowed from one-way relay channels, such as amplify-and-forward (AF) [@KattiSIGCOMM07]-[@XuICCASP10], compress-and-forward [AleksicIT09]{}[@LimIT11], and decode-and-forward (DF) [@KramerIT05]-[@GunduzAsilomar08]. These strategies generally perform well away from the channel capacity due to noise amplification and multiplexing loss [GunduzAsilomar08]{}. Recently, several relaying schemes have been proposed to support PNC in MIMO TWRCs [@HJYangIT11]-[@NazerIT11]. The basic idea is to jointly decompose the channel matrices of the two users to create multiple scalar channels, over which multiple PNC streams are transmitted. Let $n_{A}$, $n_{B}$, and $n_{R}$ denote the numbers of antennas of user $A$, user $B$, and the relay, respectively. For configurations with $%
n_{A},n_{B}\geq n_{R}$, a generalized singular-value-decomposition (GSVD) scheme was shown to achieve the asymptotic capacity of MIMO TWRCs at high SNR [@HJYangIT11]. For configurations with $n_{A},n_{B}<$ $n_{R}$, all existing schemes may perform quite far away from the capacity. Such configurations, however, are of most practical importance. For example, due to the constrained physical sizes of the user terminals, it is usually convenient to implement more antennas at the relay station than at the user ends, as suggested in the standards of next generation networks [@3GPP][@WiMAX].
In this paper, we propose a new space-division based PNC scheme for MIMO TWRCs. Specifically, we first establish a novel joint channel decomposition, which characterizes the mutual orthogonality of the channel directions of the two users seen at the relay. Based on this decomposition, the overall signal space is divided into two orthogonal subspaces. In one subspace, the channel directions of one user are orthogonal (or close to orthogonal) to those of the other user. In this subspace, the spatial streams of the two users are completely decoded. In the other subspace, the channel directions of the two users are parallel or close to parallel. In this subspace, linear functions of the corresponding spatial streams are computed, without completely decoding the individual spatial streams. These linear functions of the spatial streams are referred to as *network-coded messages*. The messages and the network-coded messages, respectively generated from the two subspaces, are jointly encoded at the relay, and then forwarded to the two users. Afterwards, the two users recover their desired messages.
We derive the achievable rates of the proposed space-division based PNC scheme for MIMO TWRCs. We analytically show that, in the high SNR regime, the proposed scheme can achieve the sum capacity of the MIMO TWRC within $%
\min \{n_{A},n_{B}\}\log (5/4)$ bits, or $\frac{1}{2}\log (5/4)\approx 0.161$ bit/user-antenna, for any antenna setup and channel realization. This gap is much smaller than (as low as 10% of) the gap for the existing best scheme. We also perform large-system analysis to derive the average achievable sum-rate of the proposed scheme in Rayleigh fading MIMO TWRCs. We show that, in the high SNR regime, the average gap between our scheme and the sum-capacity upper bound is greatest when the antenna configuration is $%
n_{A}=n_{B}=\frac{1}{2}n_{R}$, with the gap being only $0.053$ bit/relay-antenna. For all other configurations, the proposed scheme perform even closer to the capacity upper bound. Particularly, as the ratio $n_{A}/n_{R}$ (or $%
n_{B}/n_{R}$) tends to 0 or 1, the gap to the capacity upper bound vanishes. All these analytical results agree well with the simulation. Numerical results demonstrate that the proposed scheme significantly outperforms the existing schemes in the literature across the full range of SNRs.
Preliminaries
=============
Notation
--------
The following notations are used throughout this paper. We use lowercase regular letters for scalars, lowercase bold letters for vectors, and uppercase bold letters for matrices. The superscripts $(\cdot)^T$ and $(\cdot)^\dag$ denote transpose and Hermitian transpose, respectively. $\|\cdot\|$ and $\|\cdot\|_F$ represent the Euclidian norm of a vector and the Frobenius norm of a matrix, respectively. $\mathcal{C}(\mathbf{X})$ represents the columnspace of a matrix $\mathbf{X}$. $\mathcal{%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
}^{n\times m}$ and $\mathcal{%
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
}^{n\times m}$ denote the $n$-by-$m$ dimensional real space and complex space, respectively. The operation $\log (\cdot )$ denotes the logarithm with base 2, and $|\cdot |$ the determinant. $I(i)$ is the indicator function with $I(i)=1$ for $i=1$ and $I(i)=0$ for $i\neq 1$; $[\cdot ]^{+}$ represents $\max \{\cdot ,0\}$; sign$(x)$ represents the sign of $x$; $%
\mathcal{N}_{c}(\mu ,\sigma ^{2})$ denotes the circularly symmetric complex Gaussian distribution with mean $\mu $ and variable $\sigma ^{2}$.
System Model
------------
In this paper, we consider a discrete memoryless MIMO TWRC in which users $A$ and $B$ exchange information via a relay, as illustrated in Fig. [Fig\_Config\_MIMOTWRC1]{}. User $m$ is equipped with $n_{m}$ antennas, $m\in
\left\{ A,B\right\} $, and the relay with $n_{R}$ antennas. We assume that there is no direct link between the two users. The channel is assumed to be flat-fading and quasi-static, i.e., the channel coefficients remain unchanged during each round of information exchange. The channel matrix from user $m$ to the relay is denoted by $\mathbf{H}_{mR}\in \mathcal{%
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
}^{n_{R}\times n_{m}}$, and that from the relay to user $m$ is denoted by $\mathbf{H}_{Rm}\in \mathcal{%
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
}^{n_{m}\times n_{R}},m\in \left\{ A,B\right\} $. We further assume that the channel matrices are always of either full column rank or full row rank, whichever is smaller, and are globally known by both users as well as by the relay.
The system operates in a half-duplex mode. Two time-slots are employed for each round of information exchange. Following the convention in [HJYangIT11]{}-[@KhinaISIT11], we assume that the two time-slots have same duration. The extension of our results to the case of unequal duration is straightforward.
In the first time-slot (referred to as *uplink phase*), the two users transmit to the relay simultaneously and the relay remains silent. The transmit signal matrix at user $m$ is denoted by $\mathbf{X}_{m}\in \mathcal{%
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
}^{n_{m}\times T},$ $m\in \left\{ A,B\right\} $, where $T$ is the number of channel uses in one time-slot. Each column of $\mathbf{X}_{m}$ denotes the signal vector transmitted by the $n_{m}$ antennas in one channel use. The average power at each user is constrained as $\frac{1}{T}E\left[ \left\Vert
\mathbf{X}_{m}\right\Vert _{F}^{2}\right] \leq P_{m},$ $m\in \left\{
A,B\right\} $. The received signal at the relay is denoted by $\mathbf{Y}%
_{R}\in \mathcal{%
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
}^{n_{R}\times T}$ with$$\mathbf{Y}_{R}=\mathbf{H}_{AR}\mathbf{X}_{A}+\mathbf{H}_{BR}\mathbf{X}_{B}+%
\mathbf{Z}_{R}, \label{YR}$$where $\mathbf{Z}_{R}\in \mathcal{%
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
}^{n_{R}\times T}$ denotes the additive white Gaussian noise (AWGN) at the relay. We assume that the elements of $\mathbf{Z}_{R}$ are independent and identically drawn from $\mathcal{N}_{c}(0,N_{0})$. Upon receiving $\mathbf{Y}%
_{R}$, the relay generates a signal matrix $\mathbf{X}_{R}\in \mathcal{%
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
}^{n_{R}\times T}$.
In the second time-slot (referred to as *downlink phase*), $\mathbf{X}%
_{R}$ is broadcast to the two users. The average power at the relay is constrained as $\frac{1}{T}E\left[ \left\Vert \mathbf{X}_{R}\right\Vert
_{F}^{2}\right] \leq P_{R}.$ The signal matrix received by user $m$ is denoted by $\mathbf{Y}_{m}\in \mathcal{%
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
}^{n_{m}\times T},m\in \left\{ A,B\right\} $, with $$\mathbf{Y}_{m}=\mathbf{H}_{Rm}\mathbf{X}_{R}+\mathbf{Z}_{m},m\in \left\{
A,B\right\} ,$$where $\mathbf{Z}_{m}$ denotes the AWGN matrix at user $m$, with the elements independently drawn from $\mathcal{N}_{c}(0,N_{0})$.
Definition of Achievable Rates
------------------------------
For the system model described above, denote the message of user $m$ by $%
W_{m}\in \{1,2,...,2^{2TR_{m}}\}$. The cardinality of $W_{m}$ is given by $%
2^{2TR_{m}}$, where the factor of $2T$ is because each round of information exchange consists of two length-$T$ time-slots. At user $A$, the estimated message of user $B$, denoted by $\hat{W}_{B}$, is obtained from the received signal $\mathbf{Y}_{A}$ and the perfect knowledge of the self message $W_{A}$. The decoding operation at user $B$ is similar. The error probability is defined as $P_{e}\triangleq \Pr \{\hat{W}_{A}\neq W_{A}$ or $\hat{W}%
_{B}\neq W_{B}\}$**.** We say that a rate-pair $(R_{A},R_{B})$ is achievable if the error probability $P_{e}$ vanishes as $T$ tends to infinity. The achievable rate-region is defined as the closure of all possible achievable rate-pairs.
Capacity Upper Bound
--------------------
Here we briefly describe a capacity upper bound of the MIMO TWRC. Let $%
\mathbf{Q}_{m}\triangleq \frac{1}{T}E\left[ \mathbf{X}_{m}\mathbf{X}%
_{m}^{\dagger }\right] ,m\in \left\{ A,B,R\right\},$ be the input covariance matrices. For given {$\mathbf{Q}_{A}$, $\mathbf{Q}_{B}$, $\mathbf{Q}_{R}$} satisfying tr$(\mathbf{Q}_{m})\leq P_{m},m\in \{A,B,R\} $, the achievable rate-pairs $(R_{A},R_{B})$ of the MIMO TWRC is upper bounded as [HJYangIT11]{}
\[rate\]$$\begin{aligned}
R_{A} &\leq &\min \left\{ R_{A}^{UL}\left( \mathbf{Q}_{A}\right)
,R_{A}^{DL}\left( \mathbf{Q}_{R}\right) \right\} \\
R_{B} &\leq &\min \left\{ R_{B}^{UL}\left( \mathbf{Q}_{B}\right)
,R_{B}^{DL}\left( \mathbf{Q}_{R}\right) \right\}\end{aligned}$$
where
\[Outer Bound\] $$\begin{aligned}
R_{m}^{UL}\left( \mathbf{Q}_{m}\right) &=&\frac{1}{2}\log \left\vert \mathbf{%
I}_{n_{R}}+\frac{1}{N_{0}}\mathbf{H}_{mR}\mathbf{Q}_{m}\mathbf{H}%
_{mR}^{\dagger }\right\vert ,m\in \left\{ A,B\right\} \\
R_{A}^{DL}\left( \mathbf{Q}_{R}\right) &=&\frac{1}{2}\log \left\vert \mathbf{%
I}_{n_{B}}+\frac{1}{N_{0}}\mathbf{H}_{RB}\mathbf{Q}_{R}\mathbf{H}%
_{RB}^{\dagger }\right\vert , \\
R_{B}^{DL}\left( \mathbf{Q}_{R}\right) &=&\frac{1}{2}\log \left\vert \mathbf{%
I}_{n_{A}}+\frac{1}{N_{0}}\mathbf{H}_{RA}\mathbf{Q}_{R}\mathbf{H}%
_{RA}^{\dagger }\right\vert .\end{aligned}$$
Here, the superscripts UL and DL respectively represent uplink and downlink, and the factor of 1/2 is due to the two time-slots used for each round of information exchange.
A capacity-region outer bound is defined as the closure of the upper-bound rate-pairs in (\[rate\]). This outer bound can be determined by optimizing $\mathbf{Q}_{A}$, $\mathbf{Q}_{B}$, and $\mathbf{Q}_{R}$, as detailed in [@HJYangIT11] and [@YangIT11]. The goal of this paper is to develop a communication strategy that can approach this outer bound.
Relaying Strategies for TWRCs with Single-Antenna Users
=======================================================
In this section, we study efficient communications over TWRCs with single-antenna users and a multi-antenna relay, i.e., $n_{A}=n_{B}=1$ and $%
n_{R}\geq 1$. The results developed in this section will be used in our studies on general MIMO TWRCs later.
Relaying Strategies: Complete Decoding versus PNC
-------------------------------------------------
For the case of single-antenna users, the channel model of the uplink phase in (\[YR\]) can be simplified as$$\mathbf{Y}_{R}=\mathbf{h}_{AR}\mathbf{x}_{A}^{T}+\mathbf{h}_{BR}\mathbf{x}%
_{B}^{T}+\mathbf{Z}_{R} \label{Vector}$$where $\mathbf{h}_{mR}\in \mathcal{%
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
}^{n_{R}\times 1}$ is the reduced version of $\mathbf{H}_{mR}$, and $\mathbf{%
x}_{m}\in \mathcal{%
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
}^{T\times 1}$ is the transmit signal vector of user $m$, with the $i$th entry of $\mathbf{x}_{m}$ being the signal transmitted at the $i$th time interval, $m\in \left\{ A,B\right\} $. Upon receiving $\mathbf{Y}_{R}$, the relay generates a network-coded message following the spirit of network coding. This network-coded message will be forwarded to the two users in the downlink phase.
The relay operation is detailed as follows. In the uplink phase, the signal direction of user $m$ is given by $\mathbf{h}%
_{mR},$ $m\in \left\{ A,B\right\} $. On one hand, if $\mathbf{h}_{AR}$ and $%
\mathbf{h}_{BR}$ are orthogonal, both messages of the two users can be decoded free of interference from each other. The recovered messages of the two users are then network-coded and forwarded to the two users. We refer to this first strategy as the *complete-decoding (CD) strategy*. On the other hand, if $\mathbf{h}_{AR}$ and $\mathbf{h}_{BR}$ turn out to be parallel (i.e., in a same direction), then it is advantageous to compute a linear function of $\mathbf{x}_{A}$ and $\mathbf{x}_{B}$, referred to as a network-coded message, instead of completely decoding both $\mathbf{x}_{A}$ and $\mathbf{x}_{B}$. We refer to this second strategy as the *PNC strategy*.
In general, the following strategy can be adopted: if $\mathbf{h}_{AR}$ and $%
\mathbf{h}_{BR}$ tend to be orthogonal, the complete-decoding strategy is applied; if $\mathbf{h}_{AR}$ and $\mathbf{h}_{BR}$ tend to be parallel, the PNC strategy is applied. The selection between these two strategies is based on their achievable rates, as described below.
### The Complete-Decoding Strategy
For complete-decoding, the uplink channel in (\[Vector\]) becomes a multiple-access (MAC) channel. Let $R_{m}^{CD},m\in \left\{ A,B\right\} $, be the rate of user $m$ for the complete-decoding strategy. Then, the uplink rate-region of the complete-decoding strategy, denoted by $\mathcal{R}%
_{{}}^{CD}$, is given by
\[MAC\] $$\begin{aligned}
R_{A}^{CD}+R_{B}^{CD} &\leq &\frac{1}{2}\log \left\vert \mathbf{I}%
_{n_{R}}+\dsum\limits_{m\in \left\{ A,B\right\} }\frac{P_{m}}{N_{0}}\mathbf{h%
}_{mR}\mathbf{h}_{mR}^{\dagger }\right\vert \\
R_{m}^{CD} &\leq &\frac{1}{2}\log \left( 1+\frac{P_{m}}{N_{0}}\mathbf{h}%
_{mR}^{\dagger }\mathbf{h}_{mR}\right) ,m\in \left\{ A,B\right\} .\end{aligned}$$
which follows from the well-known MAC capacity region [@Cover91].
### The PNC Strategy
For the PNC strategy in [@NamIT10][@WilsonIT10], it is required that the two user-signals lie in a same spatial direction. This is not guaranteed here due to the availability of multiple antennas at the relay. To facilitate PNC, we next propose a projection-based method as follows. The signals from the two users are first projected to a common direction, denoted by a unit vector $\mathbf{p}\in \mathcal{%
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
}^{n_{R}\times 1}$. The choice of $\mathbf{p}$ will be discussed in the next subsection. This projection operation creates an aligned scalar channel given by$$\mathbf{p}^{\dag }\mathbf{Y}_{R}=\mathbf{p}^{\dag }\mathbf{h}_{AR}\mathbf{x}%
_{A}^{T}+\mathbf{p}^{\dag }\mathbf{h}_{BR}\mathbf{x}_{B}^{T}+\mathbf{p}%
^{\dag }\mathbf{Z}_{R}$$with the effective channel coefficients given by $\mathbf{p}^{\dag }\mathbf{h%
}_{AR}$ and $\mathbf{p}^{\dag }\mathbf{h}_{BR}$.
From [@HJYangIT11], if the sum of the two users’ codewords, i.e., $%
\mathbf{p}^{\dag }\mathbf{h}_{AR}\mathbf{x}_{A}^{T}+\mathbf{p}^{\dag }%
\mathbf{h}_{BR}\mathbf{x}_{B}^{T}$, is required to be decoded, an achievable rate-pair ($R_{A}^{PNC},R_{B}^{PNC}$) of the uplink phase is given by$$R_{m}^{PNC}=\frac{1}{2}\left[ \log \left( \frac{Q_{m}\left\vert \mathbf{p}%
^{\dag }\mathbf{h}_{mR}\right\vert ^{2}}{N_{0}}\right) \right] ^{+},m\in
\left\{ A,B\right\} , \label{R_PNC}$$where $Q_{m}=\frac{1}{T}E[\mathbf{x}_{m}^{\dagger }\mathbf{x}_{m}]$ represents the transmission power of user $m$. If $\mathbf{p}^{\dag }\mathbf{%
h}_{AR}\mathbf{x}_{A}^{T}+\mathbf{p}^{\dag }\mathbf{h}_{BR}\mathbf{x}%
_{B}^{T} $ is not necessarily decoded, the above rate-pair can be further improved by using the lattice-modulo operation and minimum mean-square error (MMSE) scaling [@NamIT10][@ErezIT04], with the achievable rate-pair given by$$R_{m}^{PNC}=\frac{1}{2}\left[ \log \left( \frac{Q_{m}\left\vert \mathbf{p}%
^{\dag }\mathbf{h}_{mR}\right\vert ^{2}}{Q_{A}\left\vert \mathbf{p}^{\dag }%
\mathbf{h}_{AR}\right\vert ^{2}+Q_{B}\left\vert \mathbf{p}^{\dag }\mathbf{h}%
_{BR}\right\vert ^{2}}+\frac{Q_{m}\left\vert \mathbf{p}^{\dag }\mathbf{h}%
_{mR}\right\vert ^{2}}{N_{0}}\right) \right] ^{+},m\in \left\{ A,B\right\} .
\label{R_PNC2}$$Notice that (\[R\_PNC\]) and (\[R\_PNC2\]) become identical and both approaches the uplink channel capacity at high SNR.
The uplink rate-region of the PNC scheme is given by$$\mathcal{R}_{{}}^{PNC}\triangleq \left\{ (R_{A},R_{B})|R_{m}\leq
R_{m}^{PNC},Q_{m}\leq P_{m},m\in \left\{ A,B\right\} ,\mathbf{p}^{\dag }%
\mathbf{p=1}\right\} . \label{Eq PNC SIMO}$$The boundary of $\mathcal{R}_{{}}^{PNC}$ can be found by optimizing $Q_{A}$, $Q_{B}$, and $\mathbf{p}$, as detailed in the next subsection.
Optimization of Projection Direction
------------------------------------
Here we focus on the rate-pair given in (\[R\_PNC\]).[^2] As achievable rate-regions are convex, the boundary points of $%
\mathcal{R}_{{}}^{PNC}$ can be determined by solving the weighted sum-rate maximization problem:
\[P1\] $$\begin{aligned}
\text{maximize} &&w_{A}R_{A}^{PNC}+w_{B}R_{B}^{PNC} \\
\text{subject to} &&||\mathbf{p}||=1,Q_{m}\leq P_{m},m\in \left\{ A,B\right\}\end{aligned}$$
where $w_{A}$ and $w_{B}$ are arbitrary nonnegative weighting coefficients. By inspecting (\[R\_PNC\]), the maximum of (\[P1\]) is achieved at $%
Q_{m}=P_{m},m\in \left\{ A,B\right\} $. Thus, we only need to optimize $%
\mathbf{p}$.
Suppose that $R_{A}^{PNC}=0$ (or $R_{B}^{PNC}=0$). Then, from (\[R\_PNC\]), the optimal $\mathbf{p}$ is trivially taken as $\frac{\mathbf{h}_{BR}}{%
\left\Vert \mathbf{h}_{BR}\right\Vert }$ (or $\frac{\mathbf{h}_{AR}}{%
\left\Vert \mathbf{h}_{AR}\right\Vert }$). Thus, we focus on the case of $%
R_{m}^{PNC}>0,m\in \left\{ A,B\right\} $. In this case, this weighted sum-rate maximization problem is equivalent to maximizing$$\max_{||\mathbf{p}||=1}w_{A}\log \left( \left\vert \mathbf{p}^{\dag }\mathbf{%
h}_{AR}\right\vert ^{2}\right) +w_{B}\log \left( \left\vert \mathbf{p}^{\dag
}\mathbf{h}_{BR}\right\vert ^{2}\right)$$or equivalently$$\max_{||\widetilde{\mathbf{p}}||=1}\left| \widetilde{\mathbf{h}}_{AR}^{T}\widetilde{\mathbf{%
p}}\right| ^{2w_{A}}\left| \widetilde{\mathbf{h}}_{BR}^{T}\widetilde{\mathbf{p}}\right| ^{2w_{B}}$$where $\widetilde{\mathbf{p}}=[\func{Re}[\mathbf{p}^{T}],\func{Im}[\mathbf{p}%
^{T}]]^{T}$ and $\widetilde{\mathbf{h}}_{mR}=[\func{Re}[\mathbf{h}_{mR}^{T}],%
\func{Im}[\mathbf{h}_{mR}^{T}]]^{T}$, $m\in \left\{ A,B\right\} $. By setting the derivative of the Lagrangian with respect to $\textbf{p}$ to zero, the optimal $\textbf{p}$ satisfies$$\alpha \widetilde{\mathbf{p}}=\frac{w_{A}\widetilde{\mathbf{h}}_{AR}}{%
\widetilde{\mathbf{p}}^{T}\widetilde{\mathbf{h}}_{AR}}+\frac{w_{B}\widetilde{%
\mathbf{h}}_{BR}}{\widetilde{\mathbf{p}}^{T}\widetilde{\mathbf{h}}_{BR}}, \notag$$where $\alpha$ is a scaling factor. Then, with some straightforward algebra, we obtain the optimal projection direction given by $$\widetilde{\mathbf{p}}^{opt}=\gamma \left( \frac{\widetilde{\mathbf{h}}_{AR}%
}{\left\Vert \widetilde{\mathbf{h}}_{AR}\right\Vert }+\beta \frac{\widetilde{%
\mathbf{h}}_{BR}}{\left\Vert \widetilde{\mathbf{h}}_{BR}\right\Vert }\right)
, \label{d}$$where$$\beta =\frac{\text{sign}(\widetilde{\mathbf{h}}_{AR}^{T}\widetilde{\mathbf{h}%
}_{BR})}{2}\left( \sqrt{\left( \frac{\widetilde{\mathbf{h}}_{AR}^{T}%
\widetilde{\mathbf{h}}_{BR}\left( 1-\frac{w_{B}}{w_{A}}\right) }{\left\Vert
\widetilde{\mathbf{h}}_{AR}\right\Vert \left\Vert \widetilde{\mathbf{h}}%
_{BR}\right\Vert }\right) ^{2}+4\frac{w_{B}}{w_{A}}}-\frac{\left\vert
\widetilde{\mathbf{h}}_{AR}^{T}\widetilde{\mathbf{h}}_{BR}\right\vert \left(
1-\frac{w_{B}}{w_{A}}\right) }{\left\Vert \widetilde{\mathbf{h}}%
_{AR}\right\Vert \left\Vert \widetilde{\mathbf{h}}_{BR}\right\Vert }\right)$$and $\gamma $ is a scaling factor to ensure $||\widetilde{\mathbf{p}}^{opt}||$ $=1$. Particularly, for the sum-rate case, i.e., $w_{A}=w_{B}=1$, the optimal projection direction $%
\widetilde{\mathbf{p}}$ is just the angular bisector of $\widetilde{\mathbf{h%
}}_{AR}$ and $\widetilde{\mathbf{h}}_{BR}$ if $\widetilde{\mathbf{h}}%
_{AR}^{T}\widetilde{\mathbf{h}}_{BR}>0$, or the angular bisector of $%
\widetilde{\mathbf{h}}_{AR}$ and $-\widetilde{\mathbf{h}}_{BR}$ if $%
\widetilde{\mathbf{h}}_{AR}^{T}\widetilde{\mathbf{h}}_{BR}<0$. By varying $%
w_{A}$ and $w_{B}$, $\mathcal{R}_{{}}^{PNC}$ can be determined.
The Overall Scheme
------------------
For the uplink phase, the achievable rate-regions $\mathcal{R}_{{}}^{CD}$ and $\mathcal{R}_{{}}^{PNC}$, for certain channel realizations of $\mathbf{h}%
_{AR}$ and $\mathbf{h}_{BR}$, are depicted in Fig. \[SIMO\]. The overall uplink rate-region, denoted by $\mathcal{R}_{{}}^{UL}$, is given by the convex hull of $\mathcal{R}_{{}}^{CD}$ and $\mathcal{R}_{{}}^{PNC}$. In the overall scheme, the relay will select between the complete-decoding and PNC strategies for a larger achievable rate-region, according to (\[MAC\]) and (\[Eq PNC SIMO\]).
For the downlink phase, the achievable rate-region is determined as follows. For the complete-decoding strategy, the relay jointly re-encode the decoded messages $\mathbf{x}_{A}$ and $\mathbf{x}_{B}$, and forward the resulting codeword to the two users in the downlink. For the PNC strategy, the relay forward the lattice-modulo of $\mathbf{p}^{\dag }\mathbf{h}_{AR}\mathbf{x}%
_{A}^{T}+\mathbf{p}^{\dag }\mathbf{h}_{BR}\mathbf{x}_{B}^{T}$, referred to as a network-coded message, to the two users. Then each user recovers the message of the other user with the help of the perfect knowledge of self message. From [@HJYangIT11]-[@KhinaISIT11], the downlink rate-region $\mathcal{R}_{{}}^{DL}$ for the two strategies are the same and given by $$\mathcal{R}_{{}}^{DL}\triangleq \left\{ (R_{A},R_{B})|R_{A}\leq
R_{A}^{DL},R_{B}\leq R_{B}^{DL}\right\}$$with$$R_{A}^{DL}=\frac{1}{2}\log \left( 1+\frac{P_{R}}{N_{0}}\mathbf{h}%
_{RB}^{\dagger }\mathbf{h}_{RB}\right) \text{ and }R_{B}^{DL}=\frac{1}{2}%
\log \left( 1+\frac{P_{R}}{N_{0}}\mathbf{h}_{RA}^{\dagger }\mathbf{h}%
_{RA}\right) .$$Finally, an achievable rate-region of the overall scheme is the intersection of the uplink rate-region $\mathcal{R}_{{}}^{UL}$ and the downlink rate-region $\mathcal{R}_{{}}^{DL}$.
Space-Division Approach for MIMO TWRCs
======================================
In the preceding section, we have studied the design of relaying strategies for TWRCs with single-antenna users. We have shown how to exploit the benefits of the complete-decoding and PNC strategies. In this section, we proceed to study the general case of [$n_{A}\geq 1$, $n_{B}\geq 1$]{}. We propose a new space-division based network-coding scheme, as a generalization for the case of single-antenna users.
Motivations
-----------
What motivates the proposed space-division approach is the following property of $\mathbf{H}_{AR}$ and $\mathbf{H}_{BR}$. Denote by $\mathcal{C}(\mathbf{H}_{AR})$ and $\mathcal{C}(\mathbf{H}%
_{BR})$ the columnspaces of the uplink channel matrices $\mathbf{H}_{AR}$ and $\mathbf{H}_{BR}$, respectively. In general, we can partition the columnspace $\mathcal{C}(\mathbf{H}%
_{AR})$ $\in \mathcal{%
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
}^{n_{R}}$ as the direct sum[^3] of three orthogonal subspaces: a subspace $\mathcal{S}_{A\parallel B}$ that is parallel to $%
\mathcal{C}(\mathbf{H}_{BR})$, i.e., any vector in $\mathcal{S}_{A\parallel
B}$ belongs to $\mathcal{C}(\mathbf{H}_{BR})$; a subspace $\mathcal{S}%
_{A\nparallel B}$ that is neither parallel nor orthogonal to $\mathcal{C}(%
\mathbf{H}_{BR})$; and a subspace $\mathcal{S}_{A\perp B}$ that is orthogonal to $\mathcal{C}(\mathbf{H}_{BR})$. Similarly, $\mathcal{C}(%
\mathbf{H}_{BR})$ is the direct sum of three orthogonal subspaces $\mathcal{S%
}_{B||A},\mathcal{S}_{B\nparallel A},$ and $\mathcal{S}_{B\perp A}$. Note that $\mathcal{S}_{A\parallel B}=\mathcal{S}_{B||A}$ since both represent the *common space* of $\mathcal{C}(\mathbf{H}_{AR})$ and $\mathcal{C}(%
\mathbf{H}_{BR})$.
In $\mathcal{S}_{A\parallel B}$, the signal directions of the two users can be efficiently aligned to a same set of directions, providing a platform to carry out PNC, as in [@HJYangIT11]-[@KhinaISIT11]. On the other hand, in $\mathcal{S}_{A\perp B}$ or $\mathcal{S}_{B\perp A}$, the two users do not interfere with each other, hence the complete-decoding strategy can be employed. The above treatments are similar to those for the case of single-antenna users, as discussed in the preceding section. What remains is the treatment for the signals in $\mathcal{S}_{A\nparallel B}$ and $\mathcal{%
S}_{B\nparallel A}$ that are neither parallel nor orthogonal. Heuristically, some channel directions in $\mathcal{S}_{A\nparallel B}$ and $\mathcal{S}%
_{B\nparallel A}$ may be nearly parallel to each other. For these channel directions, the PNC strategy is preferable for the related spatial streams. On the other hand, some channel directions in $\mathcal{S}_{A\nparallel B}$ and $\mathcal{S}_{B\nparallel A}$ may be nearly orthogonal to each other. Then, the complete-decoding strategy is preferable. The main challenge lies in how to identify those nearly parallel/orthogonal channel directions. To this end, we next propose a new joint channel decomposition.
Joint Channel Decomposition
---------------------------
Let the compact singular value decomposition (SVD) of $\mathbf{H}_{mR}$ be$$\mathbf{H}_{mR}=\mathbf{U}_{m}\mathbf{\Delta }_{m}\mathbf{V}_{m}^{\dag },%
\mathbf{\ }m\in \left\{ A,B\right\} \label{Channel_Decomp}$$where $\mathbf{U}_{m}$ is an $n_{R}$-by-$n_{m}$ orthonormal matrix with $%
\mathbf{U}_{m}^{\dag }\mathbf{U}_{m}=\mathbf{I}_{n_{m}}$.
Denote by $\lambda _{i}$ the $i$th eigenvalue of the matrix $\mathbf{U}_{A}%
\mathbf{U}_{A}^{\dag }+\mathbf{U}_{B}\mathbf{U}_{B}^{\dag }$, and by $%
\mathbf{u}_{i}$ the corresponding eigenvector. Without loss of generality, we arrange {$\lambda _{i}$} in the descending order. As the eigenvalues of $\mathbf{U}_{m}\mathbf{U}_{m}^{\dag }$ are either 1 or 0, the eigenvalues {$%
\lambda _{i}$} are valued between $0$ and $2$. Note that $\lambda
_{i}=2$ implies that $\mathbf{U}_{A}\mathbf{U}_{A}^{\dag }\mathbf{u}_{i}%
\mathbf{=u}_{i}$ and $\mathbf{U}_{B}\mathbf{U}_{B}^{\dag }\mathbf{u}_{i}%
\mathbf{=u}_{i}$. This means that $\mathbf{u}_{i}$ is in the common space $%
\mathcal{S}_{A\parallel B}$ of $\mathbf{H}_{AR}$ and $\mathbf{H}_{BR}$. Also note that $\lambda _{i}=1$ implies {$\mathbf{U}_{A}\mathbf{U}_{A}^{\dag }\mathbf{u}_{i}\mathbf{=u}_{i}$, $\mathbf{U}_{B}\mathbf{U}_{B}^{\dag }\mathbf{u}_{i}\mathbf{=0}$} or {$%
\mathbf{U}_{A}\mathbf{U}_{A}^{\dag }\mathbf{u}_{i}\mathbf{=0}$, $\mathbf{U%
}_{B}\mathbf{U}_{B}^{\dag }\mathbf{u}_{i}\mathbf{=u}_{i}$} (cf., Theorem 4.3.4 in [@HornTextbook]). That is, $\mathbf{%
u}_{i}$ is in $\mathcal{S}_{A\perp B}$ (or $\mathcal{S}_{B\perp A}$) which is orthogonal to the space spanned by $\mathbf{H}_{BR}$ (or $\mathbf{H}_{AR}$). In addition, the number of eigenvalues between 1 and 2 is the same as that between 0 and 1, as we will see later.
Let $k$ be the number of eigenvalues of $\mathbf{U}_{A}\mathbf{U}_{A}^{\dag
}+\mathbf{U}_{B}\mathbf{U}_{B}^{\dag }$ equal to 2; $l$ be the number of eigenvalues between 1 and 2; $d_{A}$ be the number of eigenvalues equal to 1 with {$\mathbf{U}_{A}\mathbf{U}_{A}^{\dag }\mathbf{u}_{i}\mathbf{=u}_{i}$, $%
\mathbf{U}_{B}\mathbf{U}_{B}^{\dag }\mathbf{u}_{i}\mathbf{=0}$}; $d_{B}$ be the number of eigenvalues equal to 1 with {$\mathbf{U}_{A}\mathbf{U}_{A}^{\dag }%
\mathbf{u}_{i}\mathbf{=0}$, $\mathbf{U}_{B}\mathbf{U}_{B}^{\dag }\mathbf{u}%
_{i}\mathbf{=u}_{i}$}. Also let $\mathbf{U}\in $ $\mathcal{%
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
}^{n_{R}\times (n_{A}+n_{B}-k)}$ be a matrix with the columns consisting of the eigenvectors corresponding to the $n_{A}+n_{B}-k$ largest eigenvalues of $\mathbf{U}_{A}\mathbf{U}_{A}^{\dag }+\mathbf{U}_{B}\mathbf{U}_{B}^{\dag }$ (as specified in (\[Def\_U\])). Due to the orthogonality of the eigenvectors, $\mathbf{U}$ is orthonormal, i.e., $\mathbf{U}_{{}}^{\dag }%
\mathbf{U}=\mathbf{I}_{n_{A}+n_{B}-k}$.
We are now ready to present the joint channel decomposition, with the proof given in Appendix A.
\[Theorem 1\]The channel matrices* *$\mathbf{H}_{AR}$ and $%
\mathbf{H}_{BR}$ can be jointly decomposed as$$\mathbf{H}_{mR}=\mathbf{UD}_{m}\mathbf{G}_{m},\mathbf{\ }m\in \left\{
A,B\right\} \label{Hm}$$where $\mathbf{G}_{m}\in $ $\mathcal{%
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
}^{n_{m}\times n_{m}}$ is a square matrix, and $\mathbf{D}_{m}\in \mathcal{%
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
}^{(n_{A}+n_{B}-k)\times n_{m}},m\in \left\{ A,B\right\} ,$ are orthonormal matrices with a block-diagonal structure given by
$$\mathbf{D}_{A}=\left[
\begin{array}{ccc}
\mathbf{I}_{k} & \mathbf{0} & \mathbf{0} \\
\mathbf{0} & \mathbf{E}_{A} & \mathbf{0} \\
\mathbf{0} & \mathbf{0} & \mathbf{I}_{d_{A}} \\
\mathbf{0} & \mathbf{0} & \mathbf{0}%
\end{array}%
\right] \text{ and }\mathbf{D}_{B}=\left[
\begin{array}{ccc}
\mathbf{I}_{k} & \mathbf{0} & \mathbf{0} \\
\mathbf{0} & \mathbf{E}_{B} & \mathbf{0} \\
\mathbf{0} & \mathbf{0} & \mathbf{0} \\
\mathbf{0} & \mathbf{0} & \mathbf{I}_{d_{B}}%
\end{array}%
\right] \label{Dm}$$
with $$\mathbf{E}_{m}=\left[
\begin{array}{cccc}
\mathbf{e}_{m;k+1} & \mathbf{0} & \cdots & \mathbf{0} \\
\mathbf{0} & \mathbf{e}_{m;k+2} & \ddots & \vdots \\
\vdots & \ddots & \ddots & \mathbf{0} \\
\mathbf{0} & \cdots & \mathbf{0} & \mathbf{e}_{m;k+l}%
\end{array}%
\right] \in \mathcal{%
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
}^{2l\times l}\text{, } \label{E}$$$$\mathbf{e}_{A;i}=\left[
\begin{array}{c}
\sqrt{\frac{\lambda _{i}}{2}} \\
\sqrt{\frac{2-\lambda _{i}}{2}}%
\end{array}%
\right] \text{ and }\mathbf{e}_{B;i}=\left[
\begin{array}{c}
\sqrt{\frac{\lambda _{i}}{2}} \\
-\sqrt{\frac{2-\lambda _{i}}{2}}%
\end{array}%
\right] . \label{Direction}$$
From (\[Hm\]), we see that $\mathcal{C}(\mathbf{U)}$ is the overall columnspace of the two channel matrices, i.e., $\mathcal{C}(\mathbf{U)}=%
\mathcal{C}([\mathbf{H}_{AR}$ $\mathbf{H}_{BR}]\mathbf{)}$. Moreover, $%
\mathbf{UD}_{m}$ specifies the columnspace of $\mathbf{H}_{m}$, i.e., $%
\mathcal{C}(\mathbf{UD}_{m})=\mathcal{C}(\mathbf{H}_{m})$. Note that $%
\mathbf{UD}_{m}$ is orthonormal, as $\mathbf{D}_{m}^{\dag }\mathbf{U}^{\dag }%
\mathbf{UD}_{m}=\mathbf{I}_{n_{m}}$, $m\in \left\{ A,B\right\} $. Therefore, the columns of $\mathbf{UD}_{m}$ give an orthogonal basis of $\mathcal{C}(%
\mathbf{H}_{mR})$, with the coordinates of $\mathbf{H}_{mR}$ specified in $%
\mathbf{G}_{m}$.
The column structures of $\mathbf{UD}_{A}$ and $\mathbf{UD}_{B}$ are explained as follows. In the first place, we note that $\mathbf{UD}_{A}$ and $\mathbf{UD}_{B}$ share the same first $k$ columns. Thus, the first $k$ columns of $\mathbf{UD}_{A}$ span $\mathcal{S}_{A\parallel B}$, i.e., the common space of $\mathcal{C}(\mathbf{H}_{AR})$ and $\mathcal{C}(\mathbf{H}%
_{BR})$. Second, from (\[Dm\]), the last $d_{A}$ columns of $\mathbf{UD}%
_{A}$ (obtained from multiplying $\mathbf{U}$ with the third block column of $\mathbf{D}_{A}$) are orthogonal to $\mathbf{UD}_{B}$. Hence, these columns of $\mathbf{UD}_{A}$ span the subspace $\mathcal{S}_{A\perp B}$, i.e., the subspace orthogonal to $\mathcal{C}(\mathbf{H}_{BR})$. Third, the remaining $%
l$ columns of $\mathbf{UD}_{A}$ span the subspace $\mathcal{S}_{A\nparallel
B}$, by noting the facts that $\mathcal{C}(\mathbf{H}_{AR}\mathbf{)}=%
\mathcal{C}(\mathbf{UD}_{A}\mathbf{)}$ and that $\mathcal{C}(\mathbf{H}%
_{AR}) $ is the direct-sum of three orthogonal subspaces $\mathcal{S}%
_{A\parallel B} $, $\mathcal{S}_{A\nparallel B}$, and $\mathcal{S}_{A\perp
B} $. Similarly, the first $k$ columns of $\mathbf{UD}_{B}$ span $\mathcal{S}%
_{A\parallel B}$, the next $l$ columns span $\mathcal{S}_{B\nparallel A}$, and the last $d_{B}$ columns span $\mathcal{S}_{B\perp A}$. Recall that $%
\mathcal{C}(\mathbf{H}_{AR})$ is the direct sum of $\mathcal{S}_{A\parallel
B}$, $\mathcal{S}_{A\nparallel B}$, and $\mathcal{S}_{A\perp B}$, and that $%
\mathcal{C}(\mathbf{H}_{BR})$ is the direct sum of $\mathcal{S}_{B\parallel
A}$, $\mathcal{S}_{B\nparallel A}$, and $\mathcal{S}_{B\perp A}$. Thus, the dimensions of these subspaces have the following relationship: $$k+l+p_{m}=n_{m},m\in \left\{ A,B\right\} \text{.}$$We summarize the geometrical meanings of the aforementioned subspaces and their dimensions as follows.$$\begin{tabular}{lll}
Subspace & Dimension & Property \\
$\mathcal{S}_{A\parallel B}$ & $k$ & $\text{common space of }\mathcal{C}(%
\mathbf{H}_{AR})\text{ and }\mathcal{C}(\mathbf{H}_{BR})$ \\
$\mathcal{S}_{A\nparallel B}$ & $l$ & not parallel/orthogonal to$\text{ }%
\mathcal{C}(\mathbf{H}_{BR})$ \\
$\mathcal{S}_{B\nparallel A}$ & $l$ & not parallel/orthogonal to$\text{ }%
\mathcal{C}(\mathbf{H}_{AR})$ \\
$\mathcal{S}_{A\perp B}$ & $d_{A}$ & orthogonal to$\text{ }\mathcal{C}(%
\mathbf{H}_{BR})$ \\
$\mathcal{S}_{B\perp A}$ & $d_{B}$ & orthogonal to$\text{ }\mathcal{C}(%
\mathbf{H}_{AR})$%
\end{tabular}%$$
Let $\mathbf{v}_{m;i}$ be the $i$th column of $\mathbf{UD}_{m}$, $m\in
\left\{ A,B\right\} $. We refer to $\mathbf{v}_{A;i}$ and $\mathbf{v}_{B;i}$ as the $i$th *channel direction pair*. Here, $\mathbf{v}%
_{A;i}^{^{\dag }}\mathbf{v}_{B;i}=1$ means that $\mathbf{v}_{A;i}$ and $%
\mathbf{v}_{B;i}$ are parallel, and $\mathbf{v}_{A;i}^{^{\dag }}\mathbf{v}%
_{B;i}=0$ means that they are orthogonal. Thus, $\mathbf{v}_{A;i}^{^{\dag }}%
\mathbf{v}_{B;i}$ can be regarded as a measure of the *degree of orthogonality* of $\mathbf{v}_{A;i}$ and $\mathbf{v}_{B;i}$. In the following corollary, the degree of orthogonality of each channel direction pair ($\mathbf{v}_{A;i}$, $\mathbf{v}_{B;i}$) is determined by the magnitude of $\lambda _{i}$, i.e., the $i$th eigenvalue of $\mathbf{U}_{A}\mathbf{U}%
_{A}^{\dag }+\mathbf{U}_{B}\mathbf{U}_{B}^{\dag }$.
\[Corollary 1\]For $i=1,...,k+l$, the degree of orthogonality of the $i$th channel direction pair ($\mathbf{v}_{A,i}$, $\mathbf{v}_{B,i}$) is given by $\mathbf{v}_{A,i}^{^{\dag }}\mathbf{v}_{B,i}=\lambda _{i}-1$.
For $i=1,...,k$, we see from (\[Dm\]) that $\lambda _{i}=2$ and $\mathbf{v}%
_{A;i}=\mathbf{v}_{B;i}$, and so $\mathbf{v}_{A;i}^{\dag }\mathbf{v}%
_{B;i}=\lambda _{i}-1$. For $i=k+1,...,k+l$, from (\[Dm\]) and (\[E\]), the $i$th column of $\mathbf{UD}_{m}$ is given by $$\mathbf{v}_{m;i}=\left[
\begin{array}{cc}
\widetilde{\mathbf{u}}_{2i-k-1} & \widetilde{\mathbf{u}}_{2i-k}%
\end{array}%
\right] \mathbf{e}_{m;i},\mathbf{\ }m\in \left\{ A,B\right\} ,\text{ }$$where $\widetilde{\mathbf{u}}_{i}$ represents the $i$th column of $\mathbf{U}
$. Then, we obtain $\mathbf{v}_{A;k+i}^{\dag }\mathbf{v}_{B;k+i}=\mathbf{e}%
_{A;k+i}^{\dag }\mathbf{e}_{B;k+i}=\lambda _{i}-1$, where the first step utilizes the fact that $\mathbf{U}$ is orthonormal, and the second step follows from (\[Direction\]).
\[Corollary 2\]For $i=k+l+1,...,n_{A}$, $\mathbf{v}_{A;i}$ is an eigenvector of $\mathbf{U}_{A}\mathbf{U}_{A}^{\dag }+\mathbf{U}_{B}\mathbf{U}%
_{B}^{\dag }$ corresponding to $\lambda _{i}=1$, and is orthogonal to $%
\mathcal{C}(\mathbf{H}_{BR})$; for $i=k+l+1,...,n_{B}$, $\mathbf{v}_{B;i}$ is an eigenvector corresponding to $\lambda _{i}=1$, and is orthogonal to $%
\mathcal{C}(\mathbf{H}_{AR})$.
The above corollaries show that the eigenvalue $\lambda _{i}$ is an indicator of the degree of orthogonality of the $i$th direction pair. In particular, $\lambda _{i}\approx 2$ means that the two channel directions are close to parallel; and $\lambda _{i}\approx 1$ means that the two channel directions are close to orthogonal.
\[Remark 4\]Before leaving this subsection, we emphasize that the joint channel decomposition in Theorem \[Theorem 1\] is general for any sizes of $\mathbf{H}_{AR}$ and $\mathbf{H}_{BR}$. Particularly, if $n_{m}\geq
n_{R},m\in \left\{ A,B\right\} $, then $k=n_{R}$ and $l=0$, implying that all the eigenvalues $\{\lambda _{i}\}$ are valued at $2$. In this case, $%
\mathbf{H}_{AR}$ and $\mathbf{H}_{BR}$ span the same columnspace. Channel alignment techniques have been proposed in [@HJYangIT11]-[KhinaISIT11]{} for efficient implementation of PNC. In what follows, we are mainly interested in the case of $n_{A},n_{B}<n_{R}$, i.e., there exist $%
\{\lambda _{i}\}$ valued between, but not including, $1$ and $2$.
Space-Division Approach for MIMO Two-Way Relaying
-------------------------------------------------
Based on the joint channel decomposition in Theorem \[Theorem 1\], we now propose a new space-division approach for MIMO two-way relaying. The main idea is to divide the overall signal space $\mathcal{C}([\mathbf{H}_{AR}$ $%
\mathbf{H}_{BR}]\mathbf{)}=\mathcal{C}(\mathbf{U)}$ into two orthogonal subspaces: 1) $\mathcal{S}^{PNC}$, in which the channel direction pairs ($%
\mathbf{v}_{A;i},\mathbf{v}_{B;i}$) are parallel or close to parallel, for carrying out PNC; 2) $\mathcal{S}^{CD}$ for carrying out the complete-decoding strategy. Let $l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
$ be an arbitrary integer between 0 and $l$. Recall that the channel direction pairs are ordered by the degree of orthogonality as in Corollary \[Corollary 1\]. Therefore, the first $k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
$ direction pairs have lower degree of orthogonality compared to the remaining pairs. Thus, we allocate the first $k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
$ direction pairs to form a basis of $\mathcal{S}^{PNC}$. The remaining channel directions give a basis of $\mathcal{S}^{CD}$. In this section, we assume that $l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
$ is given. The details on the optimization of $l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
$ will be discussed later in Sections V and VI.
### Space-Division Operation
Let the RQ decomposition of $\mathbf{G}_{m}$ be$$\mathbf{G}_{m}=\mathbf{R}_{m}\mathbf{T}_{m}^{\dagger },\text{ }m\in \left\{
A,B\right\}$$where $\mathbf{R}_{m}\in $ $\mathcal{%
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
}^{n_{m}\times n_{m}}$ is an upper-triangular matrix given by$$\mathbf{R}_{m}=\left[
\begin{array}{cccc}
r_{m;1,1} & r_{m;1,2} & \cdots & r_{m;1,n_{m}} \\
0 & r_{m;2,2} & \cdots & r_{m;2,n_{m}} \\
\vdots & \vdots & \ddots & \vdots \\
0 & \cdots & 0 & r_{m;n_{m},n_{m}}%
\end{array}%
\right] ,$$and $\mathbf{T}_{m}\in $ $\mathcal{%
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
}^{n_{m}\times n_{m}}$ is unitary. Together with (\[Hm\]), the channel matrices* *can be jointly decomposed as$$\mathbf{H}_{m}=\mathbf{UD}_{m}\mathbf{R}_{m}\mathbf{T}_{m}^{\dag },\mathbf{\
}m\in \left\{ A,B\right\} . \label{Decomp}$$ Then, the received signal at the relay, after left-multiplying $\mathbf{U}^{\dag }$, can be represented as$$\mathbf{Y}_{R}^{\prime }=\mathbf{U}^{\dag }\mathbf{Y}_{R}=\mathbf{D}_{A}%
\mathbf{R}_{A}\mathbf{X}_{A}^{\prime }+\mathbf{D}_{B}\mathbf{R}_{B}\mathbf{X}%
_{B}^{\prime }+\mathbf{Z}_{R}^{\prime }, \label{33}$$where $\mathbf{X}_{m}^{\prime }=\mathbf{T}_{m}^{\dag }\mathbf{X}_{m},m\in
\left\{ A,B\right\} $, and $\mathbf{Z}_{R}^{\prime }=\mathbf{U}^{\dag }%
\mathbf{Z}_{R}$ with i.i.d. elements $\sim \mathcal{N}_{c}(0,N_{0})$.
We partition $\mathbf{R}_{m}$ and $\mathbf{D}_{m}$ as$$\mathbf{R}_{m}=\left[
\begin{array}{cc}
\mathbf{R}_{m;1,1} & \mathbf{R}_{m;1,2} \\
\mathbf{0} & \mathbf{R}_{m;2,2}%
\end{array}%
\right] \text{ and }\mathbf{D}_{m}=\left[
\begin{array}{cc}
\mathbf{D}_{m;1,1} & \mathbf{0} \\
\mathbf{0} & \mathbf{D}_{m;2,2}%
\end{array}%
\right] ,m\in \left\{ A,B\right\} \label{Dm3}$$where $\mathbf{R}_{m;1,1}\in
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
^{(k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
)\times (k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
)}$ and $\mathbf{R}_{m;2,2}\in
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
^{(l-l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
+p_{m})\times (l-l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
+p_{m})},$ $m\in \left\{ A,B\right\} $, are upper triangular matrices, and $%
\mathbf{D}_{m;1,1}\in
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
^{(k+2l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
)\times (k+l)}$ and $\mathbf{D}_{m;2,2}\in
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
^{(n_{R}-k-2l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
)\times (l-l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
+p_{m})},m\in \left\{ A,B\right\} $, are block-diagonal matrices. Then, ([33]{}) can be written as$$\left[
\begin{array}{c}
\mathbf{Y}_{R}^{PNC} \\
\mathbf{Y}_{R}^{CD}%
\end{array}%
\right] =\dsum\limits_{m\in \left\{ A,B\right\} }\left[
\begin{array}{cc}
\mathbf{D}_{m;1,1}\mathbf{R}_{m;1,1} & \mathbf{D}_{m;1,1}\mathbf{R}_{m;1,2}
\\
\mathbf{0} & \mathbf{D}_{m;2,2}\mathbf{R}_{m;2,2}%
\end{array}%
\right] \left[
\begin{array}{c}
\mathbf{X}_{m}^{PNC} \\
\mathbf{X}_{m}^{CD}%
\end{array}%
\right] +\left[
\begin{array}{c}
\mathbf{Z}_{R}^{PNC} \\
\mathbf{Z}_{R}^{CD}%
\end{array}%
\right] , \label{SS}$$where $\mathbf{Y}_{R}^{\prime }$, $\mathbf{X}_{m}^{\prime }$, and $\mathbf{Z}%
_{R}^{\prime }$ are correspondingly partitioned as$$\mathbf{Y}_{R}^{\prime }=\left[
\begin{array}{c}
\mathbf{Y}_{R}^{PNC} \\
\mathbf{Y}_{R}^{CD}%
\end{array}%
\right] \text{, }\mathbf{X}_{m}^{\prime }=\left[
\begin{array}{c}
\mathbf{X}_{m}^{PNC} \\
\mathbf{X}_{m}^{CD}%
\end{array}%
\right] ,\text{ and }\mathbf{Z}_{R}^{\prime }=\left[
\begin{array}{c}
\mathbf{Z}_{R}^{PNC} \\
\mathbf{Z}_{R}^{CD}%
\end{array}%
\right] .$$Here, the superscript PNC (or CD) represents the PNC (or complete-decoding) strategy.
Based on the signal model in (\[SS\]), the proposed space-division based relaying strategy is described as follows. At user $m$, two groups of spatial streams are generated: one group, referred to as the *complete-decoding spatial streams*, form the codeword matrix $\mathbf{X}%
_{m}^{CD}$; and the other group, referred to as the *PNC spatial streams*, form the codeword matrix $\mathbf{X}_{m}^{PNC}$, $m\in \left\{
A,B\right\} $.
### Complete-Decoding Spatial Streams
Due to the block triangular structure of the channel matrices in (\[SS\]), the relay can completely decode the spatial streams $\mathbf{X}_{A}^{CD}$ and $\mathbf{X}_{B}^{CD}$ free of interference from the PNC spatial streams. Specifically, the relay completely decodes both $\mathbf{X}_{A}^{CD}$ and $%
\mathbf{X}_{B}^{CD}$ based on$$\mathbf{Y}_{R}^{CD}=\dsum\limits_{m\in \left\{ A,B\right\} }\mathbf{D}%
_{m;2,2}\mathbf{R}_{m;2,2}\mathbf{X}_{m}^{CD}+\mathbf{Z}_{R}^{CD}.
\label{S_CD}$$Then, $\mathbf{X}_{A}^{CD}$ and $\mathbf{X}_{B}^{CD}$ are canceled from the received signal in (\[SS\]).
### PNC Spatial Streams
After the cancelation of $\mathbf{X}_{A}^{CD}$ and $\mathbf{X}_{B}^{CD}$, the system model for the PNC spatial streams is given by$$\mathbf{Y}_{R}^{PNC}=\dsum\limits_{m\in \left\{ A,B\right\} }\mathbf{D}%
_{m;1,1}\mathbf{R}_{m;1,1}\mathbf{X}_{m}^{PNC}+\mathbf{Z}_{R}^{PNC}.$$From (\[Dm\]), the first $k$ columns of $\mathbf{D}_{A;1,1}$ and $\mathbf{D%
}_{B;1,1}$ are identical; however, for $i=k+1,...,k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
$, the $i$th columns of $\mathbf{D}_{A;1,1}$ and $\mathbf{D}_{B;1,1}$ are not. Following Section III, we project each column pair of $%
\mathbf{D}_{A;1,1}$ and $\mathbf{D}_{B;1,1}$ onto a common direction, so as to facilitate PNC.
By inspection, the only difference between the $i$th columns of $\mathbf{D}%
_{A;1,1}$ and $\mathbf{D}_{B;1,1}$ is given by the 2-by-1 vectors $\mathbf{e}%
_{A;i}$ and $\mathbf{e}_{B;i}$, for $i=k+1,...,k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
$. Without loss of generality, denote by $\mathbf{p}_{i}$ a 2-by-1 unit vector representing the projection direction of $\mathbf{e}_{A;i}$ and $%
\mathbf{e}_{B;i}$. The choice of $\mathbf{p}_{i}$ is similar to that described in Section III and will be detailed in the next section.
Now the projection process can be described in a matrix form as follows. Define the projection matrix$$\mathbf{P}=\left[
\begin{array}{cccc}
\mathbf{I}_{k} & \mathbf{0} & \cdots & \mathbf{0} \\
\mathbf{0} & \mathbf{p}_{k+1} & \ddots & \vdots \\
\vdots & \ddots & \ddots & \mathbf{0} \\
\mathbf{0} & \cdots & \mathbf{0} & \mathbf{p}_{k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
}%
\end{array}%
\right] \in
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
^{(k+2l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
)\times (k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
)}\text{. }$$After the projection, the resulting signal model is given by$$\widetilde{\mathbf{Y}}_{R}^{PNC}=\mathbf{P}^{T}\mathbf{Y}_{R}^{PNC}=\dsum%
\limits_{m\in \left\{ A,B\right\} }\widetilde{\mathbf{H}}_{m}^{PNC}\mathbf{X}%
_{m}^{PNC}+\widetilde{\mathbf{Z}}_{R}^{PNC} \label{S_relay5}$$where $\widetilde{\mathbf{H}}_{m}^{PNC}=\mathbf{P}^{T}\mathbf{D}_{m;1,1}%
\mathbf{R}_{m;1,1}=\mathbf{\tilde{D}}_{m;1,1}\mathbf{R}_{m;1,1}\in
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
^{(k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
)\times (k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
)}$, with$$\mathbf{\tilde{D}}_{m;1,1}=\text{diag}\left\{ 1,...,1,\mathbf{p}_{k+1}^{T}%
\mathbf{e}_{m;k+1},...,\mathbf{p}_{k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
}^{T}\mathbf{e}_{m;k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
}\right\} ,m\in \left\{ A,B\right\} , \label{ddd}$$and $\widetilde{\mathbf{Z}}_{R}^{PNC}=\mathbf{P}^{T}\mathbf{Z}_{R}^{PNC}$ $%
\in
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
^{(n_{R}-l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
)\times 1}$ with the entries being i.i.d. random variables $\sim \mathcal{N}%
_{c}(0,N_{0}\dot{)}$. Note that the equivalent channel matrices $\widetilde{%
\mathbf{H}}_{A}^{PNC}$ and $\widetilde{\mathbf{H}}_{B}^{PNC}$ are $(k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
)$-by-$(k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
)$ square matrices. For such an equivalent MIMO TWRC, efficient techniques can be employed to align the signal directions of the two user into a same set of $k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
$ directions. This provides a platform to carry out $k$ PNC streams.
So far, we have presented the signal processing techniques used in the proposed space-division scheme to manipulate the uplink channel. The encoding and decoding details of the overall scheme will be described in the next section.
An Achievable Rate-Region of MIMO TWRC
======================================
In this section, we derive an achievable rate-pair of the proposed space-division based PNC scheme. Based on that, we optimize the system parameters to determine the achievable rate-region.
Achievable Rate-Pairs
---------------------
### Complete-Decoding Spatial Streams
The equivalent channel model seen by the complete-decoding spatial streams is given in (\[S\_CD\]), with the equivalent channel matrices given by $\mathbf{D}_{m,2,2}\mathbf{R}_{m;2,2}$, $m\in \left\{
A,B\right\} $.
The signal model in (\[S\_CD\]) is a standard MIMO MAC channel. Let $%
\mathbf{Q}_{m}^{CD}=\frac{1}{T}E\left[ \mathbf{X}_{m}^{CD}(\mathbf{X}%
_{m}^{CD})^{\dag }\right] $ be the input covariance matrix of the complete-decoding spatial steams of user $m$. Then, the achievable rate-pair of the complete-decoding spatial streams satisfies [@Cover91]
\[Rate CD\] $$\begin{aligned}
R_{A}^{CD}+R_{B}^{CD} &\leq &\frac{1}{2}\log \left\vert \mathbf{I}+\frac{1}{%
N_{0}}\dsum\limits_{m\in \left\{ A,B\right\} }\mathbf{D}_{m;2,2}\mathbf{R}%
_{m;2,2}\mathbf{Q}_{m}^{CD}\mathbf{R}_{m;2,2}^{\dag }\mathbf{D}%
_{m;2,2}^{\dag }\right\vert \\
R_{m}^{CD} &\leq &\frac{1}{2}\log \left\vert \mathbf{I}+\frac{1}{N_{0}}%
\mathbf{D}_{m;2,2}\mathbf{R}_{m;2,2}\mathbf{Q}_{m}^{CD}\mathbf{R}%
_{m;2,2}^{\dag }\mathbf{D}_{m;2,2}^{\dag }\right\vert ,m\in \left\{
A,B\right\} .\end{aligned}$$
### PNC Spatial Streams
The equivalent channel seen by the PNC streams is given in (\[S\_relay5\]). Recall that $\widetilde{\mathbf{H}}_{m}^{PNC}$ is a ($k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
$)-by-($k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
$) square matrix, and the efficient design of PNC for this case has been discussed in [@HJYangIT11]-[@KhinaISIT11]. Here, we follow the GSVD-based approach in [@HJYangIT11], as briefly described below.
Applying the generalized singular-value decomposition (GSVD) [@GolubMC97] to $\widetilde{\mathbf{H}}_{m}^{PNC}$, we obtain$$\widetilde{\mathbf{H}}_{m}^{PNC}=\mathbf{B\Sigma }_{m}\mathbf{T}%
_{m}^{\dagger },m\in \left\{ A,B\right\} ,$$where $\mathbf{B}\in
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
^{(k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
)\times (k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
)}$ is a nonsingular matrix, $\mathbf{T}_{m}\in
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
^{(k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
)\times (k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
)}$ is an orthogonal matrix, $m\in \left\{ A,B\right\} $, and $\mathbf{%
\Sigma }_{m}\in
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
^{(k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
)\times (k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
)}$ is a diagonal matrix with the $i$th diagonal element denoted by $\sigma
_{m;i}$. We further take the QR decomposition to the matrix $\mathbf{B}$, yielding$$\widetilde{\mathbf{H}}_{m}^{PNC}=\mathbf{Q\tilde{R}\Sigma }_{m}\mathbf{T}%
_{m}^{\dagger },m\in \left\{ A,B\right\} , \label{Decom}$$where $\mathbf{\tilde{R}}\in
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
^{(k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
)\times (k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
)}$ is an upper triangular matrix. The transmit signal $\mathbf{X}_{m}^{PNC}$ in (\[S\_relay5\]) is designed as$$\mathbf{X}_{m}^{PNC}=\mathbf{T}_{m}\mathbf{\Psi }_{m}^{1/2}\mathbf{S}%
_{m}^{PNC},m\in \left\{ A,B\right\} , \label{SigModel}$$where $\mathbf{\Psi }_{m}^{1/2}=$ diag$\left\{ \sqrt{\psi _{m;1}},\sqrt{\psi
_{m;2}},\cdots ,\sqrt{\psi _{m;k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
}}\right\} $ is a diagonal matrix with $\psi _{m;i}\geq 0,i=1,2,...,k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
$, and $\mathbf{S}_{m}$ $\in
%TCIMACRO{\U{2102} }%
%BeginExpansion
\mathbb{C}
%EndExpansion
^{(k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
)\times T}$ is the signal matrix with each element independent and identically drawn from $\mathcal{N}_{c}(0,1)$.
Let $R_{m}^{PNC}$ be the total rate of the PNC spatial streams of user $m$. From Theorem 1 in [@HJYangIT11], the achievable rate-pair is given by$$R_{m}^{PNC}=\sum_{i=1}^{k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
}\frac{1}{2}\left[ \log \left( \frac{I(i)\sigma _{m;i}^{2}\psi _{m;i}^{{}}}{%
\sigma _{A;i}^{2}\psi _{A;i}^{{}}+\sigma _{B;i}^{2}\psi _{B;i}^{{}}}+\frac{%
\tilde{r}_{i,i}^{2}\sigma _{m;i}^{2}\psi _{m;i}^{{}}}{N_{0}}\right) \right]
^{+},m\in \left\{ A,B\right\} \label{Rate PNC}$$where $I(i)$ is the indicator function with $I(i)=1$ for $i=1$ and $I(i)=0$ for $i\neq 1$.
### The Overall Scheme
We now consider the overall achievable rate-pair of the proposed space-division based PNC scheme. Before going into details, we note that the power constraint of user $m$, i.e., $\frac{1}{T}E\left[
\left\Vert \mathbf{X}_{m}\right\Vert _{F}^{2}\right] \leq P_{m},$ $m\in
\left\{ A,B\right\} $, can be equivalently expressed as$$\text{tr}\{\mathbf{Q}_{m}^{CD}\}+\sum_{i=1}^{k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
}\psi _{m;i}^{{}}\leq P_{m},m\in \left\{ A,B\right\} . \label{PC1}$$We are now ready to present the following theorem on the achievable rates of the proposed scheme.
For given $\mathbf{Q}_{m}^{CD}$, $\mathbf{\Psi }_{m}^{{}}$, and $\mathbf{Q}%
_{R}$ satisfying (\[PC1\]) and tr$(\mathbf{Q}_{R})\leq P_{R}$, a rate-pair $(R_{A},R_{B})$ for the MIMO TWRC is achievable if$$R_{m}\leq \min \{R_{m}^{CD}+R_{m}^{PNC},R_{m}^{DL}\},m\in \left\{
A,B\right\} , \label{49}$$where $R_{A}^{CD}$ and $R_{B}^{CD}$ satisfy (\[Rate CD\]), $R_{m}^{PNC}$ is given by (\[Rate PNC\]), and $R_{m}^{DL}$ is given by (\[Outer Bound\]).
Here we provide a sketch of proof. The overall encoding and decoding process for the proposed scheme is described as follows. The messages of the user $m$ are doubly indexed as $(W_{m}^{CD},W_{m}^{PNC})$, with $W_{m}^{CD}\in
\{1,2,...,2^{2TR_{m}^{CD}}\}$ for the complete-decoding spatial streams, and $W_{m}^{PNC}\in \{1,2,...,2^{2TR_{m}^{PNC}}\}$ for the PNC spatial streams. Each $W_{m}^{CD}$ is one-to-one mapped to $\mathbf{X}_{m}^{CD}$ in (\[SS\]), and each $W_{m}^{PNC}$ is one-to-one mapped to $\mathbf{S}_{m}^{PNC}$ in (\[SigModel\]). In the uplink phase, $\mathbf{X}_{m}^{CD}$ and $\mathbf{X}%
_{m}^{PNC}=\mathbf{T}_{m}\mathbf{\Psi }_{m}^{1/2}\mathbf{S}_{m}^{PNC}$ are transmitted via the channel in (\[SS\]), with the transmit power constrained by (\[PC1\]).
Upon receiving $\mathbf{Y}_{R}$, the relay first completely decode $\mathbf{X%
}_{A}^{CD}$ and $\mathbf{X}_{B}^{CD}$ based on $\mathbf{Y}_{R}^{CD}$ in ([S\_CD]{}), with the achievable rate-pair given in (\[Rate CD\]). The decoded $\mathbf{X}_{A}^{CD}$ and $\mathbf{X}_{B}^{CD}$ are subtracted from $\mathbf{%
Y}_{R}^{PNC}$. Let $(\mathbf{s}_{m,i}^{PNC})^{T}$ be the $i$th row of $%
\mathbf{S}_{m}^{PNC}$. Then, the network-coded PNC spatial streams, i.e., $%
\tilde{r}_{i,i}^{{}}\sigma _{A;i}^{{}}\psi _{A;i}^{1/2}\mathbf{s}%
_{A,i}^{PNC}+\tilde{r}_{i,i}^{{}}\sigma _{B;i}^{{}}\psi _{B;i}^{1/2}\mathbf{s%
}_{B,i}^{PNC}$, $i=k+l{\acute{}},k+l{\acute{}}-1,...,1$, are successively recovered and canceled from $\widetilde{\mathbf{Y%
}}_{R}^{PNC}$ in (\[S\_relay5\]), with the achievable rate-pair given in (\[Rate PNC\]). The decoded messages from the complete-decoding streams, together with the network-coded messages from the PNC streams, are then jointly encoded. The new codeword is forwarded to the two users in the downlink phase, with the transmit power constrained by tr$(\mathbf{Q}%
_{R})\leq P_{R}$. Following the discussions in [@HJYangIT11]-[KhinaISIT11]{}, the achievable rate-pair of the downlink phase is given by ($%
R_{A}^{DL},R_{B}^{DL}$) in (\[Outer Bound\]). This completes the proof.
Determining Achievable Rate-Region
----------------------------------
Now we consider determining the boundary of the achievable rate-region. From (\[49\]), the downlink achievable rates are the same as the capacity upper bound in (\[Outer Bound\]). Here, we focus on the uplink rate-region.
The boundary of the uplink rate-region can be determined by solving the following weighted-sum-rate maximization problem:
\[WSR\] $$\begin{aligned}
\text{maximize} &&\dsum\limits_{m\in \left\{ A,B\right\} }w_{m}\left(
R_{m}^{CD}+R_{m}^{PNC}\right) \\
\text{subject to} &&\dsum\limits_{j=1}^{k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
}\psi _{m;j}+\text{tr}\{\mathbf{Q}_{m}^{CD}\}\leq P_{m},\mathbf{Q}%
_{m}^{CD}\succeq \mathbf{0},\psi _{m;i}\geq 0,\text{ for }i=1,...,k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
. \\
&&R_{m}^{PNC}=\sum_{i=1}^{k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
}\frac{1}{2}\left[ \log \left( \frac{I(i)\sigma _{m;i}^{2}\psi _{m;i}^{{}}}{%
\sigma _{A;i}^{2}\psi _{A;i}^{{}}+\sigma _{B;i}^{2}\psi _{B;i}^{{}}}+\frac{%
\tilde{r}_{i,i}^{2}\sigma _{m;i}^{2}\psi _{m;i}^{{}}}{N_{0}}\right) \right]
^{+}, \\
&&R_{A}^{CD}+R_{B}^{CD}\leq \frac{1}{2}\log \left\vert \mathbf{I}+\frac{1}{%
N_{0}}\dsum\limits_{m\in \left\{ A,B\right\} }\mathbf{D}_{m;2,2}\mathbf{R}%
_{m;2,2}\mathbf{Q}_{m}^{CD}\mathbf{R}_{m;2,2}^{\dag }\mathbf{D}%
_{m;2,2}^{\dag }\right\vert , \\
&&R_{m}^{CD}\leq \frac{1}{2}\log \left\vert \mathbf{I}+\frac{1}{N_{0}}%
\mathbf{D}_{m;2,2}\mathbf{R}_{m;2,2}\mathbf{Q}_{m}^{CD}\mathbf{R}%
_{m;2,2}^{\dag }\mathbf{D}_{m;2,2}^{\dag }\right\vert ,m\in \left\{
A,B\right\} .\end{aligned}$$
The above problem involves the optimization of $l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
,\{\mathbf{p}_{i}\}_{i=k+1}^{k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
},\mathbf{Q}_{A}^{CD},\mathbf{Q}_{B}^{CD},\left\{ \psi _{A;i}\right\}
_{i=1}^{k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
}$, and $\left\{ \psi _{B;i}\right\} _{i=1}^{k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
}$, as detailed below.
### Determining the Projection Directions
The optimization of the projection directions $\{\mathbf{p}%
_{i}\}_{i=k+1}^{k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
}$ to maximize the weighted sum-rate is in general difficult to solve. To simplify the problem, we consider the high SNR regime, with the weighted sum-rate given by$$\begin{aligned}
&&w_{A}R_{A}^{PNC}+w_{B}R_{B}^{PNC}\overset{(a)}{\approx }\frac{1}{2}%
\dsum\limits_{m\in \left\{ A,B\right\} }w_{m}\left( \sum_{i=2}^{k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
}\log \left( \frac{\tilde{r}_{i,i}^{2}\sigma _{m;i}^{2}\psi _{m;i}^{{}}}{%
N_{0}}\right) \right) \notag \\
&&\overset{(b)}{=}\frac{1}{2}\dsum\limits_{m\in \left\{ A,B\right\}
}w_{m}\log \left\vert \frac{P_{m}}{N_{0}n_{m}}\mathbf{\tilde{R}\Sigma }_{m}%
\mathbf{\Sigma }_{m}^{\dag }\mathbf{\tilde{R}}^{\dag }\right\vert \notag \\
&&\overset{(c)}{=}\frac{1}{2}\dsum\limits_{m\in \left\{ A,B\right\}
}w_{m}\log \left\vert \frac{P_{m}}{N_{0}n_{m}}\mathbf{\tilde{D}}_{m;1,1}%
\mathbf{R}_{m;1,1}\mathbf{R}_{m;1,1}^{\dag }\mathbf{\tilde{D}}_{m;1,1}^{\dag
}\right\vert \notag \\
&&\overset{(d)}{=}\frac{1}{2}\dsum\limits_{m\in \left\{ A,B\right\} }\left(
w_{m}\log \left\vert \frac{P_{m}}{N_{0}n_{m}}\mathbf{R}_{m;1,1}\mathbf{R}%
_{m;1,1}^{\dag }\right\vert +w_{m}\log \left\vert \mathbf{\tilde{D}}_{m;1,1}%
\mathbf{\tilde{D}}_{m;1,1}^{\dag }\right\vert \right) \label{WSR1}\end{aligned}$$where step (*a*) follows from substituting (\[Rate PNC\]), step (*b*) from the facts that $\mathbf{\tilde{R}}$ is upper-triangular and that equal power allocation is asymptotically optimal (i.e., $\psi
_{m;i}^{{}}=\frac{P_{m}}{n_{m}}$), step (*c*) by noting $\mathbf{%
\tilde{D}}_{m;1,1}\mathbf{R}_{m;1,1}=\widetilde{\mathbf{H}}_{m}^{PNC}=%
\mathbf{Q\tilde{R}\Sigma }_{m}\mathbf{T}_{m}^{\dagger }$ (cf., (\[S\_relay5\]) and (\[Decom\])), and step (d) by utilizing $\left\vert \mathbf{I}+\mathbf{AB}\right\vert =\left\vert \mathbf{I}+%
\mathbf{BA}\right\vert $. In the above, $\log \left\vert \mathbf{\tilde{D}}%
_{m;1,1}\mathbf{\tilde{D}}_{m;1,1}^{\dag }\right\vert $ is the only term related to $\mathbf{p}_{i}$. Recall from (\[ddd\]) that $\mathbf{\tilde{D}}_{m;1,1}=%
\mathbf{P}^{T}\mathbf{D}_{m;1,1}$ with $\mathbf{D}_{m;1,1}$ being the principle submatrix of $\mathbf{D}_{m}$ in (\[Dm3\]). Thus, the weighted sum-rate maximization problem over $\{\mathbf{p}_{i}\}_{i=k+1}^{k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
}$ can be decoupled into $l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
$ independent subproblems as$$\max_{||\mathbf{p}_{i}||=1}w_{A}\log \left( \left\vert \mathbf{p}_{i}^{\dag }%
\mathbf{e}_{A;i}\right\vert ^{2}\right) +w_{B}\log \left( \left\vert \mathbf{%
p}_{i}^{\dag }\mathbf{e}_{B;i}\right\vert ^{2}\right) ,\text{ for }%
i=k+1,...,k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
,$$where $\mathbf{e}_{A;i}$ and $\mathbf{e}_{B;i}$ are given in (\[Direction\]). From (\[d\]) and the discussions therein, the optimal $\mathbf{p}_{i}$ to maximize the weighted sum-rate is a real vector given by$$\mathbf{p}_{i}=\gamma _{i}\left( \mathbf{e}_{A;i}+\beta _{i}\mathbf{e}%
_{B;i}\right) ,\text{ for }i=k+1,...,k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
, \label{dd}$$where$$\beta _{i}=\frac{1}{2}\left( \sqrt{\left( \lambda _{i}-1\right) ^{2}\left( 1-%
\frac{w_{B}}{w_{A}}\right) ^{2}+4\frac{w_{B}}{w_{A}}}-\left( \lambda
_{i}-1\right) \left( 1-\frac{w_{B}}{w_{A}}\right) \right) ,$$and $\gamma _{i}$ is a scaling factor to ensure $||\mathbf{p}_{i}||$ $=1$.
### Determining $\mathbf{Q}_{A}^{CD}$ and $\mathbf{Q}_{B}^{CD}$
Given $\{\mathbf{p}_{i}\}$ in (\[dd\]), the optimization problem in ([WSR]{}) can be decoupled into two separate problems by predetermining the power allocated to the two signal subspaces. Let $P_{m}^{CD}$ be the power of user $m$ used for the complete-decoding spatial streams. Then, the power for the PNC streams is given by $P_{m}^{PNC}=P_{m}^{{}}-P_{m}^{CD},m\in
\left\{ A,B\right\} $. For given $P_{A}^{CD}$ and $P_{B}^{CD}$, the optimal $%
\mathbf{Q}_{A}^{CD}$ and $\mathbf{Q}_{B}^{CD}$ to (\[WSR\]) can be found by solving the following problem:
$$\begin{aligned}
\text{maximize} &&w_{A}R_{A}^{CD}+w_{A}R_{A}^{CD} \\
\text{subject to} &&\text{tr}\{\mathbf{Q}_{m}^{CD}\}\leq P_{m}^{CD},\mathbf{Q%
}_{m}^{CD}\succeq \mathbf{0}, \\
&&R_{A}^{CD}+R_{B}^{CD}\leq \frac{1}{2}\log \left\vert \mathbf{I}+\frac{1}{%
N_{0}}\dsum\limits_{m\in \left\{ A,B\right\} }\mathbf{D}_{m,2,2}\mathbf{R}%
_{m;2,2}\mathbf{Q}_{m}^{CD}\mathbf{R}_{m;2,2}^{\dag }\mathbf{D}%
_{m,2,2}^{\dag }\right\vert \\
&&R_{m}^{CD}\leq \frac{1}{2}\log \left\vert \mathbf{I}+\frac{1}{N_{0}}%
\mathbf{D}_{m,2,2}\mathbf{R}_{m;2,2}\mathbf{Q}_{m}^{CD}\mathbf{R}%
_{m;2,2}^{\dag }\mathbf{D}_{m,2,2}^{\dag }\right\vert ,m\in \left\{
A,B\right\} .\end{aligned}$$
The above is a standard weighted sum-rate maximization problem for a MIMO multiple-access channel with two users [@YuIT04]. This problem is convex, and the optimal solution can be numerically obtained using convex programming tools [@Boyd04].
### Determining Power Allocation for PNC Streams
Now we consider the optimization of $\left\{ \psi _{A;i}\right\} _{i=1}^{k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
}$ and $\left\{ \psi _{B;i}\right\} _{i=1}^{k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
}$. Given $P_{A}^{PNC}$ and $P_{B}^{PNC}$, the optimal $\left\{ \psi
_{A;i}\right\} _{i=1}^{k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
}$ and $\left\{ \psi _{B;i}\right\} _{i=1}^{k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
}$ can be determined by solving
$$\begin{aligned}
\text{maximize} &&\dsum\limits_{m\in \left\{ A,B\right\} }w_{m}\left(
\sum_{i=1}^{k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
}\frac{1}{2}\left[ \log \left( \frac{I(i)\sigma _{m;i}^{2}\psi _{m;i}^{{}}}{%
\sigma _{A;i}^{2}\psi _{A;i}^{{}}+\sigma _{B;i}^{2}\psi _{B;i}^{{}}}+\frac{%
\tilde{r}_{i,i}^{2}\sigma _{m;i}^{2}\psi _{m;i}^{{}}}{N_{0}}\right) \right]
^{+}\right) \\
\text{subject to} &&\dsum\limits_{i=1}^{k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
}\psi _{m;i}\leq P_{m}^{PNC},\psi _{m;i}\geq 0,\text{ for }i=1,...,k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
.\end{aligned}$$
A similar problem has been considered in [@HJYangIT11], and the optimal solution can be obtained by solving the Karush-Kuhn-Tuchker (KKT) conditions. We omit details here for simplicity.
Base on the above discussions, the weighted sum-rate problem in (\[WSR\]) is numerically solvable given the values of $l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
,P_{m}^{CD}$ and $P_{m}^{PNC}$, $m\in \left\{ A,B\right\} $. The optimal $l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
,P_{m}^{CD}$ and $P_{m}^{PNC}$, $m\in \left\{ A,B\right\} $ can be found using the exhaustive search. The complexity involved is not significant by noting $P_{m}^{CD}+P_{m}^{PNC}=P_{m},m\in \left\{ A,B\right\} $ and the fact that $l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
$ is an integer between $0$ and $l$.
Asymptotic Sum-Rate Analysis
============================
In the preceding section, we have shown the achievable rates of the proposed space-division based network-coding strategy for MIMO TWRCs. In general, it is difficult to represent the achievable rate of the optimized space-division based scheme in a closed-form. Thus, it is not easy to evaluate the gap between the achievable rate of the proposed scheme and the capacity upper bound of the MIMO TWRC. In this section, we derive a closed-form expression for the asymptotic sum-rate of the proposed strategy in the high SNR regime.
Asymptotic Sum-Rate as SNR $\rightarrow \infty $
------------------------------------------------
Here, we analyze the uplink achievable sum-rate$$R^{SD}=\dsum\limits_{m\in \left\{ A,B\right\} }R_{m}^{CD}+R_{m}^{PNC}$$as the SNRs, i.e., $\frac{P_{A}}{N_{0}}$ and $\frac{P_{B}}{N_{0}}$, tend to infinity. It is known that, in the high SNR regime, equal power allocation is asymptotically optimal. Then, the upper bound of the uplink sum-rate of the MIMO TWRC is given by (cf., (\[Outer Bound\]))$$R^{UL}\approx \frac{1}{2}\dsum\limits_{m\in \left\{ A,B\right\} }\log
\left\vert \mathbf{I}_{n_{R}}+\frac{P_{m}}{N_{0}n_{m}}\mathbf{H}_{m}\mathbf{H%
}_{m}^{\dagger }\right\vert \label{UpperBound}$$where $x\approx y$ means$$\lim_{SNR\longrightarrow \infty }\left( x-y\right) =0. \notag$$
Now, we present the following theorem on the asymptotic sum-rate of the proposed scheme. Denote by $R^{SD}$ the uplink achievable sum-rate of the proposed space-division scheme.
\[Theorem 3\]For a given $l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
$, the uplink achievable sum-rate of the proposed space-division scheme satisfies
$$\lim_{SNR\longrightarrow \infty }R^{UL}-R^{SD}=\Delta ^{SD}
\label{Inequality}$$
where$$\Delta ^{SD}\triangleq -\log \dprod\limits_{i=k+1}^{k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
}\frac{\lambda _{i}}{2}-\log \dprod\limits_{i=k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
+1}^{k+l}\sqrt{\lambda _{i}(2-\lambda _{i})}\geq 0. \label{Delta_SD}$$
The proof of Theorem 3 can be found in Appendix B. Notice that the first term in (\[Delta\_SD\]), i.e., $-\log \dprod\limits_{i=k+1}^{k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
}\frac{\lambda _{i}}{2}$, is the rate loss incurred by the PNC spatial streams, and the second term, i.e., $\log \dprod\limits_{i=k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
+1}^{k+l}\sqrt{\lambda _{i}(2-\lambda _{i})}$, is that incurred by the complete-decoding spatial streams.
For the case of $n_{A},n_{B}\geq n_{R}$, we have $l=0$ and $\lambda _{i}=2$ for $i=1,...,k$. (See Remark (\[Remark 4\]).) Then, from (\[Delta\_SD\]), we have $\Delta ^{SD}=0$, which means that the scheme is asymptotically optimal. This agrees with the fact that our proposed space-division scheme reduces to the GSVD scheme which is indeed asymptotically optimal in the high SNR regime [@HJYangIT11].
The optimal $l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
$ to minimize the rate gap $\Delta ^{SD}$ in (\[Delta\_SD\]) satisfies$$2>\lambda _{k+1}\geq ...\geq \lambda _{k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
}\geq \frac{8}{5}>\lambda _{k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
+1}\geq ...\geq \lambda _{k+l}>1. \label{Range1}$$With this choice of $l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
$, the asymptotic rate gap $\Delta ^{SD}$ is at most $l\log (5/4)$ bits, which occurs when $\lambda _{k+1}=\lambda _{k+2}=...=\lambda _{k+l}=\frac{8}{%
5}$.
From the above corollary, the asymptotic gap to the sum-capacity upper bound is $l\log (5/4)$ bits for the worst case. Noting $l\leq n_{m},m\in \left\{
A,B\right\} $, we see that the gap is at most $\min \{n_{A},n_{B}\}\log (5/4)$ bits, or $\frac{1}{2}\log (5/4)\approx
0.161$ bits per user-antenna.
Average Sum-Rate via Large-System Analysis
------------------------------------------
In this subsection, we investigate the statistical average of the rate gap $%
\Delta ^{SD}$ in fading channels. To this end, the distribution of {$\lambda _{i}$}, i.e., the eigenvalues of $%
\mathbf{U}_{A}\mathbf{U}_{A}^{\dagger }+\mathbf{U}_{B}\mathbf{U}%
_{B}^{\dagger }$, is required. However, such a distribution is difficult to obtain in general. Here, we employ the large-system analysis to find an approximation of the distribution of {$\lambda _{i}$}. The distribution obtained in this way becomes exact as the number of antennas in the system is large.
We assume Rayleigh fading, in which the channel coefficients are i.i.d. circularly symmetric complex Gaussian random variables. Then, the matrices $%
\mathbf{U}_{A}$ and $\mathbf{U}_{B}$ in (\[Channel\_Decomp\]) are truncated uniformly distributed unitary matrices, or alternatively, are asymptotically free random matrices [@TulinoTextBook]. Thus, we can use the theory of free probabilities to derive the asymptotic eigenvalue distribution (a.e.d.) of $\mathbf{U}_{A}\mathbf{U}_{A}^{\dagger }+\mathbf{U}_{B}\mathbf{U}%
_{B}^{\dagger }$ as $n_{R}$ tends to infinity, with the result given in the lemma below. Define $\eta _{m}\triangleq \frac{n_{m}}{n_{R}},m\in \left\{
A,B\right\} $.
\[Lemma 4\]As $n_{R}\rightarrow \infty $ with $\frac{n_{A}}{n_{R}}%
\rightarrow \eta _{A}$ and $\frac{n_{B}}{n_{R}}\rightarrow \eta _{B}$, the a.e.d. of $\mathbf{U}_{A}\mathbf{U}_{A}^{\dagger }+\mathbf{U}%
_{B}\mathbf{U}_{B}^{\dagger }$ is given by$$\begin{aligned}
\mathcal{F}\left( \lambda ;\eta _{A},\eta _{B}\right) &=&\left[ 1-\eta
_{A}-\eta _{B}\right] ^{+}\delta \left( \lambda \right) +\left\vert \eta
_{A}-\eta _{B}\right\vert \delta \left( \lambda -1\right) +\left[ \eta
_{A}+\eta _{B}-1\right] ^{+}\delta \left( \lambda -2\right) \notag \\
&&+\frac{1}{\pi }\func{Im}\left[ \frac{\sqrt{(1-\eta _{A}-\eta
_{B})^{2}-\left( 2\lambda -\lambda ^{2}\right) \left( 1- \left( \frac{\eta
_{A}-\eta _{B}}{\lambda -1}\right) ^{2}\right) }}{2\lambda -\lambda ^{2}}%
\right] \label{Distribution}\end{aligned}$$where $\delta \left( \cdot \right) $ is a Dirac delta function and $\func{Im}%
\left[ \cdot\right] $ is the imaginary part of a complex number.
The proof of the above lemma can be found in Appendix C. As $%
n_{R}\rightarrow \infty $, we see that for $\eta _{A}+\eta _{B}\geq 1$, the portion of eigenvalues $\left\{ \lambda _{i}\right\} $ equal to 2 is given by $\eta _{A}+\eta _{B}-1$. This portion corresponds to the dimension of the common space $\mathcal{S}_{A\parallel B}$ of $\textbf{H}_{AR}$ and $\textbf{H}_{BR}$. In addition, for $\eta _{A}\neq \eta _{B}$, the portion of eigenvalues $\left\{ \lambda _{i}\right\} $ equal to 1 is given by $\left\vert \eta _{A}-\eta _{B}\right\vert $. This portion corresponds to the dimension of $\mathcal{S}_{A\perp B}$ if $\eta_A \geq \eta_B$ or the dimension of $\mathcal{S}_{B\perp A}$ if $\eta_A < \eta_B$.
We are now ready to present the following asymptotic result.
\[Theorem 4\]As $n_{R}\rightarrow \infty $ with $\frac{n_{A}}{n_{R}}%
\rightarrow \eta _{A}$ and $\frac{n_{B}}{n_{R}}\rightarrow \eta _{B}$, the gap to the capacity upper bound satisfies $$r^{SD} \triangleq \lim_{n_{R}\rightarrow \infty }\text{ }\frac{\Delta ^{SD}}{n_{R}}\text{ }%
=-\left( \int_{1}^{\frac{8}{5}}\log \sqrt{\lambda (2-\lambda )}+\int_{\frac{8%
}{5}}^{2}\log \frac{\lambda }{2}\right) \mathcal{F}(\lambda ;\eta _{A},\eta
_{B})d\lambda . \label{Limit1}$$
The a.e.d. of $\lambda _{i}$ is given by Lemma \[Lemma 4\]. Then, letting $%
n_{R}$ tends to infinity in (\[Delta\_SD\]), we immediately obtain the theorem.
Let $\overline{R}^{UL}$ be the average sum-capacity upper bound. Then, for a large $n_{R}$, the average sum-rate of the proposed SD scheme can be first-order approximated as$$\label{approx}
\overline{R}^{SD}=\overline{R}^{UL}-n_{R}r^{SD}$$with $r^{SD}$ given in (\[Limit1\]).
We next study the symmetric case that the two users are equipped with the same number of antennas, i.e., $\eta _{A}=\eta _{B}=\eta $.
\[Corollary 5\] For $0\leq \eta \leq \frac{1}{10}$,
$$r^{SD} %
=-\int_{1}^{\lambda ^{\ast }(\eta)}\log \sqrt{\lambda (2-\lambda )}\mathcal{G}%
(\lambda ;\eta )d\lambda ; \label{Limit2}$$
for $\frac{1}{10}<\eta \leq 1$,$$r^{SD}=-\left(
\int_{1}^{\frac{8}{5}}\log \sqrt{\lambda (2-\lambda )}+\int_{\frac{8}{5}%
}^{\lambda ^{\ast }(\eta )}\log \frac{\lambda }{2}\right) \mathcal{G}%
(\lambda ;\eta )d\lambda , \label{Limit3}$$where $\lambda ^{\ast }(\eta )=1+\sqrt{1-\left( 1-2\eta \right) ^{2}}$ and$$\mathcal{G}\left( \lambda ;\eta \right) =\frac{1}{\pi }\frac{\sqrt{\left(
2\lambda -\lambda ^{2}\right) -(1-2\eta )^{2}}}{2\lambda -\lambda ^{2}}.
\label{G_lemda}$$
Letting $\eta _{A}=\eta _{B}=\eta $, we obtain that $\mathcal{F}\left(
\lambda ;\eta _{A},\eta _{B}\right) =\mathcal{G}\left( \lambda ;\eta \right)
$ for $1<\lambda <\lambda ^{\ast },$ and $\mathcal{F}\left( \lambda ;\eta
_{A},\eta _{B}\right) =0$ for $\lambda ^{\ast }<\lambda <2$. In addition, $%
\lambda ^{\ast }(\eta)=\frac{8}{5}$ implies $\eta =\frac{1}{10}$. Based on these facts and Theorem \[Theorem 4\], we obtain the corollary.
From the above, we see that, if $\eta \leq \frac{1}{10},$ the probability of $\lambda _{i}>\frac{8}{5}$ approaches zero as $n_{R}\rightarrow \infty $, implying that complete decoding achieves a higher rate than PNC for all spatial streams.
\[Corollary 6\]The asymptotic normalized rate gap $r^{SD}$ in ([Limit1]{}) is maximized at $\eta _{A}=\eta _{B}=1/2$, with the maximum given by$$-\frac{1}{\pi }\left( \int_{1}^{\frac{8}{5}}\frac{\log \sqrt{\lambda
(2-\lambda )}}{\sqrt{2\lambda -\lambda ^{2}}}d\lambda +\int_{\frac{8}{5}}^{2}%
\frac{\log \frac{\lambda }{2}}{\sqrt{2\lambda -\lambda ^{2}}}d\lambda
\right) \approx 0.053\text{ bit.}$$
We first consider optimizing $\eta _{A}$ and $\eta _{B}$ under the constraint of $\eta _{A}+\eta _{B}=2\eta $. From (\[Distribution\]), we see that, for any $\lambda \in (1,2)$, $\mathcal{F}\left( \lambda ;\eta
_{A},\eta _{B}\right) $ is maximized at $\eta _{A}=\eta _{B}=\eta $, and so is $r^{SD}$.
What remains is to optimize $\eta $. From (\[G\_lemda\]), $\mathcal{G}\left( \lambda ;\eta \right) $ is maximized at $\eta =1/2$. Therefore, $r^{SD}$ is maximized at $\eta =1/2$, which completes the proof.
Fig. \[SD\_rate\_gp\] illustrates the function of the normalized asymptotic rate gap $r^{SD}$ against $\eta $. From Fig. \[SD\_rate\_gp\], this rate gap is maximized at $%
\eta =1/2$, which verifies Corollary \[Corollary 6\]. Also, this rate gap vanishes as $\eta $ tends to 0, implying that, for any fixed $n_{A}=n_{B}$, the proposed space-division scheme can achieve the asymptotic capacity as $%
n_{R}$ tends to infinity. Moreover, this rate gap vanishes as $\eta $ tends to 1. This agrees with the fact that, for $\eta \geq 1$, or equivalently, $%
n_{A}=n_{B}\geq n_{R}$, the proposed space-division based scheme reduces to the GSVD scheme in [@HJYangIT11].
Numerical Results
=================
In this section, we provide numerical results to evaluate the performance of the proposed space-division based network-coding strategy for MIMO TWRCs. The results presented below are obtained by averaging over 1,0000 random channel realizations. Rayleigh-fading is assumed, i.e., the coefficients in the channel matrices are independently and identically drawn from $\mathcal{N%
}_{c}(0,1)$.
We first present the numerical results for a MIMO TWRC of $n_{A}=n_{B}=2$ and $n_{R}=4$ in Fig. \[Figure nT2nR4\]. The sum-capacity upper bound (UB), the proposed space-division (SD) scheme, the GSVD scheme in [HJYangIT11]{} and the complete-decoding scheme in [@YangIT11] are included for comparison. We see that, at a relatively high SNR, e.g., SNR = 25 dB, the rate gap between the proposed SD scheme and the sum-capacity upper bound is about 0.15 bit/channel-use, which is almost unnoticeable. We also plot the high-SNR analytical result in (\[approx\]) of the proposed SD scheme. We observe that our analytical result are very tight for SNRs greater than 10 dB. From this figure, it is clear that the proposed SD schemesignificantly outperforms the other schemes in the entire SNR range of interest. For example, at the rate of $14$ bits per channel use, the proposed SD scheme outperforms the complete-decoding and GSVD schemes by more than $2.4$ dB. The slope of the achievable sum-rate curve is parallel to that of the capacity upper bound, which implies that the proposed SD scheme achieves full multiplexing gain.
In Fig. \[Figure nT2nR3\], we present the numerical results for a MIMO TWRC of $n_{A}=n_{B}=2$ and $n_{R}=3$. The same set of rate curves from simulation as in Fig. \[Figure nT2nR4\] are included for comparison. Again, we see that the gap between the sum-rate of the proposed SD scheme and the sum-capacity upper bound is almost unnoticeable at a relatively high SNR, e.g., greater than 15 dB. The proposed SD scheme outperforms its counterparts throughout the SNR range of interest.
In Figures \[Figure Eda1\_2\] and \[Figure Eda2\_3\], we show the scaling effect of the antennas on the average achievable sum-rates. We see that the asymptotic rate gap between the proposed SD scheme and the sum-capacity upper bound increases linearly as the increase of $n_{R}$ for fixed $\eta _{A}$ and $\eta _{B}$. For example, for the case of $\eta _{A}=\eta
_{B}=1/2$ in Fig. \[Figure Eda1\_2\], the rate gap at SNR = 25 dB is 0.14 bits per channel use for $n_{R}=4$; 0.29 bits per channel use for $n_{R}=6$; and 0.40 bits per channel use for $n_{R}=8$. These numerical results agree well with the asymptotic results in Corollaries \[Corollary 5\] and [Corollary 6]{}.
In Fig. \[Figure CapacityRegion\], we show the achievable rate-region of the proposed SD scheme. The capacity-region outer bound and the achievable rate-region of the complete-decoding scheme are also included for comparison. From Fig. \[Figure CapacityRegion\], the difference between the achievable rate-region of the proposed SD scheme and the capacity region outer bound is negligible for a relatively high SNR. We also see that the proposed SD scheme can achieve rate-pairs that cannot be achieved by the complete-decoding scheme.
Conclusion
==========
In this paper, we developed a new joint channel decomposition for MIMO TWRCs. Based on that, we proposed a space-division based network-coding scheme with the achievable sum-rate within $\frac{1}{2}\log (5/4)\approx
0.161$ bit per user-antenna of the capacity upper bound in the high SNR regime. We also show that, for Rayleigh-fading MIMO TWRCs, the average gap between the achievable rate of the proposed scheme and the capacity upper bound is no more than $0.053$ bit per relay-antenna in the high SNR regime. We remark that this marginal gap is due to the fact that the complete-decoding and PNC strategies, collectively, fail to achieve the asymptotic capacity upper bound, even for the case of single-antenna users. To completely remove this gap, more advanced multi-dimension PNC relaying strategies may be required. Moreover, in this paper, channel state information is assumed to be globally known by both the transmitter and receiver sides. It is of theoretical, and more practical, interests to investigate how to efficiently communicate over MIMO TWRCs where only the receiver-side channel state information is available. We will look into these problems in our future research.
Proof of Theorem \[Theorem 1\]
==============================
Here we prove Theorem \[Theorem 1\]. Let $\lambda _{i}$ be an eigenvalue of $\mathbf{U}_{A}\mathbf{U}_{A}^{\dagger }+\mathbf{U}_{B}\mathbf{U}%
_{B}^{\dagger }$ and $\mathbf{u}_{i}$ be the corresponding unit-length eigenvector satisfying$$\left( \mathbf{U}_{A}\mathbf{U}_{A}^{\dagger }+\mathbf{U}_{B}\mathbf{U}%
_{B}^{\dagger }\right) \mathbf{u}_{i}=\lambda _{i}\mathbf{u}_{i}.
\label{eigen}$$We are interested in four cases of $\lambda _{i}$: (a) $\lambda _{i}=2$; (b) $1<\lambda _{i}<2;$ (c) $\lambda _{i}=1$; and (d) $0<\lambda _{i}<1$.
For case (a), $\lambda _{i}=2$ implies that$$\mathbf{U}_{A}\mathbf{U}_{A}^{\dagger }\mathbf{u}_{i}=\mathbf{u}_{i}\text{
and }\mathbf{U}_{B}\mathbf{U}_{B}^{\dagger }\mathbf{u}_{i}=\mathbf{u}_{i}%
\text{.}$$Thus, $\mathbf{u}_{i}$ lies in the common space of $\mathcal{C}(\mathbf{U}%
_{A})$ and $\mathcal{C}(\mathbf{U}_{B})$.
For case (c), we have
$$\mathbf{U}_{A}\mathbf{U}_{A}^{\dagger }\mathbf{u}_{i}=\mathbf{u}_{i}\text{
and }\mathbf{U}_{B}\mathbf{U}_{B}^{\dagger }\mathbf{u}_{i}=\mathbf{0}
\label{case c1}$$
or$$\mathbf{U}_{A}\mathbf{U}_{A}^{\dagger }\mathbf{u}_{i}=\mathbf{0}\text{ and }%
\mathbf{U}_{B}\mathbf{U}_{B}^{\dagger }\mathbf{u}_{i}=\mathbf{u}_{i}\text{.}
\label{case c2}$$
We next show that the eigenvalues in case (b) and case (d) appear in a pair-wise manner. Denote$$\mathbf{l}_{m;i}=\mathbf{U}_{m}\left( \mathbf{U}_{m}^{\dagger }\mathbf{U}%
_{m}\right) ^{-1}\mathbf{U}_{m}^{\dagger }\mathbf{u}_{i}=\mathbf{U}_{m}%
\mathbf{U}_{m}^{\dagger }\mathbf{u}_{i}. \label{Eq 3}$$Note that $\mathbf{l}_{m;i}$ is the projection of vector $\mathbf{u}_{i}$ onto the column space of $\mathbf{U}_{m}$. From (\[eigen\]), we obtain$$\mathbf{u}_{i}=\frac{1}{\lambda _{i}}\left( \mathbf{l}_{A;i}+\mathbf{l}%
_{B;i}\right) . \label{Eq 4}$$The above implies that $\mathbf{u}_{i}$, $\mathbf{l}_{A;i}$ and $\mathbf{l}%
_{B;i}$ lie on the same two-dimension plane (denoted by $\mathcal{S}_{i}$). We have the following facts.
For any $\lambda _{i}$ in case (b), the corresponding $\mathbf{u}_{i}$ is the angular bisector of $\mathbf{l}_{A;i}$ and $\mathbf{l}_{B;i}$, i.e.$$\left\Vert \mathbf{l}_{A;i}\right\Vert ^{2}=\mathbf{u}_{i}^{\dagger }\mathbf{%
l}_{A;i}=\mathbf{u}_{i}^{\dagger }\mathbf{l}_{B;i}=\left\Vert \mathbf{l}%
_{B;i}\right\Vert ^{2}. \label{Eq 5}$$
To prove the lemma, we first multiply both sides of (\[eigen\]) by $%
\mathbf{U}_{A}\mathbf{U}_{A}^{\dagger }$. Then, after some straightforward manipulations, we obtain$$\left( \lambda _{i}-1\right) \mathbf{U}_{A}\mathbf{U}_{A}^{\dagger }\mathbf{u%
}_{i}=\mathbf{U}_{A}\mathbf{U}_{A}^{\dagger }\mathbf{U}_{B}\mathbf{U}%
_{B}^{\dagger }\mathbf{u}_{i}. \label{Eq 6}$$Similarly, we have $$\left( \lambda _{i}-1\right) \mathbf{U}_{B}\mathbf{U}_{B}^{\dagger }\mathbf{u%
}_{i}=\mathbf{U}_{B}\mathbf{U}_{B}^{\dagger }\mathbf{U}_{A}\mathbf{U}%
_{A}^{\dagger }\mathbf{u}_{i}. \label{Eq 7}$$Then, $$\begin{aligned}
& \mathbf{u}_{i}^{\dagger }\mathbf{l}_{A;i}\text{ }\overset{(a)}{=}\mathbf{u}%
_{i}^{\dagger }\mathbf{U}_{A}\mathbf{U}_{A}^{\dagger }\mathbf{u}_{i}\overset{%
(b)}{=}\frac{1}{\lambda _{i}-1}\mathbf{u}_{i}^{\dagger }\mathbf{U}_{A}%
\mathbf{U}_{A}^{\dagger }\mathbf{U}_{B}\mathbf{U}_{B}^{\dagger }\mathbf{u}%
_{i} \\
& \overset{(c)}{=}\frac{1}{\lambda _{i}-1}\mathbf{u}_{i}^{\dagger }\mathbf{U}%
_{B}\mathbf{U}_{B}^{\dagger }\mathbf{U}_{A}\mathbf{U}_{A}^{\dagger }\mathbf{u%
}_{i}\overset{(d)}{=}\mathbf{u}_{i}^{\dagger }\mathbf{U}_{B}\mathbf{U}%
_{B}^{\dagger }\mathbf{u}_{i}\overset{(e)}{=}\mathbf{u}_{i}^{\dagger }%
\mathbf{l}_{B;i}\text{ }\end{aligned}$$where step $(a)$ follows from (\[Eq 3\]), $(b)$ from (\[Eq 6\]), $(c)$ from the fact that the Hermitian transpose of a real-valued scalar is itself, $(d)$ from (\[Eq 7\]), and $(e)$ again from (\[Eq 3\]). From (\[Eq 3\]), the projection of $\mathbf{u}_{i}$ onto $\mathbf{l}_{m;i}$ is just $\mathbf{l}_{m;i}$. Thus, $%
\left\Vert \mathbf{l}_{m;i}\right\Vert ^{2}=\mathbf{u}_{i}^{\dagger }\mathbf{%
l}_{m;i}$, which completes the proof.
For any $\lambda _{i}$ $\in (1,2)$ (as in case (b)), $\lambda _{i}^{\prime
}=2-\lambda _{i}$ is also an eigenvalue of $\mathbf{U}_{A}\mathbf{U}%
_{A}^{\dagger }+\mathbf{U}_{B}\mathbf{U}_{B}^{\dagger }$, and the corresponding unit-length eigenvector is given by$$\mathbf{u}_{i}^{\prime }=\frac{1}{\sqrt{\lambda _{i}\lambda _{i}^{\prime }}}%
\left( \mathbf{l}_{A;i}-\mathbf{l}_{B;i}\right) . \label{Eq 9}$$
By definition, we have$$\begin{aligned}
& \left( \mathbf{U}_{A}\mathbf{U}_{A}^{\dagger }+\mathbf{U}_{B}\mathbf{U}%
_{B}^{\dagger }\right) \mathbf{u}_{i}^{\prime } \notag \\
& \overset{(a)}{=}\frac{1}{\sqrt{\lambda _{i}\lambda _{i}^{\prime }}}\left(
\mathbf{U}_{A}\mathbf{U}_{A}^{\dagger }+\mathbf{U}_{B}\mathbf{U}%
_{B}^{\dagger }\right) \left( \mathbf{l}_{A;i}-\mathbf{l}_{B;i}\right)
\notag \\
& \overset{(b)}{=}\frac{1}{\sqrt{\lambda _{i}\lambda _{i}^{\prime }}}\left(
\mathbf{U}_{A}\mathbf{U}_{A}^{\dagger }\mathbf{u}_{i}+\left( \lambda
_{i}-1\right) \mathbf{U}_{B}\mathbf{U}_{B}^{\dagger }\mathbf{u}_{i}-\left(
\lambda _{i}-1\right) \mathbf{U}_{A}\mathbf{U}_{A}^{\dagger }\mathbf{u}_{i}-%
\mathbf{U}_{B}\mathbf{U}_{B}^{\dagger }\mathbf{u}_{i}\right) \notag \\
& =\frac{\lambda _{i}^{\prime }}{\sqrt{\lambda _{i}\lambda _{i}^{\prime }}}%
\left( \mathbf{l}_{A;i}-\mathbf{l}_{B;i}\right) =\lambda _{i}^{\prime }%
\mathbf{u}_{i}^{\prime }\end{aligned}$$where step $(a)$ follows from (\[Eq 9\]), and step $(b)$ from (\[Eq 3\]), (\[Eq 6\]) and (\[Eq 7\]).
What remains is to show that $\left\Vert \mathbf{u}_{i}^{\prime }\right\Vert
=1$. To see this, we left-multiply both sides of (\[eigen\]) by $\mathbf{u}%
_{i}^{\dag }$, yielding$$\left\Vert \mathbf{l}_{A;i}\right\Vert ^{2}+\left\Vert \mathbf{l}%
_{B;i}\right\Vert ^{2}=\lambda _{i}. \label{Eq 8}$$Together with (\[Eq 5\]), we obtain$$\left\Vert \mathbf{l}_{A;i}\right\Vert ^{2}=\left\Vert \mathbf{l}%
_{B;i}\right\Vert ^{2}=\frac{\lambda _{i}}{2}. \label{Eq 2}$$Moreover, left multiplying (\[Eq 6\]) and (\[Eq 7\]) respectively by $%
\mathbf{u}_{i}^{\dagger }$ and plugging in (\[Eq 3\]), we obtain$$\mathbf{l}_{A;i}^{\dag }\mathbf{l}_{B;i}=(\lambda _{i}-1)\left\Vert \mathbf{l%
}_{A;i}\right\Vert ^{2}=(\lambda _{i}-1)\left\Vert \mathbf{l}%
_{B;i}\right\Vert ^{2}=\mathbf{l}_{B;i}^{\dag }\mathbf{l}_{A;i}.
\label{Eq 10}$$Then$$\begin{aligned}
\mathbf{u}_{i}^{\prime \dag }\mathbf{u}_{i}^{\prime }& =\frac{1}{\lambda
_{i}\lambda _{i}^{\prime }}\left( \mathbf{l}_{A;i}-\mathbf{l}_{B;i}\right)
^{\dag }\left( \mathbf{l}_{A;i}-\mathbf{l}_{B;i}\right) \notag \\
& =\frac{1}{\lambda _{i}\lambda _{i}^{\prime }}\left( \left\Vert \mathbf{l}%
_{A;i}\right\Vert ^{2}-\mathbf{l}_{A;i}^{\dag }\mathbf{l}_{B;i}-\mathbf{l}%
_{B;i}^{\dag }\mathbf{l}_{A;i}+\left\Vert \mathbf{l}_{B;i}\right\Vert
^{2}\right) \notag \\
& \overset{(a)}{=}\frac{1}{\lambda _{i}\lambda _{i}^{\prime }}\left( \lambda
_{i}-(\lambda _{i}-1)\left\Vert \mathbf{l}_{A;i}\right\Vert ^{2}-(\lambda
_{i}-1)\left\Vert \mathbf{l}_{B;i}\right\Vert ^{2}\right) \notag\\
& \overset{(b)}{=}1\end{aligned}$$where step ($a$) follows from (\[Eq 2\]) and (\[Eq 10\]), and step ($b$) from (\[Eq 2\]) and the fact of $\lambda _{i}^{\prime }=2-\lambda _{i}$. This completes the proof.
The subspace $\mathcal{S}_{i}$ spanned by $\textbf{l}_{A;i}$ and $\textbf{l}_{B;i}$ is orthogonal to $\mathcal{S}%
_{j}$, for any $j$ $\neq i$.
From (\[Eq 4\]) and (\[Eq 9\]), we see that both $\mathbf{u}_{i}$ and $%
\mathbf{u}_{i}^{\prime }$ lie on the plane $\mathcal{S}_{i}$. As $\mathbf{u}%
_{i}$ and $\mathbf{u}_{i}^{\prime }$ are orthogonal to each other, $\mathcal{%
S}_{i}$ is spanned by $\mathbf{u}_{i}$ and $\mathbf{u}_{i}^{\prime }$. Then, the lemma holds straightforwardly by noting the orthogonality between the eigenvectors.
Now we give an overall picture of the eigenvalues and eigenvectors of $%
\mathbf{U}_{A}\mathbf{U}_{A}^{\dagger }+\mathbf{U}_{B}\mathbf{U}%
_{B}^{\dagger }$. Denote the $k$ eigenvalues in case (a) by $\lambda
_{1},\cdots ,\lambda _{k}$, and the corresponding orthogonal eigenvectors by $\mathbf{u}_{1},\cdots ,$ $\mathbf{u}_{k}$. Also denote the $l$ eigenvalues in case (b) by $\lambda _{k+1},\cdots ,\lambda _{k+l}$ in the descending order, and the corresponding eigenvectors by $\mathbf{u}_{k+1},\cdots ,$ $%
\mathbf{u}_{k+l}$. As the eigenvalues in (b) and (d) appears in a pair-wise manner, we further denote the $l$ eigenvalues in case (d) by $\lambda
_{k+1}^{\prime },\cdots ,$ $\lambda _{k+l}^{\prime }$ in the descending order, and the corresponding eigenvectors by $\mathbf{u}_{k+1}^{\prime
},\cdots ,$ $\mathbf{u}_{k+l}^{\prime }$. Moreover, we denote the $d_{A}$ orthogonal eigenvectors in case (c.1) by $\mathbf{u}_{k+l+1},\cdots ,\mathbf{%
u}_{k+l+d_{A}}$, and the $d_{B}$ orthogonal eigenvectors in case (c.2) by $%
\mathbf{u}_{k+l+d_{A}+1},\cdots ,\mathbf{u}_{k+l+d_{A}+d_{B}}$. Let$$\mathbf{U}=\left[ \mathbf{u}_{1},\cdots ,\mathbf{u}_{k},\mathbf{u}_{k+1},%
\mathbf{u}_{k+1}^{\prime },\cdots ,\mathbf{u}_{k+l},\mathbf{u}_{k+l}^{\prime
},\mathbf{u}_{k+l+1},...,\mathbf{u}_{k+l+d_{A}+d_{B}}\right] . \label{Def_U}$$It can be readily verified that $\mathbf{U}$ is an orthonormal matrix satisfying $\mathbf{U}_{{}}^{\dag }\mathbf{U}=\mathbf{I}_{n_{A}+n_{B}-k}$. Define
$$\mathbf{U}_{A}^{\prime }=\left[ \mathbf{u}_{1},\cdots ,\mathbf{u}_{k},\frac{%
\mathbf{l}_{A;k+1}}{\left\Vert \mathbf{l}_{A;k+1}\right\Vert },\cdots ,\frac{%
\mathbf{l}_{A;k+l}}{\left\Vert \mathbf{l}_{A;k+l}\right\Vert },\mathbf{u}%
_{k+l+1},...,\mathbf{u}_{k+l+d_{A}}\right]$$
and$$\mathbf{U}_{B}^{\prime }=\left[ \mathbf{u}_{1},\cdots ,\mathbf{u}_{k},\frac{%
\mathbf{l}_{B;k+1}}{\left\Vert \mathbf{l}_{B;k+1}\right\Vert },\cdots ,\frac{%
\mathbf{l}_{B;k+l}}{\left\Vert \mathbf{l}_{B;k+l}\right\Vert },\mathbf{u}%
_{k+l+d_{A}+1},...,\mathbf{u}_{k+l+d_{A}+d_{B}}\right] .$$
In the above, $\mathbf{u}_{1},\cdots ,\mathbf{u}_{k}$ are the eigenvectors in case (a); $\mathbf{u}_{k+l+1},...,\mathbf{u}_{k+l+d_{A}}$ are the eigenvectors in case (c) satisfying $\mathbf{U}_{A}\mathbf{U}_{A}^{\dag }%
\mathbf{u}_{i}=\mathbf{u}_{i},$ for $i=k+l+1,...,k+l+d_{A}$; $\mathbf{u}%
_{k+l+d_{A}+1},\cdots ,$ $\mathbf{u}_{k+l+d_{A}+d_{B}}$ are the eigenvectors in case (c) satisfying $\mathbf{U}_{B}\mathbf{U}_{B}^{\dag }\mathbf{u}_{i}=%
\mathbf{u}_{i},$ for $i=k+l+d_{A}+1,...,k+l+d_{A}+d_{B}$.
Then, from Lemmas 3 and 4, it can be verified that $\mathbf{D}_{m}$ in ([Dm]{}) satisfies$$\mathbf{U}_{m}^{\prime }=\mathbf{UD}_{m},m\in \left\{ A,B\right\} .
\label{Dm2}$$
From Lemma 3 and the fact that $\mathbf{l}_{m;i}\in \mathcal{S}_{i}$ for $%
i=k+1,\cdots ,k+l$, the columns of $\mathbf{U}_{m}^{\prime }$ are orthonormal. Together with the fact that all columns of $\mathbf{U}%
_{m}^{\prime }$ lie in the columnspace of $\mathbf{U}_{m}$ (and so in the columnspace of $\mathbf{H}_{m}$), we see that $\mathbf{H}_{m}$ and $\mathbf{%
U}_{m}^{\prime }$ share the same columnspace. Thus, there exists an $n_{m}$-by-$n_{m}$ square matrix $\mathbf{G}_{m}$ such that$$\mathbf{H}_{mR}=\mathbf{U}_{m}^{\prime }\mathbf{G}_{m}. \label{Eq 13}$$Combining (\[Dm2\]) and (\[Eq 13\]), we obtain$$\mathbf{H}_{mR}=\mathbf{UD}_{m}\mathbf{G}_{m} \label{Eq 14}$$which completes the proof of Theorem \[Theorem 1\].
Proof of Theorem \[Theorem 3\]
==============================
We first consider the sum-rate upper bound: $$\begin{aligned}
&&R^{UL}\overset{(a)}{\approx }\frac{1}{2}\dsum\limits_{m\in \left\{
A,B\right\} }\log \left\vert \mathbf{I}_{n_{R}}+\frac{P_{m}}{N_{0}n_{m}}%
\mathbf{UD}_{m}\mathbf{R}_{m}\mathbf{R}_{m}^{\dag }\mathbf{D}_{m}^{\dagger }%
\mathbf{U}^{\dagger }\right\vert \notag \\
&&\overset{(b)}{=}\frac{1}{2}\dsum\limits_{m\in \left\{ A,B\right\} }\log
\left\vert \mathbf{I}_{n_{m}}+\frac{P_{m}}{N_{0}n_{m}}\mathbf{R}_{m}\mathbf{R%
}_{m}^{\dag }\right\vert \notag \\
&&\overset{(c)}{\approx }\frac{1}{2}\dsum\limits_{m\in \left\{ A,B\right\}
}\log \left\vert \frac{P_{m}}{N_{0}n_{m}}\mathbf{R}_{m}\mathbf{R}_{m}^{\dag
}\right\vert \label{Upp}\end{aligned}$$where step (*a*) follows by substituting (\[Decomp\]) into ([UpperBound]{}), step (*b*) follows from the facts that $\mathbf{D}%
_{m}^{\dagger }\mathbf{U}^{\dagger }\mathbf{U\mathbf{D}}_{m}\mathbf{=I}%
_{n_{m}}$ and $\left\vert \mathbf{I}+\mathbf{AB}\right\vert =\left\vert
\mathbf{I}+\mathbf{BA}\right\vert $, and step (*c*) utilizes the fact that $\mathbf{R}_{m}$ is a square matrix.
Now we consider the achievable sum-rate of the proposed space-division scheme. For notational simplicity, let $\mathbf{H}_{m}^{CD}=\mathbf{D}%
_{m;2,2}\mathbf{R}_{m;2,2}$, $m\in \left\{ A,B\right\} $. From (\[Rate CD\]), the sum-rate of the complete-decoding spatial streams can be expressed as$$\begin{aligned}
R_{A}^{CD}+R_{B}^{CD}\overset{(a)}{=}\frac{1}{2}\log \left\vert \mathbf{I}%
+\dsum\limits_{m\in \left\{ A,B\right\} }\frac{P_{m}}{N_{0}n_{m}}\mathbf{H}%
_{m}^{CD}(\mathbf{H}_{m}^{CD})^{\dag }\right\vert \notag\end{aligned}$$ $$\begin{aligned}
%&&R_{A}^{CD}+R_{B}^{CD}\overset{(a)}{=}\frac{1}{2}\log \left\vert \mathbf{I}%
%+\dsum\limits_{m\in \left\{ A,B\right\} }\frac{P_{m}}{N_{0}n_{m}}\mathbf{H}%
%_{m}^{CD}(\mathbf{H}_{m}^{CD})^{\dag }\right\vert \\
&&\overset{(b)}{=}\frac{1}{2}\log \left\vert \mathbf{I}+\frac{P_{A}}{%
N_{0}n_{A}}\mathbf{H}_{A}^{CD}(\mathbf{H}_{A}^{CD})^{\dag }\right\vert +%
\frac{1}{2}\log \frac{\left\vert \mathbf{I}+\dsum\limits_{m\in \left\{
A,B\right\} }\frac{P_{m}}{N_{0}n_{m}}\mathbf{H}_{m}^{CD}(\mathbf{H}%
_{m}^{CD})^{\dag }\right\vert }{\left\vert \mathbf{I}+\frac{P_{A}}{N_{0}n_{A}%
}\mathbf{H}_{A}^{CD}(\mathbf{H}_{A}^{CD})^{\dag }\right\vert } \\
&&\overset{(c)}{\approx }\frac{1}{2}\log \left\vert \frac{P_{A}}{N_{0}n_{A}}%
\mathbf{R}_{A;2,2}\mathbf{R}_{A;2,2}^{\dag }\right\vert +\frac{1}{2}\log
\left\vert \frac{P_{B}}{N_{0}n_{B}}(\mathbf{H}_{B}^{CD})^{\dag }\left(
\mathbf{I}+\frac{P_{A}}{N_{0}n_{A}}\mathbf{H}_{A}^{CD}(\mathbf{H}%
_{A}^{CD})^{\dag }\right) ^{-1}\mathbf{H}_{B}^{CD}\right\vert \\
&&\overset{(d)}{=}\dsum\limits_{m\in \left\{ A,B\right\} }\frac{1}{2}\log
\left\vert \frac{P_{m}}{N_{0}n_{m}}\mathbf{R}_{m;2,2}\mathbf{R}%
_{m;2,2}^{\dag }\right\vert +\frac{1}{2}\log \left\vert \mathbf{D}%
_{B;2,2}^{\dag }\left( \mathbf{I}+\frac{P_{A}}{N_{0}n_{A}}\mathbf{H}%
_{A}^{CD}(\mathbf{H}_{A}^{CD})^{\dag }\right) ^{-1}\mathbf{D}%
_{B;2,2}\right\vert\end{aligned}$$where step (*a*) utilizes the fact that equal power allocation is asymptotically optimal, and step (d) follows by substituting $\mathbf{H}%
_{B}^{CD}=\mathbf{D}_{B;2,2}\mathbf{R}_{B;2,2}$. Applying the matrix inversion lemma to $\left( \mathbf{I}+\frac{P_{A}}{N_{0}n_{A}}\mathbf{H}%
_{A}^{CD}(\mathbf{H}_{A}^{CD})^{\dag }\right) ^{-1}$, we further obtain$$\begin{aligned}
&&R_{A}^{CD}+R_{B}^{CD}=\dsum\limits_{m\in \left\{ A,B\right\} }\frac{1}{2}%
\log \left\vert \frac{P_{m}}{N_{0}n_{m}}\mathbf{R}_{m;2,2}\mathbf{R}%
_{m;2,2}^{\dag }\right\vert \notag \\
&&\text{ \ \ \ \ \ }+\frac{1}{2}\log \left\vert \mathbf{I}-\frac{P_{A}}{%
N_{0}n_{A}}\mathbf{D}_{B;2,2}^{\dag }\mathbf{H}_{A}^{CD}\left( \mathbf{I}+%
\frac{P_{A}}{N_{0}n_{A}}(\mathbf{H}_{A}^{CD})^{\dag }\mathbf{H}%
_{A}^{CD}\right) ^{-1}(\mathbf{H}_{A}^{CD})^{\dag }\mathbf{D}%
_{B;2,2}\right\vert \notag \\
&&\overset{(a)}{\approx }\dsum\limits_{m\in \left\{ A,B\right\} }\frac{1}{2}%
\log \left\vert \frac{P_{m}}{N_{0}n_{m}}\mathbf{R}_{m;2,2}\mathbf{R}%
_{m;2,2}^{\dag }\right\vert +\frac{1}{2}\log \left\vert \mathbf{I}-\mathbf{D}%
_{B;2,2}^{\dag }\mathbf{D}_{A;2,2}\mathbf{D}_{A;2,2}^{\dag }\mathbf{D}%
_{B;2,2}\right\vert \notag \\
&&\overset{(b)}{=}\dsum\limits_{m\in \left\{ A,B\right\} }\frac{1}{2}\log
\left\vert \frac{P_{m}}{N_{0}n_{m}}\mathbf{R}_{m;2,2}\mathbf{R}%
_{m;2,2}^{\dag }\right\vert +\frac{1}{2}\log \dprod\limits_{i=k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
+1}^{k+l}\lambda _{i}(2-\lambda _{i}) \label{Con_Coded}\end{aligned}$$where step (*a*) follows by noting $\mathbf{I}+\frac{P_{A}}{N_{0}n_{A}%
}(\mathbf{H}_{A}^{CD})^{\dag }\mathbf{H}_{A}^{CD}\approx \frac{P_{A}}{%
N_{0}n_{A}}(\mathbf{H}_{A}^{CD})^{\dag }\mathbf{H}_{A}^{CD}$ and $\mathbf{H}_{A}^{CD}=\mathbf{D}%
_{A;2,2}\mathbf{R}_{A;2,2}$, $m\in \left\{ A,B\right\} $, and step (*b*) utilizes the definitions in (\[Dm\]) and (\[Dm3\]). Moreover, letting $w_{A}=w_{B}=1$ in (\[WSR1\]), we obtain the sum-rate of the PNC spatial streams as$$R_{A}^{PNC}+R_{B}^{PNC}=\frac{1}{2}\dsum\limits_{m\in \left\{ A,B\right\}
}\left(\log \left\vert \frac{P_{m}}{N_{0}n_{m}}\mathbf{R}_{m;1,1}\mathbf{R}%
_{m;1,1}^{\dag }\right\vert +\log \left\vert \mathbf{\tilde{D}}_{m;1,1}%
\mathbf{\tilde{D}}_{m;1,1}^{\dag }\right\vert\right) \text{.} \label{PNC_Coded}$$From (\[dd\]), $\mathbf{p}_{i}$ is the angular bisection of $\mathbf{e}%
_{A;i}$ and $\mathbf{e}_{B;i}$, or equivalently, $\mathbf{p}_{i}=[1,0]^{T}$, for the sum-rate case of $w_{A}=w_{B}=1$. Then, using the definition in ([ddd]{}), we obtain$$\log \left\vert \mathbf{\tilde{D}}_{m;1,1}\mathbf{\tilde{D}}_{m;1,1}^{\dag
}\right\vert =\dsum\limits_{i=k+1}^{k+l%
%TCIMACRO{\U{b4}}%
%BeginExpansion
{\acute{}}%
%EndExpansion
}\log \frac{\lambda _{i}}{2}. \label{DD}$$Combining (\[Upp\])-(\[DD\]), we complete the proof of Theorem 3.
Proof of Lemma \[Lemma 4\]
==========================
We prove by using the theory of free probability [@Voiculescu93]. The a.e.d. of $\mathbf{U}_{m}\mathbf{U}_{m}^{\dagger }$ is given by $$p_{m}\left( \lambda \right) =\eta _{m}\delta \left( \lambda -1\right)
+\left( 1-\eta _{m}\right) \delta \left( \lambda \right) \text{, }m\in
\left\{ A,B\right\} .$$Let $X_{m}$ be a random variable with PDF $p_{m}\left( \lambda \right) $. Its Stieltjes transform is given by (cf., (2.40) in [@TulinoTextBook])$$S_{X_{m}}\left( z\right) =E\left[ \frac{1}{X_{m}-z}\right] =\frac{\eta _{m}}{%
1-z}-\frac{1-\eta _{m}}{z}.$$Then, the inverse function of $S_{X_{m}}\left( z\right) $ is given by$$S_{X_{m}}^{-1}\left( s\right) =\frac{-\left( 1-s\right) \pm \sqrt{\left(
1-s\right) ^{2}-4s\left( \eta _{m}-1\right) }}{2s}.$$Using the relation between Stieltjes transform and R-transform (cf., (2.72) in [@TulinoTextBook]), we obtain the R-transform of $X_{m}$ as $$R_{X_{m}}\left( z\right) =S_{X_{m}}^{-1}\left( -z\right) -\frac{1}{z}=\frac{%
z-1\mp \sqrt{\left( z-1\right) ^{2}+4\eta _{m}z}}{2z}.$$From Theorem 2.64 of [@TulinoTextBook], as $\mathbf{U}_{A}\mathbf{U}%
_{A}^{\dagger }$ and $\mathbf{U}_{B}\mathbf{U}_{B}^{\dagger }$ are asymptotically free random matrices, the R-transform of the a.e.d. of $\mathbf{U}_{A}\mathbf{U}_{A}^{\dagger }+\mathbf{U}_{B}%
\mathbf{U}_{B}^{\dagger }$ is given by$$R_{AB}\left( z\right) =R_{X_{A}}\left( z\right) +R_{X_{B}}\left( z\right)
=\dsum\limits_{m\in \left\{ A,B\right\} }\frac{z-1\mp \sqrt{\left(
z-1\right) ^{2}+4\eta _{m}z}}{2z}.$$Then, the Stieltjes transform of the a.e.d. of $\mathbf{U}_{A}\mathbf{U}%
_{A}^{\dagger }+\mathbf{U}_{B}\mathbf{U}_{B}^{\dagger }$ satisfies$$S_{AB}^{-1}\left( -z\right) =1\mp \dsum\limits_{m\in \left\{ A,B\right\} }%
\frac{\sqrt{\left( z-1\right) ^{2}+4\eta _{m}z}}{2z}.$$Letting $y=S_{AB}^{-1}\left( -z\right) $ , we obtain $$\dsum\limits_{m\in \left\{ A,B\right\} }\sqrt{\left( z-1\right) ^{2}+4\eta
_{m}z}=\mp 2z(y-1).$$Multiplying $\sqrt{\left( z-1\right) ^{2}+4\eta _{A}z}-\sqrt{\left(
z-1\right) ^{2}+4\eta _{B}z}$ on both sides, we have$$\sqrt{\left( z-1\right) ^{2}+4\eta _{A}z}-\sqrt{\left( z-1\right) ^{2}+4\eta
_{B}z}=\mp \frac{2(\eta _{A}-\eta _{B})}{y-1}.$$Adding the above two equations and taking the square, we further obtain$$\left( z-1\right) ^{2}+4\eta _{A}z=\left( z(y-1)+\frac{\eta _{A}-\eta _{B}}{%
y-1}\right) ^{2}.$$Solving $z$, we obtain$$S_{AB}\left( z\right) =-\frac{1-\eta _{A}-\eta _{B}\mp \sqrt{(1-\eta
_{A}-\eta _{B})^{2}+\left( 2z-z^{2}\right) \left( \left( \frac{\eta
_{A}-\eta _{B}}{z-1}\right) ^{2}-1\right) }}{2z-z^{2}}.$$From (2.45) in [@TulinoTextBook], the a.e.d. of $\mathbf{U}_{A}\mathbf{U}%
_{A}^{\dagger }+\mathbf{U}_{B}\mathbf{U}_{B}^{\dagger }$ is given by$$\mathcal{F}\left( \lambda \right) =\lim_{\omega \rightarrow 0^{+}}\frac{1}{%
\pi }\func{Im}\left[ S_{AB}\left( \lambda +j\omega \right) \right] .$$Thus, for $0<\lambda <1$ and $1<\lambda <2$, we obtain $$\mathcal{F}\left( \lambda \right) =\frac{1}{\pi }\func{Im}\left[ \frac{\sqrt{%
(1-\eta _{A}-\eta _{B})^{2}+\left( 2\lambda -\lambda ^{2}\right) \left(
\left( \frac{\eta _{A}-\eta _{B}}{\lambda -1}\right) ^{2}-1\right) }}{%
2\lambda -\lambda ^{2}}\right] . \label{F_lemda}$$In addition, for a randomly generated pair of $\mathbf{U}_{A}$ and $%
\mathbf{U}_{B}$, there are $n_{A}+n_{B}-n_{R}$ orthogonal eigenvectors for $%
\lambda _{i}=2$, $\left\vert n_{A}-n_{B}\right\vert $ orthogonal eigenvectors for $\lambda _{i}=1$, and $n_{R}-n_{A}-n_{B}$ orthogonal eigenvectors for $\lambda _{i}=0$. Thus, as $n_{R}$ tends to infinity, the PDF $\mathcal{F}\left( \lambda \right) $ at $\lambda =2$ is given by $\left[
\eta _{A}+\eta _{B}-1\right] ^{+}\delta \left( \lambda -2\right) $; that at $%
\lambda =1$ is given by $\left\vert \eta _{A}-\eta _{B}\right\vert \delta
\left( \lambda -1\right) $; and that at $\lambda =0$ is given by $\left[
1-\eta _{A}-\eta _{B}\right] ^{+}\delta \left( \lambda \right) $. This concludes the proof of the lemma.
Acknowledgment {#acknowledgment .unnumbered}
==============
The work of Xiaojun Yuan was partially supported by a grant from the University Grants Committee (Project No. AoE/E-02/08) of the Hong Kong Special Administrative Region, China. The work of Tao Yang was supported by CSIRO OCE Postdoctoral Fellowships. It is also supported under the Australian Government’s Australian Space Research Program.
[99]{} R. Ahlswede, N. Cai, S.-Y. R. Li, and R. W. Yeung, Network information flow, *IEEE Trans. Inf. Theory*, vol. 46, pp. 1204–1216, Oct. 2000.
S. Zhang, S. Liew, and P. P. Lam, Hot topic: physical-layer network coding, in Proc. 12th Annual International Conf. Mobile Computing Networking (ACM MobiCom’06), Sept. 2006, pp. 358-365.
B. Nazer and M. Gastpar, The case for structured random codes in network communication theorems, *ITW 2007*, Lake Tahoe, California, Sept., 2007.
P. Popovski and H. Yomo, Physical network coding in two-way wireless relay channels,*IEEE ICC*, Glasgow, Scotland, June 2007.
W. Nam, S. Chung, Y. H. Lee, Capacity of the Gaussian two-way relay channel to within 1/2 bit, *IEEE Trans. Inf. Theory*, vol. 56, no. 11, pp. 5488-5494, Nov. 2010.
M. P. Wilson, K. Narayanan, H. D. Pfister and A. Sprintson, Joint physical layer coding and network coding for bidirectional relaying, *IEEE Trans. Inf. Theory*, vol. 56, no. 11, pp. 5641-5654, Nov. 2010.
G. Foschini, Layered space-time architecture for wireless communication in a fading environment when using multi-element antennas, *Bell Labs Technical Journal*, Autumn 1996, pp. 41-59.
S. Katti, S. Gollakota, and D. Katabi, Embracing wireless interference: Analog network coding, *ACM SIGCOMM’07*.
R. Zhang, Y.-C. Liang, C. C. Chai, and S. Cui, Optimal beamforming for two-way multi-antenna relay channel with analogue network coding, *IEEE J. Select. Area. Comm.*, Vol. 27, No. 5, pp. 699-712, June 2009.
S. Xu and Y. Hua, Source-relay optimization for a two-way MIMO relay system, *Proc. IEEE ICCASP 2010*, pp. 3038-3041.
M. Aleksic, P. Razaghi, and W. Yu, Capacity of a class of modulo-sum relay channels, *IEEE Trans. Inf. Theory*, vol. 55, pp. 921–930, March 2009.
S. H. Lim, Y.-H. Kim, A. El Gamal, and S.-Y. Chung, Noisy network coding, *IEEE Trans. Inf. Theory*, vol. 57, pp. 3132–3152, May 2011.
G. Kramer, M. Gastpar, and P. Gupta, Cooperative strategies and capacity theorems for relay networks, *IEEE Trans. Inf. Theory*, vol. 51, pp. 3037–3063, September 2005.
A. El Gamal, N. Hassanpour, and J. Mammen, Relay networks with delays, *IEEE Trans. Inf. Theory*, vol. 53, pp. 3413–3431, October 2007.
D. Gunduz, A. Goldsmith and H. V. Poor, MIMO two-way relay channel: diversity-multiplexing tradeoff analysis, *Asilomar 2008*.
H. J. Yang, J. Chun and A. Paulraj, Asymptotic capacity of the separated MIMO two-way relay channel, *IEEE Trans. Inf. Theory*, vol. 57, no. 11, pp. 7542-7554, Nov. 2011.
T. Yang, X. Yuan, P. Li, I. B. Collings and J. Yuan., Eigen-direction alignment based physical-layer network coding for MIMO two-way relay channels, *IEEE Trans. Comm.*, submitted, 2012. (Available at http://arxiv.org/abs/1201.2471)
A. Khina, Y. Kochman, U. Erez, Physical-layer MIMO relaying, in *Proc. ISIT’11*.
B. Nazer and M. Gasper, Compute-and-forward: Harnessing interference through structured codes, *IEEE Trans. Inf. Theory*, vol. 57, pp. 6463-6486, Oct. 2011.
3GPP, Overview of 3GPP Release 10, Release 10 V0.1.1 (2011-6), June 2011. (Available at http://www.3gpp.org/Release-10)
IEEE, IEEE Standard for Local and metropolitan area networks Part 16: Air Interface for Broadband Wireless Access Systems Amendment 3: Advanced Air Interface, IEEE Std 802.16m-2011, May 2011. (Available at http://www.ieee802.org/16/published.html)
U. Erez and R. Zamir, Achieving 1/2log(1 + SNR) on the AWGN channel with lattice encoding and decoding, *IEEE Trans. Inf. Theory*, vol. 50, pp. 2293-2314, Oct. 2004.
G. H. Golub and C. F. V. Loan, *Matrix Computation*, 3rd ed. Baltimore, MD: The Johns Hopkins University Press, 1997.
T. M. Cover, and J. A. Thomas, *Elements of Information Theory*, Wiley, New York, 1991.
S. Boyd and L. Vandenberghe, *Convex Optimization*, Cambridge University Press, 2004.
A. Tulino and S. Verdu, *Random Matrix Theory and Wireless Communications*, Now Publishers Inc, 2004, ISBN: 193301900X.
D. Voiculescu, The analogues of entropy and of Fisher’s information measure in free probability theory, I, *Communications in Math. Physics*, vol. 155, pp. 71–92, July 1993.
R. A. Horn and C. R. Johnson, *Matrix Analysis*, *Cambridge Unversity Press*, 1990.
W. Yu, W. Rhee, S. Boyd, and J. Cioffi, Iterative waterfilling for Gaussian vector multiple access channels, *IEEE Trans. Inf. Theory*, vol. 50, no. 1, 145-151, Jan. 2004.
![The function of the average normalized gap $r^{SD}$ in (\[Limit1\]) against $%
\protect\eta $.[]{data-label="SD_rate_gp"}](SD_rate_gap.eps){width="4.6in" height="3.8in"}
![Average achievable sum-rates of various schemes for the Rayleigh fading MIMO TWRC with $n_{A}=n_{B}=2$ and $n_{R}=4$.[]{data-label="Figure nT2nR4"}](Fig_nT2nR4.eps){width="4.6in" height="3.8in"}
![Average achievable sum-rates of various schemes for the Rayleigh fading MIMO TWRC with $n_{A}=n_{B}=2$ and $n_{R}=3$. []{data-label="Figure nT2nR3"}](Fig_nT2nR3.eps){width="4.6in" height="3.8in"}
![Scaling effect of the average sum-rates of various schemes for the Rayleigh fading MIMO TWRCs with $\protect\eta _{A}=\protect\eta _{B}=1/2$. []{data-label="Figure Eda1_2"}](SD_beta1over2.eps){width="4.6in" height="3.8in"}
![Scaling effect of the average sum-rates of various schemes for the Rayleigh fading MIMO TWRCs with $\protect\eta _{A}=\protect\eta _{B}=2/3$. []{data-label="Figure Eda2_3"}](SD_beta2over3.eps){width="4.6in" height="3.8in"}
![Average achievable rate-regions for the Rayleigh fading MIMO TWRC with $n_{A}=n_{B}=2$ and $n_{R}=3$. The average SNRs for all the channel links are set to $30$ dB.[]{data-label="Figure CapacityRegion"}](RateRegion_SD_2by3.eps){width="4.8in" height="4in"}
[^1]: This manuscript is submitted to IEEE Transactions on Information Theory
[^2]: The treatment for the rate-pair in (\[R\_PNC2\]) is similar, and thus omitted.
[^3]: Let $\mathcal{S}$ be a vector space, and let $\mathcal{S}_{1},\mathcal{S}%
_{2},...,\mathcal{S}_{n}$ be subspaces of $\mathcal{S}$. $\mathcal{S}$ is defined to be a direct sum of $\mathcal{S}_{1},\mathcal{S}_{2},...,%
\mathcal{S}_{n}$ when $\mathcal{S}_{1},\mathcal{S}_{2},...,\mathcal{S}_{n}$ are mutually orthogonal and for every vector $\mathbf{x}$ in $\mathcal{S}$, there is $\mathbf{x}_{i}$ in $\mathcal{S}_{i},i=1,2,...,n,$ such that $%
\mathbf{x=}\dsum\limits_{i=1}^{n}\mathbf{x}_{i}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This article describes the main mathematical researches performed, in England and in the Continent between 1742-1827, on the subject of hyperbola rectification, thereby adding some of our contributions. We start with the Maclaurin inventions on Calculus and their remarkable role in the early mid 1700s; next we focus a bit on his evaluation, 1742, of the *hyperbolic excess*, explaining the true motivation behind his research. To his geometrical-analytical treatment we attach ours, a purely analytical alternative. Our hyperbola inquiry is then switched to John Landen, an amateur mathematician, who probably was writing more to fix his priorities than to explain his remarkable findings. We follow him in the obscure proofs of his theorem on hyperbola rectification, explaining the links to Maclaurin and so on. With a chain of geometrical constructions, we attach our interpretation to Landen’s treatment. Our modern analytical proof to his hyperbolic limit excess, by means of elliptic integrals of the first and second kind is also provided, and we demonstrate why the so called Landen transformation for the elliptic integrals cannot be ascribed to him. Next, the subject leaves England for the Continent: the character of Lagrange is introduced, even if our interest concerns only his 1785 memoir on irrational integrals, where the *Arithmetic Geometric Mean*, AGM, is established by him. Nevertheless, our objective is not the AGM, but to detect the real source of the so-called Landen transformation for elliptic integrals. In fact, Lagrange’s paper encloses a differential identity stemming from the AGM: integrating it, we show how it could be possible to arrive at the well-known Legendre recursive computation of a first kind elliptic integral, which appeared in his *Traité*, 1827, much after the Lagrange’s paper.'
author:
- 'Giovanni Mingari Scarpello,Daniele Ritelli and Aldo Scimone'
title: 'The hyperbola rectification from Maclaurin to Landen and the Lagrange-Legendre transformation for the elliptic integrals'
---
Introduction {#introduction .unnumbered}
============
The reader should know we have been driven by three criteria, the first of which is to get him in touch with old Masters still capable of teaching many things, nowadays forgotten after 250 years, surmounting the main difficulties of language, notations, and often, of a completely different view of approaching questions.
The second criterion is our philological approach: thanks to the power of the Internet, we were able to keep close reference to the original prints of each work. As a matter of fact, for instance with John Landen, we realized how the difficult access to antique texts produced many mistakes and errors in the past due to inaccurate assertions being repeated without cross-checks, which then propagated over time.
The third step concerns our attached analytical research. Our Authors often write in involved, rather obscure ways and prove theorems in special cases only, even if they have a greater generality. Thus we decided to work on some statements, not only for the sake of an independent check, but also to frame an old conquest within the context of a modern outline and language. In such a way we pass by means of modern mathematical objects (special functions, elliptic integrals, successive theorems) from philology to analytical inquiries, often through the computer algebra tools: modern tools and classical mind. By comparison of our computations to original results, we highlight the quality displayed at time by the Authors, who worked with plain old tools and relied solely on their brilliance and ingenuity. To Maclaurin’s research we added some explanations and figures providing also a modern alternative formulation to the hyperbolic excess. Landen’s writing is usually obscure, and his theorems are proved in special cases only, so for his excess computations, we provide alternative proofs. For Lagrange we show how his AGM transformation can lead to the famous modular transformation for the elliptic integrals, mistakenly attributed to Landen.
Maclaurin
=========
Maclaurin’s name remains in the history of science thanks to the first textbook dealing with Newtonian Calculus: his *Treatise of Fluxions*, Edinburgh, 1742. In it Maclaurin tried to present the Calculus with the rigour of the ancients” supporting Newton in the polemics on Calculus fundaments’ tenability raised in *The Analyst, a discourse addressed to an Infidel Mathematician* [@Berk] by George Berkeley (1685-1753), who had criticized the foundations of the analysis in the works by Newton.[^1]
Maclaurin’s works and his *Fluxions*
------------------------------------
Colin Maclaurin (1698-1746) lived and worked in a remarkable scientific context. His name is commonly associated with the *Maclaurin series* $ f(x)=f(0)+f'(0)x + 1/2f''(0)x^{2} +\cdots$ despite the fact that it had been published by the Englishman Brook Taylor (1685-1731) in its more general form (Taylor series) already in 1715, and James Gregory (1638-1675) had used it in special cases. Maclaurin’s papers on journals can be divided as issued in the Philosophical Transactions (9 articles between 1718 and 1743 concerning curves, equations, and …*Cells wherein the Bees deposit their Honey*). A second group has been added to these writings, with two further publications about astronomy and in the Physical and Literary Society, Edinburgh, Vol. I.
Maclaurin’s books covered each branch of Mathematics: *Geometria Organica* London, 1720; *A Treatise of Fluxions*, 2 volumes, Edinburgh, 1742; *A Treatise of Algebra*, with an Appendix, *De Linearum Geometricarum Proprietatibus generalibus*, 1748[^2]; [*An Account of Sir Isaac Newton’s Philosophy*]{} (1748). In his *Geometria* Maclaurin dealt with conics, cubics, quartics, and general properties of curves, such as the famous *trisectrix*, he met while studying the ancient problem of the angle’s trisection and whose equation in cartesian coordinates is $y^{2}=(3+x)x^{2}/({1-x}).$ The treatise *Fluxions*, [@Fluxions], had its origin in the Maclaurin’s defense of the Newtonian doctrine of fluxions, expanding, well beyond a pamphlet, to more than 760 pages: a very complex construction where most of the time mathematics is presented, organized and applied to several physics problems. It therefore will not be described here, and we will stick only to a very superficial discussion of it. The reader can have a satisfactory idea of its internal partitions and contents by referring to Sageng’s pages, [@Sageng].
The n. 927 of *Fluxions* entitled: *The construction of the elastic curve, and of other figures, by the rectification of the conic sections*, starts with:
> The celebrated author who first resolved this as well as several other curious problems, after his account of this figure (which is commonly called the clastic curve), adds: Ob graves causas suspicor curvae nostrae constructionem a nullius sectionis conicae seu quadratura seu rectificatione pendere” Act. Lips. 1694, page 272. But it is constructed by the rectification of equilateral hyperbola.
The unmentioned author is Jakob Bernoulli, and such a motivation is by Maclaurin postponed to the theoretical treatment of hyperbola rectification, n. 805. We will concentrate on this treatment, and proceed to provide more elements on its employ in the elastica problem.
Maclaurin’s extensive interest in almost all Mathematical Physics and Calculus of his time, led him to the problem of fluents, and, not only to solve the elastica, but to rectify the curves as well. In *Fluxions*, n. 755, Maclaurin defines a research program concerning the classification of irrational fluents, to be followed by D’Alembert in his *Recherches sur le calcul integral*, 1746, published in 1748 by the Berlin Academy, see [@Dal]. Whilst D’Alembert used a purely analytical approach without any figure, doing only algebraic computations by means of several changes of variable, Maclaurin, on the contrary performed the integration of irrational differentials by means of arcs of conics and often with the help of geometrical arguments. In such a way the influence of Maclaurin induced D’Alembert to study by means of an algebraic process (*Recherches*, page 203) the fluent of $${\rm d}x\sqrt{\frac{x}{x^{2}\pm fx+b^{2}}},$$ establishing that it can be reduced to the addition of an arc of hyperbola of certain semiaxes plus an algebraic term. At n. 798 the Maclaurin classification is a bit more strict: let us give an account of it.
First class: when a fluent can be represented in a finite number of algebraic terms, like the fluent of $$\frac{{\rm d}x}{\sqrt{1\pm x}}.$$ Second class: includes fluents like $$\frac{{\rm d}x}{\sqrt{1\pm x^{2}}},$$ which can be reduced to areas of a circle and the hyperbolic areas of logarithms: they cannot be assigned in algebraic terms, but have been computed by several methods
Third class: fluents like $$\sqrt{\frac{x}{1\pm x^{2}}}\,{\rm d}x\quad\text{or}\quad\frac{{\rm d}x}{\sqrt{x}\sqrt{1\pm x}}$$ which cannot be reduced to any form and are required in some useful problems. They can only be assigned by hyperbolic and elliptic arcs; namely the computation of the length of a hyperbola leads to fluents of this type. Maclaurin realized that the elastica analysis could lead to the same (hopeless) integrals met when trying to rectify the conical sections. In such a way he judged the problem as solved, since its solution is given by a known and traceable curve; the practical computation was accomplished by expanding the function and performing a termwise integration.
Inside the *Fluxions*: the hyperbola and its excess
---------------------------------------------------
Before entering Maclaurin’s topic of our interest, we need to present shortly some definitions which precede a couple of theorems due to Apollonius, and quoted by Maclaurin himself.
### Apollonius’ theorems on hyperbola
First of all, some terms will be recalled to be used later. Let us consider the hyperbola $AP\ldots$, and $A'P'\ldots$ of equation $$\frac{x^{2}}{a^{2}}-\frac{y^{2}}{b^{2}}=1. \label{ipxya}$$ $a$ and $b$ being the *semiaxes*; the focal axis FF’, with the focal distance given by $\sqrt{a^{2}+b^{2}}$, named *transverse* is assumed as $Ox$, while $Oy$ is named *not transverse*, or *imaginary*; the hyperbola asymptotes are given by $y=\pm\,(b/a)\,x$. We will take into account also the hyperbola: $$\frac{y^{2}}{b^{2}}-\frac{x^{2}}{a^{2}}=1 \label{ipyx}$$ having the same asymptotes and focal distance as , but exchanging transverse and not transverse axes. The hyperbolæ of equations and are named *conjugate*, and a whichever straight line passing through their common center O is said to be the *diameter* for both. Two diameters (for instance $PP'$ and $P_{1}P_{1}'$, see \[f01\]) are said to be *conjugate* whenever the tangents to hyperbola at each extreme of one of them are parallel to the other one.
In such a way the four straight lines touching a couple of conjugate diameters at the extremes define a parallelogram (marked as $R_{1}WLG$ in \[f02\]) which circumscribes the couple of conjugate hyperbolæ.
Now let us introduce Apollonius’ theorems.
For any hyperbola the absolute value of the difference of the squares of any couple of conjugate diameters has a constant value, given by $\vert{a^{2}-b^{2}}\vert.$
The area of a parallelogram circumscribing two conjugate hyperbolæ of equations , , having the asymptotes as diagonals, is constant. Its value is given by $4ab$, namely the area of a rectangle whose sides are twice the semiaxes.
We make a special use of a corollary of this second theorem, corollary implicitly assumed by Maclaurin in his work on rectification. Minding \[f03\] we have:
Given a couple of conjugate hyperbolæ of center S, for whichever semidiameter $SH$, the product of its length to the distance $\overline{SP}=p$ is then constant, being P the point where the perpendicular drawn from $S$ crosses the tangent to the hyperbola parallel to $SH$: $$\overline{SH}\times p=ab \label{scorollary}$$
We refer to \[f02\].
Using the second Apollonious’ theorem we have ${\rm area} (RGDJ)={\rm area} (LNKH)=4ab,$ where $a$ and $b$ are the hyperbola semiaxses. This implies $$\displaystyle{\frac{1}{4} {\rm area} (RGDJ) = \frac{1}{4} {\rm area} (OMDU)={\rm area} (OAKB)=ab}$$ being M midpoint of JD and U midpoint of GD. Now observe that rectangle OWZU, where OW (the Maclaurin’s $p$ segment) perpendicular to the tangent line to the hyperbola in Z and OU conjugate semidiameter to the hyperbola, is equivalent to parallelogram OMDU, since they have the common base OU and the same altitude OW. Thus we have $${\rm area} (OWZU)={\rm area} (OMDU)=ab.$$ This follows from the equality $${\rm area} (OWZU)=OU\times p.$$
The hyperbola’s *pedal* equation
--------------------------------
In a cartesian orthogonal reference frame let us consider in the first quadrant the branch AQE of the hyperbola of *eccentricity* $e=({a^{2}+b^{2}}/{a^{2}})^{1/2}$ and its conjugate : such branches, see \[f03\], cut the axes at the points $A(a, 0)$ and $B(0, b)$ respectively. Let S be the common centre of both hyperbolæ where we put our origin. On the branch AQE we consider a point E marked by the radius $\overline{SE}=r$, and draw to such a branch at E the tangent straight line $\tau\tau$, crossed at P by the straight line $nn$ through S perpendicular to it; we put[^3]: $\overline{SP}=p$. It is assumed, for instance, $a>b$; then, as a consequence of the first Apollonius theorem we get: $$\label{obel}
\overline{SE}^{2}-\overline{SH}^{2}=a^{2}-b^{2}=2a\varepsilon,$$ where $$\varepsilon=a\left(1-\frac{e^{2}}{2}\right) \label{lengeps}$$ is a convenient length, and its *eccentricity* is $e.$ Putting $\overline{SE}=r$ in we get: $
\overline{SH}^{2}=r^{2}-2a\varepsilon.
$
On the other side, the Apollonius corollary, formula provides $\overline{SH}^{2}=a^{2}b^{2}/p^{2}.$ By comparison of $\overline{SH}^{2}$ expressions, Maclaurin gets: $$r^{2}-2a\varepsilon=aX,\label{rho}$$ where it has been defined[^4]: $$X=\frac{ab^{2}}{p^{2}}.\label{ics}$$ Solving to the radius $r$, taking its derivative with respect to $X$, one obtains: $${\rm d}r=\frac{a}{2\sqrt{2a\varepsilon+aX}}{\rm d}X.$$ In \[f03\], however, considering the similar triangles $\triangle{\rm E}{\rm Q}{\rm S}$ and $\triangle{\rm E}{\rm P}{\rm S}$, by equating the ratios of the hypotenuse to the greater cathetus, we have: $${\rm d}s=\frac{r}{\sqrt{r^{2}-p^{2}}}\,{\rm d}r.$$ The infinitesimal arclength ${\rm d}s$ is then obtained with the hyperbola $(a,b)$ as a function of $p$ and $r$. Plugging there the expressions and depending on $X$ alone, Maclaurin gets: $${\rm d}s=\frac{\sqrt{a}}{2}\frac{\sqrt{X}}{\sqrt{X^{2}+2\varepsilon X-b^{2}}}\,{\rm d}X%\label{dsx}$$ where $X$ is variable with the point E and then with $p$; three fixed hyperbola parameters appear, i.e.: the semiaxes $a$ and $b$ and the arclength $\varepsilon$, whose two are independent. At this point Maclaurin introduces the *length of the tangent*, namely the segment bounded by $E$ and $P$: $
\overline{EP}=\sqrt{r^{2}-p^{2}}.
$ Putting there the expressions of $r$ and $p$ as functions of $X$, we get $\overline{EP}$ as a function of $X$ alone, so that, taking the differential, one finds: $${\rm d}(\overline{EP})=\frac{\sqrt{a}}{2x\sqrt{X}}\frac{X^2+b^2}{\sqrt{X^2+2\varepsilon X-b^2}}\,{\rm d}X.$$ The *differential excess* ${\rm d}\Delta$ concerning a single $E$-point of the hyperbola is the shift between the relevant tangent segment and the arclength whenever $X$ undergoes a change ${\rm d}X$, so that the radius changes of $${\rm d}r=\frac{\sqrt{a}}{2}\frac{1}{\sqrt{X+2\varepsilon}}\,{\rm d}X.$$ Thus: $${\rm d}\Delta={\rm d}(\overline{EP})-{\rm d}s=\frac{\sqrt{a}}{2}\frac{1}{\sqrt{X^2+2\varepsilon X-b^2}}\,\frac{b^2}{X\sqrt{X}}\,{\rm d}X.\label{fond}$$ No doubt such a fluxion belongs to the third class of Maclaurin’s ranking of irrational ones.
The excess *p*-formula and its consequences
-------------------------------------------
Starting from (\[fond\]), minding the $X$-definition, Maclaurin puts[^5] $
p^{2}/{a}=b^{2}/{X}=\zeta,
$ so that $dX=-b^{2}\,{\rm d}\zeta/\zeta^{2}$. In such a way (\[fond\]) becomes: $${\rm d}\Delta=\frac{-\sqrt{a}}{2}\frac{\sqrt{\zeta }}{\sqrt{b^2+2\varepsilon \zeta -\zeta ^2}}\,{\rm d}\zeta.$$ Such a formula[^6] allows to construct the *excess p-formula*. For the purpose, plugging in $\zeta $ as a function of $p$, after a little algebra, we have: $${\rm d}\Delta=\frac{-p^2}{\sqrt{a^2 b^2+2\varepsilon ap^2-p^4}}\,\,{\rm d}p\label{dDp}$$ namely the *excess p-formula*, see *Fluxions*, page 245, row 8. Let us go to the implications of . First, let us put $a=b$, then $\varepsilon=0$, equilateral hyperbola, so that: $${\rm d}\Delta=\frac{-p^2}{\sqrt{a^4-p^4}}{\rm d}p\label{ddDp}.$$ This connection really explains why Maclaurin, D’Alembert and Landen expended considerable effort over a curve, like the hyperbola, only marginally involved in the astronomic or ballistic computations. In other words, given the hyperbola in *pedal* form $$r(p)=a\left(2\varepsilon+a\frac{b^{2}}{p^{2}}\right),$$ let us construct the excess ${\rm d}\Delta$ between the length of the tangent at a point and the arc of the hyperbola from its vertex to that $E$-point. Then the elastica curve can stem also by integrating the excess (\[ddDp\]); or, as seen, by rectifying the lemniscate.[^7] A further consequence of the $p$-formula of the differential excess consists of providing the finite excess, say $\Delta_\infty$, when the $E$-point, moving on the hyperbola, slides to infinity.[^8] In such a case, both the tangential length and the arc, are really increasing without limits, so that their difference would appear indeterminate. We will show in the next section the relevant Maclaurin evaluation of the excess by means of a series expansion.
When the E-point on the hyperbola slides from E to infinity, the perpendicular segment $p$ changes from the value $a$ to 0. Minding the $e$ and $\varepsilon$ definitions, then becomes: $${\rm d}\Delta=\frac{p^{2}}{ab}\left(1+\frac{p^{2}}{b^{2}}\right)^{-1/2}\left(-\frac{a}{\sqrt{a^{2}-p^{2}}}\,{\rm d}p\right).$$ Expanding the second factor in binomial series, Maclaurin obtains: $${\rm d}\Delta=\frac{1}{ab}\left(-\frac{a}{\sqrt{a^{2}-p^{2}}}\,{\rm d}p\right)\left(p^{2}-\frac{p^{4}}{2b^{2}}+\frac{3}{8}\frac{p^{6}}{b^{4}}-\&c\right).$$ After this -we think- he would have evaluated (putting $p=a\sin u$ in order to obtain three integrals of even powers of $\sin u$): $$\int_0^a\frac{p^2}{\sqrt{a^{2}-p^{2}}}\,{\rm d}p=\frac{\pi a^2}{4},\,\int_0^a\frac{p^4}{\sqrt{a^{2}-p^{2}}}\,{\rm d}p=\frac{3 \pi a^4}{16},\,\int_0^a\frac{p^6}{\sqrt{a^{2}-p^{2}}}\,{\rm d}p=\frac{5 \pi a^6}{32}$$ promptly leading to the final Maclaurin formula,[^9] which we can read at *Fluxions*, page 245, row 8: $$\label{maccessoappr}
\Delta_\infty=\frac{\pi a^{2}}{2b}\left(\frac{1}{2}-\frac{3a^{2}}{16b^{2}}+\frac{15a^{4}}{128b^{4}}-\&c\right).$$
What Maclaurin could not know…
------------------------------
Starting from , let us consider the *excess elliptic integral* $$\Delta_{\infty}=\int_0^a\frac{p^2}{\sqrt{a^2b^2+2\varepsilon ap^2-p^4}}\,{\rm d}p$$ but recalling relationships between $\varepsilon,\,e,\,a,$ and $b$ and changing variable by $p^2=q$ we get: $$\Delta_{\infty}=\frac12\int_0^{a^{2}}\sqrt{\frac{q}{(b^2+q)(a^2-q)}}\,{\rm d}q.$$ Let us refer to [@Grr], page 263 integral 3.141-10: then the excess is given by the difference between two complete elliptic integrals of second and first kind: $$\label{maccesso}
\Delta_{\infty}=\sqrt{a^2+b^2}\,\boldsymbol{E}(k)-\frac{b^2}{\sqrt{a^2+b^2}}\,\boldsymbol{K}(k)$$ where the elliptic modulus $k$ is given by $
k=a/\sqrt{a^2+b^2}.
$ Now by the hypergeometric series expansions for $\boldsymbol{K}(k)$ and $\boldsymbol{E}(k)$ $$\begin{split}
\boldsymbol{K}(k)&=\frac{\pi}{2}\,_{2}\mathrm{F}_{1}\left( \left.
\begin{array}{c}
\frac12;\frac12 \\[2mm]
1
\end{array}
\right| k^2\right)=\frac\pi2\,\sum_{n=0}^\infty\left[\frac{(2n)!}{2^{2n}(n!)^2}\right]^2k^{2n}\\
\boldsymbol{E}(k)&=\frac{\pi}{2}\,_{2}\mathrm{F}_{1}\left( \left.
\begin{array}{c}
-\frac12;\frac12 \\[2mm]
1
\end{array}
\right| k^2\right)=\frac\pi2\,\sum_{n=0}^\infty\left[\frac{(2n)!}{2^{2n}(n!)^2}\right]^2\frac{k^{2n}}{1-2n}
\end{split}$$ expanding around $a=0$ relation , we find exactly .
Somewhat like to was also found by John Landen in 1780 as we will see in the next section.
Landen
======
The hyperbola’s theorem publishing history
------------------------------------------
In the introduction to his *Théorie des fonctions analytiques*, 1797, Joseph-Louis Lagrange (1736-1813) refers to John Landen (1711-1790) as a *habile Géométre anglais*. Landen, famous among the most anti-academic mathematicians, had one final purpose: improving Maclaurin and D’Alembert’s results [@lan1]:
> Mr. Maclaurin, in his Treatise of fluxions, has given sundry very elegant theorems for computing the fluents of certain fluxions by means of elliptic and hyperbolic arcs; and Mr. D’Alembert, in the Memoirs of Berlin Academy, has made some improvement upon what had been written on that subject. But some of the theorems given by those gentlemen being in part expressed by the difference between an arc of an hyperbola and its tangent, and such difference being not directly attainable,…
The hyperbola papers are in substance only one in its three variants 1771, 1775, 1780. Each of them holds the hyperbola rectification theorem, some corollary about the excess, and an application to a circular pendulum.
The first paper [@lan1] where Landen faced with the problem of hyperbola rectification was read on June 6, 1771 and enclosed in vol. 61 of the *Philosophical Transactions of the Royal Society* under the title *A disquisition…* Its startup has been referred: his main criticism again of the limiting value of the excess will be analyzed later. There he refers explicitly to integration methods by means of arcs of hyperbola and ellipse and notes to have performed the hyperbola rectification, the proof of which he would show later in his second memoir. Afterwards Landen tried to apply this to the fluxion of time of a heavy bead freely descending from rest along a circular arc. The article of 1775 summarizes the conclusions of that issued four years earlier and describes his finding on page 285 with these words:
> Thus, beyond my expectation, I find that the hyperbola may in general be rectified by means of two ellipses!
After having applied again the fluxions to the time of a pendulum, Landen recalls that Maclaurin, Jakob and Johann Bernoulli, and Leibniz deemed that the elastic curve could not be constructed by the quadrature or rectification of conic sections. However, in [@lan2], page 288:
> the contents of this paper, properly applied, will evince that the elastic curve (with many others) may be constructed by the rectification of the ellipsis only, without failure in any point.
Was Euler aware of know Landen’s 1775 paper? Legendre’s answer ([@Leg1], p. 89) is negative:
> Euler n’ait rien écrit à l’occason du Mémoire de Landen, imprimé dans les Transactions philosophiques de 1775, d’ où il faut conclure que ce Mémoire n’est pas parvenu à sa conaissance; car dans l’hypothése contraire, cet illustre Géomètre aurai sans doute, suivant son usage, publié sers propres réflexions sur une découverte analytique qui devait particulièrement l’intéresser.
In any case, we do not agree with the level of importance commonly given to the 1775 article: Euler lived till to 1783 and he could have read its third version, published by Landen in *Mathematical Memoirs*, 1780. The final article [@lan3] under the simple title *Memoir of the ellipsis and the hyperbola*, was enclosed as II Memoir in the first volume of *Mathematical memoirs*. The author, perhaps aware of the obscurity of his two first editions, looked for better clarity, that was not always achieved. In any case, we shall follow this closely. Landen’s aim was to investigate the difference between an arc of an hyperbola and its tangent, since such differences were not directly attainable.
Landen’s excess formula as a consequence of Maclaurin’s
-------------------------------------------------------
In order to prove his hyperbolic theorem, Landen built as a first step, a formula for the hyperbolic excess which, either in his premises (Apollonius theorems, hyperbolic radius function of the $p$ normal length), or in the process (similar triangles, fluxions and chain of variables), tracks on the Maclaurin’s scheme, as quoted by himself.
Given the hyperbola of center S, semiaxes $a,\,b$ of equation , the hyperbola asymptotes are $y=\pm\,(b/a)\,x$. In a cartesian orthogonal reference frame we consider in the first quadrant the branch of the above hyperbola of *eccentricity* $
e=\sqrt{(a^{2}+b^{2})/a^{2}}.
$ Recalling $\varepsilon$ as introduced in formula we also recall $X$ from equation . Taking the derivative with respect to $X$, he finds: $${\rm d}(\overline{EP})=\frac{\sqrt{a}}{2x\sqrt{X}}\frac{X^2+b^2}{\sqrt{X^2+2\varepsilon X-b^2}}\,{\rm d}X.$$ The *differential excess* ${\rm d}\Delta$ concerning a single E-point of the hyperbola is the shift between the relevant tangent segment and the arclength whenever $X$ undergoes a change ${\rm d}X$, is found to be as in equation : Landen’s 1780 demonstration adds nothing to Maclaurin’s, and we will refrain from expanding on it. We prefer to display how, starting from , taken from *Fluxions*, one can quickly express ${\rm d}\Delta$ given as the square root of a ratio of two quadratic binomials of a certain variable.
Let us start from the $(x,y)$ coordinates equation which will not be the final hyperbola of the theorem. The length set by Maclaurin as $\varepsilon,$ , is named by Landen $f=(a^{2}-b^{2})/(2a).$ The perpendicular length $p$ stated by Landen as: $ p^{2}=m z$ which, compared with , provides the relationship linking the state variable $X$ by Maclaurin, to the Landen’s $z$: $$\begin{aligned}
X&=\frac{b^{2}}{z}\label{icchese}\\
{\rm d}X&=-\frac{b^{2}}{z^{2}}{\rm d}z.\label{de-icchese}\end{aligned}$$ Then, plugging and in one gets: $${\rm d}\Delta=-\frac{1}{2}\frac{\sqrt{a}\sqrt{z}}{\sqrt{b^{2}+2fz -z^{2}}} {\rm d}z,\label{delta}$$ which can be read at Landen’s *Memoirs*, page 25, row 3. Up to this point he makes a change of coefficients putting: $$a=m-n;\quad b=2\sqrt{mn},\label{semi}$$ and defines, instead of $z$ a new variable: $$t^{2}=(m-n)^{2}-p^{2},\label{ti}$$ so that $mz=m^{2}-t^{2}.$ In such a way the excess (\[delta\]) will become: $${\rm d}\Delta= \sqrt{\frac{(m-n)^{2}-t^{2}}{(m+n)^{2}-t^{2}}}\,\,{\rm d}t .\label{excess}$$ Having assumed $t$ defined by (\[ti\]) as a new independent variable in the hyperbola equation \[ipxya\], then (\[excess\]) provides the excess of the hyperbola whose semiaxes are given, see (\[semi\]), by the difference and by the double geometrical mean of the coefficients $(m,n)$ entering . Of course in there $m-n$ is the transverse semiaxis of the hyperbola, so that, wishing to compute its finite excess, $t$ shall be spanned minding . The first of \[f21\] shows the geometrical link of length segments $t,\,m-n$ and $p$.
Landen’s auxiliary ellipses
---------------------------
To deal with the last integral which was out of his capabilities, Landen followed the approach of integration by means of curves. He started from a first ellipse, $$\frac{x^{2}}{a'^{2}}+\frac{y^{2}}{b'^{2}}= 1$$ whose differential arclength, by means of the characteristic triangle, can, after little work, be written as: $${\rm d}\eta_{1}=\sqrt{\frac{a'^{2}-g'x^{2}}{a'^{2}-x^{2}}}\, {\rm d}x\label {di-etauno}$$ where $g'=(a'^{2}-b'^{2})/{b'^{2}}.$ Afterwards he specialized the curve assuming $a'=m+n,\, b'=2\sqrt{mn}$ where $m-n$ and $2\sqrt{mn}$ are the semiaxes of the hyperbola, and $
g'=\left[(m-n)/(m+n)\right]^{2}.
$ Making the change of variables $x\mapsto t$: $$x=\frac{m+n}{m-n}\,t, \label{xt}$$ the finite $\eta_{1}$ first ellipse’s arclength will be the fluent[^10] of: $
\sqrt{((m+n)^{2}-t^{2})/((m-n)^{2}-t^{2})}.%\label {etauno}
$ Let the second ellipse have semiaxes $m$ and $n$, so that $
g=(m^{2}-n^{2})/{m^{2}}.
$ We make reference to the mid \[f21\]: chosen along the ellipse a point E of abscissa $x$, we draw the tangent at this point, whose segment $t$ is limited by E and the intersection P with the perpendicular sent from the $C_{1}$ ellipse’s center.
By means of standard procedures, one can find the value of $\overline{EP}=t$ as a function of $x$: $$t=gx\sqrt{\frac{m^{2}-n^{2}}{m^{2}-gx^{2}}} .\label {ti of x}$$ Solving to $x^{2}$: $$2gx^{2}=t^{2}+gm^{2}\mp\sqrt{(m^{2}-n^{2} )^{2}-2t^{2}(m^{2}+n^{2})+t^{4}}.\label {squarex}$$ Landen discards the minus sign, so that, after differentiation, writing the radicand as $$\left[(m-n)^{2}-t^{2}\right]\left[(m+n)^{2}-t^{2}\right]$$ by he gets the fluxion: $$2gx{\rm d}x=t{\rm d}t+\frac{(m^{2}+n^{2})t\,{\rm d}t -t^{3}\;{\rm d}t}{\left[(m-n)^{2}-t^{2}\right]^{1/2}\left[(m+n)^{2}-t^{2}\right]^{1/2}}$$ as written in his *Memoirs*, page 32, row 4 from bottom. Dividing last formula[^11] to $t$ and minding , he gets: $$\frac{gx\,{\rm d}x}{t}=\sqrt{\frac{m^{2}-gx^{2}}{m^{2}-x^{2}}} {\rm d}x, \label{pippow}$$ which has the same right hand side of (\[di-etauno\]), which provides the elliptic elementary arc ${\rm d}\eta_{1}$ whose semiaxis $a$ has been changed to $m$: such a new elementary arclength we are naming ${\rm d}\eta_{2}$. Therefore: $${\rm d}\eta_{2}=\frac{gx{\rm d}x}{t}=\frac{1}{2}\, {\rm d}t+\frac{1}{4} \sqrt{\frac{(m-n)^{2}-t^{2}}{(m+n)^{2}-t^{2}}}\, {\rm d}t+\frac{1}{4} \sqrt{\frac{(m+n)^{2}-t^{2}}{(m-n)^{2}-t^{2}}}\,{\rm d}t ,\label{deta2}$$ where the element $dt$ is tangential to the $(m,n)$ ellipse.
The finite excess and Landen’s geometric Theorem
------------------------------------------------
Before integrating formula it will be observed that at its right hand side the second term means the elementary excess of hyperbola of semiaxes $m-n$ and $2\sqrt{mn}$, whilst the third term is the elementary arc (\[di-etauno\]) of the ellipse of semiaxes $m+n$ and $2\sqrt{mn}$. In such a way, by integration we will get: $$\eta^{(A_{2}E_{2})}=\frac{1}{2}\overline{PE}+\frac{1}{4}(\overline{FP}-\overline{FA})+\frac{1}{4}\eta^{(A_{1}E_{1}).}$$ Writing $\overline{PF}= t_{Hyp}$ and recalling $\overline{PE}=t$, we get: $$\overline{AF}=Hyp=t_{Hyp}+2t+\eta^{(A_{1}E_{1})}-4\eta^{(A_{2}E_{2})}.\label{llanden}$$ We now highlight to how the *Landen hyperbolic theorem itself can be read as belonging to the family of integrable combinations[^12]*: it states in fact that even if each of the three arcs of hyperbola + ellipse 1+ ellipse 2 is not algebraically integrable, nevertheless their linear combination $
Hyp-\eta^{(A_{1}E_{1})}+4\eta^{(A_{2}E_{2})}
$ is algebraically integrable and equates the addition of segments $t_{Hyp}+2t$.
The relationship displays the Landen theorem:
The arc $\overline{AF}$ of hyperbola from a vertex A to F, whose elementary excess is given by and then having semiaxes $(m-n)$ and $2\sqrt{mn}$, is computable by , where:
1. $\overline{FP}$ is the straight segment of pedal [^13] tangent to hyperbola at F;
2. $t_{Hyp}$ is the straight segment of pedal tangent to the ellipse of semiaxes $(m,n)$;
3. $\eta^{(A_{1}E_{1})}$ is the arc of the ellipse 1 of semiaxes $(m+n, 2\sqrt{mn})$;
4. $\eta^{(A_{2}E_{2})}$ is the arc of the mentioned ellipse 2
According to G. N. Watson, who has really understood [@Marquis] all the above machinery of ellipses:
> Pairs of ellipses whose semiaxes are related in the manner of two ellipses of this problem are said to be connected by Landen’s transformation. In the hands of Legendre the transformation became a most powerful method for computing elliptic integrals.
Let us add our 12-step operations’ sequence in order to achieve a complete geometrical overview of all the machinery , starting from the sole knowledge of $m$ and $n$. We refer to \[f21\].
1. By $m>0$ and $n>0$, we construct the segments $m-n$ and $2\sqrt{mn}$: they will be semiaxes of the hyperbola which we draw pointwise on a $(S, x, y)$ frame, branch $(x>0, y>0)$, vertex A. Let us draw the asymptote *asy* too.
2. We draw the vertical straight line *vert-tan* passing through A.
3. Half-circle over the diameter $\overline{AS}$ and whose centre is $C_{0}$.
4. Centre $C_{1}$, we draw the quadrant $(x> 0, y> 0)$ of the ellipse 1: $$\frac{x^{2}}{(m+n)^{2}}+\frac{y^{2}}{4mn}= 1$$ of semiaxes $m+n$ and $2\sqrt{mn}$, the last quantity being provided as ordinate of the intersection N between *vert-tan* and *asy*.
5. Let us draw $(x>0, t>0)$ the straight line of connecting the origin to Z$\equiv(t=m-n, x=m+n)$ being the $t$ reference axis heading downwards.
6. Centre $C_{2}$, we draw the quadrant $(x>0, y>0)$ of the ellipse: $$\frac{x^{2}}{m^{2}}+\frac{y^{2}}{n^{2}}= 1$$ of semiaxes $m$ and $n$.
7. Choosing a point E on the second ellipse, and the relevant E-tangent, we construct the perpendicular from $C_{2}$ to it, P is found: let it be $\overline{A_{2}E_{2}}=t$ which we transfer as $\overline{C_{1}t}$, $\overline{AH}$, and so on.
8. From the third diagram we enter the second one and, given $t$, by means of the projecting horizontal line, we get both $x$ and $\eta^{(A_{1}E_{1})}$.
9. Entering the hyperbola, let H be the intersection of the mentioned projecting horizontal line with *vert-tan*.
10. Centre in A, spread $\overline{AH}$, we get a circle which crosses that of centre $C_{0}$, finding K.
11. The triangle AKS is right in K, so that, if $\overline{AK}=t$, $\overline{AS}=m-n$, so that: $
\overline{SK}^{2}=\overline{AS}^{2}-\overline{AK}^{2}
$ then by , it shall be: $\overline{SK}=p$.
12. Known $p$, we draw the circle of centre S and radius $p$, and consider the straight line *tan-ob* touching *simultaneously* it and the branch $(x> 0, y> 0)$ of the hyperbola. Such a tangent is unique.[^14] In this way we find both points P and F.
We have then shown the way of drawing the finite hyperbolic excess between the points A and F of the $(m, n)$ hyperbola parametrized on the $t$ value of the tangent length cut on the second $(m,n)$ ellipse. Furthermore we displayed how the fixing of $t$ is equivalent to the fixing of the upper F point on the hyperbola.
Simpson’s Fluxions and his treatment of hyperbolic excess
---------------------------------------------------------
Thomas Simpson’s (1710-1761) *A New treatise of Fluxions* was issued in 1737 (five years before Maclaurin’s); in 1750 he published a new edition of it, which, however, he had wished to be considered as new work rather than a new edition of an old one. We consulted the posthumous reprint, 1776, where, page 509 of volume I, Simpson describes the problem:
> To determine the difference between the length of the arch of a semi-hyperbola infinitely produced, and its asymptote.
Simpson could not have seen any of Landen papers (1771, 1775, 1780) so this is either his personal elaboration (following Maclaurin whose treatise had went out in 1742), or a reflection of some discussion on the subject he had with Landen. Let us give some elements on his treatment. Both Maclaurin and Landen in treating such a subject make use of non-cartesian variables and for their purpose produce pedal coordinates, polar anomalies, lengths of the tangent, and so on, integration variables not all having a direct geometric visibility. On the contrary, Simpson makes use only of the cartesian orthogonal coordinates $(x,y)$. If $a$ and $b$ denote the hyperbola semiaxes, being the first over the $x$ axis, of course, $y=(b/a) \sqrt{x^{2}-a^{2}}$, for its elementary arclength he obtains: $${\rm d}s=-\frac{a}{\delta}\, \frac{\sqrt{1-\delta^{2}u^{2}}}{u^{2}\sqrt{1-u^{2}}}\,{\rm d}u$$ where $ \delta^{2}=a^{2}/(a^{2}+b^{2}).$ Expanding the power $1/2$ by means of the binomial theorem and integrating term by term, the first integrated one gives by itself the length of the tangent. In such a way:
> therefore the difference between the arc and the asymptote will be equal to the fluent of the remaining terms in the difference sought.
Finally, the hyperbola arclength is computed in some particular cases of interest.
Fagnano’s theorem on elliptic arcs whose difference is rectifiable
------------------------------------------------------------------
We have been ventured to mention the *integrable combinations* which Johann Bernoulli had introduced, 1695. He showed that on some curves two arcs whose difference was rectifiable could be found, although each separate arc could not be rectified.
That subject is quite close to the work of the Italian mathematician Giulio Carlo Toschi di Fagnano (1682-1766), Fellow of the Royal Society since 1723, who published in 1716 a paper[^15] [@Giorn], which has some connection with that of Landen’s. Fagnano proved that on some curves it is possible to find infinite arcs whose differences can be algebraically found, *even if the single arcs cannot be rectified*. Or, analytically speaking, infinite differential combinations integrable over those curves. We can read in [@Leg1]:
> Un Géomètre italien dÕune grande sagacité, ouvrit la route à des spéculations plus profondes. Il prouva que sur toute ellipse ou sur toute hyperbole donnée, on peut assigner, dÕune infinité de manières, deux arcs dont la différence soit égale à une quantité algébrique. Il démontra en mème temps que la courbe nommée lemniscate jouit de cette singulière propriété, que ses arcs peuvent ètre multipliés ou divisés algébriquement, comme les arcs de cercle, quoique chacun dÕeux soit une transcendante dÕun ordre supérieur.
Fagnano was especially successful with the cubic parabola, the lemniscate[^16] and the ellipse, but he is now better remembered in connection with the latter. On a quadrant AQPB (see \[f22\] ) of a given C-centered ellipse, Fagnano found pairs of points, like P and Q, such that the difference of the arcs $\overline{BP}$ and $\overline{AQ}$ is rectifiable by ordinary integration. It was afterwards found that if perpendiculars CM and CM’ are drawn from C, onto the tangents at P and Q, then $\overline{PM}=\overline{QM'}$; and that each of these is equal to the difference between the two arcs mentioned.
Fagnano’s work, though with methods different from his *Produzioni matematiche*, is considered quite extensively in [@Green], pages 182-189.
The Italian mathematician Francesco Siacci (1839-1907) see [@Siacci], collecting some Fagnano’s theorems and further contributions (alternative proofs, additions, details, corollaries, geometrical constructions, and so on) due to: Leonard Euler (1707-1783), Adrien Marie Legendre (1752-1833), Augustin Luis Cauchy (1789-1855), C. Küpper[^17], William Wallace (1768-1843), John Brinkley (1763-1835), Pierre Verhulst (1804-1849), Paul Serret (1827-1898), Michel Chasles (1793-1880), and Johann August Grünert (1797-1872), *surprisingly did not ever cite Landen!* On the contrary, in his memoir *Of the ellipsis and hyperbola* about the hyperbolic excess theory, in articles 4, 9 and 13, see \[f23\],
by purely fluxional means, Landen constructs a theorem (by himself not specially highlighted but) whose statement is:
> If $\overline{p'e'}$ and $\overline{p''e''}$ be equal tangents to the ellipsis $ae'e''d$; the arc $\overline{ae'}$ will be equal to the arc $\overline{de''}$+the tangent $\overline{p''e''}$
*namely nothing but the ellipse of Fagnano’s theorem*. The theorem of Comte Fagnani” was cited and explained by[^18] [@Leg4], and again in [@Leg2], page 44. On the contrary, we ignore whether Landen in 1780 was aware of Fagnano’s article editions (1716; 1750) or not; in any case he preferred to discard it, showing his own proof. In any case he did not ever cite Fagnano, and we are inclined to believe he did not know about Fagnano’s paper at all. A comparison between the proofs of Fagnano and Landen will not be performed here.
Landen and the hyperbola limit excess
-------------------------------------
### His pendulum-based motivation
In the paper of 1780, articles $2\div14$ concerning the hyperbola, are followed by articles 15 and 16 about the ellipse. At page 34 Landen meets a differential of the type: $$\frac{{\rm d}z}{\sqrt{z}\sqrt{n^{2}+2fz-z^{2}}},\label{ahhh}$$ which he solves through arcs of ellipse, namely doing an integration by means of curves. Even if the second memoir holds application to a pendulum, the third memoir entitled *Of the descent of a body in a circular arc*, is completely devoted to the heavy body’s motion *in vacuo* along a circular arc. At page 37 the fluxion of the time of descent is set as: $$\frac{1}{2}\frac{r h^{-1/2}x^{-1/2}\dot{x}}{\sqrt{2b r-b^{2}-2 \overline {r-b} x-x^{2}}},$$ $h,\,r,\,b$ being some constants and $x$ a variable distance during the bead motion.[^19] The above integral is of the same type described at , which is exactly the same of previously met as the differential hyperbolic excess of Maclaurin’s . What the above confirms the practical tendency kept by Landen, who, even when faced with theoretical questions, saw calculus as a tool to be improved on more and more, in order to solve recreational problems, rectification of curves, problems of dynamics and algebraic equations as well.
Whenever the F point along the hyperbola goes to infinity, the excess becomes $\infty-\infty$, and then indeterminate. Landen then provides different approaches for obtaining the value of the limit hyperbolic excess he calls $L$.
### The limit excess: first proof, 1771, 1775.
The calculation is in article [@lan2] of 1775, organized as it follows. Solving to the arc of the second ellipse, he gets: $$\eta^{(A_{2}E_{2})}=\frac{1}{2} t+\frac{1}{4}(t_{Hyp}-Hyp)+\frac{1}{4}\eta^{(A_{1}E_{1})}.$$ But when the hyperbola’s point tends to infinity, then, see , $p\to 0$ so that $t\to m-n$. For the first ellipse the abscissa of point E becomes, see , $m+n$ and then the arc fills the whole quadrant, say $S_{1}$, while that relevant to the second ellipse becomes: $
1/2S_{2}+1/2(m-n)
$ $S_2$ being the second ellipse quadrantal arc. Landen claims to have demonstrated the above relationship at art. 10 of his paper, 1771, an extremely long and useless subject which will be omitted here. Inserting all this in , he gets: $$\frac{1}{2}S_{2}+\frac{1}{2}(m-n)=\frac{1}{2}(m-n)+\frac{1}{4}L+\frac{1}{4}S_{1}$$ or: $$L=2S_{2}-S_{1},\label{luna}$$ so that the indetermination is solved: *the limit $L$ of the excess of a $(m,n)$ hyperbola is found as a simple combination of quadrantal arcs of the auxiliary ellipses 1 and 2 whose semiaxes closely depend on $m$ and $n$*. Two short presentations that attempt to translate the machinery of Landen’s hyperbola theorem and its auxiliary ellipses into mathematical language of the 1900s on a pure analytical basis, without reference to Maclaurin or the rectification landscape, are due to [@Marquis] and to [@Cantor4] pp. 842-847, both readable papers, but where the historical context is deliberately lost.
### The Landen limit excess final proof, 1780.
With reference to \[f21\], the starting point is: $$\overline{AF}=Hyp=t_{Hyp}+2t+\eta_{1}-4\eta_{2}\label{anden}.$$ Landen applies[^20] to the second ellipse what is merely Fagnano’s theorem on the elliptic arcs whose difference is rectifiable: $$\overline{p'e'}=\overline{p''e''}=t=\eta_{2}-\overline {de''}.\label{Fag}$$ Plugging (\[Fag\]) in (\[anden\]), he gets: $
Hyp=t_{Hyp} +\eta_{1}-2\eta_{2}-2\overline{de''}.
$ Naming $S_{2} $ the quadrantal arc $ad$, by \[f21\] we have: $
\eta_{2}=S_{2}-\overline{e'e''}-\overline{e''d}
$ so that (\[anden\]) becomes: $
Hyp=t_{Hyp}-2S_{2}+2\overline{e'e''}+\eta_{1},
$ so that the excess will be given by $2S_{2}-2\overline{e'e''}-\eta_{1}.$ When the point on the hyperbola is going to infinity, we know the variable $t$ attains its maximum value $m-n$, so that the arc of first ellipse fills all its quadrant, assuming the value $S_{1}$. Such a maximum value is unique so that it is relevant to only one arc of the second ellipse. Otherwise speaking, two different arcs $ae'$ and $e''d$ cannot exist at which $t$ attains its maximum. Then $e'\equiv e''$so that $\overline{e'e''}\to 0.$ The conclusion is then , again.
Our direct approach to the excess through elliptic integrals.
-------------------------------------------------------------
\[cess\] For the hyperbola of equation with $a\geq b>0$, the excess is: $$\label{rcesso}
\mathscr{E}(0)=\sqrt{a^2+b^2}\, \boldsymbol{E}\left(\frac{a}{\sqrt{a^2+b^2}}\right)-\frac{b^2}{\sqrt{a^2+b^2}}\,\boldsymbol{K}\left(\frac{a}{\sqrt{a^2+b^2}}\right)$$
Rotate negatively by $\pi/2-\arctan(b/a)$ the axes so that the asymptote of equation $y=(b/a)x$ will coincided with the vertical axis. Then applying the transformation of coordinates $$\begin{cases}
x'=x\cos\left(\arctan\frac{b}{a}-\frac{\pi}{2}\right)-y\sin\left(\arctan\frac{b}{a}-\frac{\pi}{2}\right)\\
y'=x\sin\left(\arctan\frac{b}{a}-\frac{\pi}{2}\right)+y\cos\left(\arctan\frac{b}{a}-\frac{\pi}{2}\right)
\end{cases}
\iff
\begin{cases}
x'=\dfrac{b x + a y}{\sqrt{a^2+b^2}}\\[3mm]
y'=\dfrac{-a x + b y}{\sqrt{a^2+b^2}}
\end{cases}$$ changing in $$\label{ha}
y=\frac{\left(a^2-b^2\right)x^2 +a^2 b^2}{2 a b x}$$ and where we will omit $^\prime$. The hyperbola vertex is: $${\rm A}=\left(\frac{a b}{\sqrt{a^2+b^2}},\frac{a^2}{\sqrt{a^2+b^2}}\right)$$ Taken $\varepsilon>0$ close enough to zero, the straight line touching hyperbola at ${\rm E}=(\varepsilon,y(\varepsilon))$ is: $$\tag{$\tau$}
y=\frac{a b}{\varepsilon }-\frac{x \left(a^2 b^2-\varepsilon ^2
\left(a^2-b^2\right)\right)}{2 a b \varepsilon ^2}$$ so that the normal to $\tau$ going out from the origin has equation: $$\tag{$\nu$}
y=\frac{2 a b \varepsilon ^2}{a^2 b^2-\varepsilon ^2 \left(a^2-b^2\right)}\,x$$ and, if ${\rm P}:=\tau\cap\nu,$ then $$\overline{{\rm PE}}=\frac{a^4 b^4-\varepsilon ^4 \left(a^2+b^2\right)^2}{2 a b \varepsilon
\sqrt{a^4 \left(b^2-\varepsilon ^2\right)^2+2 a^2 b^2 \varepsilon ^2
\left(b^2+\varepsilon ^2\right)+b^4 \varepsilon ^4}}.$$ The excess evaluation will be completed by : $$\sqrt{1+\left(\frac{{\rm d}y}{{\rm d}x}\right)^2}=\frac{\sqrt{\left(a^2+b^2\right)^2x^4 -2 a^2 b^2 \left(a^2-b^2\right)x^2 +a^4 b^4}}{2abx^2}$$ so that the excess as a function of the abscissa $\varepsilon$ will be: $$\begin{split}
\mathscr{E}(\varepsilon)&=\overline{{\rm PE}}-\overarc[.7]{\rm AE}\\%{\rm arc}{{\rm AE}}\\
&=\frac{a^4 b^4-\varepsilon ^4 \left(a^2+b^2\right)^2}{2 a b \varepsilon
\sqrt{a^2 b^2-2 a^2 b \varepsilon +a^2 \varepsilon ^2+b^2 \varepsilon ^2}
\sqrt{a^2 b^2+2 a^2 b \varepsilon +a^2 \varepsilon ^2+b^2 \varepsilon ^2}}\\
&-\int_{\varepsilon}^{\frac{a b}{\sqrt{a^2+b^2}}}\frac{\sqrt{\left(a^2+b^2\right)^2x^4 -2 a^2 b^2 \left(a^2-b^2\right)x^2 +a^4 b^4}}{2abx^2}\,{\rm d}x.
\end{split}$$
Integrating by parts: $$\begin{split}
\overarc[.7]{\rm AE}&=\frac{1}{2ab}\left\{\frac{\sqrt{a^4 b^4+\varepsilon ^4 \left(a^2+b^2\right)^2-2 a^2 b^2 \varepsilon
^2 \left(a^2-b^2\right)}}{\varepsilon }-2 a b^2\right.\\
& \left. +\int_{\varepsilon}^{\frac{a b}{\sqrt{a^2+b^2}}}\frac{2\left(a^2+b^2\right)^2 x^2 -2 a^2 b^2
\left(a^2-b^2\right)}{\sqrt{\left(a^2+b^2\right)^2x^4 -2 a^2
b^2 \left(a^2-b^2\right)x^2 +a^4 b^4}}\,{\rm d}x\right\}.
\end{split}$$ In such a way we get rid of the indetermination, so that, passing to limit for $\varepsilon\to0^{+}$ $$\label{cesso:r}
\mathscr{E}(0)=b-\frac{1}{ab}\int_{0}^{\frac{a b}{\sqrt{a^2+b^2}}}\frac{\left(a^2+b^2\right)^2 x^2 - a^2 b^2
\left(a^2-b^2\right)}{\sqrt{\left(a^2+b^2\right)^2x^4 -2 a^2
b^2 \left(a^2-b^2\right)x^2 +a^4 b^4}}\,{\rm d}x.$$ The excess will be then computed by means of a couple of elliptic integrals: $$\begin{aligned}
\mathscr{E}_1&=\int_{0}^{\frac{a b}{\sqrt{a^2+b^2}}}\frac{\left(a^2+b^2\right)^2 x^2}{\sqrt{\left(a^2+b^2\right)^2x^4 -2 a^2
b^2 \left(a^2-b^2\right)x^2 +a^4 b^4}}\,{\rm d}x,\\
\mathscr{E}_2&=\int_{0}^{\frac{a b}{\sqrt{a^2+b^2}}}\frac{- a^2 b^2
\left(a^2-b^2\right)}{\sqrt{\left(a^2+b^2\right)^2x^4 -2 a^2
b^2 \left(a^2-b^2\right)x^2 +a^4 b^4}}\,{\rm d}x.\end{aligned}$$ Using in sequence [@By] entry 361.53 page 215 and 361.53 page 215, we get $$\mathscr{E}_1=\frac{ab\sqrt{a^2+b^2}}{2}\left[\boldsymbol{K}\left(\frac{a}{\sqrt{a^2+b^2}}\right)-2\boldsymbol{E}\left(\frac{a}{\sqrt{a^2+b^2}}\right)+\frac{2 b}{\sqrt{a^2+b^2}}\right].$$ In order to compute $\mathscr{E}_2$ invoking entry 3.138-7 page 259 of [@Grr] we get$$\mathscr{E}_2=-\frac{a b \left(a^2-b^2\right)}{2 \sqrt{a^2+b^2}}\,\boldsymbol{K}\left(\frac{a}{\sqrt{a^2+b^2}}\right)$$ Combining $\mathscr{E}_1$ with $\mathscr{E}_2$ we obtain .
Comparison with Landen’s formula
--------------------------------
Landen shows synthetically the hyperbola of equation $$\frac{x^2}{(m-n)^2}-\frac{y^2}{4mn}=1$$ has an excess given by the difference between the double of quadrantal arc of the ellipse of equation $$%\tag{S$_1$}
\frac{x^2}{(m+n)^2}+\frac{y^2}{4mn}=1$$ and of the quadrantal arc of the ellipse of equation $$%\tag{S$_2$}
\frac{x^2}{m^2}+\frac{y^2}{n^2}=1.$$ This can be done in modern notation by means of complete elliptic integrals:$$\label{lcesso}
\mathscr{L}=2m\, \boldsymbol{E}\left(\sqrt{\frac{m^2-n^2}{n^2}}\right)-(m+n)\, \boldsymbol{E}\left(\frac{m-n}{m+n}\right).$$ The task will be to prove two determinations of the excess, namely that and lead to the same thing. The first step is to express the coefficients $m$ and $n$ by means of the semiaxes $a$ and $b$ $$\begin{cases}
m-n=a\\
4mn=b^2
\end{cases}
\implies
\begin{cases}
m=\dfrac{1}{2} \left(\sqrt{a^2+b^2}+a\right)\\[2mm]
n=\dfrac{1}{2} \left(\sqrt{a^2+b^2}-a\right)
\end{cases}$$ Then, by equating and we have the identity $$\label{iid}
\begin{split}
&2\sqrt{a^2+b^2}\,
\boldsymbol{E}\left(\frac{a}{\sqrt{a^2+b^2}}\right)-\left(\sqrt{a^2+b^2}+a\right)
\boldsymbol{E}\left(\frac{2 \sqrt{a\sqrt{a^2+b^2}}}{a+\sqrt{a^2+b^2}}\right)=\\
& \frac{b^2}{\sqrt{a^2+b^2}}\,\boldsymbol{K}\left(\frac{a}{\sqrt{a^2+b^2}}\right).
\end{split}$$ To check that is true, we will use theorem 1.2 (c) page 12 of [@Borweins] where is proved that: $$\label{Borw}
\boldsymbol{E}(k)=\frac{1+k}{2}\boldsymbol{E}\left(\frac{2\sqrt{k}}{1+k}\right)+\frac{1-k^2}{2}\boldsymbol{K}(k).$$ Now, dividing both sides of by $\sqrt{a^2+b^2}$ and substituting $$k=\frac{a}{a^2+b^2}$$ provided that $$\frac{2 \sqrt{a\sqrt{a^2+b^2}}}{a+\sqrt{a^2+b^2}}=\frac{2\sqrt{k}}{1+k},\quad \frac{b^2}{a^2+b^2}=1-k^2$$ which is : namely the excess obtained through the analytical argument of theorem \[cess\] coincides with Landen’s synthetic excess. Of course we could change the viewpoint and by two excess representations infer that formula of [@Borweins] is true.
Did Landen really create the [*Landen transformation*]{}?
---------------------------------------------------------
Landen is often ignored in several historical works. For instance, M. Kline (1908-1992) [@Kline], in all of 1500 pages for 24 centuries of mathematicians, never once mentioned Landen’s name. The same with [@Katz], 800 pages. An exception is G. Loria (1862-1954), [@Loriast] who covered Landen’s contribution with 26 lines in total, but where the elliptic” contribution is misunderstood, made the common attribution’s mistake regarding elliptic integrals, as we will see. Furthermore Landen is often quoted carelessly and mostly for the *Residual Analysis*, rather than for his contributions to the hyperbola or to his (almost unknown) differential solution of algebraic equations. On the contrary, Florian Cajori (1859-1930), [@Cajo2], when describing almost all treatises or mathematics pamphlets, polemics and improvements, during a restricted range of time of the British mathematics, focused for a while on Landen’s historical bearing, writing appropriately although not entering analytical details.
Finally, several authors describe Landen’s contributions regarding the elliptic integrals, contributions which do not exist. For instance in [@Arm] p. 99 formula (4.54) we read:
> In 1775, Landen gave the formula $$%\tag{4.54}
> \int_{0}^{\phi_1}\left(1-k_{1}^{2}\sin^{2}\theta_{1}\right)^{-1/2}{\rm d}\theta_{1}=(1+k')\int_{0}^{\phi}\left(1-k^{2}\sin^{2}\theta\right)^{-1/2}{\rm d}\theta$$ where $\sin(\phi_{1})=(1+k')\sin(\phi)cos(\phi)(1-k^{2}\sin^{2}(\phi) )^{-3/2}$ and $k_1=(1-k')/(1+k')$. His proof is to be found in the Philosophical Transactions of the Royal Society, LXV, (1775), page 285.
and [@Mit] had written:
> Euler’s addition theorem and the transformation theory of Landen and Lagrange were the two fundamental ideas of which the theory of elliptic function was in possession when this theory was brought up for renoved consideration by Legendre in 1786.
Both statements are wrong. Such a transformation was really initiated, unconsciously, by Lagrange, as we shall see. Mistaken references like these can be found in many treatises and/or articles on elliptic functions: their authors probably did not check on the original Landen sources. Some of them have a better criticism but not always clear ones. A viewpoint we agree with is [@Guicc]:
> Landen’s problem was not that of integrating functions of the form $$\left[\left(1-x^{2}\right)\left(1-q^{2}x^{2}\right)\right]^{-1/2}$$ and he never expressed Landen’s transformations in the form known today.
Accordingly [@Cay], in a rather cautious note writes:
> Landen’s capital discovery is that of the theorem known by his name (obtained in its complete form in the memoir of 1775, and reproduced in the first volume of the Mathematical Memoirs) for the expression of the arc of an hyperbola in terms of two elliptic arcs. To find this, he integrates a differential equation derived from the equation $$t=gx \sqrt{\frac{m^{2}-x^{2}}{m^{2}-g^{2}x^{2}}}$$ interpreting geometrically in an ingenious and elegant manner three integrals which present themselves. If, in the foregoing equation we write $m = 1$, $g = k^{2}$, and instead of $t$ consider the new variable $y = \frac{t}{1-k'}$, then $$t=(1+k')x\sqrt{\frac{1-x^{2}}{1-k^{2}x^{2}}}$$which is the form known as Landen’s transformation in the theory of elliptic functions; but his investigation does not lead him to obtain the equivalent of the resulting differential equation $$\frac{{\rm d}y}{\sqrt{(1-y^{2})(1-\lambda^{2}y^{2})}}=(1+k')\frac{{\rm d}x}{\sqrt{(1-x^{2})(1-k^{2}x^{2})}}$$ where $$\lambda=\frac{1-k'}{1+k'}$$ due it would appear to Legendre and which (over and above Landen’s own beautiful result) gives importance to the theorem as leading directly to the quadric transformation of an elliptic integral in regard to the modulus.
The same caution is shown by some historians like [@Cooke] pages 529-539, who affirm the so-called transformation of elliptic integrals to be embedded *inside* the Landen theorem. This untrue, because *referring to* his paper of 1775, there is no match of such a claim; [@Greenn] write wisely:
> While his interest and application in these directions were acute, Landen failed to realize that the whole of his analytical transformations were particular cases of one general one, now known as the Landen transformations.
[@Smadja], is ascribing the origin of such a distortion to Legendre, who credits Landen with the autorship of transformation, in such a way lessening the merits of Maclaurin, D’Alembert, and of Lagrange too[^21]. But Landen could not have had the slightest idea of handling those mathematical objects: module, amplitude, and several kinds of irrational integrals which were to be classified by Legendre many years after him. Furthermore Landen could not have had a clear understanding on the relationship between his ellipses as a transformation capable of mapping two sets of coefficients, but keeping an elliptic integral invariant. Before the role of Lagrange is shown in the next section, let us try to fix a few points here.
The first to establish a change of variables capable of generating a recursive algorithm in order to compute a particular elliptic integral, was Lagrange in a paper [@Lag], a few years after Landen’s. But Legendre, see for instance *Remarque génerale* page 87 of [@Leg1] implemented (ibidem, p. 84) a purely analytical method for rectifying the hyperbola by means of elliptic integrals. At some point he introduces in the integral: $$F(\varphi,k ):=\int_{0}^{\varphi }\frac{\mathrm{d}\theta }{\sqrt{1-k^{2}\sin
^{2}\theta }}$$ a trigonometric transformation: $\varphi \to \hat{\varphi},\, k\to q$, namely:[^22] $$\sin \left( 2\hat{\varphi} -\varphi \right) =k\sin \varphi,\quad\frac{2\sqrt{k}}{1+k}=q \label{AA2}$$ and after this he establishes the relationship: $$F(\varphi,k)=\frac{2}{1+k}F\left( \hat{\varphi} ,\frac{2\sqrt{k}}{1+k}\right) \label{AA9}$$ by Richelot (1808-1875) named [@ric], *Legendre Gleichung*, being the new amplitude $\hat{\varphi}$ computed by the old one $\varphi$ by means of . Nevertheless is referred as *Module Amplitude Transformation*. About which let us cite again [@Cay]:
> The trigonometrical form \[…\] does not occur in Landen; it is employed by Legendre, I believe, in an early paper, *Mém. de l’Acad. de Paris*, 1786, and in the *Exercices*, 1811, and also in the *Traitè des Fonctions Elliptiques*, 1825, and by means of it obtains an expression for the arc of hyperbola in terms of two elliptic functions $E(c,\phi),\, E(c',\phi ')$, showing that the arc of the hyperbola is expressible by means of two elliptic arcs, le beau théorème dont Landen a enrichi la géométrie”.
This explains why a *transformation of Lagrange-Legendre* has its own life apart from the theorem of Landen”. The [@Smadja] page 389 recent reconstruction,
> Dans le cas qui nous occupe, il est tout aussi légitime de regarder le théorème de Landen comme un aboutissement que comme un commencement, selon qu’on l’envisage dans le contexte des recherches de Maclaurin et de d’Alembert qu’il complète et perfectionne, ou dans le contexte correspondant à la lecture de Legendre. Vouloir absolument trancher, de manière décontextualisèe, la question de savoir qui de Landen, Lagrange ou Legendre est le premier” à dégager la transformation de Landen est vain et illusoire
is the best description of what we deem false. We take the opportunity to highlight almost all the theorems about elliptical objects attributed to Landen (including e.g. the jacobian Theta functions), by treatises on Calculus like: [@WW]; [@Bell]; [@dur]; [@boros]; [@Mol]; [@Mit]; [@San]; [@Cas] and so on, are *lacking of any historical value*. As a matter of fact none of formulae referred to in this section was ever written by Landen. A further distortion is provided in [@Manna], page 289 equations (1-17) and (1-18), where Landen is mistakenly credited with deriving some elliptic integrals identities.
Lagrange
========
The *Nouvelle méthode*
----------------------
In volume II (1785) of Memoirs of the Turin Royal Academy we can read Lagrange’s long paper entitled: *Nouvelle méthode de calcul integral pour les différentielles affectées d’un radical carré sous lequel la variable ne passe pas le quatriéme degré* and where he is concerned with elliptic integrals, namely of rational functions of $x$ and of the square root of a fourth degree polynomial without multiple roots. He was perfectly aware of the problem’s peculiarity:
> si la plus haute de ces puissances ne monte pas au delà du quatrième degré on peut dans plusieurs cas construire l’intégrale par les arcs des séctions coniques.
which is of poor help in trying to compute them:
> mais il n’est d’aucune utilité pour l’integration effective de ces differentielles, car la rectification des sections coniques n’est encore connue que trés-imparfaitement,
so that the series expansion is, after all:
> le seul moyen de rappeler à l’integration toutes les formules differentielles d’une forme essentiellement irrationelle,
whose truncation error can be reduced ad libitum by taking a greater number of the terms expansion.
Lagrange considers elliptic integrals whose integrand has the form $$P(x)=M(x)+\frac{N(x)}{\sqrt{a+bx+cx^{2}+ex^{3}+fx^{4}}},$$ being $M(x),\,N(x)$ rational functions of $x$, so that the differential to be integrated will be split in a rational term $M(x)$ integrable through logarithms and arcs of circle, plus an irrational term on which Lagrange concentrated his effort. By means of algebraic transformations he proved that this term can be split in two terms, rational and irrational, the last being: $$\frac{Q(y^{2})}{\sqrt{\varepsilon +\zeta y^{2}+\eta y^{4}}}$$ where $\varepsilon ,\zeta ,\eta $ are constants and $Q$ a rational function of $y^{2}$. Next, he showed his method requires only the trinomial $\varepsilon
+\zeta y^{2}+\eta y^{4}$ can be broken in two binomial like $\alpha +\beta
y^{2},\,\gamma +\delta y^{2}$ being $\alpha,\, \beta,\, \gamma,\,\delta $ real quantities. Now the problem was: how to integrate such a differential. He succeeded in reducing it to: $$\frac{\text{d}y}{\sqrt{\left( 1\pm p^{2}y^{2}\right) \left( 1\pm
q^{2}y^{2}\right) }} \label{ZZ}$$ where $p$ and $q$ are real quantities, $p>q$, so that the square root is certainly real. Formula is the starting point for the arithmetic-geometric transformation that led Lagrange to a particular approach in order to transform the elliptic differentials, but which he did not draw the conclusions from. Our next section will be devoted to this. Let us note briefly that he, from through successive reductions, arrives at: $$\frac{{\rm d}z}{\sqrt{(b^{2}\pm z^{2})^{2}-\beta ^{2}}}$$ Expanding the binomial series, and by a term by term integration, he got an infinite series of integrals of rational functions. And then:
> Est donc assuré de pouvoir integrer aussi exactement qu’on voudra toute differentielle affecté d’un radical carré où la variable sous le signe monte jusqu’à la quatrième puissance; ce qui est le cas d’un grand nombre de probèmes géomeétriques et mécaniques.
Such a triumphalism is over: no one, having to evaluate an elliptic integral, with some success would have hoped to complete and control such a calculation: Lagrange used algebraic techniques to break the 4$^{\rm th}$ degree polynomial in the root, to avoid Ferrari’s formulæ, and complicated root calculations. Lagrange of course realized that the difficult process should be put to the test:
> Comme cette méthode est d’un genre assez nouveau, et qu’on pourrait rencontrer encore quelques difficultés dans son usage, nous allons l’appliquer en détail à la rectification des arcs elliptiques et hyperboliques.
When the eccentricity is very small the elementary elliptic arc can be integrated through a convergent series of even powers of the eccentricity itself, but when it becomes
> peu differente de l’unité, ce qui est le cas d’une ellipse ou d’une hyperbole très aplatie
the things become much involved, and even worst if
> nous allons appliquer notre méthode générale à la rectification d’une ellipse e d’une hyperbole quelconque.
Lagrange then went on through 30 intricate pages: its complicated approach cannot be expressed by any formula, but by several ones, holding several parameters often stemming from series expansions.
The birth of the Arithmetic Geometric Mean
------------------------------------------
At this point he constructed a sequence of arithmetic means and a sequence of geometric means as follows, fix $p=p_0>q=q_0$ and iterate: $$p_{n}=\frac{p_{n-1}+q_{n-1}}{2},\quad q_{m}=\sqrt{p_{n-1}q_{n-1}}.$$ After this he carried out algebraic arguments concluding that, following the assumptions, the sequence $(p_{n})$ decreases while $(q_{i})$ increases and $q_n<p_n$. The convergence is very fast: a Gauss’s example [@Gauss] shows that if for instance $p_0=1$ and $q_0=0.8$, at the third step, the values of $p_{3}
$ and $q_{3}$ are different since the 12$^{\rm th}$ digit on. Sequences $(
p_{n}) $ and $(q_{n})$ converge to a common limit denoted by $M(p,q)$, say the Arithmetic-Geometric Mean between $p$ and $q$. An extensive treatment of the AGM can be found at [@Cox] and [@Borweins] and the references therein. Gauss discovered, through the AGM introduced by Lagrange, a way for computing the elliptic integrals. He provided the arclength $L$ of the lemniscate, of equation $(x^{2}+y^{2})^{2}=a^{2}(x^{2}-y^{2}),$ as: $$\label{gauslem}
L=4a\int_{0}^{1}\frac{dt}{\sqrt{1-t^{4}}}=\frac{2\pi a}{M\left( 1,\sqrt{2}
\right) }$$ so that nowadays the number $1/M\left(1,\sqrt{2}\right)$ is known as the *Gauss lemniscatic constant*. For Gauss’ proof of see [@Cox]. The Arithmetic Geometric Mean was first set forth in the Turin memoir [@Lag] published in 1785[^23]. Afterwords Lagrange proceeded from variable $y$ of , to $y_{1}$, and from $y_{1}$ to $y_{2}$, and so on: $$y=\frac{y_{1}R_{1}}{1\pm q_{1}^{2}y_{1}},\,y_{1}=\frac{y_{2}R_{2}}{1\pm
q_{2}^{2}y_{2}},\,\ldots\,,y_{i}=\frac{y_{i+1}R_{i+1}}{1\pm q_{i+1}^{2}y_{i+1}},$$ being $R_{i}=\sqrt{\left( 1\pm p_{i}^{2}y_{i}^{2}\right) \left( 1\pm
q_{i}^{2}y_{i}^{2}\right) },\quad i=1,\,2,\,\dots,n.$ Next he gave $y_{i+1}^{2}$ as a function of $y_{i}$ $$y_{i+1}^{2}=\frac{\pm q_{i+1}^{2}y_{i}^{2}-1+R_{i}}{\pm 2p_{i+1}^{2}}$$ and established the differential relationships: $$\frac{dy}{R}=\frac{dy_{1}}{R_{1}}=\frac{dy_{2}}{R_{2}}=\cdots \label{FF}$$ namely: the joint variable transformation on $p$,$q$, and on the variable $y$ of integration kept invariant the elliptic differential . Notice that Lagrange did not stand much on the differential identities whose genesis is not explained at all. Soon after he went on other subjects. Exactly at this point we can show that, starting from one arrives at the famous Legendre identity.
From Lagrange differential identity to Legendre’s formula
---------------------------------------------------------
This section is devoted to fill a gap: Lagrange did not complete his work but Legendre somehow succeeded in providing the famous transformation. What could have happened in the meantime? We do not really know, but we succeeded in filling such a gap by introducing what Legendre *could have done*. Given $0<q<p,$ let us start by analyzing in the case with negative signs $$\int \frac{\mathrm{d}y}{\sqrt{(1-p^{2}y^{2})(1-q^{2}y^{2})}}.$$ The positivity ranges for the expression under root, where one wishes to work, lead to $$(1-p^{2}y^{2})(1-q^{2}y^{2})\geq 0\iff y\in \left( -\infty ,-1/q\right] \cup
\left[ -1/p,1/p\right] \cup \left[ 1/q,\infty \right) ;$$ and then we are going to consider the definite integral $$I(p,q):=\int_{0}^{1/p}\frac{\mathrm{d}y}{\sqrt{(1-p^{2}y^{2})(1-q^{2}y^{2})}}
\label{defmeno}$$ by introducing functions’ family $$R_{p,q}(y):=\sqrt{(1-p^{2}y^{2})(1-q^{2}y^{2})}$$ and $p_{1}<p$ and $q_{1}>q$ with $p_{1}>q_{1},$ a new integration variable $y_{1}$ in function of $y$ is introduced: $$y=\frac{y_{1}}{1-q_{1}^{2}\,y_{1}^{2}}\,R_{p_{1},q_{1}}(y_{1}),$$ or: $$y=y_{1}\sqrt{\frac{1-p_{1}^{2}y_{1}^{2}}{1-q_{1}^{2}y_{1}^{2}}}
\label{stremo}$$ Let us work inside the set where $1-p_{1}^{2}y_{1}^{2}>0$ and $
1-q_{1}^{2}y_{1}^{2}>0$. Converting the former differential in terms of the new variable $y_{1}$$$\mathrm{d}y=\frac{p_{1}^{2}q_{1}^{2}y_{1}^{4}-2p_{1}^{2}y_{1}^{2}+1}{\left(
1-q_{1}^{2}y_{1}^{2}\right) \sqrt{\left( 1-p_{1}^{2}y_{1}^{2}\right) \left(
1-q_{1}^{2}y_{1}^{2}\right) }}\,\mathrm{d}y_{1}$$ so that the integral is changed in: $$\int_{0}^{s(p,q)}\frac{p_{1}^{2}q_{1}^{2}y_{1}^{4}-2p_{1}^{2}y_{1}^{2}+1}{%
\sqrt{\left( 1-p_{1}^{2}y_{1}^{2}\right) \left( 1-q_{1}^{2}y_{1}^{2}\right)
\left( p^{2}p_{1}^{2}y_{1}^{4}-(p^{2}+q_{1}^{2})y_{1}^{2}+1\right) \left(
p_{1}^{2}q^{2}y_{1}^{4}-(q^{2}+q_{1}^{2})y_{1}^{2}+1\right) }}\,\,\mathrm{d}%
y_{1} .$$ Now, assuming $p_{1}$ and $q_{1}$ to come by the above iteration of arithmetic-geometric type, $$\begin{cases}
p_1=\dfrac{p+q}{2}\\
q_1=\sqrt{pq}
\end{cases}
\iff
\begin{cases}
p=p_1+\sqrt{p_1^2-q_1^2}\\
q=p_1-\sqrt{p_1^2-q_1^2}
\end{cases}$$ we succeed in simplifying $$\begin{split}
&\int_{0}^{s(p,q)}\frac{p_{1}^{2}q_{1}^{2}y_{1}^{4}-2p_{1}^{2}y_{1}^{2}+1}{%
\sqrt{\left( 1-p_{1}^{2}y_{1}^{2}\right) \left( 1-q_{1}^{2}y_{1}^{2}\right)
\left( p_{1}^{2}q_{1}^{2}y_{1}^{4}-2p_{1}^{2}y_{1}^{2}+1\right) ^{2}}}\,\,%
\mathrm{d}y_{1}=\\
&\int_{0}^{s(p,q)}\frac{\mathrm{d}y_{1}}{\sqrt{\left(
1-p_{1}^{2}y_{1}^{2}\right) \left( 1-q_{1}^{2}y_{1}^{2}\right) }}
\end{split}$$ where the polynomial $
p_{1}^{2}q_{1}^{2}y_{1}^{4}-2p_{1}^{2}y_{1}^{2}+1$ results to be positive for $%
y_{1}\in [0,s(p,q)].$ In such a way we obtained in integral form Lagrange’s differential identity . The upper integration limit $s(p,q)$ is obtained by putting $y=p$ in , providing $p_{1}$ and $q_{1}$ as a function of $p$ and $q$ (Arithmetic Geometric Mean) and solving to $y_{1}$: $$s(p,q)=\sqrt{\frac{2}{p(p+q)}}=\frac{1}{\sqrt{p_{1}\sqrt{p_{1}^{2}-q_{1}^{2}}%
+p_{1}^{2}}} .$$ The case of positive signs is analogous, and really more simple, not needing a signs discussion. Being $0<q<p$, taking $0<x<1/p$, it is true that $$\int_{0}^{x}\frac{\mathrm{d}y}{\sqrt{(1-p^{2}y^{2})(1-q^{2}y^{2})}}%
=\int_{0}^{s(x,p,q)}\frac{\mathrm{d}y_{1}}{\sqrt{%
(1-p_{1}^{2}y_{1}^{2})(1-q_{1}^{2}y_{1}^{2})}} \label{glei}$$ with $$s(x,p,q)=\frac{\sqrt{2}}{p+q}\,\sqrt{1+pqx^{2}-\sqrt{\left(
1-p^{2}x^{2}\right) \left( 1-q^{2}x^{2}\right) }}$$ and if $x=1/p$ then $s(1/p,p,q)=s(p,q)$. Notice that the expression for $s(p,q)$ is more simple than $s(x,p,q)$ being the value $y=1/p$ the absolute maximum attained by the function used by Lagrange in : $$y=y_{1}\sqrt{\frac{1-p_{1}^{2}y_{1}^{2}}{1-q_{1}^{2}y_{1}^{2}}}=y_{1}\sqrt{%
\frac{1-\frac{1}{4}y_{1}^{2}(p+q)^{2}}{1-pqy_{1}^{2}}}:=f(y_1).$$ Noting that $$\frac{q_{1}}{p_{1}}=\frac{2\sqrt{pq}}{p+q}=\frac{2\sqrt{k}}{1+k}:=\hat{k},\quad
\frac{p}{p_{1}}=\frac{2}{1+k}$$ putting $k:=q/p$ and making use of homothetics $z=p\,y,\,z_{1}=p_{1}\,y_{1}$ at both sides of , we get: $$\int_{0}^{px}\frac{\mathrm{d}z}{\sqrt{(1-z^{2})(1-k^{2}z^{2})}}=\frac{2}{1+k}%
\int_{0}^{p_{1}\,s(x,p,q)}\frac{\mathrm{d}z}{\sqrt{(1-z_{1}^{2})(1-\hat{k}
^{2}z_{1}^{2})}} \label{gleichung}$$ published for the first time on at page 7 of the first volume in his [*Traité*]{} and by Jacobi and Richelot named as *Legendre Gleichung*. This fact, never observed before, strengthens the idea one shall refer to a Lagrange-Legendre transformation” apart from the hyperbolic Landen theorem”. Let us see the relationship between the amplitudes in two consecutive stages of iteration. For that of the integral at left hand side of one finds: $\sin \varphi =px,$ and $$\sin \hat{\varphi}=p_{1}\,s(x,p,q)=\sqrt{\frac{1+pqx^{2}-\sqrt{\left(
1-p^{2}x^{2}\right) \left( 1-q^{2}x^{2}\right) }}{2}} \label{hat}$$ defines that of the integral at right hand side of in . The wanted link between $\varphi $ and $\hat{\varphi}$ is provided by:
The Lagrange’s AGM trasformation on a first kind elliptic integral $F(k,\varphi )$ changes its parameters $k,\varphi $, defining a new amplitude $\hat{\varphi}$ such that[^24]: $$\tan \varphi =\frac{\sin (2\hat{\varphi})}{k+\cos (2\hat{\varphi})} .
\label{landen}$$
It is well known that $\varphi $ and $\hat{\varphi}$ are ranged between 0 and $\pi /2$ and then $$\tan \varphi =\frac{\sin \varphi }{\sqrt{1-\sin ^{2}\varphi }}=\frac{px}{%
\sqrt{1-p^{2}x^{2}}}.$$ On the other side $$\frac{\sin (2\hat{\varphi})}{k+\cos (2\hat{\varphi})}=\frac{2\sin \hat{%
\varphi}\sqrt{1-\sin ^{2}\hat{\varphi}}}{k+1-2\sin ^{2}\hat{\varphi}}$$ Thesis follows by the fact that, using , reminding that $k=q/p$ one can write: $$\begin{split}
\frac{\sin (2\hat{\varphi})}{k+\cos (2\hat{\varphi})}&=\frac{p\sqrt{%
x^{2}\left( p^{2}+q^{2}-2p^{2}q^{2}x^{2}+2pq\sqrt{\left( 1-p^{2}x^{2}\right)
\left( 1-q^{2}x^{2}\right) }\right) }}{q-p^{2}qx^{2}+p\sqrt{\left(
1-p^{2}x^{2}\right) \left( 1-q^{2}x^{2}\right) }}\\
&=\frac{px}{\sqrt{
1-p^{2}x^{2}}}.
\end{split}$$
The well-known link between amplitudes has then be proved starting from the Lagrange’s arithmetic-geometric transformation.
Lagrange appreciated Landen’s work (see: *Théorie des fonctions analytiques*, 1797, about the *Residual Analysis,* 1758). On the subject of hyperbolic Landen theorem, in his letter to Condorcet[^25] of January 3$^{\rm rd}$, 1777, see [@Oplag], tome XIV, page 41, we can read:
> J’ai vu, dans le dernier volume des Transactions philosophiques, un théorème de M. Landen qui me parait bien singulier. Il réduit la rectification des arcs elliptiques à celle des arcs hyperboliques. Je n’ai pas encore eu le temps d’examiner s’il n’y a pas de paralogisme dans la démonstration.
Nevertheless, given that the elliptic arc was as a primary element for computing even more difficult fluents, the true meaning of the new is the reverse, namely the reduction of one of the most difficult (hyperbola) integrations, to some more simple objects, like the elliptic arcs. However Lagrange did not quote Landen’s articles on the hyperbola in his Turin memoir [@Lag], whilst Legendre did not cite Lagrange at all, but mentioned Landen, asking himself why Euler did not write anything about Landen whilst in his time (1751) he had written about Fagnano. He concludes that most likely Euler didn’t know Landen papers. Formula was unknown to Landen; where Legendre took it, we cannot say. Jacobi, as we will see in a next paper, stayed on the subject, providing a geometric interpretation. Furthermore, not Gauss, but Lagrange, author of that paper only on the elliptic integrals, established the Arithmetic-Geometric Mean using it for building a transformation of the basic values $p$ and $q$ which. Working on it one can advance the formula: $$\frac{2\sqrt{k}}{1+k}:=\hat{k}$$ for scaling modules; and that (\[landen\]) for amplitudes, both due to Legendre[^26], who published them in first volume of his [*Traitè*]{} submitted to the Academy in 1825 and issued in 1827.
Conclusions
===========
This paper’s aim is the hyperbola rectification, 1742-1827, with all the relevant problems of analytical, historical, geometrical nature. Our conclusions have been split up to give the due room separately to each of the founders of the theory, Maclaurin, Landen, Lagrange. The Legendre contribution will be the object of a next specific treatment.
Maclaurin {#maclaurin-1 .unnumbered}
---------
In *Fluxions*, n. 755, Maclaurin, moved by his interest in all Mathematical Physics and Calculus of his time, defined a research program concerning the classification of irrational fluents, then followed by D’Alembert too, *Recherches sur le calcul integral*, through a purely analytical approach. On the contrary Maclaurin performed their integration by means of arcs of conics and often with the help of geometrical arguments. Among these problems the *elastica* was absolutely crucial. Not being possible to integrate it, such a problem can be switched either in rectifying the lemniscate or in computing the hyperbolic excess, at which then he arrived from the elastica and strain analysis. His Apollonius’s tributary scheme for measuring a hyperbola’s length is analytically explained and discussed. Finally, his procedure for evaluating the hyperbolic excess is provided, adding to it our modern treatment and comparing the results.
Landen {#landen-1 .unnumbered}
------
We try to make known his famous theorem on hyperbola rectification whose original, rather obscure proofs (1775, 1780) are far from easy. After detailed explanations going back to Maclaurin’s previous fundamental analysis and so on, we add some geometrical interpretations to all the Landen processes by means of a continuous chain of geometry constructions inferred by his treatment. We show that he studied the same irrational fluent of Maclaurin, namely the hyperbolic excess, which for him did not stem from the elastica, but from the pendulum time equation. His synthetical topic on the limit hyperbolic excess has been analytically confirmed by us through the elliptic functions which will become a standard much later. Finally, Landen is proved to be completely irrelevant to the transformation bearing his name, and that his name applies only to the hyperbolic theorem.
Lagrange {#lagrange-1 .unnumbered}
--------
In the *Nouvelle méthode*, [@Lag] defines a differential identity stemming from the AGM, established by him in the same paper and mistakenly attributed to Gauss. Integrating his identity, we arrive at the well-known Legendre formula for a recursive computation of the first kind elliptic integral. Such a transformation was completely *unknown to Landen*, as one can understand from [@lan1], [@lan2] and [@lan3]: it was envisioned briefly but not developed by Lagrange, who was ahead of his times, but not very interested in elliptic integrals. The transformation was published in 1827 by Legendre, who applied it extensively throughout the first volume of his [*Traité* ]{}.
[165]{}
Almkvist, G. and Berndt, B. 1988. Gauss, Landen, Ramanujan, the Arithmetic-Geometric Mean, Ellipses, $\pi $ and the Ladies Diary. [*American Mathematical Monthly*]{} 95: 585-608. Armitage, J.V. and Eberlein, W.F. [*Elliptic functions.*]{} Cambridge: Cambridge University Press.
Bellacchi, G. 1894. [*Introduzione storica alla teoria delle funzioni ellittiche.*]{} Florence: Barbera. Berkeley, G. 1734. [*The Analyst, a discourse addressed to an Infidel Mathematician.*]{} London: J.Tonson. Boros, G. ans Moll, V. H. 2001. Landen transformations and the integration of rational functions. [*Mathematics of computation*]{} 71: 649-668. Borwein, J.W. and Borwein:, P.B. 1987. *Pi and the AGM.* New York: John Wiley and Sons, Inc. Byrd, P.F., Friedman, M.D. 1971. [*Handbook of Elliptic Integrals for Engineers and Scientists, second edition.*]{} New York: Springer. Cajori, F. 1919. *A history of the conceptions of limits and fluxions in Great Britain from Newton to Woodhouse* Chicago and London: The Open Court publishing Company. Cantor, M. 1908. [*Vorlesungen über Geschichte der Mathematik. Vol. 4.*]{} Leipzig: B.G. Teubner. Casorati, F. 1868. [*Teorica delle variabili complesse.*]{} Pavia: Tip. dei Fratelli Fusi. Cayley, A. 1882. Note on Landen’s theorem. [*Proceedings of the London Mathematical Society*]{} 13: 47-48. Cooke, R. 1994. Elliptic integrals and functions. In [*Companion encyclopedia of the history and philosophy of the mathematical sciences. Vol. 1.*]{} Ed. I. Grattan-Guinness. London: Routledge.
Cox, D. 1984. The arithmetic geometric mean of Gauss. [*Enseignement Mathématique*]{} 30: 274-330, 1984. D’Alembert, J.L.R. 1784. Recherches sur le calcul intégral. [*Histoire de l’Académie royale des sciences et des belles-lettres de Berlin*]{} 182-224
Durège. 1887. [*Theorie der elliptischen Funktionen. Fünfte Auflage.*]{} Leipizg: Teubner. Gauss, K.F. 1876. *Nachlass. Arithmetisch geometrisches Mittel, Werke, Bd. 3.* Göttingen: Königlichen Gesell. Wiss. Grabiner, J.V. 1997. Was Newton’s Calculus a Dead End? The continental influence of Maclaurin’s Treatise of fluxions. [*American Mathematical Monthly*]{} 104: 393-410. Gradshteyn I. S., Ryzhik I. M. 2000. [*Table of integrals, series, and products. 6.th edition.*]{} New York: Academic Press. Greenhill, A.G. 1959. [*The applications of elliptic functions.*]{} (Reprint) New York: Dover.
Gwinned Green, H. and Winter, H.J.J. 1944. John Landen FRS, 1719-1790, mathematician. [*Isis*]{} 35:6-10 Guicciardini, 1989. N. *The development of Newtonian Calculus in Britain, 1700-1800.* Cambridge: Cambridge University Press. Katz, V.J. 1998. [*History of Mathematics: An Introduction.*]{} New York: Addison-Wesley. Kline, M. 1972. [*Mathematical thought from ancient to modern times.*]{} New York: Oxford University Press.
Küpper, M.C. 1857. Transformation de Landen considerée geometriquement. [*Journal für die reine und angewandte Mathematik*]{} 55: 89-93. Lagrange, J. L. 1784-1785/1786. Sur une novelle méthode de calcul intégral pour les différentielles affectées d’un radical carré sous lequel la variable ne passe pas le quatriéme degré. [*Mémoires de l’Acadèmie royale des Sciences de Turin*]{} 7: 218-290. Lagrange, J.L. 1867-1892. [*Oeuvres de Lagrange, vol. I-XIV. publiées par les soins de J. A. Serret et G. Darboux.*]{} Paris: Gauthier-Villars.
Landen, J. 1772. A disquisition concerning certain fluents, which are assignable by the arcs of the conic sections; where are investigated some new and useful theorems for computing such fluents. [*Philosophical Transactions of the Royal Society of London*]{} 61: 298-309. Landen, J. 1775. An investigation of a general theorem for finding the length of any arc of any conic hyperbola, by means of two elliptic arcs, with some new and useful theorems deduced therefrom. [*Philosophical Transactions of the Royal Society of London*]{} 65: 283-289. Landen, J. 1780. Mathematical Memoirs respecting a variety of subjects, with an Appendix containing Tables of Theorems for the Calculation of Fluents, vol. I: 23-36. London: J. Nourse. Legendre, A.M. 1825 *Traité des fonctions elliptiques et des intégrales eulériennes. Tome premier.* Paris: Imprimerie de Huzard-Courcier. Legendre A.M. 1811 *Exercices de calcul intégral sur divers ordres de transcendantes et sur les quadratures.* Paris: Courcier, 1811. Legendre A.M. 1786. Mémoire sur les intégrations par arcs d’ellipse. [*Mémoires de l’Académie Royale des Sciences, pour l’année 1786*]{}: 616-643. Legendre A.M. 1787. Second mémoire sur les intégrations par arcs d’ellipse. Et sur la comparaison de ces arcs, [*Mémoires de l’Académie Royale des Sciences, pour l’année 1786*]{}: 644-683.
Loria, G.B. 1950. *Storia delle Matematiche.* Milan: Hoepli.
MacLaurin, C. 1801. [*A Treatise on Fluxions, vol. 2.*]{} (Reprint) London: William Baynes. Manna, D.N. and Moll, V.H. 2007 Landen survey. [*Probability, Geometry and Integrable Systems.*]{} 55: 287-319. Mittag-Leffler, G. 1923. An introduction to the theory of the elliptic functions. [*The Annals of Mathematics*]{} 24: 271-351. Moll V.H, Nowalsky J.L. and Solanilla L. 2002. The story of Landen, the hyperbola and the ellipse. [*Elemente der Mathematik*]{} 57: 19-25. Richelot, F.J. 1868. *Die Landen’sche Transformation in ihrer Anwendung auf die Entwicklung der elliptischen Funktionen.* Königsberg: Verlag von Hübner & Matz. Ritt, J.F. 1948. [*Integration in finite terms.*]{} New York: Columbia University Press. Sageng, E. 2005. 1742, Colin Maclaurin, A treatise on Fluxions. In *Landmark writings in Western Mathematics 1640-1940,* ed. I. Grattan-Guinness, 143-158. Amsterdam: Elsevier.
Sansone, G. 1955. [*Funzioni di variabile complessa.*]{} Padua: Cedam. F. Siacci, F. 1870. Sul teorema del conte di Fagnano. [*Bullettino di bibliografia e di storia delle scienze matematiche e fisiche pubblicato da B. Boncompagni*]{} 3: 1-26. Smadja, I. 2001. Des mèthodes d’intègration par arcs de sections coniques aux èchelles de modules. Legendre lecteur de Landen. [*Arch. Hist. Exact Sci.*]{} 65: 343-395. Toschi di Fagnano, G.C. 1761. Teorema da cui si deduce una nuova misura degli archi ellittici, iperbolici e cicloidali. [*Giornale de’Letterati d’Italia*]{} 26: 266-279. Volterra, V., Loria, G. and Gambioli, D. 1911-12. [*Opere matematiche del marchese Giulio Carlo de’ Toschi di Fagnano pubblicate sotto gli auspici della società italiana per il progresso delle scienze dai soci V. Volterra, G. Loria, D. Gambioli.*]{} Milan: Società editrice Dante Alighieri. Watson, G.N. 1933. The Marquis and the land-agent; a tale of the eighteenth century. [*The Mathematical Gazette*]{} 17: 5-17. Whittaker, E.T. and Watson, G.N. 1927. [*A Course of Modern Analysis.*]{} Cambridge: Cambridge University Press.
Giovanni Mingari Scarpello
via Negroli, 6 Milan, Italy
[email protected]
Daniele Ritelli
Dipartimento di Statistica University of Bologna
viale Filopanti, 5 Bologna, Italy
[email protected]
Aldo Scimone
Via C. Nigra, 30 Palermo, Italy
[email protected]
[^1]: The infidel mathematician” is believed to have been (perhaps) the astronomer E. Halley, responsible for financing in 1687 the print of *Principia* by Newton.
[^2]: A milestone theorem of the planar cubics theory (probably already known to Gua de Malves, 1740), is held in this Appendix: if a straight line meets 2 real inflection points of a cubic, it will cross it again in a third real inflection point.
[^3]: Notice that $r$ and $p$ are called *pedal coordinates* of the hyperbola with respect to S. The name is coming from the *pedal curve* of a curve with respect to a fixed (pedal) point, namely the locus of the points where the successive perpendiculars through that point cross the successive tangents to the curve.
[^4]: Maclaurin used $x$, but in order to avoid mistakes with the next sections, we prefer the capital letter $X$, whilst $x$ is kept for the abscissa.
[^5]: Really Maclaurin used $z$ but we changed to $\zeta $.
[^6]: Notice that due to a print error in the 1742 edition, and not amended in the 1801 one, the factor 2” in the above formula at *Fluxions*, page 245, row 6, is missing.
[^7]: The shortest way driving to the rectification of the lemniscate of equation $\rho^{2}=R^{2}\cos (2 \theta) $ is that assuming the polar anomaly as a variable: the elementary arc is $${\rm d}s=R\frac{{\rm d}\theta}{\sqrt{\cos(2\theta)}},$$ so that the one-quarter arclength is given by: $$s_{\frac{1}{4}}=R\int_{0}^{\frac{\pi }{4}}\frac{{\rm d}\theta }{\sqrt{1-2\sin
^{2}\theta }}=\frac{R}{\sqrt{2}}\,\boldsymbol{K}\left(
\frac{1}{\sqrt{2}}\right).$$ where $\boldsymbol{K}\left(\tfrac{1}{\sqrt{2}}\right)$ is the complete elliptic integral of first kind of modulus $k=\tfrac{1}{\sqrt{2}}$.
[^8]: Even though the limit concept was a matter the of the XIX century, the notation used by Maclaurin was quite clumsy: the symbol *lim* made its first appearance in the memoir *Exposition elementaire des principes des calculs superieurs* by S. Lhuilier, Berlin, 1786.
[^9]: Maclaurin strangely writes $N$ instead of $\pi/2$. The first to use $\pi$ definitely for the ratio of circumference to radius was William Jones (1675-1749) in his *Synopsis Palmariorum Matheseos*, 1706. Euler adopted the symbol in 1737 and since then it became of general use.
[^10]: Of course what above means that the required elliptic arc has been taken between abscissa zero and $x$ and is given by the definite integral: $$\int_{0}^{\frac{m-n}{m+n}x}\sqrt{\frac{(m+n)^{2}-t^{2}}{(m-n)^{2}-t^{2}}}{\rm d}t=(m+n)\,E\left(\arcsin\frac{x}{m+n},\frac{m-n}{m+n}\right)$$ where we used entry 3.171-14 page 306 of [@Grr].
[^11]: About the above formula, we read in [@Smadja] the comment:
> La réduction de l’arc hyperbolique à deux arcs elliptiques résulte en effet de la décomposition algébrique d’une expression rationnelle en une somme de deux expressions rationnelles plus simples, en sorte que la procédure algébrique opère sur les termes fluxionnels en mettant de côté ce que Landen nomme leur figurative sense”.
[^12]: We will come back later about them.
[^13]: Such a tangent is said to be pedal” because it is bound by the P-foot of the perpendicular drawn through a fixed “pedal point” to it. The pedal curve” of a given line is the locus of all the P-points when the tangent varies continuously along the profile of the line. The pedal curve of a rectangular hyperbola with the pedal point at its focus is a circle, but with pedal point at its center, one will find a lemniscate of Bernoulli.
[^14]: Apart from each possible intuition, the thing can easily be set analytically. First, one assumes a named $(x_{P},y_{P})$ point of the circle and writes its tangent there, which, by means of the circle equation, will be parametrized on the abscissa of such a point. The same for the hyperbola on the abscissa of its generic point $(x_{F},y_{F})$. Imposing that both slopes and intercept of two straight lines shall be the same, one gets a 2-equation, 8th degree non-linear algebraic system holding the required abscissas. Then we find four straight lines, but restricting to the hyperbola’s upper half branch $(x>0,y>0)$, there will only be one bitangent, say the straight line named *tan-ob* in \[f21\].
[^15]: In 1750 this article and almost all his writings entered his collected works entitled *Produzioni matematiche*. Finally in 1911-12 a modern complete edition [@Gambioli] was issued, with better figures, intelligible formulas, letters, and a detailed biography of Fagnano.
[^16]: The general lemniscate is known as a *Cassini oval*, 1680. A special kind of it, called *hyperbolic*, was considered by Jakob Bernoulli, and investigated by Fagnano since 1750, and by Euler, 1751,1752.
[^17]: About Küpper we know that he was author of the item [@kup], but we did not succeed in finding anything else about this German professor of the 19th century.
[^18]: Siacci’s confusion is due to the Legendre relevant papers, namely [@Leg4] and [@Leg3], inserted in succession in the same volume of the *Mémoires de l’Académie Royale.*
[^19]: Landen’s above fluxion in modern terms reads as $$\frac{1}{2}\frac{r h^{-1/2}x^{-1/2}}{\sqrt{2b r-b^{2}-2({r-b})x-x^{2}}}\,{\rm d}x.$$
[^20]: There is some notation discrepancy because Landen changed it very frequently, now calling $Q$ the arc of the first ellipse, $R$ that of the second, while their quadrantal arcs are named $E$ and $E''$ respectively. For a better reading we tried to keep, as far as possible, the same symbols through Maclaurin and all of Landen’s variants: furthermore we will use a lighter notation for the incomplete arcs of ellipse, writing $\eta_{1}$ and $\eta_{2}$ instead of $\eta^{(A_{1}E_{1})}$ and $\eta^{(A_{2}E_{2})}$ respectively.
[^21]: Lagrange appreciated Landen, even if he did not cite his work on the hyperbola in the Turin memoria. On the contrary Legendre did not cite Lagrange but Landen, asking himself why Euler never wrote anything on Landen whilst, in due time in 1751, had referred to Fagnano. His conclusion is that Euler probably ignored Landen’s papers.
[^22]: Later, concerning the role of Lagrange, we will guess a possible mathematical path for arriving at this transformation starting from Lagrange’s statements. Of course this will show the feasibility of such a path, but nobody could tell which route Legendre took actually.
[^23]: In 1799 or in 1800 Gauss wrote a paper (appeared in 1866) describing his many discoveries on the Arithmetic Geometric Mean on which he had started to work (aged 14) since 1791, as by himself confided to his friend Schumacher in a letter dated April 16$^{\rm th}$, 1816. On the third volume of his works [@Gauss] we can read four entries on the subject.*We are a bit astonished that almost all authors ignore the Lagrange’s priority,* assuming Arithmetic Geometric Mean as detected by Gauss. The only valuable exception to this wrong course is due to Almkvist and Berndt [@AB]. Unfortunately neither they eluded the other wrong trend concerning the so-called Landen transformation”, probably for not having had access to Landen’s 1775 paper.
[^24]: Notice that the relationship between the amplitudes ${\varphi}$ and $\hat{\varphi}$ as in the first of is the same thing as .
[^25]: Jean-Antoine-Nicolas de Caritat, marquis De Condorcet (1743-1794), a French mathematician and economist.
[^26]: For being more precise the amplitudes scaling (\[landen\]) is introduced and used by Legendre in the form: $$\sin(2\varphi _{1}-\varphi )=k\sin \varphi$$ equivalent to the previous one.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Several recently proposed techniques achieve latency reduction by trading it off for some amount of additional bandwidth usage. But how would one quantify whether the tradeoff is actually beneficial in a given system? We develop an economic cost vs. benefit analysis for answering this question. We use the analysis to derive a benchmark for wide-area client-server applications, and demonstrate how it can be applied to reason about a particular latency saving technique — redundant DNS requests.'
author:
- |
Ashish Vulimiri\
\
\
P. Brighten Godfrey\
\
\
Sri Varsha Gorge\
\
\
- |
Zitian Liu\
\
\
Scott Shenker\
\
bibliography:
- 'benchmark.bib'
title: 'A cost-benefit analysis of low latency via added utilization'
---
Introduction
============
Many techniques for improving latency in the Internet trade off some amount of extra bandwidth consumption for reduced latency. Examples include DNS prefetching [@ChromiumPrefetching], redundant [@Vulimiri2012; @Vulimiri2013] and hedged [@Dean2013] queries, and speculative TCP loss recovery mechanisms [@Flach2013]. But what is the true cost of the added overhead, and when is it outweighed by the latency reduction achieved? In this brief note, we use an economic cost vs. benefit analysis to study these questions. We consider the tradeoff between cost and benefit in a specific class of systems: wide-area client-server applications (such as web browsing, DNS queries, etc.) involving clients using consumer-level connectivity and service providers in the cloud. The framework we develop here serves as a baseline; it can be refined or extended for other systems.
Our framework allows for various combinations of incentives at the servers and the clients. In the common scenario where both servers and clients care exclusively about their own benefit, we show that any technique that saves more than $10$ ms of latency (in the mean or the tail, depending on the metric we are concerned with) for every kilobyte of extra traffic that it sends is useful, even with very pessimistic estimates for the additional cost induced at both clients and servers. This is a conservative bound assuming the most expensive cost estimates we found; the threshold can be orders of magnitude lower in many realistic scenarios, such as when clients use DSL instead of cellular connectivity.
We develop a framework for comparing the cost of and benefit from latency-saving techniques (§\[sec:framework\]); use this framework to derive a benchmark for wide-area client-server applications (§\[sec:analysis\]); and demonstrate how the benchmark can be applied in practice via a case study (§\[sec:case-study\]).
Framework {#sec:framework}
=========
Consider DNS prefetching, where web browsers pre-emptively initiate DNS lookups for links on a webpage to save latency if the user chooses to follow the link. Prefetching adds overhead both to the client, which potentially sends DNS requests for more links than the user actually follows, and to the DNS infrastructure, which needs to service these additional requests. The corresponding benefit is the latency reduction at the client when following a prefetched link, which also translates to an increase in expected ad revenue at the server [@brutlag09]. DNS prefetching affects several entities, including clients, servers, and network operators.
We account for the cost and benefit to all the stakeholders affected by any given latency-saving technique by comparing the following five quantitities:
- $\ell$ (ms/KB): the average latency savings achieved by the technique, normalized by the volume of extra traffic it adds
- $p_s$, $p_c$ (\$/KB): the average price of processing extra traffic at the servers and the clients
- $v_s$, $v_c$ (\$/ms): the average value from latency improvement to the servers and the clients
We denote increased utilization in units of data transfer volume and measure added cost at the server and the client. Note however that these calculated costs are a proxy for *all* the costs (not just bandwidth) incurred by *all* affected entities. For instance, network operator costs are accounted for via the bandwidth costs ISPs charge servers and clients, and CDN costs are accounted for via the usage fees paid by servers. One kilobyte of added client-side traffic in a web service has server-side costs including server utilization, energy, network operations staff, network usage fees, and so on. In essence, we amortize all these diverse costs over units of client- and server-side traffic.
From the perspective of a selfish client, any latency-saving technique is useful as long as the benefit it adds outweights the cost to the client: that is, $\ell \times v_c \ge p_c$, or in other words $$\ell \ge \frac{p_c}{v_c}$$
Similarly, a selfish server would need $$\ell \ge \frac{p_s}{v_s}$$
------------------------------------------------------------- ----- ----- --------------- ---- -------------- -----
Service plan
*Server-side plans*
Amazon web services: “Common Customer” web app 2 67 5 95
Amazon Route 53 (DNS) assuming 0.5KB/query 1 40 3 12
Amazon CloudFront: U.S., 1 GB/mo, 1 KB/object 91 2 03
Amazon EC2 and Microsoft Azure: bandwidth, Brazil 25 56
NearlyFreeSpeech.net: web hosting 25 56
Amazon CloudFront: U.S., 1 GB/mo, 10 KB/object 20 45
Amazon EC2 and Microsoft Azure: bandwidth, US 12 27
MaxCDN: based on “starter” plan overage fee 08 18
DreamHost: cloud storage, object delivery 075 17
*Client-side plans*
AT&T, low volume cell plan, based on overage fees 68 27 9 55
AT&T, high volume cell plan, based on overage fees 15 00 2 10
O$_{2}$ mobile broadband, based on 1GB$\to$2GB increment 8 02 1 12
AT&T DSL 20 028
------------------------------------------------------------- ----- ----- --------------- ---- -------------- -----
We will require that both conditions be satisfied in the analysis in §\[sec:analysis\] – that is, the benchmark we develop identifies latency-saving techniques that directly benefit both servers and clients. Other combinations are possible. For instance, a server might directly value both its own benefit as well as the improvement in user experience at the client, in which case we would need $\ell \ge \max \left\{ p_s / (v_s + v_c), p_c / v_c \right\}$. The analysis can be modified to account for whatever incentives are necessary in any given application scenario.
Analysis {#sec:analysis}
========
[ **Cost estimates.**]{} To estimate server-side cost $p_s$, we use a range of advertised rates for cloud services which implement usage-based pricing, listed in the second column of Table \[table:cost-benefit\]. The most expansive (and expensive) of these is the first line, based on an Amazon Web Services sample customer profile of a web application.[^1] The profile models a 3-tier auto-scalable web application, with a load balancer, two web servers, two app servers, a high-availability database server, 30 GB of storage, and other services, which utilizes $120$ GB/month of data transfer out of EC2 and $300$ GB/month out of CloudFront. The resulting amortized cost of \$$2.67$ effectively models the cost (per transferred GB) of an average operation in this system, including the cost of all utilized services.[^2] The other services listed in the table model the cost of more limited operations, such as DNS or bandwidth alone.
On the client side, we limit this investigation to clients in which the dominant cost of incrementally added utilization is due to network bandwidth. Table \[table:cost-benefit\] lists costs $p_c$ based on several types of connectivity. For these calculations, we assume a user who has paid for basic connectivity already, and calculate the cost of bandwidth from overage charges. Client-side bandwidth costs can be substantially higher than server-side total costs in extreme (cellular) cases but are comparable or cheaper with DSL connectivity.
Of course, there are scenarios which the above range of application costs does not model. For example, a cellular client whose battery is nearly empty may value energy more than bandwidth. But in a large class of situations, bandwidth is the most constrained resource on the client.
[ **Value estimates.**]{} The value of time $v$ is more difficult to calculate, at both the client and server.
For the server, direct value may come from obtaining revenue (ads, sales). We consider the case of Google. A study by Google indicated that users experiencing an artifical $400$ ms added delay on each search performed $0.74$% fewer searches after 4-6 weeks [@brutlag09]. Google’s revenue per search has been estimated[^3] at \$$0.0231$; therefore, we can estimate a savings of $400$ milliseconds on a single search generates, on average, an additional $\$0.0231\cdot0.0074$ in revenue, or \$$1.54$ per hour of reduced latency. As another estimate, a $500$ millisecond delay in the Bing search engine reduced revenue per user by $1.2$%, or $4.3$% with a $2$-second delay [@souders09velocity]. Using the latter (smaller) figure, combined with an estimated[^4] revenue per Bing search of $\$0.0314$, we have a \$$2.43$ per hour value. We use the more pessimistic Google value of \$$1.54$/hr in our calculations.
On the client side, value may be obtained from a better or faster human experience. Among all the components of our analysis, this value is the hardest to estimate: it may be highly application-specific, and may depend on mean or tail latency in ways best quantified by a human user study of quality of experience. But as a first approximation, we assume the value of time is simply the US average earnings of \$24.54 per hour in August 2014 [@avgHourlyWage], which implies $v_c \approx 6.82 \cdot 10^{-6} ~\$/\textrm{ms}$.
[ **Finding the threshold.**]{} We can now use our cost and value estimates to solve $\ell \geq p / v$ to obtain the break-even point, in terms of the necessary latency savings per kilobyte of additional traffic.
Table \[table:cost-benefit\] shows the break-even values of $\ell$ for various scenarios. For example, the table indicates that a server replicating DNS traffic would obtain greater return in ad revenue than the cost of increased utilization with any latency-saving technique that saves more than $3.12$ milliseconds per KB of traffic that it adds. The values are divided into four quadrants, one for each cost/benefit combination:
[ ]{}
[ ]{}
[ ]{}
$p_s / v_s$ (upper-left quadrant): break-even $\ell$ for a server making a selfish decision.
$p_c / v_c$ (lower-right): a client making a selfish decision.
$p_s / v_c$ (upper-right): a server that directly values a client’s quality of service.
$p_c / v_s$ (lower-left): a client that directly values the server’s benefit. This is unrealistic: a client would not typically value the server’s ad revenue yet ignore its own benefit.
Interestingly, excluding the last unrealistic scenario, both the server’s and the client’s worst-case break-even benefit occurs in a similar range of $6$-$10$ ms/KB; the client’s higher costs are roughly balanced by its greater benefit. This analysis suggests that a given technique may be cost-effective even in the most conservative cases as long as we can save more than $\approx$ 6-10 milliseconds (in the mean or tail, depending on the goal) for each kilobyte of added traffic. Note that this is the worst-case value: the threshold can be orders of magnitude lower in many realistic scenarios. For example, if clients use DSL (instead of cellular) connectivity and servers use an external web-host (instead of managing their own website on Amazon), the required latency savings threshold drops to 0.25 milliseconds per KB of added traffic.
Case study: Redundant DNS\
requests {#sec:case-study}
==========================
As an example, we now show how this benchmark can be applied to analyze a particular latency-saving technique, targeting DNS lookups: replicate DNS requests to one or more publicly accessible DNS servers in parallel, in addition to the default local ISP DNS server, and take the first reply that arrives [@Vulimiri2012; @Vulimiri2013].
In previous work [@Vulimiri2012; @Vulimiri2013] we showed that this technique can achieve a significant reduction in DNS response times. Replicating DNS requests to up to $10$ DNS servers in total we observed a $24 - 62\%$ lower request latency in the mean, median and tail than the unreplicated baseline. The absolute improvement in all the metrics we tested was between $23$ and $761$ ms per KB of extra traffic added — compared to the $10$ ms/KB benchmark we developed in §\[sec:analysis\], this suggests that replicating DNS requests to $\le 10$ DNS servers is always cost-effective when considering raw DNS performance.
But protocol-level performance does not allow us to quantify client-side benefit. Therefore, we now evaluate the *application-level* impact of replicating DNS requests, by quantifying the improvement in total web page load times when the technique is used. We tested two deployments, Google Chrome running on an Emulab node, and Mozilla Firefox running on a laptop connected to the Internet via a residential DSL connection. On both deployments we started with a list of $10$ DNS servers, the local ISP DNS server as well as $9$ publicly accessible DNS servers, and ran the following experiment: (1) Rank the list of DNS servers in order of their average DNS request latency. (2) Repeatedly pick a random website from Alexa.com’s top-1000 list [@alexa] and a random level of DNS replication $k \in [1, 10]$, and measure the time the browser takes to complete loading the website’s homepage when every DNS request during the page load is replicated to the first $k$ DNS servers in the ranked list. We set a $30$ second timeout and dropped all requests taking longer than $30$ seconds to complete (these were typically indicative of a failed script or a popup preventing the page from loading completely).
Figure \[fig:relative\] shows the percentage improvement in mean and 95th percentile page load times (compared to the unreplicated baseline) at various levels of replication. In both deployments, we obtain a 6-15% improvement in the two metrics, translating to an absolute improvement of 200-700 ms in the mean and 500-2300 ms in the 95th percentile.
Figure \[fig:ell-abs\] normalizes the observed improvement by the traffic overhead added and compares against the $10$ ms/KB benchmark from §\[sec:analysis\]. The results show that replicating DNS queries to $10$ or more servers would be a net positive, both in the mean and the tail, in all the scenarios we analyzed in §\[sec:analysis\].
Note that we observe diminishing returns: while the improvement generally increases with the level of redundancy, the *incremental* improvement from each additional DNS server added to the system keeps decreasing. At what point does adding servers cease to be cost-effective? Figure \[fig:ell-incr\] answers this question by comparing the incremental improvement from each additional server against the $10$ ms/KB benchmark from §\[sec:analysis\]. The results suggest that while replicating DNS requests to $2$ (perhaps $3$) DNS servers is better for mean latency than not replicating at all, $4$-way or higher replication is, at best, an economic net neutral. We believe higher levels of replication may be appropriate in the tail, but at this point the data we have are too noisy to permit a similar analysis of tail latency; we are working on repeating these experiments at larger scales.
Conclusion
==========
We proposed an analytical framework for evaluating techniques that improve latency by trading it off for additional bandwidth usage. The analysis suggests a simple benchmark for client-server deployments in which both clients and servers care solely about their own benefit: any technique that improves latency by more than $10$ ms for each KB of extra traffic it adds is economically a net positive, even with very pessimistic estimates of the added costs at both servers and clients. We showed how the analysis can be applied in practice by using it to identify the choices of parameters with which a particular latency-saving technique, DNS redundancy [@Vulimiri2013; @Vulimiri2012], would prove beneficial.
[^1]: <http://calculator.s3.amazonaws.com/calc5.html#key=a-simple-3-tier-web-app>
[^2]: This is likely pessimistic since it includes, for example, the cost of increased storage which would not scale linearly with an increase in service operations.
[^3]: Based on forecasts at <http://www.trefis.com>.
[^4]: <http://www.trefis.com/company?hm=MSFT.trefis&driver=idMSFT.0817#>
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'E. Hernández-Lemus,[@inst1][^1] K. Baca-López,[@inst1] R. Lemus,[@inst1] R. García-Herrera[@inst1]'
title: The role of master regulators in gene regulatory networks
---
Introduction: Transcriptional master regulators
===============================================
Phenotypic conditions in living cells are largely determined by the interplay of a multitude of molecules; in particular, genes and their protein products. The coordinated behavior of such a large number of players is often represented by means of a gene regulatory network (GRN). In a GRN, regulatory processes between genes, transcription factors and other molecular components are represented by nodes and links. One common way of inferring this gene regulatory networks is by probabilistic analysis of whole genome gene expression data [@ehlphysa; @aracne].\
Specific, context-dependent analysis of regulatory activity of particular cellular phenotypes (say tumor cells) may also be performed with the aid of transcriptional interaction networks. Commonly, such GRNs present a complex topology, often compliant with a scale-free hierarchic nature, in which a relatively small number of key players dominate the function and dynamics of the network. Some of these key players in GRNs are transcription factors often known as *master regulators* (MRs). MRs are deemed responsible for the control of the whole regulatory program for cells under the associated phenotype [@basso; @muna]. Master regulators may, indeed, act over rather generalistic cellular processes [@mtor], but also on specific cellular phenotypes [@muna; @pax5; @gcn4p].\
For instance, it is known that the mTOR molecule is active in concerting signals regulating control growth, metabolism, and longevity. Malfunction of mTOR complexes has been associated with developmental abnormalities, autoimmune diseases and cancer [@mtor]. The main role of mTOR seems to be the regulation of protein synthesis. Detailed mechanisms remain unknown, but ribosome profiling seems to point out to translational regulation and transcriptional activation activity. Due to the multiplicity of mTOR signaling interactions, this molecule acts as a master regulator on a variety of phenotypes. More specific master regulatory activity may be exemplified by cases such as the one of VASH1 that has been identified as a master-regulator of endothelial cell apoptosis [@muna]; PAX5 is known to be a master regulator of B-cell development also involved in neoplastic processes in leukemogenesis [@pax5] and the yeast protein Gcn4Pp (that contains a conserved domain cd12193 present in human JUN proto-oncogene) that is triggered by starvation and stress signals [@gcn4p] and is an MR in the phenotypic response to such stimuli. Due to the complex mechanisms behind transcriptional regulation in eukaryotes, identification of MRs is mostly based on the (inferred or empirical) relationships between them and their downstream RNA targets in the GRN.\
In brief, MRs are transcription factor genes that are located *upstream* in the genomic regulation programme, hence they possess a high hierarchy in the GRN. They are considered to be important players behind the presence of (some) amplification cascades in transcriptional regulatory networks, and it has been hypothesized that they may coordinate the dynamic transcriptional response and phenotype (in the case of eukaryotes) of the cells.\
As it may be evident, MRs may have a big impact on cancer-related phenotypes. This is so since under genome instability conditions, the uncontrolled synthesis of these molecules may give rise to large amplification of transcriptional cascades. In Ref. [@karolplos], the role that some molecules (in particular, MEF2C) may have in processes involving metabolic deregulation and MR activity at the onset of primary breast cancer was studied. The approach followed there involved the inference of GRNs centered in a number of molecules considered to be candidate MRs associated with the breast cancer phenotypes at early stages (i.e., primary tumors). As it can be seen there, MEF2C resulted a quite promissory molecule due to the large number of (probabilistically inferred) targets it possesses, but also due to the main biological processes spanned by its targets.\
A *what if?* scenario
---------------------
Now, let us resort for a moment to a hypothetical scenario: Imagine you have a eukaryotic cell with *deregulated metabolism* (e.g., large local free energy fluctuations) and a gene with transcription factor activity that has a low activation energy threshold (i.e., a relatively low absolute value for the free energy of formation). Now imagine that this gene is located (within the regulatory network) *close* to energy transduction pathways and that it also possesses a high hierarchy on the transcriptional regulatory network as well as a relatively large relaxation time.\
To put it more clearly, the fact that a gene has low activation energies means that the amount of energy needed to activate its cellular biosynthesis is minimum [@activa1; @activa2], thus making this molecule more prone to be produced by large free energy fluctuations. The probability to have such large energy fluctuations may be increased under abnormal metabolism conditions [@epste; @karolplos; @tenant] that may enable events leading to transcriptional cascades.\
In such scenario, large local free energy fluctuations may *randomly* activate the transcription of such a gene that in turn may be able to activate long ranged transcriptional cascading before decaying, thus affecting to a large degree the whole transcriptional regulation programme of such cell. Indeed, such a gene may be acting as a transcriptional master regulator over that specific cell condition (phenotype).
MEF2C as a master regulator
===========================
In Ref. [@karolplos], we discussed the evidence that may point out to the MEF2C molecule as a candidate master regulator for the transcriptional regulator of human cells under the primary breast cancer phenotype. Regarding this molecule, we know the following: MEF2C is a transcription factor gene located (in humans) in 5q14.3 on the minus strand. This gene is 200,723 bp long and it encodes a 473-aminoacid protein weighting 51.221 kDa. MEF2C is a member of the Mef2 family that by means of controlling gene expression (MEF2 molecules are commonly acting as activator transcription factors) is able to regulate cellular differentiation and development [@mef2]. MEF2 members are highly versatile regulators since they contain both MEF2 and MADS-box DNA binding domains (see Fig. \[mef2cst\]). The MADS-box serves as the minimal DNA-binding domain, however an adjacent 29-amino acid extension called the Mef2 domain is required for high affinity DNA-binding and dimerization, hence conferring a combinatorial DNA binding repertoire through a number of transcription factor recognition marks [@meff].\
![Poisson-Boltzmann visualization of the MEF2C transcription factor protein showing the action of both, mef2 and MADS-box DNA-binding domains.[]{data-label="mef2cst"}](MEF2C.png){width="0.8\columnwidth"}
It is also known that the MEF2C protein interacts with MAPK7 (involved in proliferation and differentiation signaling) [@mapk7], EP300 (a transcription factor that regulates cell growth and cellular division) [@ep300], TEAD1 (an enhancer TF that co-regulates transcription with MEF2C), as well as with a number of histone deacetylases, most notably HDAC4, HDAC7, and HDAC9 [@hdac; @hdac2]. These protein-protein interactions, mostly with other transcription factors, enhancers or epigenomic regulators joined with their inherent binding-site transcriptional activity made MEF2C a quite functional and adaptable MR.\
Aside from this, a non-equilibrium thermodynamics analysis of the coupling between transcriptional regulation processes and metabolic de-regulation in breast cancer cells has led to some further evidence pointing out to MEF2C as an MR that may be playing an important role in carcinogenic processes. This thermodynamic evidence has been supplemented with information given by probabilistically-inferred gene regulatory networks centered around genes coordinating the coupling of transcriptional control and metabolism. The GRN was inferred by mutual information calculations [@itgrn; @casm] on a database of 1191 whole genome gene expression experiments in biopsy captured tissue from primary breast cancer patients and healthy controls [@karolplos].\
To further express that this behavior is related to the coupling between transcriptional regulation control and energy transduction pathways, let us resort to Fig. \[metab\]. Figure \[metab\] shows a gene ontology network containing the biological processes statistically enriched in a list of MEF2C regulated genes, differentially expressed in a 1191-sample database of whole genome gene expression experiments curated in our group [@karolplos]. In this figure, we may see that statistically enriched biological processes are shown as color-coded (white to red) according to a p-value calculated from a hypergeometric urn model test and corrected via the false discovery rate (FDR) measure as it is explained elsewhere [@karolplos]. We can see that the two major families of biological processes enriched are precisely those related with energy-release metabolic pathways and transcriptional regulation.
![Hierarchical network displaying statistical enriched Gene Ontology Biological Process entries related with MEF2C cascading, we may see that this evidence supports the hypothesis made in Ref. [@karolplos] regarding a non-equilibrium coupling between metabolism and transcriptional regulation.[]{data-label="metab"}](GOnet.png){width="1.0\columnwidth"}
### Transcription factor binding site analysis
In order to further validate the findings given by non-equilibrium thermodynamics and probabilistic regulatory networks, a computational analysis of databases for DNA transcription factor binding sites (TFBS) was performed. This study included a systematic TFBS analysis for MEF2C transcriptional influence by applying an algorithm (MotEvo) [@motevo] that incorporated -via Bayesian optimization- information additional to the sequence (physicochemical and electrostatic features, motif conservation and phylogeny, ChIP experiments, DeepCAGE sequencing, etc.). Such analysis was performed with a stringent statistical significance level (Response values $>$ 1.5 corresponding approx. to p\* $<$ 0.001) and showed that genome-wide MEF2C is able to regulate 200 genes directly.\
The results of the TFBS analysis showed that the set of MEF2C targets includes a number of genes that participate in oncogenic processes such as MRAS, IGFBP3, CTNND1, FOXN3, FOXP4, HGMA2, MMP19, CORO1C, JAG1, ASXL1, HSPB1, MB, RBL2, ZIC2, NR2F6, BCL-2, CBX7, DNM2, MAFA, LGALS3BP, among others. Additionally, a number of MEF2C targets are in turn TFs, which enlarges the range of transcriptional influence of MEF2C (and help it to become an MR).\
In fact, second order transcriptional interactions increase the range of influence of MEF2C to a GRN composed of 1896 genes and 2156 regulatory interactions (see Fig. \[mef2cgrn\]) that was able to further increase to 5852 genes and 18801 interactions up to third order.
![Gene regulatory network including up to second order transcriptional interactions in MEF2C targets. The network is composed of 1896 genes and 2156 interactions.[]{data-label="mef2cgrn"}](MEF2CGRN2.png){width="1.0\columnwidth"}
Dynamics of master regulator activity
=====================================
Cells sometimes present bursts or pulses of activity in their gene expression dynamic patterns. Bursting may result from a series of stochastic biochemical events and may be a source of large phenotypic heterogeneity of cell behavior and thus on cellular conditions and disease. Noise in transcriptional regulatory activity arises not only as a consequence of randomness of biochemical processes at the molecular level due to low molecule counts, it also may arise from thermodynamic fluctuations in cellular components and system level phenomena due to cooperativity. Different levels of gene regulation may be strongly coupled. When all these elements are present, we say that the cells are undergoing transcriptional bursts (TBs). Under such dynamic scenario, protein production occurs in pulses, each due to a single promoter or transcription factor binding event. It is in these instances that the phenomena can be related to the presence of master regulators in the transcriptional networks.
### Bursting and synchronization
The non-linear behavior of GRN interactions can be better understood in the light of periodic or quasi-periodic expression levels for certain groups of genes. By means of *Power Spectral Density* (PSD) calculations we may gain greater insight in such dynamic behavior. Power spectral density is useful to describe the evolution of the variance that in turn provides us with greater insight on the correlation structure of the underlying regulatory processes. The power spectral density of stochastic quasi-stationary processes can be estimated when considering non-linear time series analysis as follows.
Let us consider $\Gamma$ as a series containing a time course for intensity levels of gene expresion of a single gene, then the associated power spectral density, $I(\omega)$ is given by:
$$\label{psd2} I(\omega) = \frac{1}{N}\left|\sum_{t}^N \Gamma(t) \; \exp(-i \; \omega t) \right|^{2}; \; \; \, \omega \in [0, \pi]$$
Periodic behavior could be detected in a *linear* model for $\Gamma$:
$$\label{gammaper} \Gamma(t) = \beta \; \cos(\omega t + \phi) + \epsilon_i$$
$\beta$ being a positive constant (amplitude), $\omega \in [0, \pi]$, $\phi$ is a uniformly distributed *phase shift* ($\phi \in (-\pi, \pi]$) and {$\epsilon_i$} is a sequence of uncorrelated random variables with mean $0$ and variance $\sigma^2$ independent of $\phi$ (i.e. a Gaussian noise). Under this model, periodic behavior could be traced-off by means of looking at *significant* peaks in the power spectral density, either within an $\omega$-continuous process or more commonly with $\omega$ taking discrete values $\frac{2\pi k}{N}; \; \; k=0,1,2 \dots , [\frac{N}{2}]$, each of these discrete values is known as a Fourier frequency.\
If a time course $\Gamma$ has hidden periodic components, say with a given frequency $\omega^\star$, then the power spectral density will show a peak at $\omega^\star$. If, on the other hand, $\Gamma$ is a random, aperiodic signal, then the $I(\omega)$ Vs $\omega$ plot will be a (noisy) straight line, mapping to $\beta = 0$ in the linear model as given by Eq. \[gammaper\]. Then we may test the null hypothesis $\beta = 0$ versus the data in order to determine *significance* . An early result from Fisher applies also to finite time series, the so called g-statistic [@fisher]:
$$\label{fishg} g = \frac{\max_k \; I(\omega_k)}{\sum_{k=1}^{N/2} \, I(\omega_k)}$$
Values of $g$ larger than *expected* lead to the rejection of the null hypothesis (i.e., random processes). The exact $g$-distribution is given by:
$$\begin{aligned}
\label{gege}
P(g > x) &=& \alpha (1-x)^{\alpha-1} - \frac{\alpha (\alpha-1)}{2} (1-2x)^{\alpha-1} \nonumber \\
&+& \dots + (-1)^r \frac{\alpha !}{r! (\alpha -r)!} (1-rx)^{\alpha-1}\end{aligned}$$
$\alpha = N/2$ and $r$ is the largest integer less than $1/x$. So if the observed value of $g$ is $g^\star$, then there is a p-value $P(g > g^\star)$ to evaluate the null hypothesis.\
In Fig. \[burst\], we can see the results of the application of Fisher’s significance analysis (red line) to the power spectral density profiles of two genes that are regulated by MEF2C. We can see that both genes present some significant peaks in the power spectral density profiles, which means that some quasi-periodicities are present (i.e., some frequency bands are enriched). Further analyses have shown that the peaks for different MEF2C targets are indeed heavily correlated in some instances. This is related to a mechanism of functional synchronization. For a more detailed description of such analysis, please see Ref. [@rmfchaos].
{width="90.00000%"}
Conclusions
===========
As a consequence of the arguments just exposed here, we have been able to reach some conclusions about the role that master regulators (in this case MEF2C) may be playing in the function and dynamics of gene regulatory networks for particular phenotypes (in this case potentially related to the onset of primary breast cancer).\
In the first place, we may mention that MEF2C, as a transcriptional master regulator, associated with tumor phenotypes has a number of important physico-chemical features: such as having a low activation energy, and long decay times. MEF2C also presents TFBSs of two quite general classes (MADS and MEf2) which makes it a highly versatile transcription factor molecule. Following stringent TFBS analyses, MEF2C is potentially involved in the regulation of up to 200 genes directly, about 2000 at second order, and almost 6000 at third order (almost one fourth of the entire human genome) and also a number of its target genes are, in turn, transcription factors some of them with global activity. However, we must stress that MEF2C is not the only molecule responsible for the regulatory programme of such molecules. To what extent its coordinated action is able to induce the phenotype *in vivo* is still something to be determined by further experimental data.\
In relation to the dynamics of biological processes induced by MEF2C cascading, we have seen that MEF2C targets present ‘stochastic bursting’ and such bursts in gene expression activity are synchronized and long range correlated to a high degree (i.e., Almost 1/f correlations). Dynamic synchronization and long range correlation appear to be functional biological phenomena. However, ad hoc experimental testing is still in design.\
With regards to the biological implications of such findings (especially in the context of cancer biology), we have observed that MEF2C may be associated with tumor phenotypes (at least in primary breast cancer, according to our results). This is so because MEF2C has a number of direct targets (and also many of the indirect ones) which are molecules with known oncogenic activity. The patterns of gene expression of MEF2C targets (even for genes that are not differentially expressed) are able to induce the phenotype consistently. Also, statistical enrichment analyses, both for biological processes and biochemical pathways, showed significant hits associated with cancer related categories.\
After this, all that one has to say is that a lot of work is still to be done to understand the complex mechanisms behind the Master Regulatory control of phenotypes (especially in disease-related scenarios). Such investigations need to be multidisciplinary in nature and must be anchored in solid mathematical foundations compliant with the tenets of the theory of complex systems, while at the same time must be heavily relying on solid biological knowledge. Comprehensive integrative analyses under the systems biology paradigm will surely hence be a must at the center of discussion on these matters.
The authors gratefully acknowledge support by grant 179431/2012 (CONACYT), as well as federal funding from the National Institute of Genomic Medicine (México).
[99]{}
E Hernández-Lemus, D Velázquez-Fernández, J K Estrada-Gil, I Silva-Zolezzi, M F Herrera-Hernández, G Jiménez-Sánchez, [*Information theoretical methods to deconvolute genetic regulatory networks applied to thyroid neoplasms*]{}, Physica A [**388**]{}, 5057 (2009).
A A Margolin, I Nemenman, K Basso, C Wiggins, G Stolovitzky, A Califano, [*ARACNe: An algorithm for the reconstruction of gene regulatory networks in a mammalian cellular context*]{}, BMC Bioinformatics [**7**]{} (Suppl I), S7 (2006).
K Basso, A A Margolin, G Stolovitzky, U Klein, R Dalla-Favera, A Califano, [*Reverse engineering of regulatory networks in human B cells*]{}, Nat. Genet. [**37**]{} 382 (2005).
M Affara, D Sanders, H Araki, Y Tamada, B J Dunmore, S Humphreys, S Imoto, C Savoie, S Miyano, S Kuhara, D Jeffries, C Print, D S Charnock-Jones, [*Vasohibin-1 is identified as a master-regulator of endothelial cell apoptosis using gene network analysis*]{}, BMC Genomics [**14**]{}, 23 (2013).
R Hosking, [*mTOR: The master regulator*]{}, Cell [**149**]{}, 955 (2012).
J. Medvedovic, A Ebert, H Tagoh, M Busslinger, [*Pax5: A master regulator of b cell development and leukemogenesis*]{}, Adv. Immunol. [**111**]{}, 179 (2011).
A G Hinnebusch, K Natarajan, [*Gcn4p, a master regulator of gene expression, is controlled at multiple levels by diverse signals of starvation and stress*]{}, Eukaryot. Cell. [**1**]{}, 22 (2002).
S J Maerkl, S R Quake, [*A systems approach to measuring the binding energy landscapes of transcription factors*]{}, Science [**315**]{}, 233 (2007).
M Sawadogo, R G Roeder, [*Energy requirement for specific transcription initiation by the human RNA polymerase II system*]{}, J. Biol. Chem. [**259**]{}, 5321 (1984).
T Epstein, L Xu, R J Gillies, R A Gatenby, [*Separation of metabolic supply and demand: Aerobic glycolysis as a normal physiological response to fluctuating energetic demands in the membrane*]{}, Cancer Metabolism [**2**]{}, 7 (2014).
K Baca-López, A Hidalgo-Miranda, M Mayorga, N Gutiérrez-Nájera, E Hernández-Lemus, [*The role of master regulators in the metabolic/transcriptional coupling in breast carcinomas*]{}, PLoS ONE [**7**]{}, e42678 (2012).
D A Tennant, R V Durn, E Gottlieb, [*Targeting metabolic transformation for cancer therapy*]{}, Nat. Rev. Cancer [**10**]{}, 267 (2010).
M J Potthoff, E N Olson, [*MEF2: A central regulator of diverse developmental programs*]{}, Development [**134**]{}, 4131 (2007).
J D Molkentin, E N Olson, [*Combinatorial control of muscle development by basic helix-loop-helix and MADS-box transcription factors*]{}, P. Natl. Acad. Sci. USA [**93**]{}, 9366 (1996).
C C Yang, O I Ornatsky, J C McDermott, T F Cruz, C A Prody, [*Interaction of myocyte enhancer factor 2 (MEF2) with a mitogen-activated protein kinase, ERK5/BMK1*]{}, Nucleic Acids Res. [**26**]{}, 4771 (1998).
V Sartorelli, J Huang, Y Hamamori, L Kedes, [*Molecular mechanisms of myogenic coactivation by p300: Direct interaction with the activation domain of MyoD and with the MADS box of MEF2C*]{}, Mol. Cell. Biol. [**17**]{}, 1010 (1997).
A H Wang, N R Bertos, M Vezmar, N Pelletier, M Crosato, H H Heng, J Th’ng, J Han, X J Yang, [*DAC4, a human histone deacetylase related to yeast HDA1, is a transcriptional corepressor*]{}, Mol. Cell. Biol. [**19**]{}, 7816 (1999).
A H Wang, X J Yang, [*Histone deacetylase 4 possesses intrinsic nuclear import and export signals*]{}, Mol. Cell. Biol. [**21**]{}, 5992 (2001).
E Hernández-Lemus, C Rangel-Escareño, [*The role of information theory in gene regulatory network inference*]{}, In: Information theory: New research, Eds. P Deloumeaux, J D Gorzalka, Mathematics Research Developments Series, Nova Publishing (2011).
E Hernández-Lemus, J M Siqueiros-García, [*Information theoretical methods for complex network structure reconstruction*]{}, Complex Adap. Syst. Mod. [**1**]{}, 8 (2013).
P Arnold, I Erb, M Pachkov, N Molina, E van Nimwegen, [*MotEvo: Integrated Bayesian probabilistic methods for inferring regulatory sites and motifs on multiple alignments of DNA sequences*]{}, Bioinformatics [**28**]{}, 487 (2012).
R A Fisher, [*Test of significance in harmonic analysis*]{}, P. Roy. Soc. A [**125**]{}, 54 (1929).
E Hernández-Lemus, K Baca-López, [*Bursting and synchronization in gene regulatory dynamics*]{}, Rev. Mex. Fis. [**S 58**]{}, 63 (2012).
[^1]: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We give an elementary proof of a celebrated theorem of Cappell, Lee and Miller which relates the Maslov index of a pair of paths of Lagrangian subspaces to the spectral flow of an associated path of selfadjoint first-order operators. We particularly pay attention to the continuity of the latter path of operators, where we consider the gap-metric on the set of all closed operators on a Hilbert space. Finally, we obtain from Cappell, Lee and Miller’s theorem a spectral flow formula for linear Hamiltonian systems which generalises a recent result of Hu and Portaluri.'
author:
- 'Marek Izydorek, Joanna Janczewska and Nils Waterstraat'
title: 'The Maslov Index and the Spectral Flow - revisited'
---
Introduction
============
Let $\langle\cdot,\cdot\rangle$ be the Euclidean scalar product on $\mathbb{R}^{2n}$ and $\omega_0(\cdot,\cdot)=\langle J\cdot,\cdot\rangle$ the standard symplectic form, where
$$\begin{aligned}
\label{J}
J=\begin{pmatrix}
0&-I_n\\
I_n&0
\end{pmatrix}\end{aligned}$$
and $I_n$ denotes the identity matrix. Let us recall that an $n$-dimensional subspace $L\subset\mathbb{R}^{2n}$ is called Lagrangian if the restriction of $\omega_0$ to $L\times L$ vanishes. The set $\Lambda(n)$ of all Lagrangian subspaces in $\mathbb{R}^{2n}$ is called the *Lagrangian Grassmannian*. It can be regarded as a submanifold of the Grassmannian $G_n(\mathbb{R}^{2n})$ and so it has a canonical topology. In what follows, we denote by $I$ the unit interval $[0,1]$. The Maslov index $\mu_{Mas}(\gamma_1,\gamma_2)$ assigns to any pair of paths $\gamma_1, \gamma_2:I\rightarrow\Lambda(n)$ an integer which, roughly speaking, is the total number of non-trivial intersections of the Lagrangian spaces $\gamma_1(\lambda)$ and $\gamma_2(\lambda)$ whilst the parameter $\lambda$ travels along the interval $I$. There are several different approaches to the Maslov index and here we just want to mention [@Maslov], [@Bott], [@Duistermaat], [@Sternberg], [@RobbinMaslov] and [@Zehnder], which is far from being exhaustive. Cappell, Lee and Miller introduced in [@Cappell] four different ways to define the Maslov index and showed that they are all equivalent. They first construct the Maslov index geometrically by using a stratification of $\Lambda(n)$ and intersection theory from differential topology following [@Sternberg]. Their approach also yields a uniqueness theorem for the Maslov index characterising this invariant uniquely by six axioms. The uniqueness theorem is then used to show that the Maslov index can alternatively be defined by determinant line bundles, $\eta$-invariants and the spectral flow, respectively.\
In this paper we focus on the latter invariant and aim to give a more elementary proof of the equality of the Maslov index and the spectral flow of a path of operators as introduced by Cappell, Lee and Miller in [@Cappell]. Let us first recall that the spectral flow is a homotopy invariant for paths of selfadjoint Fredholm operators that was invented by Atiyah, Patodi and Singer in [@APS], and since then has been used in various different settings (see e.g. [@Fredholm §5.2]). The spectrum of a selfadjoint Fredholm operator consists only of eigenvalues of finite multiplicity in a neighbourhood of $0\in\mathbb{R}$ and, roughly speaking, the spectral flow of a path of such operators is the net number of eigenvalues crossing $0$ whilst the parameter of the path travels along the interval.\
Let us now consider for a pair of paths $(\gamma_1,\gamma_2)$ in $\Lambda(n)$ the differential operators
$$\begin{aligned}
\label{ops-def}
\mathcal{A}_\lambda:\mathcal{D}(\mathcal{A}_\lambda)\subset L^2(I,\mathbb{R}^{2n})\rightarrow L^2(I,\mathbb{R}^{2n}),\quad (\mathcal{A}_\lambda u)(t)=Ju'(t),\end{aligned}$$
where
$$\begin{aligned}
\label{ops-def-domain}
\mathcal{D}(\mathcal{A}_\lambda)=\{u\in H^1(I,\mathbb{R}^{2n}):\, u(0)\in\gamma_1(\lambda), u(1)\in\gamma_2(\lambda)\}.\end{aligned}$$
By an elementary computation, $\mathcal{A}_\lambda$ is symmetric, and it is also not difficult to see that it actually is a selfadjoint Fredholm operator. Note that the kernel of $\mathcal{A}_\lambda$ is isomorphic to $\gamma_1(\lambda)\cap\gamma_2(\lambda)$, which suggests that the spectral flow of the path $\mathcal{A}=\{\mathcal{A}_\lambda\}_{\lambda\in I}$ is related to the Maslov index of the pair $(\gamma_1,\gamma_2)$. As we already mentioned above, their equality is one of the main achievements of [@Cappell]. However, before we formulate this as a theorem, we want to highlight a further issue related to this problem.\
Above, we have spoken about paths of differential operators and so tacitly assumed continuity. Note that the family has the non-constant domains and so continuity is a non-trivial problem. There are different metrics on spaces of unbounded selfadjoint Fredholm operators on a Hilbert space $H$ and we recommend [@Lesch] for an exhaustive discussion (see also [@Wahl]). A classical approach is to transform unbounded selfadjoint operators $T$ by functional calculus to the bounded selfadjoint operators
$$\begin{aligned}
\label{Riesz}
(I_H+T^2)^{-\frac{1}{2}}\in\mathcal{L}(H),\end{aligned}$$
and to use the operator norm on $\mathcal{L}(H)$ for introducing a distance between unbounded operators. Actually, Atiyah, Patodi and Singer defined the spectral flow in [@APS] for bounded selfadjoint Fredholm operators and applied it to paths of differential operators by using . However, checking continuity along these lines is tedious, if possible at all (see e.g. [@Nicolaescu]), and it seems that the continuity of families of unbounded operators has sometimes been ignored in the literature.\
Every (generally unbounded) selfadjoint operator on a Hilbert space is closed, and there is a canonical metric on the set of all closed operators which is called the gap-metric (see §IV.2 in Kato’s monograph [@Kato]). It was shown in [@Nicolaescu] (see also [@Lesch Prop. 2.2]) that every path of selfadjoint Fredholm operators that is mapped to a continuous path of bounded operators under is also continuous with respect to the gap-metric. Finally, Booss-Bavnbek, Lesch and Phillips constructed in [@UnbSpecFlow] the spectral flow for paths of selfadjoint Fredholm operators in this more general setting. The main result of this paper now reads as follows (see [@Cappell Thm. 0.4]).
\[main\] If $(\gamma_1,\gamma_2)$ is a pair of paths in $\Lambda(n)$, then the family of differential operators is continuous with respect to the gap-metric and
$$\operatorname{sf}(\mathcal{A})=\mu_{Mas}(\gamma_1,\gamma_2).$$
Let us make a few comments on our proof. Firstly, we want to emphasise that we prove the gap-continuity of the family from first principles just by elementary estimates and standard facts about orthogonal projections that can all be found in the monograph [@Kato]. Secondly, our proof of the spectral flow formula in Theorem \[main\] is surprisingly simple. We assume at first that
$$\begin{aligned}
\label{introduction-transversal}
\gamma_1(0)\cap\gamma_2(0)=\gamma_1(1)\cap\gamma_2(1)=\{0\}\end{aligned}$$
and show that the Maslov index can be characterised in this case by three axioms. This uniqueness theorem needs nothing else than the elementary properties of the Maslov index and the fact that the fundamental group of $\Lambda(n)$ is infinitely cyclic, which was known already from Arnold’s classical paper [@Maslov]. Two of our axioms are trivially satisfied for the spectral flow of , and the remaining one only requires the computation of the spectra of two simple examples of differential operators as in . The general case when is not assumed, can easily be obtained from the previous case by a simple conjugation by a path of invertible operators. After a brief recapitulation of the Maslov index in Section \[section-Maslov\], and the gap-metric and spectral flow in Section \[section-spectralflow\], we explain all this in detail in Section \[section-proof\] where we prove Theorem \[main\]. Throughout the paper, we aim our presentation to be rather self-contained, and we will just use some well-known facts from Kato [@Kato].\
Finally, we review a recent spectral flow formula for linear Hamiltonian systems by Hu and Portaluri from [@Hu], which they call *a new index theory on bounded domains*. Firstly, we note that the considered families of Hamiltonian systems are continuous with respect to the gap-metric, which follows easily from our approach to Cappell, Lee and Miller’s Theorem. Secondly, we obtain a spectral flow formula in this setting by a conjugation from Cappell, Lee and Miller, and we explain that our result actually is a generalisation of Hu and Portaluri’s Theorem.
Maslov Index and Spectral Flow - a brief recap
==============================================
The Maslov Index {#section-Maslov}
----------------
The aim of this section is to briefly recall the definition of the Maslov index, where we follow [@PiccioneBook].\
Let $\operatorname{Sp}(2n,\mathbb{R})$ denote the group of symplectic matrices on $\mathbb{R}^{2n}$, i.e., those $A\in M(2n,\mathbb{R})$ satisfying $A^TJA=J$ or, alternatively, which preserve $\omega_0$. If we identify $\mathbb{R}^{2n}$ with $\mathbb{C}^n$ by $(x_1,\ldots,x_{2n})\mapsto (x_1,\ldots,x_n)+i(x_{n+1},\ldots, x_{2n})$ then the standard hermitian scalar product on $\mathbb{C}^n$ is
$$\langle x,y\rangle_{\mathbb{C}}=\langle x,y\rangle-i\omega_0(x,y).$$ Hence each unitary matrix $U\in U(n)$ preserves $\omega_0$ and so we can regard $U(n)$ as a subset of $\operatorname{Sp}(2n,\mathbb{R})$. Also, the orthogonal matrices $O(n)$ can be seen as a subgroup of $U(n)$ by complexification. Then $O(n)$ consists exactly of those $A\in U(n)$ which leave $\mathbb{R}^n\times\{0\}$ invariant.\
Obviously, $AL\in\Lambda(n)$ if $L\in\Lambda(n)$ and $A\in\operatorname{Sp}(2n,\mathbb{R})$, and it can be shown that the restriction of this action to $U(n)\times \Lambda(n)\rightarrow\Lambda(n)$ is transitive. As the stabiliser subgroup of $\mathbb{R}^n\times\{0\}\in\Lambda(n)$ is $O(n)$, we see that there is a diffeomorphism
$$\begin{aligned}
\label{diffeoU}
U(n)/O(n)\simeq\Lambda(n),\quad A\mapsto A(\mathbb{R}^n\times\{0\}).\end{aligned}$$
Let us now consider the map $d:U(n)\rightarrow S^1$, $d(A)=\det^2(A)$, which descends to the quotient by
$$\overline{d}:U(n)/O(n)\rightarrow S^1,\quad A\cdot O(n)\mapsto \det{^2}(A).$$ Note that $$\ker(d)/O(n)\hookrightarrow U(n)/O(n)\xrightarrow{\overline{d}} S^1$$ is a fibre bundle, and it is not difficult to see that $\ker(d)/O(n)\simeq SU(n)/SO(n)$, where the latter space is simply connected. It follows from the long exact sequence of a fibre bundle that the induced map
$$\overline{d}_\ast:\pi_1(U(n)/O(n))\rightarrow\pi_1(S^1)\cong\mathbb{Z}$$ is an isomorphism. Consequently, we obtain from an isomorphism
$$\mu_{Mas}:\pi_1(\Lambda(n))\rightarrow\mathbb{Z},$$ which is the *Maslov index* for closed paths in $\Lambda(n)$. Roughly speaking, given an arbitrary $L_0\in\Lambda(n)$, the Maslov index counts the total number of intersections of a loop in $\Lambda(n)$ with $L_0$. This is independent of the particular choice of $L_0$, which however is no longer the case if we extend the definition to non closed paths in $\Lambda(n)$ as follows.\
We fix $L_0\in\Lambda(n)$ and note at first that $L_0$ yields a stratification
$$\Lambda(n)=\bigcup^n_{k=0}\Lambda_k(L_0),$$ where
$$\Lambda_k(L_0)=\{L\in\Lambda(n):\, \dim(L\cap L_0)=k\}.$$ From the fact that $\Lambda_0(L_0)$ is contractible (see e.g. [@PiccioneBook Rem. 2.5.3]) and the long exact sequence of homology, we see that the inclusion induces an isomorphism
$$H_1(\Lambda(n))\rightarrow H_1(\Lambda(n),\Lambda_0(L_0)).$$ Also, as $\pi_1(\Lambda(n))$ is abelian, $H_1(\Lambda(n))$ is isomorphic to $\pi_1(\Lambda(n))$ and so we obtain a sequence of isomorphisms
$$\begin{aligned}
\label{Maslovsequence}
H_1(\Lambda(n),\Lambda_0(L_0))\rightarrow H_1(\Lambda(n))\rightarrow \pi_1(\Lambda(n))\rightarrow\pi_1(U(n)/O(n))\rightarrow\mathbb{Z}.\end{aligned}$$
Finally, every path in $\Lambda(n)$ having endpoints in $\Lambda_0(L_0)$ canonically yields an element in$H_1(\Lambda(n),\Lambda_0(L_0))$. The *Maslov index* of the path is the integer obtained from the sequence of isomorphisms .\
Let us note from the very definition the following three properties of the Maslov index:
- If $\gamma_1,\gamma_2$ are homotopic by a homotopy having endpoints in $\Lambda_0(L_0)$, then
$$\mu_{Mas}(\gamma_1,L_0)=\mu_{Mas}(\gamma_2,L_0).$$
- If $\gamma_1, \gamma_2$ are such that $\gamma_1(1)=\gamma_2(0)$, then
$$\mu_{Mas}(\gamma_1\ast\gamma_2,L_0)=\mu_{Mas}(\gamma_1,L_0)+\mu_{Mas}(\gamma_2,L_0).$$
- If $\gamma(\lambda)\in\Lambda_0(L_0)$ for all $\lambda\in I$, then $\mu_{Mas}(\gamma,L_0)=0$.
Let us point out that (iii) also follows from (i) and (ii) independently of the construction.\
The Maslov index can easily be generalised to a pair of paths in $\Lambda(n)$. To this aim let us call a pair of paths $(\gamma_1,\gamma_2)$ *admissible* if
$$\gamma_1(0)\cap\gamma_2(0)=\gamma_1(1)\cap\gamma_2(1)=\{0\}.$$ In what follows we consider $\mathbb{R}^{2n}\times\mathbb{R}^{2n}$ as a symplectic space with respect to the symplectic form $(-\omega_0)\times\omega_0$. Note that the diagonal $\Delta$ is in $\Lambda(2n)$, as well as $L_1\times L_2$ for any $L_1,L_2\in\Lambda(n)$. Moreover, $L_1\cap L_2\neq\{0\}$ if and only if $(L_1\times L_2)\cap\Delta\neq\{0\}$. Hence it is natural to define the *Maslov index* for a pair $(\gamma_1,\gamma_2)$ of admissible paths in $\Lambda(n)$ as
$$\mu_{Mas}(\gamma_1,\gamma_2)=\mu_{Mas}(\gamma_1\times\gamma_2,\Delta).$$ Note that the basic properties which we previously mentioned carry over immediately, i.e.,
- $\mu_{Mas}(\gamma_1,\gamma_2)=0$ if $\gamma_1(\lambda)\cap\gamma_2(\lambda)=\{0\}$ for all $\lambda\in I$.
- $\mu_{Mas}(\gamma_1\ast\gamma_3,\gamma_2\ast\gamma_4)=\mu_{Mas}(\gamma_1,\gamma_2)+\mu_{Mas}(\gamma_3,\gamma_4)$ if $\gamma_{1}(1)=\gamma_{3}(0)$ and $\gamma_{2}(1)=\gamma_{4}(0)$.
- $\mu_{Mas}(\gamma_1,\gamma_2)=\mu_{Mas}(\gamma_3,\gamma_4)$ if $\gamma_1\simeq\gamma_3$ and $\gamma_2\simeq\gamma_4$ are homotopic by a homotopy through admissible pairs.
Also, it is not difficult to see from the construction of the Maslov index that
- $\mu_{Mas}(\gamma_1,\gamma_2)=\mu_{Mas}(\gamma_1,L_0)$ in case that $\gamma_2(\lambda)=L_0$ for some $L_0\in\Lambda(n)$ and all $\lambda\in I$,
- $\mu_{Mas}(\gamma_1,\gamma_2)=-\mu_{Mas}(\gamma_2,\gamma_1)$ for any admissible pair $(\gamma_1,\gamma_2)$.
Finally, let us define the Maslov index for a non-admissible pair of paths. It is important to note that in this case there are different definitions in the literature. Here we follow [@Cappell], and note that given $L_1, L_2\in\Lambda(n)$ such that $L_1\cap L_2\neq\{0\}$, there is $\varepsilon>0$ such that $e^{\Theta J}L_2\in\Lambda(n)$ and $L_1\cap e^{\Theta J}L_2=\{0\}$ for all $0<|\Theta|\leq\varepsilon$. We define the Maslov index as
$$\mu_{Mas}(\gamma_1,\gamma_2)=\mu_{Mas}(\gamma_1,e^{-\Theta J}\gamma_2),$$ where $\Theta$ is such that $\gamma_1(0)\cap e^{-\Theta' J}\gamma_2(0)=\gamma_1(1)\cap e^{-\Theta' J}\gamma_2(1)=\{0\}$ for all $0<|\Theta'|\leq\Theta$. By the homotopy invariance, it is clear that this definition does not depend on the choice of $\Theta$. Also, it coincides with the previous definition in case that the pair of paths is admissible.
### The Paths $\gamma_{nor}$ and $\gamma'_{nor}$ {#section-gammanor}
The aim of this section is to compute the Maslov index for two elementary paths that will also become important in our proof of Theorem \[main\] below. The examples also show that is very convenient to obtain paths in $\Lambda(n)$ with a given Maslov index.\
Let us first consider the path
$$[0,1]\ni \lambda\mapsto A(\lambda)=\operatorname{diag}(e^{i\pi\lambda},1,\ldots,1)\in U(n)$$ and its projection $\overline{A}(\lambda):=A(\lambda)\cdot O(n)$ to the quotient $U(n)/O(n)$. Note that$A(0)\operatorname{diag}(-1,1,\ldots,1)=A(1)$ and so $\overline{A}$ is a closed curve. Also, as $\det^2(A(\lambda))=e^{2\pi i\lambda}$, we see that the Maslov index of the corresponding path in $\Lambda(n)$ is $1$. Using the identification $\mathbb{C}^n\cong\mathbb{R}^{2n}$, it is readily seen that
$$\gamma_{nor}(\lambda):=A(\lambda)(\mathbb{R}^n\times\{0\})=\mathbb{R}(\cos(\pi\lambda)e_1+\sin(\pi\lambda)e_{n+1})+\sum^{n}_{j=2}{\mathbb{R}e_j}\in\Lambda(n).$$ Hence we have found a path $\gamma_{nor}$ such that $\gamma_{nor}(0)=\gamma_{nor}(1)=\mathbb{R}^n\times\{0\}$ and $\mu_{Mas}(\gamma_{nor})=1$.\
Let us now consider
$$[0,1]\ni \lambda\mapsto B(\lambda)=\operatorname{diag}(-ie^{ i\pi\lambda},i,\ldots,i)\in U(n)$$ and note that again the projection $\overline{B}$ to $U(n)/O(n)$ is a closed path and $\det^2(B(\lambda))=(-1)^{n}e^{2\pi i\lambda}$. Hence
$$\gamma'_{nor}(\lambda):=B(\lambda)(\mathbb{R}^n\times\{0\})=\mathbb{R}(\sin(\pi\lambda)e_1-\cos(\pi\lambda)e_{n+1})+\sum^{2n}_{j=n+2}{\mathbb{R}e_j}\in\Lambda(n)$$ is such that $\gamma'_{nor}(0)=\gamma'_{nor}(1)=\{0\}\times\mathbb{R}^n$ and $\mu_{Mas}(\gamma'_{nor})=1$.
The Gap-Metric and the Spectral Flow {#section-spectralflow}
------------------------------------
Our first aim of this section is to recall the definition of the gap-metric, where we follow Kato’s monograph [@Kato].\
Let $H$ be a real Hilbert space and let $G(H)$ denote the set of all closed subspaces of $H$. For every $U\in G(H)$ there is a unique orthogonal projection $P_U$ onto $U$ which is a bounded operator on $H$. We set
$$d_G(U,V)=\|P_U-P_V\|,\quad U,V\in G(H),$$ and note that this is obviously a metric on $G(H)$. The distance between two non-trivial subspaces $U,V\in G(H)$ can also be obtained as follows. Let $S_U$ denote the unit sphere in $U$ and $d(u,V)=\inf_{v\in V}{\|u-v\|}$. Then for $\delta(U,V)=\sup_{u\in S_U} d(u,V)$,
$$\begin{aligned}
\label{gap}
d_G(U,V)=\max\{\delta(U,V),\delta(V,U)\},\end{aligned}$$
which explains why $d_G(U,V)$ is called the *gap* between $U$ and $V$.\
We now consider operators $T:\mathcal{D}(T)\subset H\rightarrow H$ which we assume to be defined on a dense domain $\mathcal{D}(T)$. Let us recall that $T$ is called *closed* if its graph $\operatorname{graph}(T)$ is closed in $H\times H$. If we denote by $\mathcal{C}(H)$ the set of all closed operators, then the gap-metric on $H\times H$ induces a metric on $\mathcal{C}(H)$ by
$$d_G(T,S)=d_G(\operatorname{graph}(T),\operatorname{graph}(S)), \quad S,T\in\mathcal{C}(H).$$ As the adjoint of a densely defined operator is closed, every selfadjoint operator on $H$ belongs to the metric space $\mathcal{C}(H)$. Moreover, let us recall that a closed operator $T$ is called *Fredholm* if its kernel and cokernel are of finite dimension. In what follows, we denote the subset of $\mathcal{C}(H)$ consisting of all $T$ which are selfadjoint and Fredholm by $\mathcal{CF}^\textup{sa}(H)$. It is well known that the spectrum $\sigma(T)$ of every selfadjoint operator is real. Moreover, if $T\in\mathcal{CF}^\textup{sa}(H)$ then $0$ is either in the resolvent set or an isolated eigenvalue of finite multiplicity (see e.g. [@Fredholm Lemma 2.2.5]).\
It was shown in [@UnbSpecFlow] that for every $T\in\mathcal{CF}^\textup{sa}(H)$ there is $\varepsilon>0$ and a neighbourhood $\mathcal{N}_{T,\varepsilon}\subset\mathcal{CF}^\textup{sa}(H)$ of $T$ such that $\pm\varepsilon\notin\sigma(S)$ and the spectral projection $\chi_{[-\varepsilon,\varepsilon]}(S)$ is of finite rank for all $S\in\mathcal{N}_{T,\varepsilon}$. Let us now consider a path $\mathcal{A}=\{\mathcal{A}_\lambda\}_{\lambda\in I}$ in $\mathcal{CF}^\textup{sa}(H)$. There are $0=\lambda_0<\lambda_1<\ldots<\lambda_N=1$ such that the restriction of the path $\mathcal{A}$ to $[\lambda_{i-1},\lambda_i]$ is entirely contained in a neighbourhood $\mathcal{N}_{T_i,\varepsilon_i}$ as above for some $T_i\in\mathcal{CF}^\textup{sa}(H)$ and some $\varepsilon_i>0$. The *spectral flow* of the path $\mathcal{A}$ is defined as
$$\begin{aligned}
\label{sfl-def}
\operatorname{sf}(\mathcal{A})=\sum^N_{i=1}{\left(\dim(\operatorname{im}(\chi_{[0,\varepsilon_i]}(\mathcal{A}_{\lambda_i}))-\dim(\operatorname{im}(\chi_{[0,\varepsilon_i]}(\mathcal{A}_{\lambda_{i-1}}))\right)}.\end{aligned}$$
It follows by an argument of Phillips [@Phillips] that $\operatorname{sf}(\mathcal{A})$ only depends on the path $\mathcal{A}$, and that the following fundamental property holds (see also [@UnbSpecFlow]).
1. Let $h:I\times I\rightarrow\mathcal{CF}^\textup{sa}(H)$ be a homotopy such that the dimensions of the kernels of $h(s,0)$ and $h(s,1)$ are constant for all $s\in I$. Then $$\operatorname{sf}(h(0,\cdot))=\operatorname{sf}(h(1,\cdot)).$$
Moreover, it is easily seen from the definition of the spectral flow that
1. if the dimension of the kernel of $\mathcal{A}_\lambda$ is constant for all $\lambda\in I$, then $\operatorname{sf}(\mathcal{A})=0$;
2. if $\mathcal{A}^1$ and $\mathcal{A}^2$ are two paths in $\mathcal{CF}^\textup{sa}(H)$ such that $\mathcal{A}^1_1=\mathcal{A}^2_0$, then
$$\operatorname{sf}(\mathcal{A}^1\ast\mathcal{A}^2)=\operatorname{sf}(\mathcal{A}^1)+\operatorname{sf}(\mathcal{A}^2).$$
Let us finally note two further elementary properties of the spectral flow which play a crucial role in our proof of Theorem \[main\] below. The first of them has been used, e.g., in [@Pejsachowicz §7].
\[lemma-sflperturbation\] Let $\mathcal{A}:I\rightarrow\mathcal{CF}^\textup{sa}(H)$ be gap-continuous and set $\mathcal{A}^\delta=\mathcal{A}+\delta I_H$ for $\delta\in\mathbb{R}$. Then, for any sufficiently small $\delta>0$, $\mathcal{A}^\delta$ is a gap-continuous path in $\mathcal{CF}^\textup{sa}(H)$ and
$$\begin{aligned}
\label{sfldelta}
\operatorname{sf}(\mathcal{A})=\operatorname{sf}(\mathcal{A}^\delta).\end{aligned}$$
We note at first that the operators $\mathcal{A}^\delta_\lambda$ are selfadjoint and Fredholm for $\delta$ sufficiently small, which follows from standard stability theory (see e.g. [@Kato]). Moreover, the path $\mathcal{A}^\delta$ is gap-continuous by [@Kato Thm. IV.2.17], and so $\operatorname{sf}(\mathcal{A}^\delta)$ is well defined.\
To show , let $0=\lambda_0<\ldots<\lambda_N=1$ be a partition of the unit interval and $\varepsilon_i>0$, $i=1,\ldots,N$, for $\mathcal{A}$ as in . Let $\mathcal{N}_{T,\varepsilon_i}$ be an open neighbourhood of some $T\in\mathcal{CF}^{sa}(H)$ as in the construction of the spectral flow such that $\mathcal{A}_\lambda\in\mathcal{N}_{T,\varepsilon_i}$ for all $\lambda\in[\lambda_{i-1},\lambda_i]$. Now there is $\delta_i>0$ such that $\mathcal{A}^{s\delta_i}_\lambda\in\mathcal{N}_{T,\varepsilon_i}$ for all $s\in[0,1]$ and all $\lambda\in [\lambda_{i-1},\lambda_i]$, i.e. the spectral projections $\chi_{[-\varepsilon_i,\varepsilon_i]}(\mathcal{A}^{s\delta_i}_{\lambda})$ are of the same finite rank. Moreover, by choosing $\delta_i>0$ smaller, we can assume that
$$\sigma(\mathcal{A}_{\lambda_i})\cap[-\delta_i,0)=\sigma(\mathcal{A}_{\lambda_{i-1}})\cap[-\delta_i,0)=\{0\}.$$ Then, as $\sigma(\mathcal{A}^{\delta_i}_\lambda)=\sigma(\mathcal{A}_\lambda)+\delta_i$, we see that
$$\dim(\operatorname{im}(\chi_{[0,\varepsilon_i]}(\mathcal{A}_{\lambda})))=\dim(\operatorname{im}(\chi_{[0,\varepsilon_i]}(\mathcal{A}^{\delta_i}_{\lambda}))),\quad \lambda=\lambda_{i-1},\lambda_i.$$ If we now set $\delta=\min\{\delta_1,\ldots,\delta_N\}>0$, then $$\dim(\operatorname{im}(\chi_{[0,\varepsilon_i]}(\mathcal{A}_{\lambda})))=\dim(\operatorname{im}(\chi_{[0,\varepsilon_i]}(\mathcal{A}^{\delta}_{\lambda}))),\quad \lambda=\lambda_{i-1},\lambda_i$$ holds simultaneously for this $\delta$ and all $i=1,\ldots,N$, and so the assertion follows from the definition .
Finally, let us note the following stability of the spectral flow under conjugation by invertible operators, where we denote by $M^T$ the adjoint of an operator in the real Hilbert space $H$.
\[lemma-conjuagtion\] Let $\mathcal{A}:I\rightarrow\mathcal{CF}^\textup{sa}(H)$ be a gap-continuous path and $M:I\rightarrow GL(H)$ a continuous family of bounded invertible operators. Then $\{M^T_\lambda\mathcal{A}_\lambda M_\lambda\}_{\lambda\in I}$ is gap-continuous and
$$\operatorname{sf}(M^T\mathcal{A}M)=\operatorname{sf}(\mathcal{A}).$$
Note that
$$\begin{aligned}
\operatorname{graph}(M^T_\lambda\mathcal{A}_\lambda M_\lambda)&=\{(u,M^T_\lambda\mathcal{A}_\lambda M_\lambda u):\, u\in M^{-1}_\lambda(\mathcal{D}(\mathcal{A}_\lambda))\}=\{(M^{-1}_\lambda v,M^T_\lambda\mathcal{A}_\lambda v):\,v\in\mathcal{D}(\mathcal{A}_\lambda)\}\\
&=\begin{pmatrix}
M^{-1}_\lambda&0\\
0&M^T_\lambda
\end{pmatrix}\,\operatorname{graph}(\mathcal{A}_\lambda)=:N_\lambda\operatorname{graph}(\mathcal{A}_\lambda)\subset H\times H,\end{aligned}$$
and so $\{N_\lambda P_{\operatorname{graph}(\mathcal{A}_\lambda)}N^{-1}_\lambda\}_{\lambda\in I}$ is a continuous family of oblique projections onto$\{\operatorname{graph}(M^T_\lambda\mathcal{A}_\lambda M_\lambda)\}_{\lambda\in I}$ in $\mathcal{L}(H\times H)$. By [@Kato Thm. I.6.35], we have for the corresponding orthogonal projections $P_{\operatorname{graph}(M^T_\lambda\mathcal{A}_\lambda M_\lambda)}$ onto $\operatorname{graph}(M^T_\lambda\mathcal{A}_\lambda M_\lambda)$ the inequality
$$\|P_{\operatorname{graph}(M^T_\mu\mathcal{A}_\mu M_\mu)}-P_{\operatorname{graph}(M^T_\lambda\mathcal{A}_\lambda M_\lambda)}\|\leq \|N_\mu P_{\operatorname{graph}(\mathcal{A}_\mu)}N^{-1}_\mu-N_\lambda P_{\operatorname{graph}(\mathcal{A}_\lambda)}N^{-1}_\lambda\|,\quad \mu,\lambda\in I.$$ Consequently, $\{P_{\operatorname{graph}(M^T_\lambda\mathcal{A}_\lambda M_\lambda)}\}_{\lambda\in I}$ is continuous, which shows that $M^T\mathcal{A}M$ is gap-continuous.\
For the equality of the spectral flows, we just need to note that $M$ is homotopic inside $GL(H)$ to the constant path given by the identity $I_H$. Let us point out that this does not even require Kuiper’s Theorem as we just need to shrink $M$ to a constant path and use that $GL(H)$ is connected. As the conjugation preserves kernel dimensions, we obtain by the homotopy invariance (i) from above
$$\begin{aligned}
\operatorname{sf}(M^T\mathcal{A}M)=\operatorname{sf}(\mathcal{A}).\end{aligned}$$
Proof of Theorem \[main\] {#section-proof}
=========================
The proof of Theorem \[main\] falls naturally into two parts. In the first part we deal with the continuity of families of the type , where we actually consider a slightly more general setting. In the second part we show the spectral flow formula in Theorem \[main\].
Continuity
----------
To simplify notation, we set $E=L^2(I,\mathbb{R}^{2n})$ and $H=H^1(I,\mathbb{R}^{2n})$. The aim of this step is to prove the following proposition, which we will later apply in the cases $X=I$ and $X=I\times I$.
\[prop-continuity\] Let $X$ be a metric space and $\gamma_1,\gamma_2:X\rightarrow\Lambda(n)$ two families of Lagrangian subspaces in $\mathbb{R}^{2n}$. Then
$$\mathcal{A}:X\rightarrow \mathcal{CF}^\textup{sa}(E),\quad (\mathcal{A}_\lambda u)(t)=Ju'(t),$$ where $$\mathcal{D}(\mathcal{A}_\lambda)=\{u\in H: u(0)\in\gamma_1(\lambda), u(1)\in\gamma_2(\lambda)\},$$ is continuous with respect to the gap-metric on $\mathcal{CF}^\textup{sa}(E)$.
We want to use and consider
$$\delta(\operatorname{graph}(\mathcal{A}_\lambda),\operatorname{graph}(\mathcal{A}_{\lambda_0})).$$ Note at first that for $u\in\mathcal{D}(\mathcal{A}_\lambda)$ and $v\in\mathcal{D}(\mathcal{A}_{\lambda_0})$
$$\begin{aligned}
\label{gapcontI}
\begin{split}
\|(u,\mathcal{A}_\lambda u)-(v,\mathcal{A}_{\lambda_0}v)\|_{E\oplus E}&=\|(u-v,J(u'-v'))\|_{E\oplus E}\\
&\leq \left(\|u-v\|^2_{E}+\|J\|\|u'-v'\|^2_{E}\right)^\frac{1}{2}\\
&=\|u-v\|_{H},
\end{split}\end{aligned}$$
where we have used that $\|J\|=1$. Let us recall that the topology of $G_n(\mathbb{R}^{2n})$ is induced by the metric $d(L,M)=\|P_L-P_M\|$, where $P_L, P_M\in M(2n,\mathbb{R})$ are the orthogonal projections onto $L$ and $M$, respectively. Hence, by the continuity of $\gamma_1$ and $\gamma_2$, there are two families of orthogonal projections $\hat{P},\tilde{P}:X\rightarrow M(2n,\mathbb{R})$ such that
$$\operatorname{im}(\hat{P}_\lambda)=\gamma_1(\lambda),\quad \operatorname{im}(\tilde{P}_\lambda)=\gamma_2(\lambda),\quad\lambda\in X.$$ We define for $w\in H$
$$(P_\lambda w)(t)=w(t)-(1-t)(I_{2n}-\hat{P}_\lambda)w(0)-t(I_{2n}-\tilde{P}_\lambda)w(1).$$ It is easily seen that $P^2_\lambda w=P_\lambda w$, as well as $P_\lambda w\in\mathcal{D}(\mathcal{A}_\lambda)$ for all $w\in H$ and $\lambda\in X$, which shows that
$$\begin{aligned}
\label{gapcontII}
\inf_{v\in\mathcal{D}(\mathcal{A}_{\lambda_0})}\|u-v\|_{H}\leq\|u-P_{\lambda_0}u\|_{H}.\end{aligned}$$
As
$$u(t)-(P_{\lambda_0}u)(t)=(1-t)(I_{2n}-\hat{P}_{\lambda_0})u(0)+t(I_{2n}-\tilde{P}_{\lambda_0})u(1),$$ it follows for $u\in\mathcal{D}(\mathcal{A}_\lambda)$ that
$$\begin{aligned}
\label{gapcontIII}
\begin{split}
\|u-P_{\lambda_0}u\|_{H}&\leq 2(\|(I_{2n}-\hat{P}_{\lambda_0})u(0)\|+\|(I_{2n}-\tilde{P}_{\lambda_0})u(1)\|)\\
&= 2(\|(I_{2n}-\hat{P}_{\lambda_0})\hat{P}_\lambda u(0)\|+\|(I_{2n}-\tilde{P}_{\lambda_0})\tilde{P}_\lambda u(1)\|)\\
&\leq 2(\|(I_{2n}-\hat{P}_{\lambda_0})\hat{P}_\lambda\|\|u(0)\|+\|(I_{2n}-\tilde{P}_{\lambda_0})\tilde{P}_\lambda\|\|u(1)\|),
\end{split}\end{aligned}$$
where we have used that $u\in\mathcal{D}(\mathcal{A}_\lambda)$ and so $\hat{P}_\lambda u(0)=u(0)$ and $\tilde{P}_\lambda u(1)=u(1)$. Let us note that the factor $2$ appears in the previous estimate as we are dealing with the norm on $H$ and so we also need to take into account the derivatives of $u-P_{\lambda_0}u$ with respect to $t$.\
Since the point evaluation is continuous in $H$, there is a constant $\alpha>0$ such that for $t=0$ and $t=1$
$$\begin{aligned}
\label{gapcontIV}
\|u(t)\|\leq\alpha\|u\|_{H}=\alpha\left(\|u\|^2_{E}+\|u'\|^2_{E}\right)^\frac{1}{2}=\alpha\left(\|u\|^2_{E}+\|Ju'\|^2_{E}\right)^\frac{1}{2},\end{aligned}$$
where we use that $J$ is an isometry on $\mathbb{R}^{2n}$. Hence, by –,
$$\begin{aligned}
d((u,\mathcal{A}_\lambda u),\operatorname{graph}(\mathcal{A}_{\lambda_0}))&=\inf_{v\in\mathcal{D}(\mathcal{A}_{\lambda_0})}\|(u,\mathcal{A}_\lambda u)-(v,\mathcal{A}_{\lambda_0}v)\|_{E\oplus E}\\
&\leq \inf_{v\in\mathcal{D}(\mathcal{A}_{\lambda_0})}\|u-v\|_{H}\leq\|u-P_{\lambda_0}u\|_{H}\\
&\leq 2(\|(I_{2n}-\hat{P}_{\lambda_0})\hat{P}_\lambda\|\|u(0)\|+\|(I_{2n}-\tilde{P}_{\lambda_0})\tilde{P}_\lambda\|\|u(1)\|)\\
&\leq 2\alpha(\|(I_{2n}-\hat{P}_{\lambda_0})\hat{P}_\lambda\|+\|(I_{2n}-\tilde{P}_{\lambda_0})\tilde{P}_\lambda\|)(\|u\|^2_{E}+\|Ju'\|^2_{E})^\frac{1}{2}. \end{aligned}$$
As the unit sphere in $\operatorname{graph}(\mathcal{A}_\lambda)$ is given by
$$\{(u,\mathcal{A}_\lambda u):\, u\in\mathcal{D}(\mathcal{A}_\lambda),\, \|u\|^2_{E}+\|Ju'\|^2_{E}=1\},$$ we finally get
$$\begin{aligned}
\label{gapfinalI}
\begin{split}
\delta(\operatorname{graph}(\mathcal{A}_\lambda),\operatorname{graph}(\mathcal{A}_{\lambda_0}))&=\sup\{d((u,\mathcal{A}_\lambda u),\operatorname{graph}(\mathcal{A}_{\lambda_0})):\,u\in\mathcal{D}(\mathcal{A}_\lambda),\,\|u\|^2+\|Ju'\|^2=1\}\\
&\leq 2\alpha(\|(I_{2n}-\hat{P}_{\lambda_0})\hat{P}_\lambda\|+\|(I_{2n}-\tilde{P}_{\lambda_0})\tilde{P}_\lambda\|).
\end{split}\end{aligned}$$
Note that if we swap $\lambda$ and $\lambda_0$ and repeat the above argument, we also have
$$\begin{aligned}
\label{gapfinalII}
\delta(\operatorname{graph}(\mathcal{A}_{\lambda_0}),\operatorname{graph}(\mathcal{A}_{\lambda}))\leq 2\alpha(\|(I_{2n}-\hat{P}_{\lambda})\hat{P}_{\lambda_0}\|+\|(I_{2n}-\tilde{P}_{\lambda})\tilde{P}_{\lambda_0}\|).\end{aligned}$$
To finish the proof, we need the following well-known theorem that can be found, e.g., in [@Kato I.6.34].
Let $E$ be a Hilbert space and $P,Q$ orthogonal projections in $E$. If
$$\|(I_E-P)Q\|<1\,\,\text{and } \|(I_E-Q)P\|<1,$$ then
$$\|(I_E-P)Q\|=\|(I_E-Q)P\|=\|P-Q\|.$$
Now, as $(I_{2n}-\hat{P}_{\lambda})\hat{P}_{\lambda_0}=(I_{2n}-\hat{P}_{\lambda_0})\hat{P}_{\lambda}=0$ for $\lambda=\lambda_0$, we have for all $\lambda$ in a neighbourhood of $\lambda_0$
$$\|(I_{2n}-\hat{P}_{\lambda})\hat{P}_{\lambda_0}\|=\|(I_{2n}-\hat{P}_{\lambda_0})\hat{P}_{\lambda}\|=\|\hat{P}_{\lambda}-\hat{P}_{\lambda_0}\|$$ and likewise $$\|(I_{2n}-\tilde{P}_{\lambda})\tilde{P}_{\lambda_0}\|=\|(I_{2n}-\tilde{P}_{\lambda_0})\tilde{P}_{\lambda}\|=\|\tilde{P}_\lambda-\tilde{P}_{\lambda_0}\|.$$ Consequently, we obtain from , and for all $\lambda$ sufficiently close to $\lambda_0$
$$\begin{aligned}
d_G(\mathcal{A}_\lambda,\mathcal{A}_{\lambda_0})&=\max\{\delta(\operatorname{graph}(\mathcal{A}_\lambda),\operatorname{graph}(\mathcal{A}_{\lambda_0})),\delta(\operatorname{graph}(\mathcal{A}_{\lambda_0}),\operatorname{graph}(\mathcal{A}_{\lambda}))\}\\
&\leq 2\alpha(\|\hat{P}_\lambda-\hat{P}_{\lambda_0}\|+\|\tilde{P}_\lambda-\tilde{P}_{\lambda_0}\|),\end{aligned}$$
which shows that $\mathcal{A}=\{\mathcal{A}_\lambda\}_{\lambda\in X}$ is indeed continuous in $\mathcal{CF}(E)$. Hence Proposition \[prop-continuity\] is shown.
The Spectral Flow Formula
-------------------------
We now prove the spectral flow formula in Theorem \[main\] in two steps.
Step 1: Theorem \[main\] for admissible paths {#step-1-theorem-main-for-admissible-paths .unnumbered}
---------------------------------------------
We begin this first step of our proof with the following elementary observation.
\[lemma-connected\] The set of all transversal pairs in $\Lambda(n)$, i.e.
$$\begin{aligned}
\label{set}
\{(L_1,L_2)\in\Lambda(n)\times\Lambda(n):\, L_1\cap L_2=\{0\}\}\subset\Lambda(n)\times\Lambda(n),\end{aligned}$$
is path-connected.
Let us first recall the well-known fact that $\Lambda_0(L_0)$ is contractible, and hence path-connected, for any $L_0\in\Lambda(n)$ (see [@PiccioneBook Rem. 2.5.3]). Now let $(L_1,L_2)$ and $(L_3,L_4)$ be two transversal pairs. As in the construction of the Maslov index in Section \[section-Maslov\], $L'_1=e^{\Theta J}L_1$ is transversal to $L_2$ and $L_4$ for any sufficiently small $\Theta>0$. In particular, we obtain a path connecting $(L_1,L_2)$ and $(L'_1,L_2)$ inside . Also, as $\Lambda_0(L'_1)$ is path-connected, there is a path connecting $(L'_1,L_2)$ and $(L'_1,L_4)$ inside . Finally, there is a path from $(L'_1,L_4)$ to $(L_3,L_4)$ inside as $\Lambda_0(L_4)$ is path-connected.
This step of the proof is based on the following proposition in which we denote by $\Omega^2$ the set of all admissible pairs of paths in $\Lambda(n)$ (see ). Let us note that by Section \[section-gammanor\] and (v’) in Section \[section-Maslov\], $\mu_{Mas}(\gamma_{nor},L_1)=1$ and $\mu_{Mas}(L_0,\gamma'_{nor})=-1$, where $L_0=\mathbb{R}^n\times\{0\}$ and $L_1=\{0\}\times\mathbb{R}^n$.
\[prop-Maslov\] Let
$$\mu:\Omega^2\rightarrow\mathbb{Z}$$ be a map such that the same properties (i’)-(iii’) from Section \[section-Maslov\] are satisfied, as well as
- $\mu(\gamma_{nor},L_1)=1$ and $\mu(L_0,\gamma'_{nor})=-1$, where $L_0=\mathbb{R}^n\times\{0\}$ and $L_1=\{0\}\times\mathbb{R}^n$.
Then $\mu=\mu_{Mas}$ on $\Omega^2$.
We note at first that we have by the properties (ii’) and (iii’) homomorphisms
$$\begin{aligned}
\label{homomorphisms}
\mu, \mu_{Mas}:\pi_1(\Lambda(n)\times\Lambda(n),(L_0,L_1))\rightarrow\mathbb{Z}\end{aligned}$$
and we now claim that they coincide.\
We first note that
$$\pi_1(\Lambda(n)\times\Lambda(n),(L_0,L_1))\cong\pi_1(\Lambda(n),L_0)\times\pi_1(\Lambda(n),L_1)\cong\mathbb{Z}\oplus\mathbb{Z},$$ where the first isomorphism is induced by the projections onto the components and the second one is given by the Maslov index. As $\gamma_{nor}(0)=L_0$, $\gamma'_{nor}(0)=L_1$ and $\mu_{Mas}(\gamma_{nor})=\mu_{Mas}(\gamma'_{nor})=1$, we see that the pairs of paths
$$\{(\gamma_{nor},L_1), (L_0,\gamma'_{nor})\}$$ define a basis of $\pi_1(\Lambda(n)\times\Lambda(n),(L_0,L_1))$. Since the homomorphisms in coincide on this basis by (N), it follows that $\mu$ and $\mu_{Mas}$ are indeed equal for closed paths based at $(L_0,L_1)$.\
Let us now assume that $(\gamma_1,\gamma_2)\in\Omega^2$ is an arbitrary admissible pair of paths. We connect $(L_0,L_1)$ to $(\gamma_1(0),\gamma_2(0))$ by a pair of paths $(\gamma_3,\gamma_4)$ and $(\gamma_1(1),\gamma_2(1))$ to $(L_0,L_1)$ by a pair of paths $(\gamma_5,\gamma_6)$, where we can assume by Lemma \[lemma-connected\] that $\gamma_3(\lambda)\cap\gamma_4(\lambda)=\gamma_5(\lambda)\cap\gamma_6(\lambda)=\{0\}$ for all $\lambda$. Then by (i’), (ii’) and the first step of our proof
$$\begin{aligned}
\mu(\gamma_1,\gamma_2)&=\mu(\gamma_3,\gamma_4)+\mu(\gamma_1,\gamma_2)+\mu(\gamma_5,\gamma_6)\\
&=\mu((\gamma_3,\gamma_4)\ast(\gamma_1,\gamma_2)\ast(\gamma_5,\gamma_6))=\mu_{Mas}((\gamma_3,\gamma_4)\ast(\gamma_1,\gamma_2)\ast(\gamma_5,\gamma_6))\\
&=\mu_{Mas}(\gamma_3,\gamma_4)+\mu_{Mas}(\gamma_1,\gamma_2)+\mu_{Mas}(\gamma_5,\gamma_6)=\mu_{Mas}(\gamma_1,\gamma_2),\end{aligned}$$
which proves the proposition.
Let $(\gamma_1,\gamma_2)$ be a pair of paths in $\Lambda(n)$ as in (i’), i.e. $\gamma_1(\lambda)\cap\gamma_2(\lambda)=\{0\}$ for all $\lambda\in I$. Then $(\gamma_1,\gamma_2)$ is homotopic to the constant pair of paths $(\widetilde{\gamma}_1(\lambda),\widetilde{\gamma}_2(\lambda))=(\gamma_1(0),\gamma_2(0))$, $\lambda\in I$, by a homotopy of admissible pairs. Hence $\mu(\gamma_1,\gamma_2)=\mu(\widetilde{\gamma}_1,\widetilde{\gamma}_2)$ by (iii’). As
$$\mu(\widetilde{\gamma}_1,\widetilde{\gamma}_2)=\mu((\widetilde{\gamma}_1,\widetilde{\gamma}_2)\ast(\widetilde{\gamma}_1,\widetilde{\gamma}_2))=\mu(\widetilde{\gamma}_1,\widetilde{\gamma}_2)+\mu(\widetilde{\gamma}_1,\widetilde{\gamma}_2)$$ by (ii’), we see that $\mu(\gamma_1,\gamma_2)=\mu(\widetilde{\gamma}_1,\widetilde{\gamma}_2)=0$ and so (i’) follows from (ii’) and (iii’). Consequently, Proposition \[prop-Maslov\] actually characterises the Maslov index by the three axioms (ii’), (iii’) and (N).
We now define
$$\mu:\Omega^2\rightarrow\mathbb{Z},\quad \mu(\gamma_1,\gamma_2)=\operatorname{sf}(\mathcal{A}),$$ where $\mathcal{A}$ is the path of differential operators for the pair $(\gamma_1,\gamma_2)$. We aim to use Proposition \[prop-Maslov\] to show Theorem \[main\] and so we need to check the properties (i’), (ii’), (iii’) and (N). Let us first note that (i’) follows immediately from (ii) in Section \[section-spectralflow\] and the fact that $\ker(\mathcal{A}_\lambda)=\gamma_1(\lambda)\cap\gamma_2(\lambda)$. Also, (ii’) follows from (iii) in Section \[section-spectralflow\]. Finally, (ii’) is an immediate consequence of the homotopy invariance (i) of the spectral flow and Proposition \[prop-continuity\].\
Hence it remains to show that $\mu(\gamma_{nor},L_1)=1$ and $\mu(L_0,\gamma'_{nor})=-1$, which will be a direct consequence of the following lemma.
The spectra of the operators $\mathcal{A}_\lambda$ in are
- for $(\gamma_1,\gamma_2)=(\gamma_{nor},L_1)$ $$\sigma(\mathcal{A}_\lambda)=\left\{\pi\lambda-\frac{\pi}{2}+\pi k:\,k\in\mathbb{Z}\right\}\cup\left\{\frac{\pi}{2}+k\pi:\,k\in\mathbb{Z} \right\},$$
- for $(\gamma_1,\gamma_2)=(L_0,\gamma'_{nor})$
$$\sigma(\mathcal{A}_\lambda)=\left\{-\pi\lambda+\frac{\pi}{2}+\pi k:\,k\in\mathbb{Z}\right\}\cup\left\{\frac{\pi}{2}+k\pi:\,k\in\mathbb{Z}\right\}.$$
We consider $Ju'=\mu u$ and note that the solutions of this equation are
$$u(t)=\exp(-\mu tJ)v,\quad t\in I, v\in\mathbb{R}^{2n}.$$ Let us first consider the path in (i). Then $u$ belongs to the domain of $\mathcal{A}_\lambda$ if and only if
$$\begin{aligned}
\label{equ}
u(0)=v\in\gamma_{nor}(\lambda),\quad u(1)=\exp(-\mu J)v\in L_1.\end{aligned}$$
As $\exp(-\mu J)v\in L_1$ if and only if $v\in\exp(\mu J)L_1$, and $\exp(\mu J)=\cos(\mu)I_{2n}+\sin(\mu)J$, we see that is equivalent to
$$(\cos(\mu)I_{2n}+\sin(\mu)J)(\{0\}\times\mathbb{R}^n)\cap\left(\mathbb{R}(\cos(\pi\lambda)e_1+\sin(\pi\lambda)e_{n+1})+\sum^{n}_{j=2}{\mathbb{R}e_j}\right)\neq\{0\}.$$ There are two different cases where these spaces intersect non-trivially. Firstly, if $\cos(\mu)=0$, i.e. $\mu=\frac{\pi}{2}+k\pi$ for $k\in\mathbb{Z}$. Secondly, if there is an $\alpha\neq 0$ such that $\sin(\pi\lambda)e_{n+1}=\alpha\cos(\mu)e_{n+1}$ and $\cos(\pi\lambda)e_1=-\alpha\sin(\mu)e_1$, where we use that $Je_{n+1}=-e_1$. Of course, the latter equations are equivalent to $\sin(\pi\lambda)=\alpha\cos(\mu)$ and $\cos(\pi\lambda)=-\alpha\sin(\mu)$, which can be rewritten as
$$\begin{aligned}
e^{i\pi\lambda}&=\cos(\pi\lambda)+i\sin(\pi\lambda)=\alpha (-\sin(\mu)+i\cos(\mu))=\alpha i e^{i\mu}=\alpha e^{i(\mu+\frac{\pi}{2})}.\end{aligned}$$
Hence $|\alpha|=1$, and this equation holds if and only if $\pi\lambda=\mu+\frac{\pi}{2}+k\pi$, or equivalently $\mu=\pi\lambda-\frac{\pi}{2}-k\pi$.\
In (ii), $u$ belongs to the domain of $\mathcal{A}_\lambda$ if and only if
$$\begin{aligned}
u(0)=v\in L_0,\quad u(1)=\exp(-\mu J)v\in\gamma'_{nor},\end{aligned}$$
which is equivalent to
$$(\cos(\mu)I_{2n}-\sin(\mu)J)(\mathbb{R}^n\times\{0\})\cap\left(\mathbb{R}(\sin(\pi\lambda)e_1-\cos(\pi\lambda)e_{n+1})+\sum^{2n}_{j=n+2}{\mathbb{R}e_j}\right)\neq\{0\}.$$ Again, there are two cases in which this intersection is non-trivial. Firstly, $\mu=k\pi+\frac{\pi}{2}$ where $\cos(\mu)=0$. Secondly, if there is some $\alpha\neq 0$ such that $\alpha\cos(\mu)e_1=\sin(\pi\lambda)e_1$ and $\alpha\sin(\mu)e_{n+1}=\cos(\pi\lambda)e_{n+1}$, which is equivalent to
$$e^{i\pi\lambda}=\cos(\pi\lambda)+i\sin(\pi\lambda)=\alpha(sin(\mu)+i\cos(\mu))=\alpha i e^{-i\mu}=\alpha e^{i(\frac{\pi}{2}-\mu)}.$$ Hence $|\alpha|=1$, and the latter equation holds if and only if $\pi\lambda=\frac{\pi}{2}-\mu+k\pi$ which finally shows that $\mu=-\pi\lambda+\frac{\pi}{2}+k\pi$.
We see from the previous lemma that in both cases there is only one eigenvalue of $\mathcal{A}_\lambda$ that crosses the axis whilst the parameter $\lambda$ travels from $0$ to $1$. It is now an immediate consequence of the definition of the spectral flow that $\operatorname{sf}(\mathcal{A})=1$ for $(\gamma_1,\gamma_2)=(\gamma_{nor},L_1)$ and $\operatorname{sf}(\mathcal{A})=-1$ for $(\gamma_1,\gamma_2)=(L_0,\gamma'_{nor})$. Hence Theorem \[main\] is shown in the admissible case.
Step 2: The general case {#step-2-the-general-case .unnumbered}
------------------------
Let $(\gamma_1,\gamma_2)$ be a pair of paths in $\Lambda(n)$ which is not necessarily admissible, and let $\mathcal{A}$ be the path . Let $\delta>0$ be as in Lemma \[lemma-sflperturbation\] such that
$$\begin{aligned}
\operatorname{sf}(\mathcal{A})=\operatorname{sf}(\mathcal{A}^{\delta_0})\end{aligned}$$
for all $0\leq \delta_0\leq\delta$.\
We consider the solution $\Psi:I\rightarrow\operatorname{Sp}(2n,\mathbb{R})$ of the differential equation
$$\begin{aligned}
\begin{cases}
J\Psi'(t)+\delta_0\Psi(t)=0,\quad t\in I\\
\Psi(0)=I_{2n},
\end{cases}\end{aligned}$$
and the operator $M\in GL(L^2(I,\mathbb{R}^{2n}))$ given by $(Mu)(t)=\Psi(t)u(t)$, $t\in I$. Then, as $\mathcal{D}(\mathcal{A}^{\delta_0}_\lambda)=\mathcal{D}(\mathcal{A}_\lambda)$, $M^T\mathcal{A}^{\delta_0}_\lambda M$ is defined on the domain
$$\begin{aligned}
\mathcal{D}(M^T\mathcal{A}^{\delta_0}_\lambda M)&=M^{-1}(\mathcal{D}(\mathcal{A}^{\delta_0}_\lambda))=\{\Psi(\cdot)^{-1}u\in H^1(I,\mathbb{R}^{2n}):u(0)\in\gamma_1(\lambda),\, u(1)\in\gamma_2(\lambda)\}\\
&=\{v\in H^1(I,\mathbb{R}^{2n}):\, v(0)\in\gamma_1(\lambda), v(1)\in\Psi(1)^{-1}\gamma_2(\lambda)\}\end{aligned}$$
and given by
$$\begin{aligned}
(M^T\mathcal{A}^{\delta_0}_\lambda Mu)(t)&=M^T(J\Psi'(t)u(t)+J\Psi(t)u'(t)+\delta_0\Psi(t)u(t))\\
&=M^T(-\delta_0\Psi(t)u(t)+J\Psi(t)u'(t)+\delta_0\Psi(t)u(t))=\Psi(t)^TJ\Psi(t)u'(t)=Ju'(t).\end{aligned}$$
As $\Psi(t)=\exp(\delta_0 Jt)$, $t\in I$, we see that $\Psi(1)^{-1}=\exp(-\delta_0 J)$. Finally, if $\delta_0>0$ is sufficiently small, we obtain by Step 1, Proposition \[prop-continuity\] and the definition of the Maslov index for non-admissible paths in Section \[section-Maslov\],
$$\operatorname{sf}(\mathcal{A})=\operatorname{sf}(\mathcal{A}^{\delta_0})=\mu_{Mas}(\gamma_1,e^{-\delta_0 J}\gamma_2)=\mu_{Mas}(\gamma_1,\gamma_2),$$ which proves Theorem \[main\] in the general case.
A Spectral Flow Formula for Hamiltonian Systems
===============================================
Let $\gamma_1,\gamma_2:I\rightarrow\Lambda(n)$ be two paths of Lagrangian subspaces in $\mathbb{R}^{2n}$. We note for later reference the following two standard properties of the Maslov index (see e.g. [@RobbinMaslov])
1. If $\Psi:I\rightarrow\operatorname{Sp}(2n,\mathbb{R})$ is a path of symplectic matrices, then
$$\begin{aligned}
\label{sympinv}
\mu_{Mas}(\Psi\gamma_1,\Psi \gamma_2)=\mu_{Mas}(\gamma_1,\gamma_2).\end{aligned}$$
2. If $\gamma'_1,\gamma'_2:I\rightarrow\Lambda(n)$ denote the reverse paths defined by $\gamma'_1(\lambda)=\gamma_1(1-\lambda)$ and $\gamma'_2(\lambda)=\gamma_2(1-\lambda)$, then
$$\begin{aligned}
\label{reverse}
\mu_{Mas}(\gamma'_1,\gamma'_2)=-\mu_{Mas}(\gamma_1,\gamma_2).\end{aligned}$$
Moreover, we need below the following homotopy invariance property which is an immediate consequence of (iii’) in Section \[section-Maslov\] and the definition of the Maslov index for non-admissible pairs of paths:
1. $\mu_{Mas}(\gamma_1,\gamma_2)=\mu_{Mas}(\gamma_3,\gamma_4)$ if $\gamma_1\simeq\gamma_3$ and $\gamma_2\simeq\gamma_4$ are homotopic by homotopies with fixed endpoints.
Let now $S:I\times I\rightarrow M(2n,\mathbb{R})$ be a two parameter family of symmetric matrices and let us consider
$$\label{Hamiltonian}
\left\{
\begin{aligned}
Ju'(t)+S_\lambda(t)u(t)&=0,\quad t\in I\\
(u(0),u(1))\in \gamma_1(\lambda)&\times \gamma_2(\lambda),
\end{aligned}
\right.$$
as well as the differential operators
$$\begin{aligned}
\label{A-hamiltonian}
\mathcal{A}_\lambda:\mathcal{D}(\mathcal{A}_\lambda)\subset L^2(I,\mathbb{R}^{2n})\rightarrow L^2(I,\mathbb{R}^{2n}),\quad (\mathcal{A}_\lambda u)(t)=Ju'(t)+S_\lambda(t)u(t) \end{aligned}$$
on the domains $\mathcal{D}(\mathcal{A}_\lambda)=\{u\in H^1(I,\mathbb{R}^{2n}):\, u(0)\in \gamma_1(\lambda),\, u(1)\in \gamma_2(\lambda)\}$.\
We denote for $\lambda\in I$ by $\Psi_\lambda:I\rightarrow\operatorname{Sp}(2n,\mathbb{R})$ the matrices defined by
$$\label{Psi}
\left\{
\begin{aligned}
J\Psi'_\lambda(t)+S_\lambda(t)\Psi_\lambda(t)&=0,\quad t\in I\\
\Psi_\lambda(0)&=I_{2n},
\end{aligned}
\right.$$
and we set $(\Psi\gamma_1)(\lambda)=\Psi_\lambda(1) \gamma_1(\lambda)$. The aim of this final section is to obtain the following spectral flow formula from Theorem \[main\].
\[thm-Hamiltonian\] Under the assumptions above, $\mathcal{A}$ is a gap-continuous path of selfadjoint Fredholm operators on $L^2(I,\mathbb{R}^{2n})$ and
$$\operatorname{sf}(\mathcal{A})=\mu_{Mas}(\Psi\gamma_1,\gamma_2).$$
We define a continuous family of bounded invertible operators on $L^2(I,\mathbb{R}^{2n})$ by $(M_\lambda u)(t)=\Psi_\lambda(t) u(t)$, $t\in I$. Then
$$\begin{aligned}
(M^T_\lambda\mathcal{A}_\lambda M_\lambda u)(t)&= \Psi^T_\lambda(t)(J\Psi'_\lambda(t)u(t)+J\Psi_\lambda(t)u'(t)+S_\lambda(t)\Psi_\lambda(t)u(t))\\
&=\Psi^T_\lambda(t)(-S_\lambda(t)\Psi_\lambda(t)u(t))+\Psi^T_\lambda(t)J\Psi_\lambda(t)u'(t)+\Psi^T_\lambda(t)S_\lambda(t)\Psi_\lambda(t)u(t)\\
&=Ju'(t)\end{aligned}$$
and
$$\begin{aligned}
\mathcal{D}(M^T_\lambda\mathcal{A}_\lambda M_\lambda)&= M^{-1}_\lambda(\mathcal{D}(\mathcal{A}_\lambda))=\{u\in H^1(I,\mathbb{R}^{2n}):\, u(0)\in\gamma_1(\lambda), u(1)\in \Psi_\lambda(1)^{-1}\gamma_2(\lambda)\}.\end{aligned}$$
By Theorem \[main\], $M^T\mathcal{A}M$ is gap-continuous, and it follows from Lemma \[lemma-conjuagtion\] that $\mathcal{A}$ is gap-continuous as well. Moreover, we obtain from Theorem \[main\] and that
$$\operatorname{sf}(\mathcal{A})=\operatorname{sf}(M^T\mathcal{A}M)=\mu_{Mas}(\gamma_1,\Psi_{(\cdot)}(1)^{-1}\gamma_2)=\mu_{Mas}(\Psi\gamma_1,\gamma_2).$$
Note that we obtain from $\Psi$ and $\gamma_1$ further pairs of paths in $\Lambda(n)$ by
$$I\ni t\mapsto \Psi_0(t)\gamma_1(0)\in\Lambda(n),\quad I\ni t\mapsto\Psi_1(t)\gamma_1(1)\in\Lambda(n).$$ The following corollary is an easy reformulation of the previous theorem.
\[cor-Hamiltonian\] Under the previous assumptions,
$$\operatorname{sf}(\mathcal{A})=\mu_{Mas}(\Psi_1(\cdot)\gamma_1(1),\gamma_2(1))+\mu_{Mas}(\gamma_1,\gamma_2)-\mu_{Mas}(\Psi_0(\cdot)\gamma_1(0),\gamma_2(0)).$$
We consider the family $\Gamma:I\times I\rightarrow\Lambda(n)\times\Lambda(n)$ defined by $\Gamma(\lambda,t)=(\Psi_\lambda(t)\gamma_1(\lambda),\gamma_2(\lambda))$. We set
$$\eta_1(t)=\Gamma(0,t), \eta_2(\lambda)=\Gamma(\lambda,1), \eta_3(t)=\Gamma(1,1-t), \eta_4(\lambda)=\Gamma(1-\lambda,0).$$ As $I\times I$ is contractible, $\eta_1\ast\eta_2\ast\eta_3\ast\eta_4$ is homotopic to a constant path by a homotopy with fixed endpoints. Hence the Maslov index of $\eta_1\ast\eta_2\ast\eta_3\ast\eta_4$ vanishes by (viii’).\
As $\mu_{Mas}(\eta_4)=-\mu_{Mas}(\gamma_1,\gamma_2)$, $\mu_{Mas}(\eta_3)=-\mu_{Mas}(\Psi_1(\cdot)\gamma_1(1),\gamma_2(1))$ and $\Psi_\lambda(0)=I_{2n}$ for all $\lambda\in I$, it follows that
$$\mu_{Mas}(\Psi \gamma_1,\gamma_2)=-\mu_{Mas}(\Psi_0(\cdot)\gamma_1(0),\gamma_2(0))+\mu_{Mas}(\gamma_1,\gamma_2)+\mu_{Mas}(\Psi_1(\cdot)\gamma_1(1),\gamma_2(1)).$$ The corollary is now an immediate consequence of the previous theorem.
Note that if $S_0(t)=S_1(t)$ for all $t\in I$, then $\Psi_0(t)=\Psi_1(t)$, $t\in I$. If, moreover, $\gamma_1(0)=\gamma_1(1)$ and $\gamma_2(0)=\gamma_2(1)$, then we obtain from the previous corollary that
$$\operatorname{sf}(\mathcal{A})=\mu_{Mas}(\gamma_1,\gamma_2).$$ Consequently, under these assumptions the paths and have the same spectral flow and so the spectral flow of does not depend on the family of matrices $S$. Note that each $S_\lambda$ is $\mathcal{A}_\lambda$-compact, i.e. $S_\lambda:\mathcal{D}(\mathcal{A}_\lambda)\rightarrow L^2(I,\mathbb{R}^{2n})$ is compact with respect to the graph norm of $\mathcal{A}_\lambda$ on $\mathcal{D}(\mathcal{A}_\lambda)$. Let us point out that for closed paths of bounded selfadjoint Fredholm operators, the spectral flow is invariant under perturbations by compact selfadjoint operators (see [@FPR Prop. 3.8]).\
Let us now consider again the general setting of Corollary \[cor-Hamiltonian\], let $\alpha, \beta:[0,1]\rightarrow[0,1]$ be two continuous functions such that
$$\begin{aligned}
\label{alphabeta}
\beta(\lambda)=\alpha(\lambda)+\lambda,\quad \lambda\in[0,1].\end{aligned}$$
Our final result generalises Theorem 2 of [@Hu], where the following spectral flow formula was shown for a particular class of functions $\alpha, \beta$ that satisfy .
Under the assumptions of Corollary \[cor-Hamiltonian\],
$$\begin{aligned}
\operatorname{sf}(\mathcal{A})&=\mu_{Mas}(\Psi_0(\alpha(\cdot))\gamma_1(0),\Psi_0(\beta(\cdot))\Psi_0(1)^{-1}\gamma_2(0))+\mu_{Mas}(\gamma_1,\gamma_2)\\
&-\mu_{Mas}(\Psi_1(\alpha(\cdot))\gamma_1(1),\Psi_1(\beta(\cdot))\Psi_1(1)^{-1}\gamma_2(1)).\end{aligned}$$
We define maps $h_1,h_2:I\times I\rightarrow I$ by
$$h_1(s,\lambda)=(1-s)\alpha(\lambda)+s(1-\lambda),\qquad h_2(s,\lambda)=(1-s)\beta(\lambda)+s,\qquad$$ and consider for $i=1,2$ the homotopies
$$H_i:I\times I\rightarrow\Lambda(n)\times\Lambda(n),\quad H_i(s,\lambda)=(\Psi_i(h_1(s,\lambda))\gamma_1(i),\Psi_i(h_2(s,\lambda))\Psi_i(1)^{-1}\gamma_2(i)).$$ As $\alpha(0)=\beta(0)$, we see that $h_1(s,0)=h_2(s,0)$ and so
$$H_i(s,0)=(\Psi_i(h_1(s,0))\gamma_1(i),\Psi_i(h_2(s,0))\Psi_i(1)^{-1}\gamma_2(i))=(\gamma_1(i),\Psi_i(1)^{-1}\gamma_2(i))$$ is independent of $s$, where we have used (vi’). Moreover, since $\alpha(1)=0$, $\beta(1)=1$, and $\Psi_i(0)=I_{2n}$,
$$H_i(s,1)=(\Psi_i(0)\gamma_1(i),\Psi_i(1)\Psi_i(1)^{-1}\gamma_2(i))=(\gamma_1(i),\gamma_2(i)),$$ and so $H_i$ is a homotopy with fixed endpoints. Hence $\mu_{Mas}(H_i(0,\cdot))=\mu_{Mas}(H_i(1,\cdot))$ by (viii’) from above.\
Finally, we note that
$$\begin{aligned}
H_i(0,\lambda)&=(\Psi_i(\alpha(\lambda))\gamma_1(i),\Psi_i(\beta(\lambda))\Psi_i(1)^{-1}\gamma_2(i)),\\
H_i(1,\lambda)&=(\Psi_i(1-\lambda)\gamma_1(i),\Psi_i(1)\Psi_i(1)^{-1}\gamma_2(i))=(\Psi_i(1-\lambda)\gamma_1(i),\gamma_2(i))\end{aligned}$$
for all $\lambda\in I$, and
$$\mu_{Mas}(\Psi_i(1-\cdot)\gamma_1(i),\gamma_2(i))=-\mu_{Mas}(\Psi_i(\cdot)\gamma_1(i),\gamma_2(i)),$$ where we have used (vii’). Now the assertion of the corollary follows from Corollary \[cor-Hamiltonian\].
Finally, let us briefly point out that a version of the Morse Index Theorem in semi-Riemannian geometry from [@PejsachowiczGeod] can easily be derived from Theorem \[thm-Hamiltonian\] as well. We do not intend to explain the geometric content of the theorem, but just mention that it deals with non-trivial solutions of boundary value problems of the type
$$\label{Jacobi}
\left\{
\begin{aligned}
Ju'(t)+S_\lambda(t)u(t)&=0,\quad t\in I\\
u(0),u(1)\in\{0\}&\times\mathbb{R}^n
\end{aligned}
\right.$$
where $J$ is as in and $S_\lambda$ is again a family of symmetric $2n\times 2n$ matrices. If we consider the operators $\mathcal{A}_\lambda$ in for the equations , then
$$\begin{aligned}
\label{semiR}
\operatorname{sf}(\mathcal{A})=\mu_{Mas}(\Psi(\{0\}\times\mathbb{R}^n),\{0\}\times\mathbb{R}^n)\end{aligned}$$
by Theorem \[thm-Hamiltonian\], where $\Psi=\{\Psi_\lambda(1)\}_{\lambda\in I}$ is the path in $\operatorname{Sp}(2n,\mathbb{R})$ obtained as in . This is Proposition 6.1 in [@PejsachowiczGeod]. Note that in this setting the path $\mathcal{A}=\{\mathcal{A}_\lambda\}_{\lambda\in I}$ has the constant domain $\mathcal{D}(\mathcal{A}_\lambda)=\{u\in H^1(I,\mathbb{R}^{2n}):\, u(0),u(1)\in \{0\}\times\mathbb{R}^n\}$, which allows to compute its spectral flow by crossing forms (see [@Robbin] and [@WaterstraatHomoclinics]) and yields the different proof of given in [@PejsachowiczGeod].
V.I. Arnold, **On a characteristic class entering into conditions of quantization**, Funkcional. Anal. i Prilozen. **1**, **1967**, 1–14
M.F. Atiyah, V.K. Patodi, I.M. Singer, **Spectral asymmetry and Riemannian geometry III**, Math. Proc. Cambridge Philos. Soc. **79**, 1976, 71–99
B. Booss-Bavnbek, M. Lesch, J. Phillips, **Unbounded Fredholm operators and spectral flow**, Canad. J. Math. **57**, 2005, 225–250
R. Bott, **On the iteration of closed geodesics and the Sturm intersection theory**, Comm. Pure Appl. Math. **9**, 1956, 171–206
S.E. Cappell, R. Lee, E.Y. Miller, **On the Maslov index**, Comm. Pure Appl. Math. **47**, 1994, 121–186
J.J. Duistermaat, **On the Morse index in variational calculus**, Advances in Math. **21**, 1976, 173–195
P.M. Fitzpatrick, J. Pejsachowicz, L. Recht, **Spectral Flow and Bifurcation of Critical Points of Strongly-Indefinite Functionals-Part I: General Theory**, J. Funct. Anal. **162**, 1999, 52–95
V. Guillemin, S. Sternberg, **Geometric asymptotics**, Mathematical Surveys **14**, American Mathematical Society, Providence, R.I., 1977
X. Hu, A. Portaluri, **Index theory for heteroclinic orbits of Hamiltonian systems**, Calc. Var. Partial Differential Equations **56**, 2017
T. Kato, **Perturbation theory for linear operators**, Reprint of the 1980 edition, Classics in Mathematics, Springer-Verlag, Berlin, 1995
M. Lesch, **The uniqueness of the spectral flow on spaces of unbounded self-adjoint Fredholm operators**, Spectral geometry of manifolds with boundary and decomposition of manifolds, 193–224, Contemp. Math., 366, Amer. Math. Soc., Providence, RI, 2005
M. Musso, J. Pejsachowicz, A. Portaluri, **A Morse Index Theorem for Perturbed Geodesics on Semi-Riemannian Manifolds**, Topol. Methods Nonlinear Anal. **25**, 2005, 69-99
L. Nicolaescu, **On the space of Fredholm operators**, An. Stiint. Univ. Al. I. Cuza Iasi. Mat. (N.S.) **53**, 2007, 209–227
J. Pejsachowicz, N. Waterstraat, **Bifurcation of critical points for continuous families of $C^2$-functionals of Fredholm type**, J. Fixed Point Theory Appl. **13**, 2013, 537–560
J. Phillips, **Self-adjoint Fredholm operators and spectral flow**, Canad. Math. Bull. **39**, 1996, 460–467
P. Piccione, D. V. Tausk, **A student’s guide to symplectic spaces, Grassmannians and Maslov index**, IMPA Mathematical Publications, Rio de Janeiro, 2008
J. Robbin, D. Salamon, **The Maslov index for paths**, Topology **32**, 1993, 827–844
J. Robbin, D. Salamon, **The spectral flow and the Maslov index**, Bull. London Math. Soc. **27**, 1995, 1–33
D. Salamon, E. Zehnder, **Morse theory for periodic solutions of Hamiltonian systems and the Maslov index**, Comm. Pure Appl. Math. **45**, 1992, 1303–1360
C. Wahl, **A new topology on the space of unbounded selfadjoint operators, $K$-theory and spectral flow**, $C^*$-algebras and elliptic theory II, 297–309, Trends. Math., Birkh., Basel, 2008
N. Waterstraat, **Fredholm Operators and Spectral Flow**, Rend. Semin. Mat. Univ. Politec. Torino **75**, 2017, 7–51
N. Waterstraat, **Spectral flow, crossing forms and homoclinics of Hamiltonian systems**, Proc. Lond. Math. Soc. (3) **111**, 2015, 275–304
Marek Izydorek\
Gdansk University of Technology\
Narutowicza 11/12\
80-233 Gdansk\
Poland\
E-mail: [email protected]
Joanna Janczewska\
Gdansk University of Technology\
Narutowicza 11/12\
80-233 Gdansk\
Poland\
E-mail: [email protected]
Nils Waterstraat\
School of Mathematics,\
Statistics & Actuarial Science\
University of Kent\
Canterbury\
Kent CT2 7NF\
UNITED KINGDOM\
E-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider a partially hinged rectangular plate and its normal modes. The dynamical properties of the plate are influenced by the spectrum of the associated eigenvalue problem. In order to improve the stability of the plate, it seems reasonable to place a certain amount of stiffening material in appropriate regions. If we look at the partial differential equation appearing in the model, this corresponds to insert a suitable weight coefficient inside the equation. A possible way to locate such regions is to study the eigenvalue problem associated to the aforementioned weighted equation. In the present paper we focus our attention essentially on the first eigenvalue and on its minimization in terms of the weight. We prove the existence of minimizing weights inside special classes and we try to describe them together with the corresponding eigenfunctions.'
address:
-
-
-
-
author:
- Elvise BERCHIO
- Alessio FALOCCHI
- Alberto FERRERO
- Debdip GANGULY
title: |
On the first frequency of reinforced\
partially hinged plates
---
Introduction {#1}
============
Following [@fergaz] one may view a bridge as a long narrow rectangular thin plate $\Omega$ hinged at two opposite edges and free on the remaining two edges: this plate well describes decks of footbridges and suspension bridges which, at the short edges, are supported by the ground. We refer to the monograph [@bookgaz] for a detailed survey either of old and new mathematical models for suspension bridges. Up to scaling, we may assume that the plate has length $\pi$ and width $2\ell$ with $2\ell\ll\pi$ so that $$\Omega=(0,\pi)\times(-\ell,\ell)\subset{{\mathbb R}}^2\, .$$ There is a growing interest of engineers on the shape optimization for the design of bridges and, in particular, on the sensitivity analysis of certain eigenvalue problems, see [@jhnm Chapter 6]. As pointed out by Banerjee [@banerjee], [*the free vibration analysis is a fundamental pre-requisite before carrying out a flutter analysis*]{}. Whence, in the the stability analysis of the plate a central role is played by the following eigenvalue problem: $$\label{weight0}
\begin{cases}
\Delta^2 u=\lambda\, u & \qquad \text{in } \Omega \\
u(0,y)=u_{xx}(0,y)=u(\pi,y)=u_{xx}(\pi,y)=0 & \qquad \text{for } y\in (-\ell,\ell) \\
u_{yy}(x,\pm\ell)+\sigma
u_{xx}(x,\pm\ell)=u_{yyy}(x,\pm\ell)+(2-\sigma)u_{xxy}(x,\pm\ell)=0
& \qquad \text{for } x\in (0,\pi)\, ,
\end{cases}$$ where $\sigma$ denotes the Poisson ratio of the material forming the plate. For most elastic materials one has $0<\sigma<0.5$; since we aim to model the deck of a bridge, which is a mixture of concrete and steel, one may take $\sigma=0.2$. The boundary conditions on the short edges tell that the plate is hinged; these conditions are named Navier since their first appearance in [@navier]. We refer to [@bebuga2] for the derivation of from the total energy of the plate. Note that in [@fergaz] the whole spectrum of was determined, while in [@bfg] the results were exploited to study the so-called torsional stability of suspension bridges for small energies. Furthermore, in [@bebuga2] the variation of the eigenvalues, under domain deformations, which may not preserve the area, was investigated, see also [@lamberti] for related results about Dirichlet polyharmonic eigenvalue problems.
In order to improve the stability of the plate, one may think to place a certain amount of stiff material within the plate. In mathematical terms this can be modelled by inserting into the equation a weight $p$, properly chosen to describe the action of the reinforcement and we end up with the *weighted* eigenvalue problem: $$\label{weight}
\begin{cases}
\Delta^2 u=\lambda\, p(x,y) u & \qquad \text{in } \Omega \\
u(0,y)=u_{xx}(0,y)=u(\pi,y)=u_{xx}(\pi,y)=0 & \qquad \text{for } y\in (-\ell,\ell)\\
u_{yy}(x,\pm\ell)+\sigma
u_{xx}(x,\pm\ell)=u_{yyy}(x,\pm\ell)+(2-\sigma)u_{xxy}(x,\pm\ell)=0
& \qquad \text{for } x\in (0,\pi)\, ,
\end{cases}$$ where, for $0<\alpha\leq\beta$ fixed, $p$ belongs to the following family of weights
$$\label{eq:famiglia}
P_{\alpha, \beta}:=\left\{p\in L^\infty(\Omega): \alpha{\leqslant}p\leq\beta \ \text{a.e. in } \Omega \
\text{ and }\int_{\Omega}p\,dxdy=|\Omega| \, \right\} \, .$$
The spectral analysis of should indicate where to place the stiff material within the plate. In this respect, the condition on the integral of $p$ is posed in order to make the comparison with the case $p\equiv 1$ consistent. It’s worth mentioning that a related linear problem has been recently treated in [@bbgza], by studying the equation $$\Delta^2 u=\frac{f(x,y)}{1+d\chi_D(x,y)}\,\quad \text{in } \Omega$$ subject to the boundary conditions in , where $\chi_D$ is the characteristic function of $D \subset \Omega$ and $d>0$ is a constant measuring the strength of the stiffening material. The solution $u$ of this equation describes the vertical displacement of the plate under the action of a load $f$ while the weight $p$ is here explicitly given by $p(x,y)=1/(1+d\chi_D(x,y))$. In particular, $p$ can be seen as an “aerodynamic damper” placed in $D$ in order to reduce the action of the external force $f$. Hence, the lowest are the values of $p$ in some region of the rectangle $\Omega$, the highest is the amount of stiffening material placed in that region, and the lower bound $p{\geqslant}\alpha>0$ in appears reasonable since it corresponds to an upper bound on the rigidity of the plate. The spectral analysis of can help to complete and enrich the results obtained in [@bbgza].
Coming back to , the natural starting point of the study is the investigation of the effect of $p$ on the fundamental frequency $\lambda_1(p)$, namely to study: $$\inf_{p \in P_{\alpha, \beta}} \, \lambda_1(p).$$ When $ \lambda_1(p)$ is the first weighted eigenvalue of $-\Delta$ under Dirichlet boundary conditions, the above problem coincides with the so-called *composite membrane* problem, see [@chanillo1]-[@chanillo3],[@sha], while if $ \lambda_1(p)$ is the first weighted eigenvalue of $\Delta^2$ under Dirichlet or Navier boundary conditions, it becomes the *composite plate* problem, see [@anedda1],[@anedda2],[@CV]-[@cuccu2]. In this field of research, typical results are existence of optimal pairs and their qualitative properties, such as symmetry or symmetry breaking. From this point of view a crucial obstruction, when passing from the membrane to the plate problem, namely from the second to the fourth order case, is represented by the loss of maximum and comparison principles which usually enter either in the study of the simplicity of the first eigenvalue and in the techniques applied to prove symmetry results, such as reflections methods or moving planes techniques. Nevertheless, a suitable choice of the boundary conditions (e.g. Navier or Steklov b.c.) or of the geometry of the domain (e.g. small perturbations of balls) may yield the validity of so-called *positivity preserving property* which basically means that solutions, of the associated linear problem, maintain the sign of data. This property generally allows to extend some of the results known in the second order to the higher order case. As concerns problem , the difficulties when passing to the higher order, are even increased by the choice of the unusual boundary conditions for which no positivity preserving property is known. Note that, problem with $p\not \equiv 1$ has never been studied in literature, hence the present paper represents the first contribution on this topic. However, in our analysis we take advantage of the fact that $\Omega$ is a planar domain and, when restricting the class of weights, some explicit computations can be performed. On the other hand, we exploit a sort of restricted positivity preserving property with respect to the $y$ variable, that we prove in Theorem \[corPPP\] below, having its own theoretical interest. We note that the above mentioned restriction on admissible weights is also justified by the applicative origin of our problem. Indeed, it is known that minimization problems, like the composite membrane problem, naturally lead to homogenization [@murat], see also [@ksw] for a stiffening problem for the torsion of a bar. Homogenization would lead to optimal designs with reinforcements scattered throughout the structure, namely designs impossible to implement for engineers. Hence, to avoid homogenization, the class of admissible reinforcements should be sufficiently small. See also Nazarov-Sweers-Slutskij [@nazarov], where only “macro” reinforcements are considered, although in a fairly different setting.
The paper is organised as follows. Section \[setting\] is devoted to the description of the notations and of some results about the case $p\equiv 1$. In Section \[s:main\] one can find the main results of the paper which are proved in Sections \[s:proofs-1\] and \[spectrum\]. In Section \[numerics\] we show some numerical results on the behaviour of the eigenvalues which complement our theoretical analysis. Finally, in Section \[pppproof\] we show the validity of a positivity preserving property for a one dimensional fourth order problem, coming from a suitable Fourier decomposition of solutions to the plate problem.
Notations and known results when $p\equiv 1$ {#setting}
============================================
The natural functional space where to set problem is $$H^2_*(\Omega)=\big\{u\in H^2(\Omega): u=0\mathrm{\ on\ }\{0,\pi\}\times(-\ell,\ell)\big\}\,.$$ For any $\sigma \in (0,1)$, $H^2_*(\Omega)$ is a Hilbert space when endowed with the scalar product $$(u,v)_{H^2_*}:=\int_\Omega \left[\Delta u\Delta v+(1-\sigma)(2u_{xy}v_{xy}-u_{xx}v_{yy}-u_{yy}v_{xx})\right]\, dx \, dy \,$$ and associated norm $$\|u\|_{H^2_*(\Omega)}^2=(u,u)_{H^2_*(\Omega)} \, ,$$ which is equivalent to the usual norm in $H^2(\Omega)$, see [@fergaz Lemma 4.1]. From now onward we assume $\sigma \in (0,1)$ fixed. Then problem may also be formulated in the following weak sense $$\label{eigenweak}
(u,v)_{H^2_*(\Omega)} =\lambda\int_{\Omega}p(x,y)uv\,dx\,dy \qquad\forall v\in H^2_*(\Omega),$$ where, for $0<\alpha {\leqslant}\beta$ fixed, $p$ belongs to the family of weights $P_{\alpha,\beta}$ defined in . Clearly, the constant weight $p\equiv 1$ belongs to the family $P_{1,1}$. Since the bilinear form $(u,v)_{H^2_*}$ is continuous and coercive and $p\in L^\infty(\Omega)$ is positive a.e. in $\Omega$, standard spectral theory of self-adjoint operators then shows that the eigenvalues of may be ordered in an increasing sequence of strictly positive numbers diverging to $+\infty$ and that the corresponding eigenfunctions form a complete system in $H^2_*(\Omega)$.
Since $p\in L^{\infty}(\Omega),$ by elliptic regularity the eigenfunctions are at least in $C^{2}(\overline \Omega).$ Furthermore, the first eigenvalue is characterized by $$\label{lambdaP}
\lambda_1(p) := \, \inf_{u \in H^2_*(\Omega) \setminus\{0\}} \frac{\|u\|_{H^2_*}^2}{\|\sqrt{p}\,u\|_{2}^2}.$$
When $p\equiv 1$ the spectrum of has been completely characterized. We recall the following statement from [@fergaz], including some refinements on the eigenvalues estimates proved in [@bebuga2].
\[eigenvalue\] Let $p\equiv 1$ in . The set of eigenvalues of may be ordered in an increasing sequence of strictly positive numbers diverging to $+\infty$ and any eigenfunction belongs to $C^\infty(\overline\Omega)$; the set of eigenfunctions of is a complete system in $H^2_*(\Omega)$. Moreover:
$(i)$ for any $m\ge1$, there exists a unique eigenvalue $\lambda=\mu_{m,1}\in((1-\sigma^2) m^4,m^4)$ with corresponding eigenfunction $$\left[\big[\mu_{m,1}^{1/2}-(1-\sigma)m^2\big]\, \tfrac{\cosh\Big(y\sqrt{m^2+\mu_{m,1}^{1/2}}\Big)}{\cosh\Big(\ell\sqrt{m^2+\mu_{m,1}^{1/2}}\Big)}+
\big[\mu_{m,1}^{1/2}+(1-\sigma)m^2\big]\, \tfrac{\cosh\Big(y\sqrt{m^2-\mu_{m,1}^{1/2}}\Big)}{\cosh\Big(\ell\sqrt{m^2-\mu_{m,1}^{1/2}}\Big)}\right]\sin(mx)\, ;$$ $(ii)$ for any $m\ge1$ and any $k\ge2$ there exists a unique eigenvalue $\lambda=\mu_{m,k}>m^4$ satisfying
$\left(m^2+\frac{\pi^2}{\ell^2}\left(k-\frac{3}{2}\right)^2\right)^2<\mu_{m,k}<\left(m^2+\frac{\pi^2}{\ell^2}\left(k-1\right)^2\right)^2$
and with corresponding eigenfunction $$\left[\big[\mu_{m,k}^{1/2}-(1-\sigma)m^2\big]\, \tfrac{\cosh\Big(y\sqrt{\mu_{m,k}^{1/2}+m^2}\Big)}{\cosh\Big(\ell\sqrt{\mu_{m,k}^{1/2}+m^2}\Big)}
+\big[\mu_{m,k}^{1/2}+(1-\sigma)m^2\big]\, \tfrac{\cos\Big(y\sqrt{\mu_{m,k}^{1/2}-m^2}\Big)}{\cos\Big(\ell\sqrt{\mu_{m,k}^{1/2}-m^2}\Big)}\right]\sin(mx)\, ;$$
$(iii)$ for any $m\ge1$ and any $k\ge2$ there exists a unique eigenvalue $\lambda=\nu_{m,k}>m^4$ with corresponding eigenfunctions $$\left[\big[\nu_{m,k}^{1/2}-(1-\sigma)m^2\big]\, \tfrac{\sinh\Big(y\sqrt{\nu_{m,k}^{1/2}+m^2}\Big)}{\sinh\Big(\ell\sqrt{\nu_{m,k}^{1/2}+m^2}\Big)}
+\big[\nu_{m,k}^{1/2}+(1-\sigma)m^2\big]\, \tfrac{\sin\Big(y\sqrt{\nu_{m,k}^{1/2}-m^2}\Big)}{\sin\Big(\ell\sqrt{\nu_{m,k}^{1/2}-m^2}\Big)}\right]\sin(mx)\, ;$$ $(iv)$ for any $m\ge1$ satisfying $\ell m\sqrt 2\, \coth(\ell m\sqrt2 )>\left(\frac{2-\sigma}{\sigma}\right)^2$ there exists a unique eigenvalue $\lambda=\nu_{m,1}\in(\mu_{m,1},m^4)$ with corresponding eigenfunction $$\left[\big[\nu_{m,1}^{1/2}-(1-\sigma)m^2\big]\, \tfrac{\sinh\Big(y\sqrt{m^2+\nu_{m,1}^{1/2}}\Big)}{\sinh\Big(\ell\sqrt{m^2+\nu_{m,1}^{1/2}}\Big)}
+\big[\nu_{m,1}^{1/2}+(1-\sigma)m^2\big]\, \tfrac{\sinh\Big(y\sqrt{m^2-\nu_{m,1}^{1/2}}\Big)}{\sinh\Big(\ell\sqrt{m^2-\nu_{m,1}^{1/2}}\Big)}\right]
\sin(mx)\, .$$ Finally, if $$\label{c0}
\text{the unique positive solution $s>0$ of: }\tanh(\sqrt{2}s\ell)=\left(\frac{\sigma}{2-\sigma}\right)^2\, \sqrt{2}s\ell\quad \text{is not an integer,}$$ then the only eigenvalues are the ones given in $(i)-(iv)$.
In the following, to avoid too many distinctions, we will always assume that holds.
By Proposition \[eigenvalue\] and [@fergaz Section 7] it is readily deduced that the first eigenvalue of problem with $p\equiv1$ is $\mu_{1,1}$, namely $ \lambda_1(1)=\mu_{1,1}$, it is simple and up to constant multiplier the first eigenfunction is given by $$\label{u1}
u_1(x,y)= \left[\big[\mu_{1,1}^{1/2}-(1-\sigma)\big]\, \tfrac{\cosh\Big(y\sqrt{1+\mu_{1,1}^{1/2}}\Big)}{\cosh\Big(\ell\sqrt{1+\mu_{1,1}^{1/2}}\Big)}+
\big[\mu_{1,1}^{1/2}+(1-\sigma)\big]\, \tfrac{\cosh\Big(y\sqrt{1-\mu_{1,1}^{1/2}}\Big)}{\cosh\Big(\ell\sqrt{1-\mu_{1,1}^{1/2}}\Big)}\right] \sin x\,.$$ Hence, $u_1$ is positive in $\Omega$, convex in the $y-$variable and concave in the $x-$variable.
Main results {#s:main}
============
Let $0<\alpha<\beta$ be two fixed constants and let $P_{\alpha,\beta}$ be the class of admissible weights defined in Section \[1\]. Then, clearly $\alpha{\leqslant}1$ and $\beta {\geqslant}1.$ Recalling , we focus on the double infimum problem $$\label{CP}
\lambda_{\alpha,\beta} :=\inf_{p \in P_{\alpha,\beta}} \, \lambda_1(p)= \inf_{p \in P_{\alpha,\beta}} \, \inf_{u \in H^2_*(\Omega) \setminus\{0\}}
\frac{\|u\|_{H^2_*(\Omega)}^2}{\|\sqrt{p}\,u\|_{2}^2}.$$
A couple $ (u_{\overline p}, \overline{p}) \in H^2_*(\Omega) \times P_{\alpha,\beta}$ which realises the double infimum in is called an [*optimal pair*]{}.
Adapting to our case [@chanillo Theorem 13] and [@CV Theorem 1.4], it can be shown that there exists an [optimal pair]{} $ (u_{\overline p}, \overline{p})$ for problem and $u_{\overline p}$ and $\overline{p}$ are suitably related.
\[thm-exist-qual\] For every $0<\alpha<\beta$, there exists and optimal pair $(u_{\overline p}, \overline{p}) \in H^2_*(\Omega) \times P_{\alpha,\beta}$. Furthermore, $u_{\overline p}$ and $\overline{p}$ are related as follows $$\label{pS}
\overline{p}(x,y) = \alpha \chi_{S} (x,y)+ \beta \chi_{\Omega \setminus S}(x,y)\,\quad \text{for a.e. }\quad (x,y)\in \Omega\,,$$ where $\chi_{S}$ and $ \chi_{\Omega \setminus S}$ are the characteristic functions of the sets $S$ and $\Omega \setminus S$ and $S\subset \Omega$ is such that $|S|=\frac{\beta-1}{\beta-\alpha}\,|\Omega|$ and $S = \{ (x,y)\in \Omega\,:\,u_{\overline p}^2(x,y) {\leqslant}t \}$ for some $t{\geqslant}0$.
Note that since $\Omega$ is planar, $u_{\overline p}\in C^0(\overline{\Omega})$ and the set $S$ is closed. The above result suggests that the plate can be made out of two materials but it gives no informations about the location of the materials and hence, no practical informations on how to built the plate. To this aim, a more explicit suggestion, even if more rought, is provided by the following
\[comparison\_esy\] Let $0<\alpha<\beta$ and $p\in P_{\alpha,\beta}$ satisfy one of the following assumptions
- $p=p(y)$ is even and there exists $z\in (0, \ell)$ such that $$p(y){\leqslant}1 \quad \text{for } y\in [0,z] \quad \text{ and } \quad p(y){\geqslant}1 \quad \text{for } y\in [z,\ell)\,.$$
- $p=p(x)$ is symmetric with respect to the line $x=\frac{\pi}{2}$ and there exists $s \in (0, \frac{\pi}{2})$ such that $$p(x){\leqslant}1 \quad \text{for } x\in (0,s] \quad \text{ and } \quad p(x){\geqslant}1 \quad \text{for } x\in [s,\frac{\pi}{2}]\,.$$
Then, $$\label{comparison}
\lambda_1(p){\leqslant}\lambda_1(1)=\mu_{1,1}\,,$$ where the $\mu_{1,1}$ is as defined in Proposition \[eigenvalue\]-(i).
It’s worth noting that the same idea of the proof of Proposition \[comparison\_esy\]-(i) can be repeated to prove that holds if $p\in P_{\alpha,\beta}$ satisfies
- $p=p(y)$ is even and there exist $2N+2$ points $0=y_0<y_1<y_2<...<y_{2N+2}=\ell$ such that $$p(y){\leqslant}1 \quad \text{for } y\in [y_{2h}, y_{2h+1}]\,, \quad p(y){\geqslant}1 \quad \text{for } y\in [y_{2h+1}, y_{2h+2}] \quad \text{ and } \int_{y_{2h}}^{y_{2h+2}}(p-1)\, dy=0 \,,$$ for all $h=0,...,N$.
Since the weights considered in Proposition \[comparison\_esy\] prove to be effective in decreasing the first frequency of , by combining Proposition \[comparison\_esy\] with Theorem \[thm-exist-qual\], it is reasonable to include in the list of candidate solutions to problem the weights: $$\label{opt}
\overline{p}(y) = \alpha \chi_{(-\frac{\ell(\beta-1)}{\beta-\alpha},\frac{\ell(\beta-1)}{\beta-\alpha})}(y)
+ \beta \chi_{ (-\ell, \ell)\setminus (-\frac{\ell(\beta-1)}{\beta-\alpha},\frac{\ell(\beta-1)}{\beta-\alpha})}(y) \qquad y\in(-\ell, \ell)$$ and $$\label{opt2}
\overline{p}(x)=\beta \chi_{(\frac{\pi}{2} \frac{1-\alpha}{\beta-\alpha},\frac{\pi}{2} \frac{2\beta-1-\alpha}{\beta-\alpha})}(x)
+\alpha \chi_{ (0,\pi)\setminus (\frac{\pi}{2} \frac{1-\alpha}{\beta-\alpha},\frac{\pi}{2} \frac{2\beta-1-\alpha}{\beta-\alpha})}(x)\qquad x\in(0, \pi) \, .$$
[![On the left, plot of the eigenfunction $u_{1,\overline p}^2(x,y)$, corresponding to $\lambda_1(\overline p)$ with $\overline{p}(y)$ as in , intersected with $t>0$. On the right, plot of $\overline{p}(y)$ (top) and plot of the sublevel set $S = \{ (x,y)\in \Omega\,:\,u_{1,\overline p}^2(x,y) {\leqslant}t \}$ (bottom).[]{data-label="planar"}](eig22.jpg "fig:")]{}
In Section \[numerics\] we obtained numerically a positive eigenfunction, denoted by $u_{1,\overline p}(x,y)$, corresponding to $\lambda_1(\overline p)$ with $\overline{p}(y)$ as in . In Figure \[planar\] on the left, we plot $z=u_{1,\overline p}^2(x,y)$ and we use it to determine qualitatively what should be the set $S$ predicted by Theorem \[thm-exist-qual\]. A comparison between the weight $\overline p(x,y)$ in , with this choice of the set $S$, and the weight $\overline{p}(y)$ in is shown in Figure \[planar\] on the right. From these plots we infer that $(u_{1,\overline p}(x,y),\overline{p}(y))$ cannot belong to a theoretical optimal pair of .
On the other hand, when restricting the class of admissible weights to a suitable subset of $P_{\alpha,\beta}$, in Theorem \[min\] below we prove that indeed $\overline{p}(y)$ belongs to an optimal pair provided that the constant $\beta$ satisfies a suitable upper bound. Note that the numerical results we state in Section \[numerics\] suggests that this upper bound is merely a technical condition.
\[min\] Let $0<\alpha<\beta<\min\{1/\mu_{1,1}\,, (1-\sigma^2) 2^4\}$ and denote
$$\begin{aligned}
& \overline P_{\alpha,\beta} =\{p\in P_{\alpha,\beta}\,: \, p=p(y) \text{ is even, } p \text{ is piecewise continuous in } (-\ell,\ell) \\[8pt]
& \qquad \text{ and }
\exists \, z\in (0, \ell)\,: p(y){\leqslant}1 \text{ in } [0,z]\,, \, p(y){\geqslant}1 \text{ in } [z,\ell)\} \, .\end{aligned}$$
The following statements hold:
- if $p_1, p_2\in \overline P_{\alpha,\beta}$ and there exists $z \in (0, \ell)$ such that $$p_1(y){\leqslant}p_2(y) \quad \text{in } [0,z]\quad \text{ and }\quad p_1(y){\geqslant}p_2(y) \quad \text{in } [z,\ell)\,,$$ then $$\lambda_1(p_1){\leqslant}\lambda_1(p_2)\,;$$
- we have
$$\min_{p\in \overline P_{\alpha,\beta}} \lambda_1(p)=\lambda_1(\overline p)\,,$$ where $\overline p$ is as defined in .
It is worth noting that, in order to lower the first eigenvalue of $\Delta^2$ under Dirichlet or Navier boundary conditions, since the eigenfunctions vanish on the boundary, one expects that the weight is more effective if it achieves its lowest value close to the boundary, see e.g. [@CV Theorem 1.5]. Theorem \[min\] shows that the partially hinged boundary conditions lead to a complete different situation since the weight $\overline p(y)$ achieves its lowest value $\alpha$ far from the free long edges, see Figure \[planar\] on the right (top). This behaviour is somehow related to the monotonicity of the first eigenfunction, as shown by Theorem \[signTH\] below, cfr. Figure \[eig1\].
\[signTH\] Let $0<\alpha<\beta<\min\{1/\mu_{1,1}\,,(1-\sigma^2) 2^4\}$ and let $\overline P_{\alpha,\beta}$ be the family of weights defined in Theorem \[min\]. Then, for any $p\in \overline P_{\alpha,\beta}$ the first eigenvalue $\lambda_1(p)$ of is simple. Furthermore, if $u_{1,p}$ is an eigenfunction of $\lambda_1(p)$ then $u_{1,p}$ is of one sign in $\Omega$ and moreover $u_{1,p}$ can be written as $u_{1,p}(x,y)=\varphi_{1,p}(y)\sin (x)$ with $\varphi_{1,p}(y)$ even and strictly monotone in $(0,\ell)$.
[![Qualitative plot of $u_{1,p}(x,y)=\varphi_{1,p}(y)\sin (x)$.[]{data-label="eig1"}](eig1.jpg "fig:")]{}
Unfortunately, the above statement does not carry over to all weights $p\in P_{\alpha,\beta}$. This is related to the well-know loss of comparison principles for higher order elliptic operators. Indeed, the proof of Theorem \[signTH\] highly relies on a sort of restricted positivity preserving property with respect to the $y$ variable that we prove by separating variables. More precisely, we have
\[corPPP\] Let $m{\geqslant}1$ be an integer and $\sigma\in(0,1)$. Furthermore, let $u\in H^2_*(\Omega)$ be a weak solution to the problem $$\label{linear}
\begin{cases}
\Delta^2 u= f(y)\, \sin(mx) & \qquad \text{in } \Omega\, \\
u(0,y)=u_{xx}(0,y)=u(\pi,y)=u_{xx}(\pi,y)=0 & \qquad \text{for } y\in (-\ell,\ell)\,\\
u_{yy}(x,\pm\ell)+\sigma
u_{xx}(x,\pm\ell)=u_{yyy}(x,\pm\ell)+(2-\sigma)u_{xxy}(x,\pm\ell)=0
& \qquad \text{for } x\in (0,\pi)\, ,
\end{cases}$$ namely $$(u,v)_{H^2_*} =\int_{\Omega} f(y)\sin(mx)\, v \qquad\forall v\in H^2_*(\Omega)\,.$$ Then, $u(x,y)=w_m(y)\sin(mx)$ and the following implication holds $$f{\geqslant}0\text{ in } (-\ell, \ell)\,(f\not \equiv 0)\quad \Rightarrow \quad w_m(y)>0 \text{ in } [-\ell, \ell]\,.$$
Proof of Theorem \[thm-exist-qual\] and Proposition \[comparison\_esy\] {#s:proofs-1}
=======================================================================
Proof of Theorem \[thm-exist-qual\]
-----------------------------------
We start with the existence issue.
\[exist\] For every $0<\alpha<\beta$, the double infimum in is achieved.
Let $\{ p_m \}_m \subset P_{\alpha,\beta}$ be a minimizing sequence for $ \lambda_{\alpha,\beta}$, i.e.
$$\lambda_1(p_m)= \lambda_{\alpha,\beta}+o(1) \quad \text{as } m \rightarrow \infty\,.$$ Let now $u_{p_m} \in H^2_*(\Omega)$ be a (normalized) eigenfunction to $\lambda_1(p_m)$, namely $\lambda_1(p_m) = ||u_{p_m}||^2_{H^2_*(\Omega)}$ and $\int_{\Omega} p_m \, u_{p_m}^2 \, dx\,dy = 1.$ This immediately implies $||u_{p_m}||_{H^2_*} {\leqslant}C,$ for some positive constant $C.$ Therefore, using the compact embedding of $H^2_*(\Omega) \hookrightarrow L^2(\Omega),$ we can extract two subsequences, still denoted by $u_{p_m},$ such that
$$u_{p_m} \rightharpoonup \overline{u} \quad \mbox{in} \ H^2_*(\Omega) \quad \mbox{as} \ m \rightarrow \infty,$$
$$u_{p_m} \rightarrow \overline{u} \quad \mbox{in} \ L^2(\Omega) \quad \mbox{as} \ m \rightarrow \infty.$$
Moreover, $p_m \in P_{\alpha,\beta}$ implies $||p_{m}||_{L^{\infty}(\Omega)} {\leqslant}\beta$ and therefore up to a subsequence, $p_m \rightharpoonup \overline{p}$ in $L^2(\Omega)$ as $m \rightarrow \infty.$ By this we have that $|\Omega|=\int_{\Omega} p_m\,1\,dx\,dy=\int_{\Omega} \overline{p}\, 1\,dx\,dy+o(1)$ as $m \rightarrow \infty$ and, since strongly closed convex sets are weakly closed, that $\alpha {\leqslant}\overline{p}{\leqslant}\beta$ a.e. in $\Omega$. Hence, $\overline{p} \in P_{\alpha, \beta}$. On the other hand, we obtain
$$\begin{aligned}
\left| \int_{\Omega}( p_m \, u_{p_m}^2 - \overline{p} \,\overline{u}^2 ) dx\, dy \right|
& = \left| \int_{\Omega} p_m ( u_{p_m}^2 - \overline{u}^2 ) \, dx\, dy + \int_{\Omega} \overline{u}^2( p_m - \overline{p}) \, dx\, dy \right| \notag \\
& {\leqslant}\beta \int_{\Omega} |(u_{p_m}-\overline{u})(u_{p_m}+\overline{u}) | \, dx\, dy + \| \overline{u}\|_{L^{\infty}(\Omega)}
\int_{\Omega} |\overline{u}|\, |p_m - \overline{p}| \, dx\, dy\\
&=2\beta \, \| \overline{u}\|_{ L^2(\Omega)}\,\|u_{p_m}- \overline{u}\|_{ L^2(\Omega)}+ o(1)=o(1) \quad \mbox{as} \ m \rightarrow \infty\,, \notag\end{aligned}$$
where we have exploited the fact that $H^2_*(\Omega) \subset L^{\infty}(\Omega)$ since $\Omega$ is a planar domain. Hence, we conclude that $\int_{\Omega} \overline{p} \, \overline{u}^2 \, dx\, dy = 1.$ Furthermore,
$$\lambda_1(\overline{p}) {\leqslant}||\overline{u}||^2_{H^2_*} {\leqslant}\liminf_{m \rightarrow \infty} ||u_{p_m}\|_{H^2_*}^2 = \lambda_{\alpha,\beta} \, .$$ Hence
$$\lambda_{\alpha,\beta} {\leqslant}\lambda_1(\overline{p}) = ||\overline{u}||^2_{H^2_*} {\leqslant}\lambda_{\alpha,\beta}.$$ Therefore, the couple $(\overline{p},\overline{u})$ is an optimal pair. Hence, $u_{\overline{p}}=\overline{u}$ and this completes the proof.
To problem we associate the following double infimum problem
$$\label{ACP}
\Lambda_{\alpha,\beta} := \inf_{\eta \in N_{\alpha,\beta}} \, \inf_{u \in H^2_*(\Omega)\setminus \{0\}}
\frac{||u||_{H^2_*(\Omega)}^2 \, + \, \lambda_{\alpha,\beta}(\beta-\alpha) \int_{\Omega} \eta \, u^2 \, dx\, dy}{\int_{\Omega} u^2 \, dx\, dy},$$
where $\lambda_{\alpha,\beta}$ is as in and
$$N_{\alpha,\beta} = \left\{ \eta\in L^\infty(\Omega):\, 0 {\leqslant}\eta {\leqslant}1 \ \text{a.e. in } \Omega \ \mbox{and} \ \int_{\Omega} \eta \, dx\, dy
= \frac{\beta -1}{\beta-\alpha} |\Omega| \right\} \, .$$
The proof of Lemma \[exist\] with minor changes shows that also problem admits an optimal pair $(u_{\overline \eta}, \overline{\eta}) \in H^2_*(\Omega) \times N_{\alpha,\beta}$. Furthermore, there is an one-to-one correspondence between problems and . Indeed, to any $\eta \in N_{\alpha,\beta}$ we can associate $p_{\eta} \in P_{\alpha,\beta}$ by setting
$$p_{\eta}=\beta-\eta(\beta-\alpha).$$
Clearly $\alpha {\leqslant}p_{\eta} {\leqslant}\beta$ and
$$\int_{\Omega} p_{\eta} \, dx\, dy = \beta |\Omega| - (\beta - \alpha) \int_{\Omega} \eta \, dx\, dy = |\Omega|.$$ Viceversa to any $p \in P_{\alpha,\beta}$ we can associate $\eta_{p} \in N_{\alpha,\beta}$ by setting
$$\eta_p = \dfrac{\beta - p}{\beta - \alpha}.$$ Clearly $0 {\leqslant}\eta {\leqslant}1$ and $\int_{\Omega} \eta_{p} \, dx\, dy = \frac{\beta - 1}{\beta - \alpha} |\Omega|.$ Furthermore, we have
\[thm-1\] Let $\lambda_{\alpha, \beta}$ and $\Lambda_{\alpha, \beta}$ be as defined in and in . There holds $$\Lambda_{\alpha, \beta} = \lambda_{\alpha, \beta}\,\beta.$$
We shall prove the lemma in two steps.
[**Step 1 :**]{} Let $\overline{p} \in P_{\alpha, \beta}$ and $u_{\overline{p}} \in H^2_*(\Omega)$ such that $\lambda_{\alpha, \beta}$ is achieved for this optimal pair and let $\eta_{\overline{p}} = \frac{\beta - \overline{p}}{\beta - \alpha} \in N_{\alpha, \beta}$. Clearly we have
$$\begin{aligned}
\Lambda_{\alpha, \beta}& {\leqslant}\inf_{u\in H^2_*(\Omega)\setminus \{0\}} \, \dfrac{||u||_{H^2_*(\Omega)}^2 + \lambda_{\alpha, \beta} (\beta-\alpha)
\int_{\Omega} \eta_{\overline{p}} \, u^2 \, dx\, dy}{\int_{\Omega} u^2 \, dx\, dy} \notag \\[9pt]
& = \inf_{u\in H^2_*(\Omega)\setminus \{0\}} \, \dfrac{ ||u||_{H^2_*(\Omega)}^2 - \lambda_{\alpha, \beta} \int_{\Omega} \overline{p} \, u^2 \, dx\, dy
+ \lambda_{\alpha, \beta}
\beta \int_{\Omega} u^2 \, dx\, dy }{\int_{\Omega} u^2 \, dx\, dy} \notag \\[9pt]
& {\leqslant}\underbrace{ \frac{||u_{\overline{p}}||_{H^2_*(\Omega)}^2 - \lambda_{\alpha,\beta}
\int_{\Omega} \overline{p} \, u_{\overline{p}}^2 \, dx\, dy}{\int_{\Omega} u_{\overline{p}}^2 \, dx\, dy}}_{= 0}
+ \lambda_{\alpha, \beta} \beta=\lambda_{\alpha, \beta}\, \beta.\end{aligned}$$
[**Step 2 :**]{} Let now $\eta \in N_{\alpha, \beta}$ and $p_{\eta} \in P_{\alpha, \beta}$ with $\eta = \frac{\beta - p_{\eta}}{\beta - \alpha},$ i.e., $p_{\eta} = \beta - \eta (\beta - \alpha).$ Then for any $u\in H^2_*(\Omega)\setminus \{0\}$
$$\begin{aligned}
\label{estimate-2}
\frac{||u||_{H^2_*(\Omega)}^2 + \lambda_{\alpha, \beta} (\beta - \alpha) \int_{\Omega} \eta \, u^2 \, dx\, dy}{\int_{\Omega} u^2 \, dx\, dy}
=\frac{ ||u||_{H^2_*(\Omega)}^2 - \lambda_{\alpha, \beta} \int_{\Omega} {p_{\eta}} \, u^2 \, dx\, dy +\lambda_{\alpha, \beta}
\beta \int_{\Omega} u^2 \, dx\, dy }{\int_{\Omega} u^2 \, dx\, dy}.\end{aligned}$$
Since, $p_{\eta} \in P_{\alpha, \beta}$ implies $\lambda_{\alpha, \beta} {\leqslant}\frac{||u||_{H^2_*(\Omega)}^2}{\int_{\Omega} p_{\eta}\, u^2 \, dx\, dy}$ for any $u \in H^2_*(\Omega)\setminus \{0\}$ and $\eta \in N_{\alpha, \beta}$, passing to the infima, yields $$\Lambda_{\alpha, \beta} {\geqslant}\lambda_{\alpha, \beta}\, \beta.$$ This completes the proof.
Finally, we prove that the optimal pair of problem can be characterised as follows
\[Lambda\] For every $0<\alpha<\beta$, let $(\overline u, \overline{\eta}) \in H^2_*(\Omega) \times N_{\alpha, \beta}$ be an optimal pair of problem . Then, $\overline u$ and $\overline{\eta}$ are related as follows $$\overline{\eta}(x,y) = \chi_{S_{\overline u}} (x,y) \qquad \text{for a.e. } (x,y)\in \Omega \, ,$$ where $\chi_{S_{\overline u}}$ is the characteristic function of a set $S_{\overline u}\subset \Omega$ such that $|S_{\overline u}|=\frac{\beta-1}{\beta -\alpha}\,|\Omega|$ and $$S_{\overline u}=\{(x,y)\in \Omega\,: \overline u^2(x,y) {\leqslant}t \}$$ for some $t> 0$.
The proof is along the line of [@CV Proposition 3.3]. For the sake of completeness we shall outline the main ideas.
**Step 1.** Let $u \in H^2_*(\Omega)$ be such that $||u||_2 = 1$ and consider the functional $I : N_{\alpha, \beta} \rightarrow \mathbb{R}$ $$I(\eta) := \int_{\Omega} \eta \, u^2 \, dx\, dy\,.$$ We prove that the infimum problem $$I_{\alpha, \beta} := \inf_{\eta \in N_{\alpha, \beta}} I(\eta)$$
admits a solution $\eta = \chi_{S_u},$ where $S_u \subset \Omega$ is such that $|S_u| = \frac{\beta -1}{\beta - \alpha} |\Omega|$ and satisfies one of the following $$\label{S}
S_u=\{(x,y)\in \Omega \, : u^2(x,y)=0 \} \quad \text{ or }\quad \{(x,y)\in \Omega\,:u^2(x,y) < t \} \subseteq S_u \subseteq \{(x,y)\in \Omega\,: u^2(x,y){\leqslant}t\},$$ where $t$ is defined as $$\label{eq:def-t}
t:= \sup \left\{ s > 0 : |\{(x,y)\in \Omega\,: u^2(x,y) < s \}| < \frac{\beta -1}{\beta-\alpha} |\Omega| \right\}.$$ Let $S_u \subset \Omega$ be as above, then $\chi_{S_u} \in N_{\alpha,\beta}$ and one obtains $$I_{\alpha, \beta} {\leqslant}I(\chi_{S_u}) = \int_{S_u} u^2 \, dx\, dy.$$ On the other hand we claim that the following inequality holds $$I(\eta) {\geqslant}I(\chi_{S_u}) \qquad \text{for any } \eta\in N_{\alpha,\beta} \,.$$ If this is true then one immediately obtain $I_{\alpha, \beta} = I (\chi_{S_u})$ and this concludes the proof of step 1.
We prove the validity of the claim by considering the cases $t>0$ and $t=0$ separately.
If $t>0$, we argue as follows $$\begin{aligned}
\label{eq:optimal}
& \int_{\Omega} u^2 (\chi_{S_u} - \eta) \, dx\, dy \\[8pt]
\notag & \qquad =\int_{\{ u^2 < t\}} u^{2} \, (\chi_{S_u}-\eta) \, dx\, dy + \int_{\{ u^2 > t\}} u^{2} \, (\chi_{S_u} - \eta) \, dx\, dy +
\int_{\{ u^2 = t\}} u^{2} \, (\chi_{S_u}-\eta) \, dx\, dy \\[8pt]
\notag & \qquad {\leqslant}t \int_{\{ u^2 < t\}} (\chi_{S_u} - \eta) \, dx\, dy - t \int_{\{ u^2 > t\}} \eta \, dx\, dy
+t \int_{\{ u^2 = t\}} \, (\chi_{S_u} - \eta) \, dx\, dy \\[8pt]
& \notag \qquad = t \int_{\Omega} (\chi_{S_u}-\eta) \, dx\, dy=0 \, .\end{aligned}$$ If $t=0$ the proof follows with minor changes.
**Step 2.** We prove that if $(\overline u,\overline \eta)$ is an optimal pair as in the statement of the lemma and if $S_{\overline u}$ is the corresponding set defined according to Step 1, then $(\overline u,\chi_{S_{\overline u}})$ is still an optimal pair.
Set $$S_{\alpha, \beta}:=\left\{S \subset \Omega\,:|S| = \frac{\beta -1}{\beta - \alpha} |\Omega|\right\} \, .$$ Since $\{ \chi_S : S \in S_{\alpha, \beta} \} \subset N_{\alpha, \beta},$ we have $$\begin{aligned}
\Lambda_{\alpha,\beta}
{\leqslant}\inf_{S \in S_{\alpha, \beta}} \, \inf_{u \in H^2_*(\Omega)\setminus \{0\}} \, \frac{||u||_{H^2_*(\Omega)}^2 + \lambda_{\alpha, \beta}(\beta - \alpha)
\int_{\Omega} \chi_{S} \, u^2 \, dx\, dy}{\int_{\Omega} u^2 \, dx\, dy} \, .\end{aligned}$$ On the other hand, letting $(\overline u,\overline \eta)$ an optimal pair as in the statement of the lemma, from Step 1 we have $$||\overline u||_{H^2_*(\Omega)}^2 +\lambda_{\alpha, \beta}(\beta - \alpha) \int_{\Omega} \, \overline{\eta} \, \overline{u}^2 \, dx\, dy
{\geqslant}||\overline u||_{H^2_*(\Omega)}^2 + \lambda_{\alpha, \beta}(\beta - \alpha) \int_{\Omega} \chi_{S_{\overline u}} \, \overline u^2 \, dx\, dy$$ and therefore $$\begin{aligned}
\Lambda_{\alpha,\beta}&=\frac{||\overline u||_{H^2_*(\Omega)}^2 +\lambda_{\alpha, \beta}(\beta - \alpha) \int_{\Omega} \,
\overline \eta \, \overline u^2 \, dx\, dy}
{\int_\Omega \overline u^2 \, dx\, dy}\ge
\frac{||\overline u||_{H^2_*(\Omega)}^2 + \lambda_{\alpha, \beta}(\beta - \alpha) \int_{\Omega} \chi_{S_{\overline u}} \, \overline u^2 \, dx\, dy}
{\int_\Omega \overline u^2 \, dx\, dy} \\[8pt]
& \ge\inf_{S \in S_{\alpha, \beta}} \, \inf_{u \in H^2_*(\Omega)\setminus \{0\}}
\, \frac{||u||_{H^2_*(\Omega)}^2 + \lambda_{\alpha, \beta}(\beta - \alpha) \int_{\Omega} \chi_{S} \, u^2 \, dx\, dy}{\int_{\Omega} u^2 \, dx\, dy} \, .\end{aligned}$$ This proves that $$\begin{aligned}
\Lambda_{\alpha,\beta}=
\inf_{S \in S_{\alpha, \beta}} \, \inf_{u \in H^2_*(\Omega)\setminus \{0\}} \, \frac{||u||_{H^2_*(\Omega)}^2 + \lambda_{\alpha, \beta}(\beta - \alpha)
\int_{\Omega} \chi_{S} \, u^2 \, dx\, dy}{\int_{\Omega} u^2 \, dx\, dy}\end{aligned}$$ and in particular that $(\overline u,\chi_{S_{\overline u}})$ is an optimal pair.
**Step 3.** Let $(\overline u,\chi_{S_{\overline u}})$ be the optimal pair introduced in Step 2 and let $\overline t$ be the number $t$ in corresponding to $\overline u$. Let $$A_{\overline t}=\{ (x,y)\in \Omega\,: \overline u^2(x,y) = \overline t\} \, .$$ We prove that $\overline t>0$ and that $|A_{\overline t}\setminus S_{\overline u}|=0$.
Suppose by contradiction that $\overline t=0$. Since $\overline u \in H^4(\Omega)$ we can write the Euler-Lagrange equation related to almost everywhere and we have $$\Lambda_{\alpha, \beta} \, \overline u=\Delta^2 \overline u+ \lambda_{\alpha,\beta} (\beta-\alpha) \chi_{S_{\overline u}} \, \overline u=\Delta^2 \overline u \quad \text{a.e. in } \Omega \, .$$ Since $\overline u$ satisfies the partially hinged boundary conditions this means that it must be one of the eigenfunctions listed in Proposition \[eigenvalue\] which is impossible since the set of zeroes of any of the eigenfunctions of Proposition \[eigenvalue\] has zero measure thus contradicting the definition of $S_{\overline u}$ which forces $S_{\overline u}$ to be a set of positive measure. This proves that $\overline t>0$.
Suppose now by contradiction that $|A_{\overline t}\setminus S_{\overline u}|>0$, we have that $$\Delta^2 \overline u+ \lambda_{\alpha, \beta} (\beta-\alpha) \chi_{S_{\overline u}} \, \overline u=\Lambda_{\alpha, \beta} \overline u\quad
\text{a.e. in } \Omega \, .$$ Now, exploiting the fact that $\overline u$ is constant in $A_{\overline t}$ and $\overline t>0$, we infer $$\Lambda_{\alpha,\beta}=\lambda_{\alpha, \beta} (\beta-\alpha) \chi_{S_{\overline u}} \qquad \text{a.e. in } A_{\overline t} \, .$$ and hence, since $\lambda_{\alpha, \beta} (\beta-\alpha) \chi_{S_{\overline u}}=0$ a.e. in $A_{\overline t}\setminus S_{\overline u}$ and $|A_{\overline t}\setminus S_{\overline u}|>0$, we obtain $\Lambda_{\alpha,\beta}=0$ and this is absurd.
**Step 4.** We complete the proof of the lemma. First of all, we observe that by Step 3, it is not restrictive, up to a set of zero measure, to assume that $A_{\overline t}\setminus S_{\overline u}=\emptyset$ in such way that $A_{\overline t}\subseteq S_{\overline u}$ and, in turn, $$\label{eq:Su}
S_{\overline u}=\{(x,y)\in \Omega: \overline u^2(x,y){\leqslant}\overline t\} \, .$$ It remains to prove that $\overline \eta=\chi_{S_{\overline u}}$ a.e. in $\Omega$. Since $(\overline u,\overline \eta)$ and $(\overline u,\chi_{S_{\overline u}})$ are both optimal pairs we have
$$\begin{aligned}
& \Delta^2 \overline u+ \lambda_{\alpha, \beta} (\beta-\alpha) \chi_{S_{\overline u}} \, \overline u=\Lambda_{\alpha, \beta} \overline u \quad
\text{a.e. in } \Omega \, , \\[7pt]
& \Delta^2 \overline u+ \lambda_{\alpha, \beta} (\beta-\alpha) \overline \eta \, \overline u=\Lambda_{\alpha, \beta} \overline u \quad
\text{a.e. in } \Omega \, ,\end{aligned}$$
thus implying that $$(\chi_{S_{\overline u}}-\overline \eta) \, \overline u=0 \qquad \text{a.e. in } \Omega \, .$$ It is easy to check that $\overline \eta=\chi_{S_{\overline u}}$ a.e. in $\{(x,y)\in \Omega:\overline u^2(x,y){\geqslant}\overline t\}$ being $\overline t>0$. In order to prove that $\overline \eta=\chi_{S_{\overline u}}$ a.e. in $\{(x,y)\in \Omega:\overline u^2(x,y)<\overline t\}$, we apply to $\overline u$, $\chi_{S_{\overline u}}$ and $\overline \eta$ observing that the inequality is an equality being $(\overline u,\overline \eta)$ and $(\overline u,\chi_{S_{\overline u}})$ both optimal pairs. In particular we have that $$\int_{\{\overline u^2<\overline t\}} \overline u^{2} \, (\chi_{S_{\overline u}}-\overline\eta) \, dx\, dy
=\overline t \int_{\{ \overline u^2<\overline t\}} (\chi_{S_{\overline u}}-\overline\eta) \, dx\, dy \, ,$$ which implies $$\label{eq:=0}
\int_{\{\overline u^2<\overline t\}} (\overline u^{2}-\overline t) \, (\chi_{S_{\overline u}}-\overline\eta) \, dx\, dy=0
$$ But the function $(\overline u^{2}-\overline t) \, (\chi_{S_{\overline u}}-\overline\eta){\leqslant}0$ in $\{(x,y)\in \Omega:\overline u^2(x,y)<\overline t\}$, as one can deduce by , and hence by we conclude that $\chi_{S_{\overline u}}=\overline\eta$ a.e. in the same set.
We have so proved that $\chi_{S_{\overline u}}-\overline\eta=0$ a.e. in $\Omega$ and this completes the proof of the lemma.
[**Proof of Theorem \[thm-exist-qual\] completed.**]{}
For every $0<\alpha<\beta$, the existence of an optimal pair $(\overline u,\overline{p}) \in H^2_*(\Omega) \times P_{\alpha, \beta}$ follows from Lemma \[exist\]. If we put $\overline \eta:=\frac{\beta-\overline p}{\beta-\alpha}$ by Lemma \[thm-1\] we deduce that $(\overline u,\overline \eta)$ is an optimal pair for $\Lambda_{\alpha, \beta}=\lambda_{\alpha, \beta}\, \beta$. Moreover by Lemma \[Lambda\] we also have that $\overline \eta=\chi_{S_{\overline u}}$ a.e. in $\Omega$ with $S_{\overline u}=\{(x,y)\in \Omega:\overline u^2(x,y){\leqslant}\overline t\}$ and $\overline t$ as in . Hence we conclude that $${\overline p}=\beta-{\overline \eta}(\beta-\alpha)=\alpha \chi_{S_{\overline u}}+\beta \chi_{S_{\overline u}^c} \, .$$
Proof of Proposition \[comparison\_esy\].
-----------------------------------------
We prove the two statements separately.
**Proof of Proposition \[comparison\_esy\]-(i).**
We know that the function $u_1(x,y)=\varphi_1(y) \sin x$ introduced in is an eigenfunction corresponding to the least eigenvalue of with $p\equiv 1$. Furthermore, $$\inf_{u\in H^2_*(\Omega)\setminus \{0\}} \frac{\|u\|_{H^2_*(\Omega)}^2}{\|u\|_{2}^2}=\frac{\|u_1\|_{H^2_*(\Omega)}^2}{\|u_1\|_{2}^2}=\mu_{1,1} \, .$$ Now, by exploiting the fact that $\varphi_1$ is even and increasing in $(0,\ell)$ and $p=p(y)$ is even, we deduce that $$\begin{aligned}
&\int_{\Omega} (1-p(y))\,u_1^2(x,y)\,dx\,dy=2 \int_0^{\pi} \int_{0}^{\ell}(1-p(y))\varphi_1^2(y)\, \sin^2 x\,dx\,dy\\
&{\leqslant}2 \varphi_1^2(z) \int_0^{\pi} \int_{0}^{z}(1-p(y))\, \sin^2 x\, dx\,dy+2 \varphi_1^2(z) \int_0^{\pi} \int_{z}^{\ell}(1-p(y))\, \sin^2 x\,dx\,dy\\
&= \varphi_1^2(z) \,\pi \, \int_{0}^{\ell} (1-p(y))\,dy= 0\,,\end{aligned}$$ where in the last step we have exploited the fact that $\int_{\Omega} p(y)\,dx\,dy=|\Omega|$, therefore $ \int_{0}^{\ell}p(y)\,dy =\ell$. Hence, $$\int_{\Omega} u_1^2(x,y)\,dxdy {\leqslant}\int_{\Omega} p(y) \,u_1^2(x,y)\,dxdy \, .$$ From the above inequality we infer $$\mu_{1,1}= \frac{\|u_1\|_{H^2_*(\Omega)}^2}{\|u_1\|_{2}^2}{\geqslant}\inf_{u\in H^2_*(\Omega)\setminus \{0\}} \frac{\|u\|_{H^2_*(\Omega)}^2}{\|\sqrt{p}u\|_{2}^2}=\lambda_1(p) \, ,$$ and the proof of the statement follows.
**Proof of Proposition \[comparison\_esy\]-(ii).**
The idea of the proof is similar to that applied to prove statement $(i)$. By exploiting the fact that $\sin(\pi-x)=\sin(x)$ and $p(\pi-x)=p(x)$ for all $x\in (0, \frac{\pi}{2})$, we deduce that $$\begin{aligned}
&\int_{\Omega} (1-p(x))\,u_1^2(x,y)\,dx\,dy =2 \int_{-\ell}^{\ell} \int_0^{\frac{\pi}{2}}(1-p(x))\varphi_1^2(y)\, \sin^2 x\,dx\,dy\\
&{\leqslant}2 \sin^2 (s) \int_{-\ell}^{\ell} \int_0^{s} (1-p(x))\, \varphi_1^2(y)\,dx\, \,dy+2 \sin^2 (s) \int_{-\ell}^{\ell} \int_s^{\frac{\pi}{2}}(1-p(x))\,\varphi_1^2(y) \,dx\,dy\\
&= 2 \sin^2 (s) \, \left( \int_{-\ell}^{\ell} \varphi_1^2(y)\,dy\right) \, \left(\int_{0}^{\frac{\pi}{2}}(1-p(x))\,dx\right)= 0\,,\end{aligned}$$ where in the last step we have exploited the assumption $\int_{\Omega} p(x)\,dx\, dy=|\Omega|$, hence $\int_{0}^{\frac{\pi}{2}}p(x)\,dx=\frac{\pi}{2}$. From the above inequality the proof follows as for statement $(i)$.
Proof of Theorem \[min\] and Theorem \[signTH\] {#spectrum}
===============================================
Let $0<\alpha<\beta$. In this section we restrict the admissible weights to the family
$\overline P_{\alpha,\beta}$ defined in Theorem \[min\].
Clearly, $\int_{0}^{\ell}p\,dy=\ell$ for all $p\in \overline P_{\alpha,\beta}$. Let $m$ be a positive integer, we consider the following scalar product in $H^2(-\ell, \ell)$: $$\langle\varphi,\phi\rangle_m:=\int_{-\ell}^{\ell}\left( \varphi'' \phi''+2m^2(1-\sigma) \varphi' \phi'-\sigma m^2 (\varphi''\phi+\varphi\phi'')+m^4 \varphi \phi\right) \,dy \, .$$ For every $m{\geqslant}1$ it defines an equivalent norm in $H^2(-\ell, \ell)$ that we will denote by ${\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert}\phi{\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}_{m}^2 =(\phi,\phi)_m$. Let $u$ be an eigenfunction of , its Fourier expansion reads $$u(x,y)=\sum_{m=1}^{+\infty} \varphi_m(y)\sin(mx)$$ with $\varphi_m \in C^2([-\ell,\ell])$ since $u\in H^4(\Omega)$ (at least). Inserting $u$ in , we get that, for every $m{\geqslant}1$ fixed, $\varphi_m$ satisfies the equation $$\label{weight1dweak}
\langle \varphi,\phi \rangle_m=\lambda\int_{-\ell}^{\ell} p(y) \varphi \phi \,dy \quad \text{for all }\phi \in H^2(-\ell, \ell)\,$$ which is the weak formulation of the problem $$\label{weight1d}
\begin{cases}
\varphi''''(y)-2m^2\varphi''(y)+m^4\varphi(y)=\lambda p(y)\varphi(y) & \qquad \text{in } (-\ell,\ell) \\
\varphi''(\pm\ell)-\sigma m^2\varphi(\pm\ell)=0 & \qquad \\
\varphi'''(\pm\ell)-(2-\sigma)m^2\varphi'(\pm\ell)=0\,.& \qquad \,
\end{cases}$$ Notice that, by elliptic regularity, any solution $\varphi \in H^2(-\ell, \ell)$ of , lies in $H^4(-\ell, \ell)\subset C^3([-\ell, \ell])$. Hence, the boundary conditions in are satisfied pointwise. Since the bilinear form $\langle\varphi,\phi\rangle_m$ is continuous and coercive the eigenvalues of problem may be ordered in an increasing sequence of strictly positive numbers diverging to $+\infty$ and the corresponding eigenfunctions form a complete system in $H^2(-\ell, \ell)$. Whence, for what remarked so far, when $p=p(y)$ there is a one to one correspondence between eigenvalues of and eigenvalues of . In particular, if we denote by $\lambda_1(p)$ the first eigenvalue of and by $\overline\lambda_1(p,m)$ the first eigenvalue of with $m{\geqslant}1$ fixed, namely $$\lambda_1(p):= \inf_{u\in H^2_*(\Omega)\setminus \{0\}} \frac{\|u\|_{H^2_*(\Omega)}^2}{\|\sqrt{p}u\|_{2}^2}\quad \text{and} \quad \overline\lambda_1(p,m)
:=\inf_{\varphi\in H^2(-\ell, \ell)\setminus \{0\}}\frac{{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert}\varphi{\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}_{m}^2}{\|\sqrt{p} \varphi\|_{2}^2} \, ,$$ it is natural to conjecture that $$\lambda_1(p)=\min_{m{\geqslant}1}\left\{\overline\lambda_1(p,m)\right\}=\overline\lambda_1(p,1) \, .$$ Unfortunately, for $p\in \overline P_{\alpha,\beta}$ fixed, due to the negative terms in the norm ${\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert}\, \cdot{\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}_{m}$, the monotonicity of $m\mapsto \overline\lambda_1(p,m)$ is not easy to detect and we do not have a proof of the above equality for general $p$; in Section \[numerics\] we give some suggestions through numerical experiments. Nevertheless, we have the following partial result
\[comp1\] If $p\in \overline P_{\alpha,\beta}$ then $$\overline\lambda_1(p,m){\leqslant}\mu_{m,1}<m^4 \, ,$$ where the $\mu_{m,1}$ are the numbers defined in Proposition \[eigenvalue\]-(i).
If furthermore $\beta{\leqslant}(1-\sigma^2) 2^4$, then $$\label{monot}
\overline\lambda_1(p,m) {\geqslant}\overline\lambda_1(p,1) \quad \text{ for all } m{\geqslant}2 \,.$$
Let $$\varphi_m(y):=\left[\big[\mu_{m,1}^{1/2}-(1-\sigma)m^2\big]\, \tfrac{\cosh\Big(y\sqrt{m^2+\mu_{m,1}^{1/2}}\Big)}{\cosh\Big(\ell\sqrt{m^2+\mu_{m,1}^{1/2}}\Big)}+
\big[\mu_{m,1}^{1/2}+(1-\sigma)m^2\big]\, \tfrac{\cosh\Big(y\sqrt{m^2-\mu_{m,1}^{1/2}}\Big)}{\cosh\Big(\ell\sqrt{m^2-\mu_{m,1}^{1/2}}\Big)}\right]\,,$$ From Proposition \[eigenvalue\] it is readily deduced that $\varphi_m(y)$ is an eigenfunction corresponding to the least eigenvalue of with $p\equiv 1$ and $m{\geqslant}1$ fixed (otherwise we will find an eigenvalue of not included in those listed in Proposition \[eigenvalue\]). Furthermore, $$\inf_{\varphi\in H^2(-\ell, \ell)\setminus \{0\}}\frac{{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert}\varphi{\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}_{m}^2}{\|\varphi\|_{2}^2}
=\frac{{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert}\varphi_m{\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}_{m}^2}{\|\varphi_m\|_{2}^2}=\mu_{m,1} \, .$$ Now, by exploiting the fact that $\varphi_m$ is even and increasing in $(0,\ell)$, the first part of the proof follows with the same argument of Proposition \[comparison\_esy\]-(i), hence we omit it.
Next we turn to the second estimate. Let $\varphi_{m,p}(y)$ be an eigenfunction corresponding to the least eigenvalue of , with $m{\geqslant}2$ fixed and with $p\in \overline P_{\alpha,\beta}$ satisfying the assumption of Lemma \[comp1\]. In particular, $\varphi_{m,1}=\varphi_{m}$, with $\varphi_{m}$ as given above. Since $p(y){\leqslant}(1-\sigma^2)2^4$ for every $y\in(-\ell, \ell)$, we get $$\overline\lambda_1(p,m)=\frac{{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert}\varphi_{m,p}{\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}_{m}^2}{\|\sqrt{p} \varphi_{m,p}\|_{2}^2}{\geqslant}\frac{1}{(1-\sigma^2)m^4}
\frac{{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert}\varphi_{m,p}{\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}_{m}^2}{\| \varphi_{m,p}\|_{2}^2}{\geqslant}\frac{\mu_{m,1}}{(1-\sigma^2)m^4} \, .$$ Then, the thesis follows by recalling that, from Proposition \[eigenvalue\]-(i), $\mu_{m,1}\in ((1-\sigma^2) m^4, m^4)$ for every $m{\geqslant}1$ and from the first part of the proof $\overline\lambda_1(p,1)<1$.
Hence, under the assumptions of Lemma \[comp1\], we have $$\lambda_1(p)=\overline\lambda_1(p,1){\leqslant}\mu_{1,1}=\lambda_1(1)\,.$$ In particular, the weights considered in Lemma \[comp1\] prove to be effective in decreasing the first frequency of , which is one of the main goal of the present analysis. In the following we refine the result by carrying on a more deeper analysis. First we note that, from above, if $\varphi_{1,p}(y)$ is an eigenfunction of $\overline\lambda_1(p,1)$, then $u_{1,p}(x,y):=\varphi_{1,p}(y)\sin (x)$ is an eigenfunction of $\lambda_1(p)$. Therefore, $\varphi_{1,p}(y)$ and $u_{1,p}(x,y)$ have the same sign.
We discuss now the sign of $\varphi_{1,p}(y)$ and the simplicity of $\lambda_1(p)$ in
\[sign\] Let $m{\geqslant}1$ integer fixed, $\sigma\in (0,1)$ and let $p\in \overline P_{\alpha,\beta}$. Then, the first eigenvalue $\overline\lambda_1(p,m)$ of problem is simple and the first eigenfunction $\varphi_{m,p}(y)$ is of one sign in $[-\ell,\ell]$.
Furthermore, if the assumptions of Lemma \[comp1\] holds, the same conclusion holds for the first eigenvalue $\lambda_1(p)$ of , namely it is simple and the corresponding eigenfunction is given by $u_{1,p}(x,y)=\varphi_{1,p}(y)\sin (x)$, hence of one sign in $\Omega$.
We apply the decomposition with respect to dual cones technique, see [@book Chapter 3] suitable combined with Theorem \[ppp\] below. We start by recalling some basic facts concerning the just mentioned decomposition. Let $H$ be a Hilbert space with scalar product $(.,.)_{H}$. Let $\mathcal{K} \subset H$ be a closed nonempty cone and let $\mathcal{K}^*$ be its dual cone, namely $$\mathcal{K}^*:=\{\psi\in H:(\psi,\phi)_{H}{\leqslant}0\quad \text{ for
all }\phi \in \mathcal{K}\} \,.$$ Then, for any $\varphi\in H$ there exists a unique $(\chi, \psi)\in \mathcal{K} \times
\mathcal{K}^*$ such that $$\varphi=\chi+\psi\,, \quad(\chi, \psi)_H=0 \,.$$ Now we turn to the proof of Lemma \[sign\]. We apply the above decomposition with $H=H^2(-\ell, \ell)$, $(.,.)_{H}=\langle.,.\rangle_{m}$ and $\mathcal{K}=\{\varphi\in H:\varphi{\geqslant}0 \text{ in }(-\ell,
\ell)\}$. We know that $$\overline\lambda_1(p,m)=\inf_{\varphi\in H^2(-\ell, \ell)\setminus \{0\}}\frac{{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert}\varphi{\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}_{m}^2}{\|\sqrt{p} \varphi\|_{2}^2}
=\frac{{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert}\varphi_{m,p}{\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}_{m}^2}{\|\sqrt{p} \varphi_{m,p}\|_{2}^2} \, .$$ For contradiction, assume that $\varphi_{m,p}$ changes sign. Then, we may decompose $\varphi_{m,p}=\chi_{m,p}+\psi_{m,p}$ with $\chi_{m,p} \in \mathcal{K}\setminus\{0\}$ and $\psi_{m,p} \in
\mathcal{K}^* \setminus\{0\}$.
In the remaining part of this proof we need some results on a positivity preserving property which is treated in Section \[pppproof\].
>From Corollary \[pppweak\], we deduce that $\psi_{m,p}<0$ in $(-\ell, \ell)$. Then, replacing $\varphi_{m,p}$ with $\chi_{m,p}-\psi_{m,p}$, exploiting the fact that $\chi_{m,p}-\psi_{m,p}>\varphi_{m,p}$ in $(-\ell, \ell)$ and the orthogonality of $\chi_{m,p}$ and $\psi_{m,p}$ in $H^2(-\ell,\ell)$, we infer
$$\frac{{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert}\chi_{m,p}-\psi_{m,p}{\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}_{m}^2}{\|\sqrt{p}
(\chi_{m,p}-\psi_{m,p})\|_{2}^2}
<\frac{{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert}\varphi_{m,p}{\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}_{m}^2}{\|\sqrt{p}
\varphi_{m,p}\|_{2}^2}\,,$$ a contradiction. Hence $\varphi_{m,p}{\geqslant}0$ in $(-\ell, \ell)$ and since $\varphi_{m,p}$ solves , by Theorem \[ppp\] with $f=\overline\lambda_1(p,m)\, p(y)\, \varphi_{m,p}$, we conclude that $\varphi_{m,p}> 0$ in $[-\ell, \ell]$.
As concerns the simplicity, it follows by noting that if $\varphi_{m,p}$ and $\bar
\varphi_{m,p}$ are two linearly independent positive minimizers, then $\varphi_{m,p}+t \bar \varphi_{m,p}$ is a sign-changing minimizer for some $t<0$ suitably chosen, a contradiction.
Next we focus on the sign of $\varphi_{1,p}'(y)$ and we prove
\[convex\] Let $\sigma\in(0,1)$. If $p\in \overline P_{\alpha,\beta}$ is such that $\beta<1/\mu_{1,1}$ and if $\varphi_{1,p}$ is a positive eigenfunction of with $m=1$ corresponding to the first eigenvalue $\overline \lambda_1(p,1)$, then $\varphi_{1,p}$ is increasing in $(0,\ell)$.
For shortness we will write $\varphi_1$ instead of $\varphi_{1,p}$. Since $p$ is even, being $\varphi_1$ positive, we infer that it is an even function. Hence, since $\varphi_1\in C^3([-\ell, \ell])$ it satisfies $\varphi_1'(0)=0=\varphi_1'''(0)$.
If $p$ is continuous, then $\varphi_{1}\in C^4([-\ell, \ell])$ and it satisfies the equation in pointwise. We recall that the boundary conditions in are satisfied pointwise also when $p$ is not continuous. Since $\varphi_1$ is positive, $\beta<1/\mu_{1,1}$ and, by Lemma \[comp1\], we know that $\overline \lambda_1(p,1){\leqslant}\mu_{1,1}$, from the equation we infer $$\label{crucial}
\varphi_{1}''''(y)-2\varphi_{1}''(y)=( \overline \lambda_1(p,1) p(y)-1)\varphi_{1}(y){\leqslant}(\mu_{1,1} p(y)-1)\varphi_{1}(y)<0\quad \text{ in }(-\ell, \ell)$$ If $p$ is not continuous, since only a finite number of points of jump discontinuity are allowed in $(-\ell,\ell)$, say $\{\tau_j\}_{j=1}^r$ for some integer $r$, the above inequality holds in each interval $(\tau_j, \tau_{j+1})$. Furthermore, for any $j=1,...,r$, the right and left fourth order derivative at $\tau_j$ exists and they are given by $(\varphi_1)''''_{\pm}(\tau_j)=\lim_{y\rightarrow
\tau_j^{\pm}}\varphi_{1}''''(y)$.
First we show that $$\label{der-sign}
\varphi_{1}' \text{ never vanishes in }(0, \ell)\,.$$
By contradiction, let $y_1\in (0,\ell)$ be such that $\varphi_{1}'(y_1)=0$. Since $\varphi_{1}'(0)=0$ and $\varphi_1\in C^3([-\ell, \ell])$, there exists $y_0\in (0, y_1)$ such that $\varphi_{1}''(y_0)=0$ and, by , $(\varphi_1)''''_{+}(y_0)<0$. Next the following two cases may occur.
$\bullet $ CASE 1: $\varphi_{1}'''(y_0){\leqslant}0$. From above, $\varphi_{1}'''$ is negative and, in turn, also $\varphi_{1}''$ is negative in a right neighborhood of $y_0$. Since the boundary conditions in yield $\varphi_{1}''(\ell)=\sigma \varphi_{1}(\ell)>0$, we infer that there exists $y_2>y_0$ such that $\varphi_{1}''(y_2)=0$, $\varphi_{1}'''(y_2) {\geqslant}0$ and $\varphi_{1}''(y){\leqslant}0$ in $(y_0,y_2)$. Whence, by , $\varphi_{1}''''(y)< 0$ in $(y_0,y_2)$ or in each of the subintervals $(\tau_j, \tau_{j+1})$ contained in $(y_0,y_2)$. Since $\varphi_{1}'''$ is continuous in $[y_0,y_2]$, in any case, we have that it is strictly decreasing in $[y_0,y_2]$, hence $\varphi_{1}'''(y)< 0$ in $(y_0,y_2]$ in contradiction with $\varphi_{1}'''(y_2){\geqslant}0$.
$\bullet $ CASE 2: $\varphi_{1}'''(y_0)>0$. We distinguish two further cases.
CASE 2a: $\varphi_{1}''(0){\leqslant}0$. By , $(\varphi_1)''''_{+}(0) < 0$, hence $\varphi_{1}'''(y)<0$ in a right neighborhood of $0$. Then, since $\varphi_{1}'''(y_0)>0$, there exists $y_3\in (0,y_0)$ such that $\varphi_{1}'''(y)<0$ in $(0,y_3)$ and $\varphi_{1}'''(y_3)=0$. In turn, $\varphi_{1}''< 0$ in $(0,y_3)$ and by $\varphi_{1}''''(y)<0$ in $(0,y_3)$ (or in each of the subintervals $(\tau_j, \tau_{j+1})$ contained in $(y_0,y_3)$). Since $\varphi_{1}'''$ is continuous this lead that it is strictly decreasing in $[0,y_3]$. Since $\varphi_{1}'''(0)=0$, we infer $\varphi_{1}'''(y_3)<0$, a contradiction.
CASE 2b: $\varphi_{1}''(0)> 0$. From $\varphi_{1}'''(y_0)>0$ and $\varphi_{1}''(y_0)=0$ we infer that $\varphi_{1}''$ is negative in a left neighborhood of $y_0$. Then, since $\varphi_{1}''(0)> 0$, there exists $y_4\in (0, y_0)$ such that $\varphi_{1}''(y) > 0$ in $(0,y_4)$ and $\varphi_{1}''(y_4)=0$. Consecutively, recalling that $\varphi_{1}''(y_0)=0$, there exists $y_5\in (y_4,y_0)$ such that $\varphi_{1}'''(y_5)= 0$ and, by , we infer that $\varphi_{1}'''(y)< 0$ in $(y_5,y_0)$, in contradiction with $\varphi_{1}'''(y_0)>0$.
Next we come back to the proof of the statement. By we know that either $\varphi_{1}'(y)<0$ in $(0, \ell)$ or $\varphi_{1}'(y)>0$ in $(0, \ell)$.
Assume that $\varphi_{1}'(y)<0$ in $(0, \ell)$, then $\varphi_{1}''(0){\leqslant}0$. Indeed, if $\varphi_{1}''(0)>0$, since $\varphi_{1}'(0)=0$, then $\varphi_{1}'$ is positive in a right neighborhood of $0$, a contradiction. From $\varphi_{1}''(0){\leqslant}0$, together with and $\varphi_{1}'''(0)=0$, it follows that $\varphi_{1}'''$ is negative in a right neighborhood of $0$ and, in turn, also $\varphi_{1}''$ is negative in a right neighborhood of $0$. Since, from the boundary conditions $\varphi_{1}''(\ell)=\sigma \varphi_{1}(\ell)>0$, we deduce that there exists $\overline y\in(0, \ell)$ such that $\varphi_{1}''(\overline y)=0$, $\varphi_{1}'''(\overline y) {\geqslant}0$ and $\varphi_{1}''(y){\leqslant}0$ in $(0,\overline y)$. But then, from , $\varphi_{1}'''$ is strictly decreasing in $[0,\overline y]$ and, recalling that $\varphi_1'''(0)=0$ we reach a contradiction.
All the above statements yield the proof of Theorem \[min\].
**Proof of Theorem \[min\] completed.**
The key point is to note that, by Lemma \[convex\], we have
$$\label{claim}
p\in \overline P_{\alpha,\beta} \quad \Rightarrow \quad \varphi_{1,p} \text{ increasing in } (0, \ell).$$
Indeed, since by $\varphi_{1,p_2}$ is increasing in $(0, \ell)$, to prove $(i)$ we may argue as in the proof of the first part of Lemma \[comp1\] with $\varphi_{1,p_2}$ instead of $\varphi_{m}$. In particular, we readily infer that $\overline \lambda_1(p_1,1){\leqslant}\overline \lambda_1(p_2,1)$ and since, from Lemma \[comp1\], $\overline \lambda_1(p,1)=\lambda_1(p)$ for all $p\in \overline P_{\alpha,\beta}$, the proof of $(i)$ follows.
Next we prove $(ii)$. Set $\overline y:=\frac{\ell(\beta-1)}{\beta-\alpha}$, for every $p\in P_{\alpha,\beta}$ there holds $$p(y){\geqslant}\overline p(y) \quad \text{in } [0,\overline y]\quad \text{ and }\quad p(y){\leqslant}\overline p(y) \quad \text{in } [\overline y,\ell)\,.$$ Then, we may argue again as in the proof of the first part of Lemma \[comp1\] with $\varphi_{1,p}$ instead of $\varphi_{m}$ and conclude $$\overline\lambda_1( p,1)\geq\overline\lambda_1(\overline p,1)\,.$$ Once more, from Lemma \[comp1\], $\overline \lambda_1(p,1)=\lambda_1(p)$ for all $p\in \overline P_{\alpha,\beta} $ and the statement of Theorem \[min\] follows.
**Proof of Theorem \[signTH\].**
The proof readily follows by combining the statements of Lemma \[sign\] and Lemma \[convex\].
Numerical Results {#numerics}
=================
In this section, for any $m{\geqslant}1$, we compute numerically the first eigenvalue $\overline{\lambda}_1(p,m)$ of problem when $p$ is as defined in . More precisely, we take $$\label{load}
\overline p_{\alpha, \beta}(y)=
\begin{cases}
\beta\qquad y\in (-\ell,-\overline{y})\cup (\overline{y},\ell)\\
\alpha \qquad y\in (-\overline{y},\overline{y})
\end{cases}$$ where $\beta>1>\alpha>0$ and $\overline{y}=\frac{\ell(\beta-1)}{\beta-\alpha}$, so that $\int_{0}^{\ell}pdy=\ell$.
In terms of engineering applications, this means that we are dealing with a weight given by the pairing of two materials having different rigidities $\alpha$ and $\beta$, properly placed on rectangular strips, having the length of the whole plate.
Note that, since $\overline p_{\alpha, \beta}(y)$ is an even function, to determine all eigenvalues of , we may focus on even and odd eigenfunctions. Indeed, if $\varphi(y)$ is an eigenfunction which is neither odd or even, it is readily verified that also $\varphi^{ev}(y):=\frac{\varphi(y)+\varphi(-y)}{2}$ and $\varphi^{od}(y):=\frac{\varphi(y)-\varphi(-y)}{2}$ are eigenfunctions, respectively even and odd, corresponding to the same eigenvalue of $\varphi(y)$. On the other hand, since by Lemma \[sign\], the first eigenvalue of is simple and the corresponding eigenfunctions is of one sign in $[-\ell,\ell]$, we infer that it must be an even function, whence to compute $\overline{\lambda}_1(p,m)$ we may concentrate on even eigenfunctions that we named $\varphi^{ev}$. For any $m{\geqslant}1$ we have that $$\label{phieven}
\varphi^{ev}(y)=\begin{cases}
h_1(-y)\qquad&{\rm on}\hspace{2mm}[-\ell,-\overline{y}]\\
h_2(y)\qquad&{\rm on}\hspace{2mm}(-\overline{y},\overline{y})\\
h_1(y)\qquad&{\rm on}\hspace{2mm}[\overline{y},\ell]\\
\end{cases}$$ where $h_1$ and $h_2$ satisfy: $$\label{num1}
\begin{cases}
h_1''''(y)-2m^2h_1''(y)+m^4h_1(y)=\lambda \beta h_1(y)\qquad {\rm on}\hspace{2mm}(\overline{y},\ell)\\
h_2''''(y)-2m^2h_2''(y)+m^4h_2(y)=\lambda \alpha h_2(y)\qquad {\rm on}\hspace{2mm}[0,\overline{y})\vspace{2mm}\\
h_1''(\ell)-\sigma m^2h_1(\ell)=0, \qquad
h_1'''(\ell)-(2-\sigma)m^2h_1'(\ell)=0,\\
h_2'(0)=0, \hspace{28mm}
h_2'''(0)=0,
\\ h_1(\overline{y})=h_2(\overline{y}),\hspace{21mm} h'_1(\overline{y})=h'_2(\overline{y}),\\h''_1(\overline{y})=h''_2(\overline{y}),\hspace{21mm} h'''_1(\overline{y})=h'''_2(\overline{y})\,.
\end{cases}$$ Note that the compatibility conditions between the functions $h_1$ and $h_2$, ensure that $\varphi^{ev} \in C^3([-\ell,\ell])$, while $h_2'(0)=h_2'''(0)=0$ come from $\varphi^{ev}(-y)=\varphi^{ev}(y)$ and its regularity. Clearly, the analytical expression of $h_1(y)$ and $h_2(y)$ depends on the roots of the characteristic polynomials related to the first two equations in ; we denote them respectively by $\zeta_1$ and $\zeta_2$ and we find that they satisfy $$\zeta_1^2=m^2\pm\sqrt{\lambda\beta}\qquad\zeta_2^2=m^2\pm\sqrt{\lambda\alpha}.$$ Therefore, the sign of $m^2-\sqrt{\lambda\beta}$ and $m^2-\sqrt{\lambda \alpha}$ determines different kinds of solutions. We introduce the following notations $$\eta_\alpha:=\sqrt{m^2+\sqrt{\lambda \alpha}},\qquad\eta_\beta:=\sqrt{m^2+\sqrt{\lambda \beta}},\qquad\omega_\alpha
:=\sqrt{|m^2-\sqrt{\lambda \alpha}|},\qquad\omega_\beta:=\sqrt{|m^2-\sqrt{\lambda \beta}|} \, ,$$ and we distinguish five cases:
- $m^4>\lambda\beta>\lambda\alpha$, implying $\lambda<m^4/\beta$ and\
$$\begin{split}
&h_1(y)=a_1\cosh\big(\eta_\beta y\big)+b_1\sinh\big(\eta_\beta y\big)+c_1\cosh\big(\omega_\beta y\big)+d_1\sinh\big(\omega_\beta y\big)\,,\\
&h_2(y)=a_2\cosh\big(\eta_\alpha y\big)+c_2\cosh\big(\omega_\alpha y\big)\,,
\end{split}$$
- $m^4=\lambda\beta$, so that $\eta_\alpha=m\sqrt{1+\sqrt{\alpha/\beta}}$, $\omega_\alpha=m\sqrt{1-\sqrt{\alpha/\beta}}$ and\
$$\begin{split}
&h_1(y)=a_1\cosh\big(\sqrt{2}m y\big)+b_1\sinh\big(\sqrt{2}m y\big)+c_1y+d_1\,,\\
&h_2(y)=a_2\cosh\big(\eta_\alpha y\big)+c_2\cosh\big(\omega_\alpha y\big)\,,
\end{split}$$
- $\lambda\alpha<m^4<\lambda\beta$, implying $m^4/\beta<\lambda<m^4/\alpha$ and\
$$\begin{split}
&h_1(y)=a_1\cosh\big(\eta_\beta y\big)+b_1\sinh\big(\eta_\beta y\big)+c_1\cos\big(\omega_\beta y\big)+d_1\sin\big(\omega_\beta y\big)\,,\\
&h_2(y)=a_2\cosh\big(\eta_\alpha y\big)+c_2\cosh\big(\omega_\alpha y\big)\,,\end{split}$$
- $m^4=\lambda\alpha$, so that $\eta_\beta=m\sqrt{1+\sqrt{\beta/\alpha}}$, $\omega_\beta=m\sqrt{\sqrt{\beta/\alpha}-1}$ and\
$$\begin{split}
&h_1(y)=a_1\cosh\big(\eta_\beta y\big)+b_1\sinh\big(\eta_\beta y\big)+c_1\cos\big(\omega_\beta y\big)+d_1\sin\big(\omega_\beta y\big)\,,\\
&h_2(y)=a_2\cosh\big(\sqrt{2}m y\big)+c_2\,,
\end{split}$$
- $m^4<\lambda\alpha<\lambda\beta$, implying $\lambda>m^4/\alpha$ and\
$$\begin{split}
&h_1(y)=a_1\cosh\big(\eta_\beta y\big)+b_1\sinh\big(\eta_\beta y\big)+c_1\cos\big(\omega_\beta y\big)+d_1\sin\big(\omega_\beta y\big)\,,\\
&h_2(y)=a_2\cosh\big(\eta_\alpha y\big)+c_2\cos\big(\omega_\alpha y\big)\,.
\end{split}$$
The six coefficients involved in the definition of $h_1$ and $h_2$ can be determined, in each of the five cases, by imposing the boundary and compatibility conditions. We present here only case $\textbf{c)}$, since the others cases can be treated similarly.
First of all we assume that $h_1$ satisfies the boundary conditions, i.e. $$(BCs)
\begin{cases}
h_1''(\ell)-\sigma m^2h_1(\ell)=0\\
h_1'''(\ell)-(2-\sigma)m^2h_1'(\ell)=0
\end{cases}
\qquad\Rightarrow\qquad
\begin{cases}
(\eta_\beta^2-\sigma m^2)[a_1\cosh(\eta_\beta\ell)+b_1\sinh(\eta_\beta\ell)]+\\-(\omega_\beta^2+\sigma m^2)[c_1\cos(\omega_\beta\ell)+d_1\sin(\eta_\beta\ell)]=0\\
(\eta_\beta^2+(\sigma-2) m^2)\eta_\beta[a_1\sinh(\eta_\beta\ell)+b_1\cosh(\eta_\beta\ell)]+\\
(\omega_\beta^2-(\sigma-2) m^2)\omega_\beta[c_1\sin(\omega_\beta\ell)-d_1\cos(\omega_\beta\ell)]=0,
\end{cases}$$ then we impose the compatibility conditions, i.e. $$\begin{split}
i)\\ii)\\iii)\\iv)
\end{split}
\begin{cases}
h_1(\overline{y})=h_2(\overline{y})\\ h'_1(\overline{y})=h'_2(\overline{y})\\h''_1(\overline{y})=h''_2(\overline{y})\\ h'''_1(\overline{y})=h'''_2(\overline{y})
\end{cases}
\qquad\Rightarrow\qquad
\begin{cases}
a_1\cosh\big(\eta_\beta \overline{y}\big)+b_1\sinh\big(\eta_\beta \overline{y}\big)+c_1\cos\big(\omega_\beta \overline{y}\big)
+d_1\sin\big(\omega_\beta \overline{y}\big)+\\-a_2\cosh\big(\eta_\alpha \overline{y}\big)-c_2\cosh\big(\omega_\alpha \overline{y}\big)=0
\\
a_1\eta_\beta\sinh\big(\eta_\beta \overline{y}\big)+b_1\eta_\beta\cosh\big(\eta_\beta \overline{y}\big)
-c_1\omega_\beta\sin\big(\omega_\beta \overline{y}\big)+d_1\omega_\beta\cos\big(\omega_\beta \overline{y}\big)+\\
-a_2\eta_\alpha\sinh\big(\eta_\alpha \overline{y}\big)-c_2\omega_\alpha\sinh\big(\omega_\alpha \overline{y}\big)=0
\\
a_1\eta_\beta^2\cosh\big(\eta_\beta \overline{y}\big)+b_1\eta_\beta^2\sinh\big(\eta_\beta \overline{y}\big)
-c_1\omega_\beta^2\cos\big(\omega_\beta \overline{y}\big)-d_1\omega_\beta^2\sin\big(\omega_\beta \overline{y}\big)+\\
-a_2\eta_\alpha^2\cosh\big(\eta_\alpha \overline{y}\big)-c_2\omega_\alpha^2\cosh\big(\omega_\alpha \overline{y}\big)=0\\
a_1\eta_\beta^3\sinh\big(\eta_\beta \overline{y}\big)+b_1\eta_\beta^3\cosh\big(\eta_\beta \overline{y}\big)
+c_1\omega_\beta^3\sin\big(\omega_\beta \overline{y}\big)-d_1\omega_\beta^3\cos\big(\omega_\beta \overline{y}\big)+
\\-a_2\eta_\alpha^3\sinh\big(\eta_\alpha \overline{y}\big)-c_2\omega_\alpha^3\sinh\big(\omega_\alpha \overline{y}\big)=0.
\end{cases}$$ We should solve a system of six equations and six unknowns; through some algebraic manipulations, we reduce it to a system of four equations and four unknowns $\textbf{v}=(a_1,b_1,c_1,d_1)^T$. More precisely, we get [$$\label{num2}
\begin{cases}
(BCs)\\
[\eta_\alpha^2 (h_1(\overline{y})-h_2(\overline{y}))-(h''_1(\overline{y})-h''_2(\overline{y}))]\omega_\alpha\sinh(\omega_\alpha\overline{y})=[\eta_\alpha^2(h'_1(\overline{y})-h'_2(\overline{y}))-(h'''_1(\overline{y})-h'''_2(\overline{y}))]\cosh(\omega_\alpha\overline{y})\\
[\omega_\alpha^2 (h_1(\overline{y})-h_2(\overline{y}))-(h''_1(\overline{y})-h''_2(\overline{y}))]\eta_\alpha\sinh(\eta_\alpha\overline{y})=[\omega_\alpha^2(h'_1(\overline{y})-h'_2(\overline{y}))-(h'''_1(\overline{y})-h'''_2(\overline{y}))]\cosh(\eta_\alpha\overline{y}).
\end{cases}$$]{} To system we associate a square matrix depending on the eigenvalues $\textbf{M}(\lambda)\in \mathbb{M}_4(\mathbb{R})$, hence rewrites $\textbf{M}(\lambda) \textbf{v}=\textbf{0}$; since we are interested in not trivial solutions we end up with the equation $$\label{num3}
f(\lambda):=\det\textbf{M}(\lambda)=0 \quad \text{with } \lambda >0.$$ In this way, for any $m{\geqslant}1$ fixed, the zeroes of the function $f(\lambda)$ in the interval $m^4/\beta<\lambda<m^4/\alpha$, if they exist, are the eigenvalues corresponding to eigenfunctions $\varphi^{ev}$ as in with $h_1$ and $h_2$ as in $\textbf{c)}$. This procedure can be applied to each of the five cases $\textbf{a)}-\textbf{e)}$.
The computation by hand of is very involved, thus we perform it numerically in all the five cases listed above. Our experiments reveal that cases $\textbf{b)}$ and $\textbf{d)}$ do not occur for $1{\leqslant}m{\leqslant}M$, for a suitable $M$ which, varying $\alpha$ and $\beta$, always satisfies $M\approx6/\ell$. This implies large $M$ for small $\ell$, as common in plates for bridges. Therefore, we focus on cases **a)**-**c)**-**e)**.
![Plot of $f(\lambda)$ in the cases a) (dashed), c) and e). Here $\overline{\lambda}^{ev}_{m,k}:=\overline{\lambda}_k^{ev}(\overline p_{\alpha, \beta},m)$.[]{data-label="eigplot"}](eigplot.jpg)
We point out that the plot of $f(\lambda)$ we get, see Figure \[eigplot\], is qualitatively the same for each $1{\leqslant}m{\leqslant}M$ and for all $0<\alpha<\beta$ taken. As Figure \[eigplot\] shows: we do not find eigenvalues in case $\textbf{a)}$, since $f(\lambda)>0$ for all $\lambda\in(0,m^4/\beta)$; the first eigenvalue $\overline{\lambda}^{ev}_1(\overline p_{\alpha, \beta},m)$ falls always in case $\textbf{c)}$; all the other eigenvalues corresponding to even functions fall in case $\textbf{e)}$. Furthermore, our numerical results yield the following bounds on eigenvalues corresponding to even eigenfunctions: $$\dfrac{m^4}{\beta}<\overline{\lambda}^{ev}_1(\overline p_{\alpha, \beta},m)=\overline{\lambda}_1(\overline p_{\alpha, \beta},m)<m^4,\qquad \overline{\lambda}^{ev}_k(\overline p_{\alpha, \beta},m)>\dfrac{m^4}{\alpha}\qquad{\rm for}\hspace{2mm}k\geq2\,.$$
We are now interested in checking if holds when the assumptions of Lemma \[comp1\] are not satisfied, i.e. if $$\overline{\lambda}_1^{ev}(\overline p_{\alpha, \beta},m) {\geqslant}\overline{\lambda}_1^{ev}(\overline p_{\alpha, \beta},1) \quad \text{for } m{\geqslant}2$$ when $\beta\gg(1-\sigma^2)2^4$. To this aim we study the behaviour of the maps $\beta\mapsto \overline{\lambda}^{ev}_1(\overline p_{\alpha, \beta},m)$ and $m \mapsto \overline{\lambda}^{ev}_1(\overline p_{\alpha, \beta},m)$. In Figure \[eigplot1\] we plot some points of the map $\beta\mapsto\overline{\lambda}_1^{ev}(\overline p_{\alpha, \beta},1) $ for $\alpha=0.5$, we register a similar behaviour for $\overline{\lambda}_1^{ev}(\overline p_{\alpha, \beta},m) $ with $m{\geqslant}2$. On the other hand, in Table \[num4\] we put the values of $\overline{\lambda}_1^{ev}(\overline p_{\alpha, \beta},m) $ for $m=1,\dots,10$, computed taken $\alpha=\beta=1$, i.e. $p\equiv1$, and for two suitable choices of $\alpha$ and $\beta$ with $\beta$ satisfying or not satisfying the smallness assumption of Lemma \[comp1\].
[![Plot of $\beta\mapsto \overline{\lambda}_1^{ev}(\overline p_{\alpha, \beta},1) $ with $\ell=\frac{\pi}{150}$ ($\alpha=0.5$).[]{data-label="eigplot1"}](eigplot1.png "fig:")]{}
All the numerical experiments performed suggest that $$\text{the map }\beta \mapsto \overline{\lambda}_1^{ev}(\overline p_{\alpha, \beta},m) \text{ is decreasing} \quad \text{and } \quad \overline{\lambda}_1^{ev}(\overline p_{\alpha, \beta},m) {\geqslant}(m-1)^4 \text{ for all } \beta>1\,$$ and the trend does not change varying $\ell$ and $\alpha$. In particular, the above lower bound for $\overline{\lambda}_1^{ev}(\overline p_{\alpha, \beta},m) $ does not depend on $\beta$ and, jointly with the fact that $\overline{\lambda}_1^{ev}(\overline p_{\alpha, \beta},m)<m^4$, supports the conjecture that $$\text{the map }m\mapsto \overline{\lambda}_1^{ev}(\overline p_{\alpha, \beta},m) \text{ is increasing}$$ for any $\beta>1$, hence the assumption $\beta\gg(1-\sigma^2)2^4$ of Lemma \[comp1\] seems a merely technical condition.
A positivity preserving property and proof Theorem \[corPPP\] {#pppproof}
=============================================================
In this section we state and prove some results about a positivity preserving property for the fourth order differential operator $$\label{eq:Lm}
L_m \varphi=\varphi''''-2m^2\varphi''+m^4 \varphi \, , \qquad m\in {{\mathbb N}}, \ m{\geqslant}1 \, , \ \varphi:[-\ell,\ell]\to {{\mathbb R}}\, ,$$ subject to the boundary conditions introduced in .
\[ppp\] Let $m{\geqslant}1$ be an integer, $\sigma\in(0,1)$ and let $f\in L^2(-\ell, \ell)$. Furthermore, assume that $w_m\in H^2(-\ell, \ell)$ is a weak solution to the problem $$\label{linODE}
\begin{cases}
w''''(y)-2m^2w''(y)+m^4w(y)=f (y) & \qquad y\in (-\ell,\ell) \\
w''(\pm\ell)-\sigma m^2w(\pm\ell)=0 & \qquad \\
w'''(\pm\ell)-(2-\sigma)m^2w'(\pm\ell)=0\,.& \qquad \,
\end{cases}$$ namely $$\label{eqppp}
\langle w,\phi\rangle_m=\int_{-\ell}^{\ell} f \phi \quad \text{ for all } \phi\in H^2(-\ell, \ell)\,.$$ Then the following implication holds $$f{\geqslant}0\text{ in } (-\ell, \ell)\,(f\not \equiv 0)\quad \Rightarrow \quad w(y)>0 \text{ in } [-\ell, \ell]\,.$$ Hence, the operator $L_m$ defined in , under the boundary conditions in , satisfies the positivity preserving property.
As a consequence of Theorem \[ppp\] we have
\[pppweak\] Let $m{\geqslant}1$ and $0<\sigma < 1$. Furthermore, set $\mathcal{K}:=\{\phi\in H^2(-\ell, \ell): \phi {\geqslant}0 \text{ in }(-\ell, \ell)\}$ and assume that $w\in H^2(-\ell, \ell)$ satisfies $$\label{ineqweak}
\langle w,\phi\rangle_m {\leqslant}0 \quad \text{ for all } \phi\in\mathcal{K} \,.$$ Then $$\text{either}\quad w \equiv 0 \quad \text{ or } \quad w<0 \text{ in } (-\ell, \ell)\,.$$
Let $f\in \mathcal{K}$ and let $\phi_{f}$ be the unique solution to $$\langle\phi_f,\psi\rangle_m=\int_{-\ell}^{\ell} f\psi\,dy \quad \text{ for all } \psi\in H^2(-\ell, \ell)\,.$$ By Theorem \[ppp\], $\phi_{f}\in \mathcal{K}$. Inserting $\phi_f$ in we infer $$\int_{-\ell}^{\ell} f w\,dy =\langle w,\phi_f\rangle_m {\leqslant}0 \quad \text{ for all } f\in \mathcal{K}\,.$$ Hence, $w {\leqslant}0$ in $(-\ell, \ell)$. By contradiction, assume that $w \not < 0$ in $(-\ell, \ell)$. Then, if $Z:=\{y\in (-\ell, \ell): w(y)= 0\}$, we have that the characteristic function of $Z$ satisfies $\chi_Z{\geqslant}0$ and $\chi_Z \not \equiv 0$. Let now $\phi_0\in H^2(-\ell, \ell)$ satisfy $$\langle\phi_0,\psi\rangle_m=\int_{-\ell}^{\ell} \chi_Z \psi\,dy \quad \text{ for all } \psi\in H^2(-\ell, \ell)\,.$$ Since, by elliptic regularity, $\phi_0\in C^3([-\ell, \ell])$ and, by Theorem \[ppp\], $\phi_0>0$ in $[-\ell, \ell]$, we deduce that for every $\phi\in H^2(-\ell, \ell)$ there exist $t_1{\leqslant}0{\leqslant}t_0$: $\phi +t_0\phi_0{\geqslant}0$ and $\phi+t_1\phi_0{\leqslant}0$ in $[-\ell, \ell]$. Furthermore, by definition of $\phi_0$ we have $$\langle\phi_0,w\rangle_m=\int_{-\ell}^{\ell} \chi_Z w\,dy=0\,.$$ Combining this with , we deduce $$0{\geqslant}\langle \phi+t_0\phi_0,w\rangle_m=\langle\phi,w\rangle_m$$ and $$0{\leqslant}\langle \phi+t_1\phi_0,w\rangle_m=\langle\phi,w\rangle_m \,.$$ Namely, $$\langle\phi,w\rangle_m=0 \quad \text{ for all } \phi\in H^2(-\ell, \ell)\,.$$ Taking $\phi=w$ in the above inequality we conclude $w \equiv 0$ in $(-\ell, \ell)$ and the proof follows.
We conclude this section with the proof of Theorem \[ppp\].
**Proof of Theorem \[ppp\].**
The proof follows by a direct inspection of the sign of the unique solution to . First we note that, for $m{\geqslant}1$ fixed and $f\in L^2(-\ell, \ell)$, all solutions to the equation $$w''''(y)-2m^2w''(y)+m^4w(y)=\overline f \quad \text{in } \mathcal{D}'({{\mathbb R}})\,,$$ where $\overline f$ denotes the trivial extension of $f$ to ${{\mathbb R}}$, write $$w(y)=c_1 \cosh(my)+c_2 \sinh(my)+c_3 y\cosh(my) +c_4y\sinh(my)+w_p(y)\,,$$ with $c_1,c_2,c_3,c_4 \in {{\mathbb R}}$ and $$w_p(y)=(q_m*\overline f)(y)=\int_{-\infty}^{+\infty}q_m(t) \overline f(y-t)\,dt$$ where $$q_m(y)=\frac{(1+m|y|)e^{-m|y|}}{4m^3}\,.$$ Exploiting the regularity of $q_m$, it follows that all the above solutions belong to $C^3({{\mathbb R}})$ (the regularity can be improved by increasing the regularity of $f$); the thesis can be reached proving that $$\widetilde{w}(y)=c_1\cosh(my)+c_2 \sinh(my)+c_3 y\cosh(my)+c_4y\sinh(my)>0$$ since $w_p(y){\geqslant}0$.
If we fix the constants $c_1,c_2,c_3,c_4 \in {{\mathbb R}}$ in such a way that: $$\begin{cases}
w''(\pm\ell)-\sigma m^2w(\pm\ell)=0 & \qquad \\
w'''(\pm\ell)-(2-\sigma)m^2w'(\pm\ell)=0\,,& \qquad \,
\end{cases}$$ then the restriction of $w$ to $[-\ell, \ell]$, that we will still denote with $w$, is the unique solution to . More precisely, by imposing the above conditions we obtain the system which decouples in the following two systems By setting $$F_m(\ell):=(3+\sigma)\sinh(m\ell)\cosh(m\ell)-m \ell (1-\sigma)>0\,,$$ $$\overline F_m(\ell):=(3+\sigma)\sinh(m\ell)\cosh(m\ell)+m \ell (1-\sigma)>0\,,$$ $$A_m(\ell):=(1+\sigma)\sinh(m\ell)-(1-\sigma)m\ell \cosh(m\ell)\,, \quad B_m(\ell):= 2\cosh(m\ell)+(1-\sigma)m\ell\sinh(m\ell)\,,$$ $$\overline A_m(\ell):=(1+\sigma)\cosh(m\ell)-(1-\sigma)m\ell \sinh(m\ell)\,,\quad \overline B_m(\ell):= 2\sinh(m\ell)+(1-\sigma)m\ell\cosh(m\ell)\,,$$ $$V_m(\ell):=\sigma m^2 w_p(\ell)-w''_p(\ell)\,, \quad W_m(\ell):=(\sigma-2) m^2 w_p'(\ell)+w'''_p(\ell)\,,$$
$$V_m(-\ell):=\sigma m^2 w_p(-\ell)-w''_p(-\ell)\,, \quad W_m(-\ell):=(\sigma-2) m^2 w_p'(-\ell)+w'''_p(-\ell)\,,$$
the solutions to the above systems write $$c_1=\frac{m A_m(\ell)[V_m(\ell)+V_m(-\ell)]+B_m(\ell)[W_m(\ell)-W_m(-\ell)]}{2m^3(1-\sigma) F_m(\ell)}$$ $$c_2=\frac{m \overline A_m(\ell)[V_m(\ell)-V_m(-\ell)]+\overline B_m(\ell)[W_m(\ell)+W_m(-\ell)]}{2m^3(1-\sigma) \overline F_m(\ell)}$$ $$c_3=\frac{m \cosh(m\ell)[V_m(\ell)-V_m(-\ell)]-\sinh(m\ell)[W_m(\ell)+W_m(-\ell)]}{2m^2 \overline F_m(\ell)}$$ $$c_4=\frac{m \sinh(m\ell)[V_m(\ell)+V_m(-\ell)]-\cosh(m\ell)[W_m(\ell)-W_m(-\ell)]}{2m^2 F_m(\ell)}\,.$$
By exploiting the symmetry of $w_m$, for $i=0$ and $i=2$, we have $$w_p^{(i)}(\ell)=\int_{0}^{2\ell}q_m^{ (i) }(t) f(\ell-t)\,dt\,, \quad w_p^{(i)}(-\ell)=\int_{0}^{2\ell}q_m^{ (i) }(t) f(-\ell+t)\,dt\,,$$ while, for $i=1$ and $i=3$, we have $$w_p^{(i)}(\ell)=\int_{0}^{2\ell}q_m^{ (i) }(t) f(\ell-t)\,dt\,, \quad w_p^{(i)}(-\ell)=-\int_{0}^{2\ell}q_m^{ (i) }(t) f(-\ell+t)\,dt\,.$$ Hence, $$V_m(\ell)=\int_{0}^{2\ell}\frac{e^{-mt}}{4m}(1+\sigma-mt(1-\sigma)) f(\ell-t)\,dt\,,\quad \quad W_m(\ell)=\int_{0}^{2\ell}\frac{e^{-mt}}{4}(2+mt(1-\sigma)) f(\ell-t)\,dt\,$$ and $$V_m(-\ell)=\int_{0}^{2\ell}\frac{e^{-mt}}{4m}(1+\sigma-mt(1-\sigma)) f(-\ell+t)\,dt\,,\quad \quad W_m(-\ell)=-\int_{0}^{2\ell}\frac{e^{-mt}}{4}(2+mt(1-\sigma)) f(-\ell+t)\,dt\,.$$ First of all we study the sign of the coefficients $c_1$ and $c_4$. Since $F_m(\ell)>0$, $c_1$ has the same sign of $$m A_m(\ell)[V_m(\ell)+V_m(-\ell)]+B_m(\ell)[W_m(\ell)-W_m(-\ell)]=$$ $$\int_{0}^{2\ell}\frac{e^{-mt}}{4}\,\left[A_m(\ell)(1+\sigma-m(1-\sigma)t)+B_m(\ell)(2+m(1-\sigma)t) \right] \left[f(\ell-t)+ f(-\ell+t)\right]\,dt$$
$$\label{c11}
=[\sinh(m\ell)((1+\sigma)^2+2m\ell(1-\sigma))+\cosh(m\ell)(4-(1-\sigma^2)m\ell)]\int_{0}^{2\ell}\frac{e^{-mt}}{4}\,[f(\ell-t)+ f(-\ell+t)]\,dt+$$
$$\label{c12}
+m(1\!-\!\sigma)\!\left[2\cosh(m\ell)\!-\!(\!1+\!\sigma)\sinh(m\ell)\!+\!(1\!-\!\sigma)m\ell(\cosh(m\ell)\!+\!\sinh(m\ell))\right]
\int_{0}^{2\ell}\dfrac{e^{-mt}t}{4}\, [f(\ell-t)+ f(-\ell+t)]\,dt\,.$$
We observe that $$\label{diseq1}
2\cosh(z)>(1+\sigma)\sinh(z)$$ for $z>0$ and for all $\sigma\in (0,1)$, implying that is positive; about the sign of we introduce the map $$z\mapsto g(z):=\sinh(z)((1+\sigma)^2+2z(1-\sigma))+\cosh(z)(4-(1-\sigma^2)z)$$ and we compute its derivative $$g'(z)=2\sigma(1+\sigma)\cosh(z)+2(3-\sigma)\sinh(z)+z(1-\sigma)[2\cosh(z)-(1+\sigma)\sinh(z)].$$ Thanks to , for $z>0$ we obtain $g'(z)>0$ so that $g(z)$ is always positive ($g(0)=4$) and in particular $c_1>0$. The sign of $c_4$ depends on $$m \sinh(m\ell)[V_m(\ell)+V_m(-\ell)]-\cosh(m\ell)[W_m(\ell)-W_m(-\ell)]=$$ $$\begin{split}
\int_{0}^{2\ell}\frac{e^{-mt}}{4}\big[(1+\sigma)\sinh(m\ell)-2\cosh(m\ell)-mt(1-\sigma)(\sinh(m\ell)+\cosh(m\ell))\big][f(\ell-t)+ f(-\ell+t)]\,dt
\end{split}$$ that, applying again , gives $c_4<0$ for all $\sigma\in (0,1)$ and $m\ell>0$.
For our purposes we need to compare the absolute value of $c_4$ and $c_3$; since the sign of $c_3$ is not known a priori, we study the sign of $2m^2(|c_4|\pm c_3)$, i.e. $$\begin{aligned}
& \int_{0}^{2\ell}\frac{e^{-mt}}{4 F_m(\ell)}\big[2\cosh(m\ell)-(1+\sigma)\sinh(m\ell)+mt(1-\sigma)(\sinh(m\ell)+\cosh(m\ell))\big][f(\ell-t)+ f(-\ell+t)]\,dt \\
&\pm\int_{0}^{2\ell}\frac{e^{-mt}}{4 \overline{F}_m(\ell)}\big[(1+\sigma)\cosh(m\ell)-2\sinh(m\ell)-mt(1-\sigma)(\sinh(m\ell)+\cosh(m\ell))\big]
[f(\ell-t)-f(-\ell+t)]\,dt \, .\end{aligned}$$ Recalling that $0<F_m(\ell)<\overline{F}_m(\ell)$, we obtain the positivity of $$\begin{aligned}
& m(1\!\!-\!\!\sigma)[\sinh(m\ell)\!+\!\cosh(m\ell)] \bigg\{\!\bigg[\!\frac{1}{F_m(\ell)}\!\!\mp\!\!\frac{1}{\overline{F}_m(\ell)}\bigg]\!\!\int_{0}^{2\ell}
\frac{e^{-mt}t}{4}f(\ell\!\!-\!\!t)\,dt
\!+\!\bigg[\frac{1}{F_m(\ell)}\!\!\pm\!\! \frac{1}{\overline{F}_m(\ell)}\bigg]\!\!\int_{0}^{2\ell}\dfrac{e^{-mt}t}{4}f(-\ell+t)\,dt\bigg\};\end{aligned}$$ thus $2m^2(|c_4|\pm c_3)>0$ if $$\frac{2\cosh(m\ell)-(1+\sigma)\sinh(m\ell)}{F_m(\ell)}\pm\frac{(1+\sigma)\cosh(m\ell)-2\sinh(m\ell)}{\overline{F}_m(\ell)}>0;$$ the achievement follows from the positivity of $\big(\cosh(z)\mp\sinh(z)\big)\big(2\pm(1+\sigma)\big)$ for all $z>0$ and $\sigma\in(0,1)$.
Fixed $m\geq1$, we set $$\widetilde{\psi}(t):=m\,\widetilde{w}(t/m)=c_1m\cosh t+c_2m\sinh t+c_3t\cosh t+c_4t\sinh t$$ and we focus on the qualitative behaviour of $\widetilde{\psi}$ where, from above, $c_1>0$, $c_4<0$ and $c_4<c_3<-c_4$. Clearly, $\widetilde{\psi}(t)$ is continuous and differentiable on $\mathbb{R}$, moreover $$\widetilde{\psi}(0)=m\,c_1>0\,;\quad \widetilde{\psi}(t)\sim \dfrac{c_3\pm c_4}{2}te^{|t|}\rightarrow-\infty \hspace{2mm}{\rm for }
\hspace{2mm}t\rightarrow\pm\infty \quad.$$ This fact implies that $\widetilde{\psi}(t)$ has at least two zeros of opposite sign on $\mathbb{R}$; we prove now that $\widetilde{\psi}(t)$ has exactly two distinct zeros on $\mathbb{R}$.
We know that $\widetilde{\psi}(t)=0$ if and only if $$\alpha(t):=(c_2m+c_4t)\tanh t+c_3t+c_1m=0 \, .$$ Computing $\alpha'(t)=\frac{1}{2\cosh^2(t)}(2c_3\cosh^2(t)+c_4\sinh(2t)+2c_4t+2c_2m)$ we observe that $$\label{zeros}
\exists!\hspace{1mm} \overline{t}\in\mathbb{R}: \alpha'(\overline{t})=0 \, .$$ This follows because $\beta(t):=2c_3\cosh^2(t)+c_4\sinh(2t)+2c_4t+2c_2m$ is always decreasing on $\mathbb{R}$; indeed $c_4<0$, $|c_4|>|c_3|$ so that $\beta'(t)=2(c_3\sinh(2t)+c_4\cosh(2t)+c_4)<0$. Moreover $\beta(t)\sim\dfrac{c_3\pm c_4}{2}e^{2|t|}\rightarrow\mp\infty$ for $t\rightarrow\pm\infty$. Now let us suppose for contradiction that $\widetilde{\psi}(t)$ has more than two zeros on $\mathbb{R}$, for instance it has 3 distinct zeros $t_1<t_2<t_3$; this implies that $\alpha(t)$ has 3 distinct zeros, then, the Rolle’s Theorem applied to $\alpha(t)$ in the intervals $[t_1,t_2]$ and $[t_2,t_3]$ ensures the existence of at least two points in which $\alpha'(t)=0$ on $\mathbb{R}$ and this contradicts . Hence, $\widetilde{\psi}$, and in turn also $\widetilde{w}$, has exactly two zeros of opposite sign on $\mathbb{R}$.
Since $\widetilde{w}(y)$ has exactly two zeros of opposite sign on $\mathbb{R}$ and $\widetilde{w}(0)>0$, if we prove that $\widetilde{w}(\pm\ell)>0$ the thesis follows. To this aim we study the sign of $\widetilde{w}(\pm\ell)=c_1\cosh(m\ell)\pm c_2\sinh(m\ell)\pm c_3\ell\cosh(m\ell)+c_4\ell\sinh(m\ell)$, in particular we consider $$\begin{aligned}
& 2m^2\widetilde{w}(\ell)=\int_{0}^{2\ell}\frac{e^{-mt}}{4m}[C_m(\ell)f(\ell-t)+\overline{C}_m(\ell)f(-\ell+t)]\,dt
+\int_{0}^{2\ell}\frac{e^{-mt}t}{4}[D_m(\ell)f(\ell-t)+\overline{D}_m(\ell)f(-\ell+t)]\,dt \\[8pt]
& 2m^2\widetilde{w}(-\ell)=\int_{0}^{2\ell}\frac{e^{-mt}}{4m}[\overline{C}_m(\ell)f(\ell-t)+C_m(\ell)f(-\ell+t)]\,dt
+\int_{0}^{2\ell}\frac{e^{-mt}t}{4}[\overline{D}_m(\ell)f(\ell-t)+D_m(\ell)f(-\ell+t)]\,dt\end{aligned}$$ where $$\begin{aligned}
& C_m(\ell)\!\!=\!\!\frac{4}{1-\sigma}\bigg(\frac{\cosh^2(m\ell)}{F_m(\ell)}\!+\!\frac{\sinh^2(m\ell)}{\overline{F}_m(\ell)}\bigg)
\!+\!\frac{(1+\sigma)^2}{2(1-\sigma)}\sinh(2m\ell)\bigg(\frac{1}{F_m(\ell)}\!+\!\frac{1}{\overline{F}_m(\ell)}\bigg)
\!-\!m\ell(1+\sigma)\bigg(\frac{1}{F_m(\ell)}\!-\!\frac{1}{\overline{F}_m(\ell)}\bigg)\\
& D_m(\ell)=2\bigg(\frac{\cosh^2(m\ell)}{F_m(\ell)}+\frac{\sinh^2(m\ell)}{\overline{F}_m(\ell)}\bigg)-\frac{1+\sigma}{2}\sinh(2m\ell)\bigg(\frac{1}{F_m(\ell)}
+\frac{1}{\overline{F}_m(\ell)}\bigg)+m\ell(1-\sigma)\bigg(\dfrac{1}{F_m(\ell)}-\frac{1}{\overline{F}_m(\ell)}\bigg) \\
& \overline{C}_m(\ell)\!\!=\!\!\frac{4}{1-\sigma}\bigg(\frac{\cosh^2(m\ell)}{F_m(\ell)}\!-\! \frac{\sinh^2(m\ell)}{\overline{F}_m(\ell)}\bigg)
+\frac{(1+\sigma)^2}{2(1-\sigma)}\sinh(2m\ell)\bigg(\frac{1}{F_m(\ell)}\!-\!\frac{1}{\overline{F}_m(\ell)}\bigg)\!-\!m\ell(1+\sigma)\bigg(\frac{1}{F_m(\ell)}
\!+\!\frac{1}{\overline{F}_m(\ell)}\bigg) \\
&\overline{D}_m(\ell)=2\bigg(\frac{\cosh^2(m\ell)}{F_m(\ell)}-\frac{\sinh^2(m\ell)}{\overline{F}_m(\ell)}\bigg)
-\frac{1+\sigma}{2}\sinh(2m\ell)\bigg(\frac{1}{F_m(\ell)}-\frac{1}{\overline{F}_m(\ell)}\bigg)
+m\ell(1-\sigma)\bigg(\frac{1}{F_m(\ell)}+\frac{1}{\overline{F}_m(\ell)}\bigg).\end{aligned}$$ The final part of the proof is devoted to prove that the coefficients $C_m(\ell), D_m(\ell), \overline{C}_m(\ell)$ and $\overline{D}_m(\ell)$ are positive. We recall that $$\frac{1}{F_m(\ell)}+\dfrac{1}{\overline{F}_m(\ell)}=\dfrac{(3+\sigma)\sinh(2m\ell)}{F_m(\ell)\overline{F}_m(\ell)}>0
\hspace{10mm}
\frac{1}{F_m(\ell)}-\dfrac{1}{\overline{F}_m(\ell)}=\dfrac{2m\ell(1-\sigma)}{F_m(\ell)\overline{F}_m(\ell)}>0 \, ,$$ and we introduce four maps related respectively to the previous coefficients $$\begin{aligned}
& z\mapsto p(z):=\dfrac{2(3+\sigma)}{1-\sigma}\sinh(2z)\cosh(2z)+4z+\dfrac{(1+\sigma)^2(3+\sigma)}{2(1-\sigma)}\sinh^2(2z)-2(1-\sigma^2)z^2 \\[8pt]
& z\mapsto q(z):=\dfrac{3+\sigma}{2}\sinh(2z)[2\cosh(2z)-(1+\sigma)\sinh(2z)]+2(1-\sigma)z+2(1-\sigma)^2z^2 \\[8pt]
& z\mapsto r(z):=\dfrac{2(3+\sigma)}{1-\sigma}\sinh(2z)+z[4\cosh(2z)-2(1+\sigma)\sinh(2z)] \\[8pt]
& z\mapsto s(z):=(3+\sigma)\sinh(2z)+(1-\sigma)z[2\cosh(2z)-(1+\sigma)\sinh(2z)]+(1-\sigma)(3+\sigma)z\sinh(2z).\end{aligned}$$ Thanks to $q(z)$, $r(z)$ and $s(z)$ are always positive for $z>0$ and for all $\sigma\in(0,1)$. The same conclusion holds for the maps $p(z)$, due to the following inequality $$\frac{(1+\sigma)^2(3+\sigma)}{1-\sigma}\sinh^2(2z)>3\sinh^2(2z)>(1-\sigma^2)(2z)^2 \, .$$ This completes the proof.
**Proof of Theorem \[corPPP\].**
The proof readily follows as a corollary of Theorem \[ppp\] by exploiting the same separation of variables performed in the Proof of Theorem \[min\].
**Acknowledgments.** The first three authors are members of the Gruppo Nazionale per l’Analisi Matematica, la Probabilità e le loro Applicazioni (GNAMPA) of the Istituto Nazionale di Alta Matematica (INdAM) and are partially supported by the INDAM-GNAMPA 2019 grant “Analisi spettrale per operatori ellittici con condizioni di Steklov o parzialmente incernierate” and by the PRIN project “Direct and inverse problems for partial differential equations: theoretical aspects and applications” (Italy). The third author was partially supported by the INDAM-GNAMPA 2018 grant “Formula di monotonia e applicazioni: problemi frazionari e stabilità spettrale rispetto a perturbazioni del dominio” and by the research project “Metodi analitici, numerici e di simulazione per lo studio di equazioni differenziali a derivate parziali e applicazioni” Progetto di Ateneo 2016, Università del Piemonte Orientale “Amedeo Avogadro”. The research of the fourth author is partially supported by an INSPIRE faculty fellowship (India).
[10]{} C. Anedda and F. Cuccu, [*Steiner symmetry in the minimization of the first eigenvalue in problems involving the p-Laplacian*]{}, Proc. Amer. Math. Soc. 144, (2016), 3431–3440.
C. Anedda, F. Cuccu and G. Porru, [*Minimization of the first eigenvalue in problems involving the bi-laplacian*]{}, Rev. Mate. Teor. Appl. 16, (2009), 127–136.
J. R. Banerjee, [*A simplified method for the free vibration and flutter analysis of bridge decks*]{}, J. Sound Vibration 260 (2003), 829–845.
E. Berchio, D. Buoso, F. Gazzola, [*On the variation of longitudinal and torsional frequencies in a partially hinged rectangular plate*]{}, ESAIM Control Optim. Calc. Var. 24 (2018), 63–87.
E. Berchio, A. Ferrero, F. Gazzola, [*Structural instability of nonlinear plates modelling suspension bridges: mathematical answers to some long-standing questions*]{}, Nonlin. Anal. Real World Appl. 28 (2016), 91–125.
E. Berchio, D. Buoso, F. Gazzola, D. Zucco, [*A Minimaxmax Problem for Improving the Torsional Stability of Rectangular Plates*]{}, J. Optim. Theory Appl. 177 (2018), 64–92.
S. Chanillo, [*Conformal geometry and the composite membrane problem*]{}, Anal. Geom. Metr. Spaces 1 (2013), 31–35.
S. Chanillo, D. Grieser, K. Kurata, [*The free boundary problem in the optimization of composite membranes*]{}, Differential geometric methods in the control of partial differential equations (Boulder, CO, 1999), 61-81, Contemp. Math., 268, Amer. Math. Soc., Providence, RI, 2000.
S. Chanillo, D. Grieser, M. Imai, K. Kurata and I. Ohnishi, [*Symmetry breaking and other phenomena in the optimization of eigenvalues for composite membranes*]{}, Comm. Math. Phys. 214, (2000), 315–337.
S. Chanillo, C. Kenig, T. To, [*Regularity of the minimizers in the composite membrane problem in ${{\mathbb R}}^2$*]{}, J. Funct. Anal. 255 (2008), no. 9, 2299–2320.
F. Colasuonno, E. Vecchi, [*Symmetry in the composite plate problem*]{}, Commun. Contemp. Math. (2018), in press.
F. Colasuonno, E. Vecchi, [*Symmetry and rigidity for the hinged composite plate problem*]{}, J. Differential Equations 266 (2019), no. 8, 4901–4924.
F. Cuccu, B. Emamizadeh and G. Porru, [*Optimization problems for an elastic plate*]{}, J. Math. Phys. 47, (2006), 12 pp.
F. Cuccu, B. Emamizadeh and G. Porru, [*Optimization of the first eigenvalue in problems involving the p-Laplacian*]{}, Proc. Amer. Math. Soc. 137, (2009), 1677–1687.
A. Ferrero, F. Gazzola, [*A partially hinged rectangular plate as a model for suspension bridges*]{}, Disc. Cont. Dyn. Syst. A. 35 (2015), 5879–5908.
F. Gazzola, [*Mathematical models for suspension bridges*]{}, MS&A Vol.15, Springer, 2015.
F. Gazzola, H. C. Grunau, G. Sweers, [*Polyharmonic boundary value problems*]{}, LNM 1991 Springer, 2010.
J. A. Jurado, S. Hernández, F. Nieto, A. Mosquera, [*Bridge aeroelasticity: sensitivity analysis and optimum design (high performance structures and materials)*]{}, WIT Press - Computational Mechanics, Southampton, 2011.
B. Kawohl, J. Stará, G. Wittum, [*Analysis and numerical studies of a problem of shape design*]{}, Arch. Rat. Mech. Anal. 114, (1991) 349–363.
P. D. Lamberti, D. Buoso, [*Eigenvalues of polyharmonic operators on variable domains*]{}, ESAIM: Control, Optim. Calc. Var., 19, (2013), n.4, 1225–1235.
F. Murat, L. Tartar, [*Calculus of variations and homogenization*]{}, Topics in the Math. Modelling of Composite Materials 31, Progr. Nonlin. Diff. Eq. Appl. (1997) 139–173.
C. L. Navier, [*Extraits des recherches sur la flexion des plans élastiques*]{}, Bulletin des Sciences de la Société Philomathique de Paris (1823), 92–102.
S. A. Nazarov, G. H. Sweers, A. S. Slutskij, [*Homogenization of a thin plate reinforced with periodic families of rigid rods*]{}, Sbornik Mathematics 202, (2011), 1127–1168.
H. Shahgholian, [*The singular set for the composite membrane problem*]{}, Comm. Math. Phys. 271, (2007), 93–101.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We review the Batalin-Tyutin approach of quantising second class systems which consists in enlarging the phase space to convert such systems into first class. The quantisation of first class systems, it may be mentioned, is already well founded. We show how the usual analysis of Batalin-Tyutin may be generalised, particularly if one is dealing with nonabelian theories. In order to gain a deeper insight into the formalism we have considered two specific examples of second class theories– the massive Maxwell theory (Proca model) and its nonabelian extension. The first class constraints and the involutive Hamiltonian are explicitly constructed. The connection of our Hamiltonian approach with the usual Lagrangian formalism is elucidated. For the Proca model we reveal the importance of a boundary term which plays a significant role in establishing an exact identification of the extra fields in the Batalin-Tyutin approach with the Stückelberg scalar. Some comments are also made concerning the corresponding identification in the nonabelian example.'
---
——————————————————————————-
[**Quantisation of second class systems in the\
Batalin-Tyutin formalism**]{}\
N. Banerjee[^1]\
Saha Institute of Nuclear Physics, 1/AF Bidhannagar\
Calcutta 700064, India\
R. Banerjee\
S. N. Bose National center for Basic Sciences\
DB-17, Salt Lake, Calcutta 700064, India\
Subir Ghosh\
Gobardanga Hindu College, North 24 Parganas\
West Bengal, India\
**Introduction**
================
Canonical quantisation of systems with first class constraints was formulated along general lines by Dirac [@r1]. The corresponding analysis in the path integral approach was initiated by Faddeev [@r2] for gauge theories. It was extended by Fradkin and collaborators [@r3; @r4] within the broader framework of preserving Becchi-Rouet-Stora-Tyutin (BRST) [@r5] invariance. The quantisation of systems with second class constraints, on the contrary, poses problems. In this case it is necessary to replace the canonical Poisson brackets (P.B) by their corresponding Dirac brackets (D.B). The conversion of the D.B’s to quantum commutators is, in general, plagued with severe factor ordering problems. Moreover, the abstraction of the canonically conjugate variables is highly nontrivial. Consequently the quantisation of second class systems, either in the canonical or in the path integral formalisms, is problematic. It may be mentioned that ther is a factor ordering problem of a different nature in first class systems. There are three varities of this problem. First, the constraints which close under P.B may do so with coefficients which are functions of the canonical variables rather than just structure constants. Second, the P.B of the constraints with the Hamiltonian may yield a combination of constraints with phase space coefficients. Finally, the Poisson algebra of the constraints with any physical variable must yield a combination of constraints which may involve structure functions instead of structure constants. In all these cases a suitable factor ordering has to be found such that the structure function operators [*precede*]{} the correctly ordered constraints [@r1; @r6]. It is only then possibly to develop a consistent quantisation program. This has been done by Kuchar [@r6] in a covariant manner. The factor ordering concerning the interpretation of the D.B as a quantum commutator remains an open issue.
In view of the above discussion it becomes natural to formulate the quantisation of second class systems without invoking D.B. A possible way would be to embed such a system in an extended space so that it gets converted into first class. One can then apply the well established machinery [@r3; @r4; @r5] for quantising first class systems. This philosophy has been recently adopted by Batalin-Fradkin [@r7] and Batalin-Tyutin [@r8]. The phase space is extended by introducing new variables which transform the original second class system into first class. It is worthwhile to mention that this idea is a logical follow up of the original notion of Stückelberg [@r9] who extended the configuration space to convert second class theories into first class. The approach of Stückelberg [@r9] is in the Lagrangian formulation which should be contrasted with the Hamiltonian formulation of Batalin-Fradkin [@r7] or Batalin-Tyutin [@r8]. A similar (Lagrangian) functional integral approach has been used by Faddeev and Satashvili [@r10] to introduce Wess-Zumino scalars [@r11] to interpret anomalous gauge theories as true ([*i.e.*]{} first class) gauge systems. Recently there have been suggestions [@r12; @r13] that the extra fields introduced in the Hamiltonian formalism may be identified with the Stückelberg scalar or the Wess-Zumino fields. We will return to this point later.
The purpose of this paper is to make a thorough investigation into the Hamiltonian formulation of Batalin-Tyutin [@r8] to convert second class systems into first class, by referring to two examples– the abelian (Proca) model and its nonabelian extension. In the course of this analysis the precise connection with the Lagrangian formulation as well as the identification of the extra fields with the Stückelberg [@r9] scalar will also be elucidated.
In section 2 we first briefly review the analysis of Batalin-Tyutin [@r8]. This will serve a twofold purpose; to set up the notation and familiarise the reader with the basics of the formalism. This formalism [@r8] is ideal for discussing abelian (second class) systems. To appreciate this point it may be recalled that, for abelian first class systems, the usual algebra of constraints among themselves and with the Hamiltonian is strongly involutive. The method developed in [@r8] also yields a strongly involutive algebra for first class systems. For nonabelian (first class) theories, on the other hand, it is well known that the corresponding algebra is only weakly involutive. We, therefore, generalise the approach of [@r8] to include this possibility. This has been discussed in great details. Its utility is made transparent when we actually convert a nonabelian second class theory into first class. We are able to reproduce the algebra which occurs in usual nonabelian first class theories. If the standard proceedure of [@r8] were adopted we would obviously fail to generate this algebra. This will also have some implications in connecting the Hamiltonian formalism with the Lagrangian version.
Section 3 is devoted to an application of the ideas discussed in section 2 to the abelian Proca model [@r9]. The original second class system is transformed into first class by initially converting the second class constraints to first class and then changing the Hamiltonian into the corresponding involutive form. The phase space partition function is constructed and explicitly evaluated in two special gauges. In the unitary gauge [@r7; @r8] which corresponds to taking the initial second class constraints as the gauge fixing conditions, the original theory is reproduced. Nontrivial consequences are obtained in the Faddeev-Popov [@r14] like gauges ([*i.e.*]{} gauges not involving the momenta). It leads to a Lagrangian embedding which reveals the first class nature of the theory. The connection with the conventional Stückelberg [@r9] Lagrangian mechanism is established. It is shown how a boundary term plays a crucial role in making a one-to-one correspondence between the extra field in the Hamiltonian formalism with the Stückelberg [@r9] scalar.
In section 4 we introduce the nonabelian version of the Proca model. We first point out a flaw in the conventional Dirac analysis performed by Senjanovic [@r15]. The correct D.B’s among the canonical variables are worked out. These are found to be field dependent. We next exploitthe generalised version of the Batalin-Tyutin [@r8] approach developed in section 2 to convert thye nonabelian second class system into first class. It is interesting to observe that while one of the (first class) constraints has a closed form, the other constraint and the involutive Hamiltonian are expressed by a power series in terms of the new fields and do not have any closed expressions. Specialising to the case where the gauge group is $SU(2)$, we are able to reproduce the gauge algebra given in the conventional Lagrangian formulation [@r16] obtained by the nonabelian Stückelberg mechanism. Although the identification of the extra fields in the Hamiltonian formulation with the Stckelberg fields (which now appear as the Euler angles [@r16]) in the Lagrangian approach is not so direct as in the abelian example, nevertheless there is no conceptual difficulty in understanding this correspondence. Finally we show that contrary to claims in the literature [@r17], the nonabelian extension of the usual Stückelberg [@r9] mechanism fails to convert the nonabelian second class theory into first class.
Our concluding observations are given in section 5.
**General formalism**
=====================
In this section we first review the abelian conversion of a second class system into first class as developed by Batalin and Tyutin (BT) [@r8]. We then discuss a natural extension of this approach which is suitable for analysing nonabelian models. Simultaneously our notations and conventions will also be specified.
Let us assume that the canonical variables, $(\phi_i(x),\pi^i(x))$ with $|i|=n$ and a Grassman parity $\epsilon(\phi_i) = \epsilon(\pi^i) = \epsilon_i$ define the initial phase space of a dynamical system. We further suppose that this system contains a set of linearly independent bosonic second class constraints $\Theta_\alpha$ which are some functions of the original canonical variables. $$\Theta_\alpha = \Theta_\alpha (\phi,\pi),~ \epsilon_\alpha = 0,~ |\alpha| =
m<2n \label{eq:2.1}$$ so that the matrix, $$\Delta_{\alpha\beta} (x,y) =
\left\{\Theta_\alpha
(x),\Theta_\beta(y)\right\}\label{eq:2.2}$$ has a nonvanishing determinant. The inclusion of other constraints ([*i.e.*]{} fermion or first class) is a matter of technical detail and poses no problems in developing the formalism.
We now convert the second class system into first class. The initial step is to obtain the first class constraints starting from (\[eq:2.1\]). Foliowing the general philosophy, new dynamical fields, $$\Phi^\alpha,~~\epsilon_\alpha = 0,~~|\alpha| = m\label{eq:2.3}$$ with the basic P.Bs, $$\left\{\Phi^\alpha(x),~~\Phi^\beta(y)\right\} ~=
\omega^{\alpha\beta}(x,y)\label{eq:2.4}$$ where $\omega$ is an invertible field independent antisymmetric matrix, $$\omega^{\alpha\beta}(x,y) = -\omega^{\beta\alpha}(y,x),
{}~~\epsilon(\omega^{\alpha\beta})
= \epsilon_\alpha+\epsilon_\beta,\label{eq:2.5}$$ are introduced to extend the original phase space, $$(\phi,\pi)\oplus (\Phi).\label{eq:2.6}$$
The new first class constraints of the system in the extended phase space (\[eq:2.6\]) are denoted by $\tau_\alpha$: $$\tau_\alpha = \tau_\alpha (\phi, \pi, \Phi);~~\epsilon(\tau_\alpha) =
\epsilon_\alpha,
{}~|\alpha| = m.\label{eq:2.7}$$ Then the abelian conversion of BT [@r8] implies that these constraints are strongly involutive, $$\left\{\tau_\alpha(x)~,~\tau_\beta(y)\right\} = 0\label{eq:2.8}$$ subject to the boundary condition, $$\tau_\alpha(\phi,\pi,0) = \Theta_\alpha(\phi, \pi)\label{eq:2.9}$$ where the right hand side of (\[eq:2.9\]) is just the original set of second class constraints (\[eq:2.1\]).
In ref. [@r8] it has been shown that it is possible to explicitly construct the solution to the algebraic problem (\[eq:2.8\]) and (\[eq:2.9\]) in the form of a power series expansion, $$\tau_\alpha (\phi, \pi, \Phi) = \sum_{n=0}^\infty \tau_\alpha^{(n)},~~
\tau_\alpha^{(n)}
\sim \Phi^n.\label{eq:2.10}$$ The first term yields (\[eq:2.9\]) $$\tau_\alpha^{(0)} = \Theta_\alpha\label{eq:2.11}$$ The remaining terms are found by inserting (\[eq:2.10\]) in (\[eq:2.8\]), using the relation (\[eq:2.2\]) and finally identifying contributions with identical powers of $\Phi^{\alpha}$. This leads to the following recursion relations, $$\begin{aligned}
\{\tau_\alpha^{(0)}(x),\tau_\beta^{(0)}(y)\}_{(\phi,\pi)} +
\{\tau_\alpha^{(1)}(x),
\tau_\beta^{(1)}(y)\}_{(\Phi)} &= 0\\\label{eq:2.12}
\{\tau_{[\alpha}^{(1)}(x),\tau_{\beta ]}^{(n+1)}(y)\}_{(\Phi)} +
B_{\alpha\beta}^{(n)}
(x,y) &=0,~~n\geq 1\label{eq:2.13}\end{aligned}$$ with, $$B_{\alpha\beta}^{(1)}(x,y)\equiv
\{\tau_{[\alpha}^{(0)}(x),\tau_{\beta]}^{(1)}(y)\}_{(\phi,
\pi)}\label{eq:2.14}$$ $$\begin{aligned}
B_{\alpha\beta}^{(n)}(x,y) \equiv \frac{1}{2}B_{[\alpha\beta]}^{(n)}(x,y)
&\equiv&
\sum_{m=0}^n
\{\tau_\alpha^{(n-m)}(x),\tau_\beta^{(m)}(y)\}_{(\phi,\pi)}\nonumber\\
%% FOLLOWING LINE CANNOT BE BROKEN BEFORE 80 CHAR
&+&\sum_{m=0}^{n-2}\{\tau_\alpha^{(n-m)}(x),\tau_\beta^{(m+2)}(y)\}_{(\Phi)},~(n\geq2)
\label{eq:2.15}\end{aligned}$$ where the suffix $(\phi,\pi)$ or $(\Phi)$ denotes the variables with respect to which the corresponding P.B. are evaluated. It is now straightforward to explicitly construct the individual terms in the series (\[eq:2.10\]). For example, expanding $\tau_\alpha^{(1)}$ as, $$\tau_\alpha^{(1)} (x)=\int dy X_{\alpha\beta}(x,y)\Phi^\beta(y)\label{eq:2.16}$$ and exploiting (\[eq:2.2\]), (\[eq:2.11\]) and (\[eq:2.13\]) it is found that $X_{\alpha\beta}(x,y)$ is given by, $$\int X_{\alpha\mu}(x,z)\omega^{\mu\nu}(z,z') X_{\beta\nu}(y,z')dz dz'
=
-\Delta_{\alpha\beta}(x,y)\label{eq:2.17}$$ where $\Delta_{\alpha\beta}$ is defined in (\[eq:2.2\]). This determines $\tau_\alpha^{(1)}$ (\[eq:2.16\]).
In order to obtained the complete series (\[eq:2.10\])it is essential to introduce the matrix $\omega_{\alpha\beta}$ and $X^{\alpha\beta}$ which are inverse to $\omega^{\alpha\beta}$ and $X_{\alpha\beta}$ respectively, $$\begin{aligned}
\int \omega^{\alpha\beta}(x,y)\omega_{\beta\gamma}(y,z) dy &=&
\delta^\alpha_\gamma \delta(x-z)\nonumber\\
\int X^{\alpha\beta}(x,y)X_{\beta\gamma}(y,z) dy &=&
\delta^\alpha_\gamma \delta(x-z)\label{eq:2.18}\end{aligned}$$ Then the particular solution of the inhomogeneous equation (\[eq:2.13\]) is given by, $$\tau_\alpha^{(n+1)}(x) = -\frac{1}{n+2}\int \Phi^\mu(z)\omega_{\mu\nu}(z,z')
X^{
\nu\rho}(z',z'')B^{(n)}_{\rho\alpha}(z'',x)dz dz' dz''~~(n\geq1)\label{eq:2.19}$$ The general solution to (\[eq:2.13\]) can be obtained from (\[eq:2.19\]) by adding to it a term containing the solution of the homogeneous equation corresponding to (\[eq:2.13\]). It has been shown [@r8] that any arbitrariness in these solutions corresponds to canonical transformations in the extended phase space. For actual computational purpose, therefore, it suffices to work with the solution (\[eq:2.16\]) and (\[eq:2.19\]).
It may be remarked that this arbitrariness in constructing the solution of the first class constraints is reminiscent of an analogous feature in the conventional ‘splitting’ approach of regarding a second class system as a gauged fixed first class one. The latter, it may be recalled, consists in interpreting one half of the second class constraints (\[eq:2.1\]) as first class ones and the other half as the corresponding gauge fixing conditions. Clearly there is an arbitrariness in this splitting which is also manifested in the ensuing symplectic structure.
The construction of the strongly involutive set of constraints (\[eq:2.10\]) completes the first part of the program. It is now necessary to construct the corresponding involutive Hamiltonian $\tilde H$ which will be defined in the extended phase space, $$\tilde H = \tilde H (\phi,\pi, \Phi) \label{eq:2.20}$$ In the spirit of BT’s approach, strong involution (or ‘abelianisation) is required, $$\left\{\tau_\alpha (x)~,~\tilde H\right\} = 0\label{eq:2.21}$$ subject to the boundary condition, $$\tilde H (\phi, \pi, 0) = H_0 (\phi, \pi)\label{eq:2.22}$$ where $H_0(\phi, \pi)$ is the original Hamiltonian, and $\tau_\alpha$ in (\[eq:2.21\]) are the involutive constraints. There is a degree of freedom in choosing the original Hamiltonian $H_0$. This may just be the usual canonical Hamiltonian. Alternatively, it may be the total Hamiltonian obtained from the canonical part by adding terms proportional to the second class constraints with the Lagrange multipliers ([*i.e.*]{} the proportionality constants) determined by the usual manner of Dirac [@r1]. The particular choice of $H_0$ is crucial for technical reasons resulting in considerable algebraic simplifications, and depends on the specific model being studied.
Expressing the solution to (\[eq:2.21\]) as a power series, $$\tilde H = \sum_{n=0}^\infty \tilde H^{(n)},~~ \tilde H^{(n)} \sim
\Phi^n
\label{eq:2.23}$$ with, $$\tilde H^{(0)} (\phi, \pi; \Phi) = \tilde H (\phi, \pi; 0) = H_0(\phi, \pi)
\label{eq:2.24}$$ Substituting (\[eq:2.23\]) and (\[eq:2.10\]) in (\[eq:2.21\]), the following set of recursion relations are obtained [@r8], $$\{ \tau_\alpha^{(1)}(x) , \tilde H^{(n+1)}\}_{(\Phi)} + G_\alpha^{(n)} (x)
=0,~~
n\geq 0 \label{eq:2.25}$$ where, $$G_\alpha^{(0)}\equiv \{\tau_\alpha^{(0)}, \tilde H^{(0)}\} \nonumber$$ $$G_\alpha^{(1)} \equiv \{\tau_\alpha^{(1)}, \tilde H^{(0)}\} +
\{\tau_\alpha^{(0)}, \tilde H^{(1)}\} +
\{\tau_\alpha^{(2)}, \tilde H^{(1)}\}_{(\Phi)}\nonumber$$ $$G_\alpha^{(n)} \equiv \sum_{m=0}^n\{\tau_\alpha^{(n-m)}, \tilde H^{(m)}\} +
\sum_{m=0}^{n-2}\{\tau_\alpha^{(n-m)}, \tilde H^{(m+2)}\}_{(\Phi)} +
\{\tau_\alpha^{(n+1)}, \tilde H^{(1)}\}_{(\Phi)},~~n\geq 2 \label{eq:2.26}$$ It is now possible to give the solution to the inhomogeneous equation (\[eq:2.25\]), $$\tilde H^{(n+1)} = -\frac{1}{n+1} \int \Phi^\mu(x) \omega_{\mu\nu}(x,y)
X^{\nu
\rho}(y,z) G_\rho^{(n)} (z) dx dy dz,~~(n\geq 0) \label{eq:2.27}$$ which gives the involutive Hamiltonian. This completes the BT [@r8] construction of the first class system which is strongly involutive.
We now analyse the characteristics of this construction and the need for some modifications. Note that the strongly involutive algebra characterises a rank zero theory [@r4]. This is rather restrictive since, in certain cases, a more natural representation could be done in terms of higher rank theories. Although the rank of a theory may not be unique, yet there may be some conceptual or technical problems in - say - expressing a rank one theory as a rank zero type. Let us illuminate by an example. It is usual to regard the Yang-Mills theory as of rank one since the algebra of the first class (Gauss) constraint is only weakly involutive, expressing the standard closure property. Now it may be possible to write down a linear combination of the first class constraints (which would also be first class constraint) so that these are strongly involutive thereby rendering the Yang-Mills theory as rank zero. In that case, however, apart from arcane algebraic structures the natural identity of the first class constraint with the zero component of the equation of motion gets lost and its simple interpretation as the generator of the gauge transformation may also be difficult to preserve. Keeping these points in view, therefore, it is desirable to extend the BT formalism to represent higher rank theories by modifying the involutive algebra (\[eq:2.8\]) as , $$\{\tau_\alpha^a(x),\tau_\beta^b(y)\}=f^{abc}\int
\Gamma_{\alpha\beta\gamma}(x,y,z) \tau_\gamma^c(z) dz\label{eq:2.28}$$ where the structure functions are field independent. This characterises a rank one theory. A rank two theory would occur if the structure functions are field dependent. This happens in the case of gravity which is kept outside the realm of the present paper. The algebra (\[eq:2.28\]) is ideal for discussing nonabelian theories where $a$, $b$, $c$ represent the corresponding group indices. Note that, in the abelian case, (\[eq:2.28\]) goes over to (\[eq:2.8\]).
Following similar techniques as BT [@r8] by expressing $\tau_\alpha^a(x)$ as a power series analogous to (\[eq:2.10\]), $$\tau_\alpha^a = \sum_{n=0}^\infty \tau_\alpha^{a(n)};
{}~~\tau_\alpha^{a(n)}\sim
(\Phi^{\alpha a})^n \label{eq:2.29}$$ we obtain the following relations corresponding to (\[eq:2.12\])–(\[eq:2.15\]) with the appropriate group indices, $$\begin{aligned}
\{ \tau_\alpha^{a(0)} (x), \tau_\beta^{b(0)}(y) \} &+&
\{ \tau_\alpha^{a(1)} (x), \tau_\beta^{b(1)}(y) \}_{(\Phi)}\nonumber\\
&=& f^{abc} \int \Gamma_{\alpha\beta\gamma}(x,y,z) \tau_\gamma^{c(0)}(z) dz
\label{eq:2.30a}\end{aligned}$$ $$\{\tau_{[\alpha}^{a(1)}(x), \tau_{\beta]}^{b(n+1)}(y)\}_{(\Phi)}
+
B_{\alpha\beta}^{ab(n)}(x,y) = 0,~~n\geq 1 \label{eq:2.30b}$$ where, $$B_{\alpha\beta}^{ab(1)}(x,y)=
\{ \tau_{[\alpha}^{a(0)} (x), \tau_{\beta]}^{b(1)}(y) \}
- f^{abc} \int \Gamma_{\alpha\beta\gamma}(x,y,z) \tau_\gamma^{c(1)}(z) dz
\label{eq:2.30c}$$ $$\begin{aligned}
B_{\alpha\beta}^{ab(n)}(x,y)&=&
\sum_{m=0}^n\{ \tau_\alpha^{a(n-m)} (x), \tau_\beta^{b(m)}(y) \} +
\sum_{m=0}^{n-2}\{ \tau_\alpha^{a(n-m)} (x), \tau_\beta^{b(m+2)}(y)
\}_{(\Phi)}\nonumber\\
&-& f^{abc} \int \Gamma_{\alpha\beta\gamma}(x,y,z) \tau_\gamma^{c(n)}(z)
dz,~~
n\geq2
\label{eq:2.30d}\end{aligned}$$ It may be checked that the required solution of the inhomogeneous equation (\[eq:2.30b\]) is given by, $$\tau_\alpha^{a(0)} = \Theta_\alpha^a \nonumber$$ $$\tau_\alpha^{a(1)} (x)=\int dy
X_{\alpha\beta}^{ab}(x,y)\Phi^{b\beta}(y)\label{eq:2.31}$$ with, $$\begin{aligned}
\int X_{\alpha\mu}^{ac}(x,z)\omega^{\mu\nu}_{cd}(z,z') &
&X_{\beta\nu}^{bd}(y,z')dz dz'\nonumber\\
&-& f^{abc}\int \Gamma_{\alpha\beta\gamma}(x,y,z) \Theta_\gamma^c(z)=
-\Delta_{\alpha\beta}^{ab}(x,y)\label{eq2.32}\end{aligned}$$ and, $$\begin{aligned}
\tau_\alpha^{a(n+1)}(x) = -\frac{1}{n+2}\int \Phi^{\mu
b}(z)\omega_{\mu\nu}^{bc}(z,z') X^{
\nu\rho}_{cd}(z'&,&z'')B^{da(n)}_{\rho\alpha}(z'',x)dz dz' dz'' \nonumber\\
& &(n\geq1)\label{eq:2.33}\end{aligned}$$ where $\Delta_{\alpha\beta}^{ab}$, $\omega_{\mu\nu}^{ab}$, $X_{ab}^{\mu\nu}$ are the analogues of (\[eq:2.2\]), (\[eq:2.18\]) with the appropriate group indices. This completes the constructions of the involutive constraints (\[eq:2.29\]) satisfying the algebra (\[eq:2.28\]). A similar analysis can be done for the Hamiltonian which is straightforward and not repeted.
**An abelian example– the Proca model**
=======================================
It is instructive to apply the general ideas discussed in the previous section to specific models. Such model based analysis clarifies several issues and gives a deeper insight into the general formalism. Let us first consider a simple abelian example which is the Proca model whose dynamics is governed by the Lagrangian density, $${\cal L} = -\frac{1}{4} F_{\mu\nu}F^{\mu\nu} + \frac{1}{2} m^2 A^\mu
A_\mu
\label{eq:3.1}$$ Note the introduction of an explicit mass term which breaks the gauge invariance of the usual Maewell’s theory. Consequently (\[eq:3.1\]) represents a second class system which is easily confirmed by the standard constraint analysis. There is one primary constraint, $$\Theta_1 = \pi_0\approx 0 \label{eq:3.2a}$$ and one secondary constraint, $$\Theta_2 = \partial_i\pi^i + m^2 A_0\approx 0 \label{eq:3.2b}$$ obtained by conserving $\Theta_1$ with the total Hamiltonian, $$H_T = H_c + \int d^3x \lambda \Theta_1 \label{eq:3.3}$$ where $H_c$ is the canonical Hamiltonian, $$H_c = \int \left[\frac{1}{2} \pi_i^2 + \frac{1}{4} F_{ij}^2 + \frac{1}{2}
(A_0^2
+ A_i^2) - A_0\Theta_2 \right] \label{eq:3.4}$$ and $\lambda$ is a Lagrange multiplier, while $\pi_\mu$ is the momenta canonically conjugate to $A^\mu$. The algebra, $$\Delta_{\alpha\beta} (x,y) = \{\Theta_\alpha, \Theta_\beta\} =
m^2
\epsilon_{\alpha\beta} \delta(x-y); ~~\alpha, \beta = 1,2 \label{eq:3.5}$$ where the antisymmetric tensor $\epsilon_{\alpha\beta}$ is normalised as, $$\epsilon_{12} = - \epsilon^{12} = -1 \label{eq:3.6}$$ reveals the second class nature of the constraints $\Theta_\alpha (x)$.
In order to convert this system into first class, the first objective is to transform $\Theta_\alpha$ into first class by extending the phase space. Following the general approach we have to introduce the matrix (\[eq:2.5\]), which is chosen as, $$\omega^{\alpha\beta} (x,y) = 2\epsilon^{\alpha\beta}\delta(x-y) \label{eq:3.7}$$ Then the other matrix $X_{\alpha\beta}$ in the game is obtained by solving (\[eq:2.17\]) with $\Delta_{\alpha\beta}$ given by (\[eq:3.5\]), $$X_{\alpha\beta} (x,y) = \left(
\begin{array}{lr}
1&0\nonumber\\ 0& \frac{m^2}{2}\nonumber
\end{array}
\right) \delta(x-y) \label{eq:3.8}$$ There is an arbitrariness in choosing $\omega^{\alpha\beta}$ (\[eq:3.7\]) which would naturally be manifested in (\[eq:3.8\]). This just corresponds to canonical transformations in the extended phase space. However, as has also been evidenced in other calculations [@r18; @r19; @r20], the choice (\[eq:3.7\]) brings about remarkable algebraic simplifications.
Using (\[eq:2.10\]), (\[eq:2.16\]) and (\[eq:3.8\]) the new set of constraints are found to be, $$\begin{aligned}
\tau_1 &=& \Theta_1 + \Phi^1\nonumber\\
\tau_2 &=& \Theta_2 + \frac{m^2}{2} \Phi^2 \label{3.9}\end{aligned}$$ which are strongly involutive, $$\{\tau_\alpha, \tau_\beta\} = 0 \label{3.10}$$ Recall the $\Phi^1$, $\Phi^2$ are the new variables satisfying the algebra (\[eq:2.4\]) with $\omega^{\alpha\beta}$ given by (\[eq:3.7\]). Observe further that only $\tau_\alpha^{(1)}$ (\[eq:2.16\]) contributes in the series (\[eq:2.10\]) defining the first class constraint. All higher order terms given by (\[eq:2.19\]) vanish as a consequence of our intelligent choice (\[eq:3.7\]).
The next step is to obtain the involutive Hamiltonian. The generating functional $G_\alpha^{(n)}$ which determines this Hamiltonian is obtained from (\[eq:2.26\]), $$G_\alpha^{(0)} = \{\Theta_\alpha,H_c\} \nonumber$$ $$G_\alpha^{(n)} = \{\tau_\alpha^{(1)},\tilde H^{(n-1)}\} +
\{\Theta_\alpha,\tilde H^{(n)}\},~~(n\geq 1) \label{3.11}$$ where $\tilde H^{(n)}$ is given in (\[eq:2.27\]) and the original Hamiltonian $\tilde H^{(0)}$ is taken to be the canonical Hamiltonian (\[eq:3.4\]). It is noteworthy that the general expression (\[eq:2.26\]) reduces to the remarkably simple form (\[3.11\]) since there are [*only*]{} two terms $\Theta_\alpha$ and $\tau_\alpha^{(1)}$ in the expansion (\[eq:2.10\]), which is the result of the judicious choice (\[eq:3.7\]). Explicit calculations yield, $$G_1^{(0)} = \Theta_2 \nonumber$$ $$G_2^{(0)} = m^2 \partial_iA^i. \label{3.12}$$ which is substituted in (\[eq:2.27\]) to obtain $\tilde H^{(1)}$, $$\tilde H^{(1)} = \int \left[ (\partial_iA^i) \Phi^1 -
\frac{1}{2}
(\partial_i\pi^i + m^2 A_0) \Phi^2\right] d^3x \label{3.13}$$ This is inserted back in (\[3.11\]) to deduce $G_\alpha^{(1)}$ which then yields $\tilde H^{(2)}$ from (\[eq:2.27\]), $$\tilde H^{(2)} = \int d^3x \left[ -\frac{m^2}{8} (\Phi^2)^2 -
\frac{1}{2m^2}
(\partial_i\Phi^1)(\partial^i\Phi^1)\right] \label{3.14}$$ after which the series terminates. Thus the complete expression for the desired Hamiltonian is, $$\tilde H = H_c + \tilde H^{(1)} + \tilde H^{(2)} \label{3.15}$$ which, by construction, is strongly involutive, $$\{\tilde H, \tau_\alpha\} = 0 \label{3.16}$$ This completes the operatorial (‘abelian’) conversion of the original second class system (with Hamiltonian $H_c$ and constraints $\Theta_\alpha$) into first class (with Hamiltonian $\tilde H$ and constraints $\tau_\alpha$).
It is well known, however, that there exists the Stückelberg [@r9] mechanism whereby the second class Lagrangian (\[eq:3.1\]) can be embedded into the corresponding first class theory by extending the configuration space, $${\cal L}^\prime = -\frac{1}{4} F_{\mu\nu}^2 + \frac{1}{2}m^2 (A_\mu
+
\partial_\mu\theta)(A^\mu + \partial^\mu\theta) \label{3.17}$$ where $\theta$, the new field, is the Stückelberg scalar. The Lagrangian (\[3.17\]) is invariant under the gauge transformations, $$\begin{aligned}
A_\mu &\to& A_\mu - \partial_\mu\alpha\nonumber\\
\theta &\to& \theta + \alpha \label{3.18}\end{aligned}$$ and characterises a first class theory.
We now unravel the correspondence of the Hamiltonian approach with Stückelberg’s formalism. The first step is to identify the new variables $\Phi^1$, $\Phi^2$ (in the Hamiltonian formalism) as a canonically conjugate pair ($\rho$, $\pi_\rho$), $$\begin{aligned}
\Phi^1 &\to& m^2 \rho\nonumber\\
\Phi^2 &\to& \frac{2}{m^2}\pi_\rho \label{3.19}\end{aligned}$$ as may be easily checked from (\[eq:2.4\]) and (\[eq:3.7\]). The phase space partition function is then given by the Faddeev formula [@r4], $$Z= \int {\cal D}(A_\mu \pi^\mu \rho \pi_\rho) \prod_{\alpha,\beta}
\delta(
\tau_\alpha) \delta(\Gamma_\beta) \det\{
\tau_\alpha,\Gamma_\beta\}e^{iS}
\label{3.20a}$$ where, $$S = \int \left( \pi_\mu {\dot A}^\mu + \pi_\rho {\dot \rho} - \tilde H
\right)
\label{3.20b}$$ with $\tilde H$, the involutive Hamiltonian (\[3.15\]), now expressed in terms of $(\rho, \ \pi_\rho)$ instead of $\Phi^1$, $\Phi^2$. The gauge conditions $\Gamma_\beta$ are chosen so that the determinant occurring in the functional measure is nonvanishing. Moreover $\Gamma_\beta$ are assumed to be independent of the momenta so that these may be considered as Faddeev-Popov type gauge conditions. We now perform the momentum integrations to pass on to the configuration space partition function. The $\pi_0$ integral is trivially performed by exploiting the delta function $\delta(\tau_1) = \delta(\pi_0 +
m^2
\rho)$ in (\[3.20a\]). The other delta function $\delta(\tau_2)$ is expressed by its Fourier transform (with Fourier variable $\xi$) and the Gaussian integral over $\pi_\rho$ performed. This yields the action, $$\begin{aligned}
S &=& \int [\pi_i {\dot A}^i - \xi \partial_i\pi^i + \frac{m^2}{2} ({\dot
\rho}^2
+\xi^2 - 2{\dot \rho}\xi) - \frac{1}{2}\pi_i^2 + \frac{1}{4} F_{ij}^2
\nonumber
\\ &+& \frac{1}{2}m^2 A_i^2 +m^2 \partial_iA^i\rho -
\frac{m^2}{2}
\partial_i\rho \partial^i\rho ] \label{3.21}\end{aligned}$$ Note that the $A_0$ term cancelled out. The integral over $A_0$, therefore, reduces to a trivial identity since one of the gauge conditions $\Gamma_\beta$ must involve $A_0$ to have a nonvanishing P.B. with the constraint $\tau_1$. The Gaussian integral over $\pi_i$ is finally performed and the Fourier variable $\xi$ relabelled as ($-A_0$) to express the action in a covariant form, $$S=\int[- \frac{1}{4} F_{\mu\nu}^2
+ \frac{1}{2}m^2 A_\mu^2 -m^2 \partial_\mu A^\mu\rho + \frac{m^2}{2}
\partial_\mu\rho \partial^\mu\rho ] \nonumber$$ $$=\int[- \frac{1}{4} F_{\mu\nu}^2
+ \frac{1}{2}m^2 (A_\mu+\partial_\mu\rho)^2 -m^2 \partial_\mu(A^\mu\rho)]
\label{3.22}$$ Ignoring the last (boundary) term we find that the Lagrangian corresponding to (\[3.22\]) agrees with the usual Stückelberg form (\[3.17\]) by identifying $\rho$ with the Stückelberg field $\theta$. There are, however, some interesting points concerning this correspondence between the Hamiltonian and Lagrangian embeddings which are now elaborated.
The first observation is that Stückelberg Lagrangian (\[3.17\]) (or, equivalently, considering (\[3.22\]) without the boundary term) leads to the primary constraint, $$T_1'= \pi_0\approx 0 \label{3.23}$$ The canonical Hamiltonian obtained by a Legendre transform is, $$H_c' = \int [ \frac{1}{2} \pi_i^2 + \frac{\pi_\rho^2}{2m^2} +
\frac{1}{4}
F_{ij}^2 + \frac{m^2}{2}(A_i + \partial_i\rho)^2 - A_0 T_2'] \label{3.24}$$ where $T_2'$ is the secondary constraint, $$T_2' = \partial_i\pi^i + \pi_\rho \approx 0 \label{3.25}$$ No further constraints are present. It simple to see that $T_1'$ and $T_2'$ form a pair of first class constraints which are in (weak) involution with the total Hamiltonian. The system is first class which is precisley what one expects. The point to be emphasised is, however, that the set of constraints $(T_1', T_2')$ and the Hamiltonian (\[3.24\]) do [*not*]{} agree with the corresponding structures (\[3.9\], \[3.15\], \[3.19\]) obtained in the Hamiltonian approach. In the latter it is necessary to modify the structure of [*all*]{} the original second class constraints in (\[3.9\]) to make them first class. Here, however, only the ‘Gauss’ constraint is modified (\[3.25\]) while the other (\[3.23\]) is not. Indeed it can be checked that in the approach of BT [@r8] a partial modification of (cf. (\[3.23\]), (\[3.25\])) would lead to an algebraic inconsistency. More specifically, the matrix $X_{\alpha\beta}$ (\[eq:2.17\]) cannot be constructed. Consequently it is found that although it is possible to obtain the Stückelberg form (\[3.17\]) starting from the BT construction, the converse is not true.
The next question, naturally, is to find the Lagrangian compatible with the BT construction. Indeed we show that it is just the first expression in (\[3.22\]), $${\cal L}''=- \frac{1}{4} F_{\mu\nu}^2
+\frac{1}{2}m^2 A_\mu^2 -m^2 \partial_\mu
A^\mu\rho
+\frac{m^2}{2}(\partial_\mu\rho)^2
\label{3.26}$$ The primary constraint is, $$T_1'' = \pi_0 + m^2 \rho \approx 0 \label{3.27}$$ The canonical Hamiltonian obtained from (\[3.26\]) is $$\begin{aligned}
H &=& + \frac{1}{2}\pi_i^2+\frac{\pi_\rho^2}{2m^2}
+ \frac{1}{4} F_{ij}^2 - \frac{1}{2}m^2 A_\mu^2\nonumber\\
&+& m^2 \partial_iA^i\rho - \frac{m^2}{2}
\partial_i\rho \partial^i\rho -A_0\partial_i\pi^i \label{3.28}\end{aligned}$$ Time conserving $T_1''$ (\[3.27\]) leads to the secondary constraint, $$T_2'' = \partial_i\pi^i + m^2 A_0 + \pi_\rho \approx 0 \label{3.29}$$ It can be checked that no further constraints are generated by this iterative scheme. Furthermore $T_1''$, $T_2''$ are a pair of first class constraints in involution with the Hamiltonian (\[3.28\]). It is important to note further that $T_1''$, $T_2''$ are exactly identical to the set of first class constraints (\[3.9\], \[3.19\]) obtained in the Hamiltonian formalism. Moreover the first class Hamiltonian (\[3.28\]) differs from the involutive Hamiltonian (\[3.15\]) by a term proportional to the first class constraint $\tau_2$ (\[3.9\]), $$H = \tilde H + \frac{\pi_\rho}{m^2} \tau_2 \label{3.30}$$ Acting on physical states, this difference is trivial since such states are annihilated by the first class constraints. Similarly the equations of motion for observable ([*i.e.*]{} gauge invariant variables) will also be unaffected by this difference since $\tau_2$ can be regarded as the generator of the gauge transformations. In the construction of the functional integral this difference is inconsequential since the constraint $\tau_2$ is strongly implemented by the delta function $\delta(\tau_2)$ (\[3.20a\]). Thus $H$ and $\tilde H$ may be regarded as canonically equivalent. This completes our demonstration of the compatibility of the Lagrangian (\[3.26\]) with the Hamiltonian description of BT.
To summarise, we emphasise the role played by the apparantly innocuous boundary term in (\[3.22\]). If we drop it, the resulting Lagrangian is exactly identical to the Stückelberg form (\[3.17\]) with the BT field identified with the Stückelberg scalar. In that case, however, the Hamiltonian and the set of constraints obtained from this Lagrangian are completely inequivalent to the original BT construction. If, on the contrary, the boundary term is retained it yields the embedded Lagrangian (\[3.26\]). The constraints and Hamiltonian following from this Lagrangian are completely equivalent to the original Hamiltonian embedding. In this case, therefore, the cycle - Hamiltonian to the Lagrangian to the Hamiltonian - closes.
**A nonabelian example – The nonabelian Proca model**
=====================================================
In this section we consider the implication of the generalised canonical approach for a nonabelian model whose dynamics is governed by the Lagrangian density, $${\cal L} = - \frac{1}{4} F^a_{\mu\nu} F^{\mu\nu,a} + \frac{1}{2}m^2
A_\mu^a
A^{\mu,a} \label{4.1}$$ which is just the nonabelian extension of the usual Proca model (\[eq:3.1\]). The corresponding gauge group may be arbitrarily taken whose structure constants will be denoted by $f^{abc}$. There is a primary constraint, $$T_1^a = \pi_0^a \approx 0\label{4.2}$$ which, conserved with the total Hamiltonian, $$H_T = H_c + \int u^a(x)\pi_0^a(x) \label{4.3}$$ where $H_c$ is the canonical Hamiltonian, $$H_c = \int [ \frac{1}{2} (\pi_i^a)^2 + \frac{1}{2}m^2(A_i^a)^2 + \frac{1}{4}
(F_{ij}^a)^2 + \frac{m^2}{2}(A_0^a)^2 - A_0^a T_2^a ] \label{4.4}$$ yields a secondary constraint, $$T_2^a = \partial_i\pi^{ia} - g f^{abc}\pi^{ib}A_i^c +m^2 A_0^a \label{4.5}$$ The Poisson algebra of constraints, $$\{T_1^a(x),T_1^b(y)\} = 0\label{4.6}$$ $$\{T_1^a(x),T_2^b(y)\} = -m^2\delta^{ab}\delta(x-y)\label{4.7}$$ $$\{T_2^a(x),T_2^b(y)\} = g f^{abc}T_2^c\delta(x-y) - gm^2f^{abc}
A_0^c
\delta(x-y)\label{4.8}$$ clearly illustrates that $T_1^a$, $T_2^a$ are a set of second class constraints. Note, particularly the equation (\[4.8\]). This algebra has been given incorrectly by Senjanovic [@r15] where the second term in the right hand side of (\[4.8\]) is missing. Cosequently the corresponding Dirac Brackets (involving $A_0^a$) have been [*incorrectly*]{} evaluated. The correct Dirac brackets may be computed from (\[4.6\]-\[4.8\]) and are found to be, $$\{\pi^{ia}(x), A_0^b(y)\}_{DB} = -\frac{g}{m^2} f^{abc}\pi^{ic}(x)
\delta(x-y)
\nonumber$$ $$\{\pi^{ia}(x), A_j^b(y)\}_{DB} = -\delta^{ab}\delta^i_j \delta(x-y)\nonumber$$ $$\{\pi^{ia}(x), \pi_j^b(y)\}_{DB} = 0\nonumber$$ $$\{A^{ia}(x), A_j^b(y)\}_{DB} = 0\nonumber$$ $$\{A^a_0(x), A_j^b(y)\}_{DB} = \frac{1}{m^2}\partial_j^x\delta^{ab} \delta
(x-y)
-\frac{g}{m^2}f^{abc} A_j^c(x)\delta (x-y)\nonumber$$ $$\{A_0^a(x) , A_0^b(y)\}_{DB} = -\frac{g}{m^2}f^{abc} A_0^c(x)\delta
(x-y).
\label{4.n}$$
There is one important difference in the Dirac brackets of the nonabelian theory (\[4.1\]) contrasted with the corresponding abelian version (\[eq:3.1\]). In the former case these brackets are [*field dependent*]{}. Consequently a transition to quantum theory where these DBs are replaced by commutators will be problematic [@r1]. this was also mentioned earlier in our introduction. In this case, therefore, the generalised canonical formalism provides a viable alternative to quantisation. By converting the theory into first class, the need for Dirac brackets is eliminated and quantisation can proceed by using the canonical Poisson brackets.
As analysed in the previous section, the initial step is to obtain the first class constraints from the second class ones, (\[4.2\]) and (\[4.5\]). In order to construct the first class constraints $\tau_\alpha^a$ satisfying the involutive algebra (\[eq:2.28\]) we have to specify the matrices $\omega^{
\alpha\beta}_{ab}$, $X_{\alpha\beta}^{ab}$ which are the nonabelian analogues of (\[eq:3.7\]) and (\[eq:3.8\])respectively. We make the following choice, $$\omega^{\alpha\beta}_{ab}(x,y)=
m^2\epsilon^{\alpha\beta}\delta^{ab}\delta(x-y)
\label{4.10}$$ $$X_{\alpha\beta}^{ab}(x,y) = \left(
\begin{array}{lr}
2\delta^{ab} & 0\nonumber \\ gf^{abc}A_0^c & \frac{1}{2}\delta^{ab}\nonumber
\end{array} \right)\delta(x-y)\label{4.11}$$ The corresponding inverse matrices are, $$%% FOLLOWING LINE CANNOT BE BROKEN BEFORE 80 CHAR
\omega_{\alpha\beta}^{ab}(x,y)=\frac{1}{m^2}\epsilon_{\alpha\beta}\delta_{ab}\delta(x-y)
\label{4.12}$$ $$X^{\alpha\beta}_{ab}(x,y) = \left(
\begin{array}{lr}
\frac{1}{2}\delta^{ab} & 0\nonumber \\ -gf^{abc}A_0^c & 2\delta^{ab}\nonumber
\end{array} \right)\delta(x-y)\label{4.13}$$ We further take the following form for $\Gamma_{\alpha\beta\gamma}(x,y,z)$ appearing in (\[eq:2.28\]), $$\Gamma_{\alpha\beta\gamma}(x,y,z) = g \delta_{\alpha 2}\delta_{\beta 2}
\delta_{
\gamma 2}\delta(x-y)\delta(y-z)\label{4.14}$$ so that (\[eq:2.28\]) with (\[4.14\]) yields the conventional involutive algebra valid for the first class Yang-Mills theory.
With the choice (\[4.10\]), (\[4.11\]), (\[4.14\]) one can compute the generating functions $B^{ab}_{\alpha\beta}$ (\[eq:2.30a\]-\[eq:2.30d\]). From this knowledge the various terms (\[eq:2.33\]) in the power series expansion (\[eq:2.29\]) may be obtained. We give below the final expressions for the first class constraints $\tau^a_1$, $\tau_2^a$ satisfying (\[eq:2.28\]), (\[4.14\]) valid upto the second power in the additional fields $\Phi^{\alpha ,a}$: $$\tau_1^a = T_1^a + 2 \Phi^{1,a} \label{4.15}$$ $$\begin{aligned}
\tau_2^a = T_2^a &+& \frac{1}{2} \Phi^{2,a} + g f^{abc}
\Phi^{1,b}A_0^c -
\frac{g}{2m^2} f^{abc}\Phi^{2,b}\Phi^{1,c}\nonumber\\
&+& \frac{g^2}{3m^2} f^{adb}f^{bce}A_0^e\Phi^{1,c}\Phi^{1,d} +
O(\Phi\Phi\Phi)
\label{4.16}\end{aligned}$$ where $T_1^a$, $T_2^a$ are the original second class constraints. Note further that $\tau_1^a$ is an [*exact*]{} result while $\tau_2^a$ gets modified in the different powers of the new fields $\Phi$.
The next step is to compute the involutive Hamiltonian. To complete the analogy with the usual Yang-Mills system we demand the following algebra to be satisfied by the involutive Hamiltonian $\tilde H$ and the constraints, $$\begin{aligned}
\{\tau_1^a (x), \tilde H\} &=& \tau_2^a(x)\nonumber\\
\{\tau_2^a(x), \tilde H\} &=& gf^{abc}A_0^b\tau_2^c\label{4.17}\end{aligned}$$ As usual $\tilde H$ is expressed as a power series (\[eq:2.23\]), with the original Hamiltonian $H_0$ taken to be the canonical piece. For algebraic considerations we mention that the old (second class) constraints and the original (canonical) Hamiltonian have identical brackets as (\[4.17\]), [*i.e.*]{} $$\begin{aligned}
\{T_1^a (x), H_c\} &=& T_2^a(x)\nonumber\\
\{T_2^a(x), H_c\} &=& gf^{abc}A_0^b(x)T_2^c(x)\label{4.18}\end{aligned}$$
We now wish to compute the generating function $G_\alpha^{(n)}$ (\[eq:2.26\]) which will yield the desired involutive Hamiltonian. Recall that the structure given in (\[eq:2.26\]) leads to a strongly involutive Hamiltonian satisfying (\[eq:2.21\]). Since our algebra (\[4.17\]) is only weakly involutive, suitable modifications must be made in the analysis given in section 2. Let us first express (\[4.17\]) in a covariant notation, $$\{ \tau_\alpha^a(x), \tilde H\} = \int dy
V^{ab}_{\alpha\beta}(x,y)\tau_\beta^b(y)
\label{4.19}$$ so that, $$\begin{aligned}
V_{12}^{ab}(x,y) &=& \delta^{ab}\delta(x-y)\nonumber\\
V_{22}^{ab} (x,y) &=& -gf^{abc}A_0^c\delta(x-y) \label{4.20}\end{aligned}$$ with all other $V$-coefficients being zero.
Proceeding as was done in sec 2, the modified expressions for the generating functions (\[eq:2.26\]) are found to be, $$G_\alpha^{a(0)}\equiv \{\tau_\alpha^{a(0)}, \tilde H^{(0)}\}
-\int dy V_{\alpha\beta}^{ab}(x,y) \tau_\beta^{b(0)}(y)\nonumber$$ $$\begin{aligned}
G_\alpha^{a(1)} \equiv \{\tau_\alpha^{a(1)}, \tilde H^{(0)}\} &+&
\{\tau_\alpha^{a(0)}, \tilde H^{(1)}\} +
\{\tau_\alpha^{a(2)}, \tilde H^{(1)}\}_{(\Phi)}\nonumber\\
&-&\int dy V_{\alpha\beta}^{ab}(x,y) \tau_\beta^{b(1)}(y)\nonumber
\end{aligned}$$ $$\begin{aligned}
G_\alpha^{a(n)} \equiv \sum_{m=0}^n\{\tau_\alpha^{a(n-m)}, \tilde H^{(m)}\} &+&
\sum_{m=0}^{n-2}\{\tau_\alpha^{a(n-m)}, \tilde H^{(m+2)}\}_{(\Phi)} +
\{\tau_\alpha^{a(n+1)}, \tilde H^{(1)}\}_{(\Phi)}\nonumber\\
&-&\int dy V_{\alpha\beta}^{ab}(x,y) \tau_\beta^{b(n)}(y), ~~n\geq 2
\label{4.21}
\end{aligned}$$ The final solution for $\tilde H$ satisfying (\[4.19\]) is given by the power series (\[eq:2.23\]) with, $$\tilde H^{(n+1)} = -\frac{1}{n+1} \int
\Phi^{\mu,a}\omega^{ab}_{\mu\nu}
X^{\nu\rho}_{bc}G^{c(n)}_{\rho},~~(n\geq 0) \label{4.22}$$ where $\omega^{ab}_{\mu\nu}$, $X^{\nu\rho}_{bc}$ are defined in (\[4.12\]), (\[4.13\]). The result for the involutive Hamiltonian now follows by a straightforward algebra, $$\tilde H = H_c -\frac{1}{m^2}\int \Phi^{1a}\Phi^{2a} - \frac{2g}{m^2}
f^{abc}
f^{cge} \int A_0^b A_0^g\Phi^{1a}\Phi^{1e} + O(\Phi\Phi\Phi) \label{4.23}$$ correct to second powers in the new fields.
It is now possible to establish a connection of our analysis with the Lagrangian formalism [@r16] of converting (\[4.1\]) into first class. In [@r16] the gauge group has been chosen as $SU(2)$ for algebraic simplification. The configuration space is extended by introducing three additional fields which are just the Eulerian angles [@r16]. This implies the addition of six new fields in the phase space (the fields of the configuration space and their canonically conjugate momenta). The new fields that we have introduced are precisely six in number for the group $SU(2)$. Three are given by $\Phi^{1,a}$ and three by $\Phi^{2,a}$ ($a$=1, 2, 3). The involutive algebra in the Lagrangian formulation [@r16] is exactly identical to (\[eq:2.28\]) with $\Gamma_{\alpha\beta\gamma}$ specified by (\[4.14\]), and (\[4.17\]). Conceptually, therefore, the first class system found here by the Hamiltonian formalism corresponds to the Lagrangian scheme [@r16]. To make an exact correspondence which will equate some combination of the $\Phi^{1,a}$, $\Phi^{2,a}$ fields with the Eulerian fields (and their conjugates) is more of a technical problem. We will have to isolate the terms in the power series (\[4.16\]), (\[4.23\]) which will yield the corresponding terms in the trigonometric (sine, tan etc) expansions given in [@r16].
We will close this section by pointing out certain erroneous observations made in the literature [@r17] regarding the conversion of (\[4.1\]) to a first class system by adopting the usual Stückelberg [@r9] mechanism. Just as the abelian (Proca) model (\[eq:3.1\]) could be converted into first class by making a ‘gauge transformation’ (see ), $$A_\mu \to A_\mu + \partial_\mu \theta \label{4.24}$$ it has been claimed [@r17] that a ‘nonabelian gauge transformation’ $$A_\mu^a \to A_\mu^a + (D_\mu\theta)^a \label{4.25}$$ is able to transform (\[4.1\]) into a first class system. This is wrong as we promptly demonstrate.
The Lagrangian (\[4.1\]) modified by (\[4.25\]) is, $${\cal L} = -\frac{1}{4} (F_{\mu\nu}^a)^2 + \frac{m^2}{2}(A_\mu^a+ (D_\mu\theta
)^a)^2 \label{4.26}$$ There is one primary constraint, $$T_1^a= \pi_0^a\approx 0\label{4.27}$$ and one secondary constraint, $$T_2^a = \partial_i\pi_i^a + gf^{abc}\pi_i^bA_i^c -m\pi_\phi^a
-gf^{abc}\pi_\phi^b\phi^c
\label{4.28}$$ obtained by time conserving (\[4.27\]) with the total Hamiltonian, $$H_T = H_c + \int u^a \pi_0^a \label{4.29}$$ where $H_c$ is the canonical Hamiltonian, $$\begin{aligned}
H_c &=&\frac{1}{2}(\pi_i^a)^2 + \frac{1}{2}(A_i^a)^2 +\frac{1}{4}(F_{ij}^a)^2 +
\frac{1}{2}(\partial_i\phi^a)^2
\frac{1}{2}(\pi_\phi^a)^2+mA_i^a\partial_i\phi^a \nonumber\\
&-&mf^{abc}A_i^b\phi^c\partial_i\phi^a + \frac{1}{2}g^2
f^{abc}f^{ade}A_i^bA_i^d\phi^c\phi^e + A_0^aT_2^a
\label{4.30}\end{aligned}$$ The algebra of constraints, $$\{T_1^a(x), T_1^b(y)\} = \{T_1^a(x), T_2^b(y)\} = 0\label{4.31a}$$ $$\{T_2^a(x), T_2^b(y)\} = g f^{abc}T_2^c\delta(x-y) -
gm^2f^{abc}\pi_\theta^c
\delta(x-y) \label{4.31}$$ clearly shows that $T_2^a$ is a second class constraint. Hence the model (\[4.26\]) is not first class. In fact, it is an example of a mixed class system because $T_1^a$ is first class. The relation in (\[4.31\]) was not computed in [@r17] which led its authors to incorrectly observe that (\[4.26\]) is a first class theory.
**Conclusion**
==============
In the preceeding sections we have made a detailed investigation of the Batalin- Tyutin [@r8]method of converting second class systems into first class. Since the method has been developed quite recently it may not be familiar even among particle physicists. We have, therefore, reviewed this analysis and also shown the possibility of extending it in other directions. In particular, the method can be generalised to yield weakly involutive systems originating from second class systems. These weakly involutive systems occur naturally in nonabelian gauge theories and in gravity. The original paper [@r8] is more suited for discussing abelian models.
The systematic application of this approach to specific models– both abelian and nonabelian– has revealed several interesting features and helps in giving new insights than would be possible by merely discussing the general formalism. The Proca model, which is an example of an abelian second class system, is systematically converted into first class. Interestingly, we find that the choice of the matrices $X_{ij}$ (\[eq:3.7\]), $\omega^{ij}$ (\[eq:3.8\]), which was earlier used by us [@r18; @r19] in other contexts, considerably simplifies the algebra. A direct connection with the usual Lagrangian embedding of Stückelberg [@r9] can be made by explicitly evaluating the momentum integrals in the phase space partition function using Faddeev-Popov [@r14]-like gauges. An exact identification of the extra field introduced in our Hamiltonian formalism with the conventional Stückelberg scalar [@r9] is possible provided one is careful with boundary terms (\[3.22\]). In this connection our analysis may be compared with [@r12], which is not only unsystematic (so may not conform to the strict mathematical rigour of quantising first class systems [@r3; @r4; @r5]) but whose correspondence with the Stückelberg mechanism is rather circuitous. We find that the present approach is more elegant and conceptually clean.
Coming next to the nonabelian extension of the Proca model we have first corrected[^2] the familiar Dirac analysis given in [@r15]. Using the generalised canonical approach, this model has been coverted into a first class system. Contrary to the abelian example, closed expressions for the Gauss constraint and the involutive Hamiltonian cannot be found. Conceptually it is possible to interpret the extra fields as the analogue of the Euler angles (regarded as field variables) introduced in the nonabelian ($SU(2)$) Stückelberg formalism [@r16]. It is also possible to obtain first class structure (of constraints, Hamiltonian etc) within the Hamiltonian formulation which are different from the conventional nonabelian Stückelberg mechanism [@r16]. Our analysis clearly reveals that, contrary to existing claim[^3] [@r17], the ususal (abelian) Stückelberg formalism [@r9] cannot transform the nonabelian Proca model into first class.
We feel that our nonabelian exercise may provide fresh insights in the fields of quantum gravity. The point is that it is possible, by suitably altering the generating fuctions, to express the constraint algebra as rank zero, rank one or rank two. It may be recalled that a strongly involutive algebra has rank zero, an algebra which closes with structure constants has rank one; and an algebra which closes with structure functions of the phase space variables has rank two. In the case of quantum gravity Ashtekar [@r20] and Witten [@r21] have given algebras which fall in these clases. It would be tempting to discuss these aspects within the generalised canonical framework.
[99]{} P. A. M. Dirac, “Lectures on quantum mechanics" (Belfer Graduate School, Yeshiva University Press, New York 1964). L. D. Faddeev, Theor. Math. Phys. [**1**]{} (1970), 1. E. S. Fradkin and G. A. Vilkovisky, Phys. Lett. [**B55**]{} (1975) 224; CERN Report TH-2332 (1977) (unpublished). M. Henneaux, Phys. Rep. [**C126**]{} (1985), 1, and “Classical foundations of BRST symmetry" (Bibliopolis, Naples, 1988). C. Becchi, A. Rouet and R. Stora, Ann. Phys. \[N.Y.\][**98**]{} (1976), 287\
I. V. Tyutin, Lebedev Preprint 39 (1975). K. Kuchar, Phys. Rev. [**D34**]{} (1986), 3044. I. A. Batalin and E. S. Fradkin, Nucl. Phys. [**B279**]{} (1987), 514; Phys. Lett. [**B180**]{} (1986), 157. I. A. Batalin and I. V. Tyutin, Int. J. Mod. Phys. [**A6**]{} (1991), 3255. E. C. G. Stückelberg, Helv. Phys. Act. [**30**]{} (1957), 209. L. D. Faddeev and shatashvili, Phys. Lett. [**B167**]{} (1986), 225. J. Wess and B. Zumino, Phys. Lett. [**B37**]{} (1971), 95. T. Fujiwara, Y. Igarashi and J. Kubo, Nucl. Phys. [**B341**]{} (1990), 695. Y. Kim, S. Kim, W. Kim, Y. Park, K. Kim and Y. Kim, Phys. Rev. [**D46**]{} (1992), 4574\
R. Banerjee, H. J. Rothe and K. D. Rothe, Phys. Rev. [**D48**]{} (1994) R5467, (in Press) L. D. Faddeev and V. N. Popov, Phys. Lett. [**B25**]{} (1967), 29. P. Senjanovic, Ann. Phys. \[N. Y.\] [**100**]{} (1976), 227. T. Kunimasa and T. Goto, Prog. Theo. Phys. [**37**]{} (1967), 452. C. Grosse-Knetter, Phys. Rev. [**D48**]{} (1993), 2854. R. Banerjee, Phys. Rev. [**D 48**]{} (1993), R5467. N. Banerjee, S. Ghosh and R. Banerjee, Nucl. Phys. [**B**]{} (1994), (in Press). N. Banerjee, S. Ghosh and R. Banerjee, Phys. Rev. [**D 49**]{} (1994), 1996. A. Ashtekar, Phys. Rev. [**D36**]{} (1987), 1587. E. Witten, Nucl. Phys. [**B311**]{} (1988), 46.
[^1]: e-mail address: [email protected]
[^2]: Equation (\[4.8\]) given in [@r15] is wrong leading to an incorrect evaluation of the Dirac brackets involving $A_0^a$. For the correct forms, see (\[4.n\])
[^3]: The algebra for $G^a$ immediately exposes the second class nature since it does not close, $\{G^a,G^b\}
=
\epsilon^{abc}G^c + \cdots$.
| {
"pile_set_name": "ArXiv"
} |
Introduction
============
Consider a linear chain of $N+1$ interacting fermions described by the Hamiltonian $$\begin{gathered}
\hat H= \sum_{k=0}^{N-1} J_k\big(a_k^\dagger a_{k+1}+a_{k+1}^\dagger a_k\big) .
\label{Ham2}\end{gathered}$$ The lattice fermions $\{a_{k}, a_{k}^{\dagger}\, |\, k = 0, 1, \ldots, N\}$ obey the common anticommutation relations, and $J_k$ expresses the coupling strength between sites $k$ and $k+1$. The Hamiltonian describes a system of $N+1$ fermions on a chain with nearest-neighbour interaction (hopping between adjacent sites of the chain) subject to a zero background magnetic field.
Hamiltonians of the type appear in various contexts. In particular, spin chains of this type are popular as channels for short distance quantum communication, and were introduced by Bose [@Bose2003; @Bose2005; @Bose2007]. The system then originates from a linear qubit chain with nearest neighbour interaction described by a Heisenberg $XY$ Hamiltonian, and is being mapped into by a Jordan–Wigner transformation [@Lieb1968; @Jordan1928]. In such models, the communication is achieved by state dynamical evolution in the spin chain, which does not require any on/off switches of the interactions between the spins, nor any modulation of external fields. Many articles dealing with such spin chains in the context of Bose’s scheme focus on perfect transmission (or perfect state transfer) in these chains [@Christandl2004; @Albanese2004; @Christandl2005; @Yung2005; @Karbach2005; @Kay2010].
By far the most elegant and simplest scheme to realize perfect state transfer (over an arbitrary long chain) was proposed by Christandl et al. [@Christandl2004; @Albanese2004; @Christandl2005]. Their choice of the modulation of the coupling strengths is given by: $$\begin{gathered}
J_k= J_{N-k-1}=\sqrt{(k+1)(N-k)}, \qquad k=0,1,\ldots,N-1.
\label{J}\end{gathered}$$ The simplicity of Christandl’s model follows from the following observation. Consider first the single-fermion states of the system: in a single-fermion basis, the Hamiltonian $\hat H$ takes the matrix form $$\begin{gathered}
M=\left(
\begin{array}{ccccc}
0 & J_0& 0 & \cdots & 0 \\
J_0 & 0 & J_1 & \cdots & 0\\
0 & J_1 & 0 & \ddots & \\
\vdots & \vdots & \ddots & \ddots& J_{N-1}\\
0 & 0 & & J_{N-1} & 0
\end{array}
\right).
\label{Ham-M}\end{gathered}$$ The dynamics (time evolution) of the system is completely determined by the eigenvalues $\epsilon_j$ and eigenvectors $\varphi_j$ of this interaction matrix. It is indeed a standard technique [@Lieb1968; @Albanese2004] to describe the $n$-fermion eigenstates of $\hat H$ ($n\leq N$) using the single-fermion eigenstates $\varphi_j$ and Slater determinants. In Christandl’s case, determined by , the eigenvalues and eigenvectors of $M$ are explicitly known. In particular, the eigenvalues are given by $\epsilon_j=-N+2j$ ($j=0,1,\ldots,N$) and the eigenvectors are given in terms of Krawtchouk polynomials.
In the model of Christandl there is, for arbitrary $N$, an analytic (closed form) expression for the eigenvalues and eigenvectors of $M$. Such spin chains are [*analytically solvable*]{} [@Chakrabarti2010; @Jafarov2010]. It implies, in particular, that the correlation function at time $t$, $$\begin{gathered}
f_{r,s}(t) = (r| \exp(-it\hat H) |s),
$$ where $r$ and $s$ are site labels belonging to $\{0,1,\ldots,N\}$ and $|r)$, $|s)$ denote the corresponding single spin states at the ‘receiver’ and ‘sender’ sites $r$ and $s$, can be computed explicitly [@Chakrabarti2010; @Jafarov2010].
Christandl’s spin chain model allows [*perfect state transfer*]{}, essential for using the spin chain as a transmission channel. Perfect state transfer at time $t=T$ from one end of the chain to the other end is expressed by $|f_{N,0}(T)|=1$. The topic of perfect state transfer in spin chains has received a lot of attention [@Kay2010]. Fairly easy sufficient conditions have been formulated in order to achieve perfect state transfer (such as mirror symmetry [@Albanese2004; @Kay2010]). Shi et al. showed that the “spectrum parity matching condition” is necessary and sufficient for perfect state transfer [@Shi2005]. According to this condition, they found a one-parameter extension of Christandl’s model in the case there is an even number of fermion sites in the chain; in our notation this means that $N$ is odd, i.e. $$\begin{gathered}
N=2m+1, \qquad m\in{\mathbb Z}_+.
$$ The couplings in Shi’s model [@Shi2005] are determined by ($k=0,1,\ldots,N-1$) $$\begin{gathered}
J_k= \begin{cases}
\sqrt{(k+1)(N-k)}, & \text{if $k$ is odd,}\\
\sqrt{(k+2\alpha+2)(N-k+2\alpha+1)}, & \text{if $k$ is even.}
\end{cases}
\label{Ja}\end{gathered}$$ Herein, $\alpha$ is a real parameter satisfying $\alpha>-1$ (the case of Shi actually corresponds to half-integer $\alpha$, but for our purposes $\alpha$ can be any real number greater than $-1$). Note that for $\alpha=-\frac12$, Shi’s model reduces to Christandl’s model (i.e. reduces to ), at least when $N=2m+1$ is odd. In Shi et al. [@Shi2005], the spectrum of the single fermion states (i.e. the eigenvalues of $M$ with data determined by ) was found; however no closed form expressions for the eigenvectors were obtained.
In the present paper, we will show that the eigenvectors can be expressed in terms of Hahn polynomials. In fact, we will first work with a two-parameter extension of Christandl’s model: $$\begin{gathered}
J_k= \begin{cases}
\sqrt{(k+1)(N-k)}, & \text{if $k$ is odd,}\\
\sqrt{(k+2\alpha+2)(N-k+2\beta-1)}, & \text{if $k$ is even.}
\end{cases}
\label{Jab}\end{gathered}$$ Now $\alpha$ and $\beta$ are real parameters satisfying $\alpha>-1$ and $\beta>0$. The case of Shi corresponds to $\beta=\alpha+1$, and the case of Christandl to $\alpha=-\frac12$, $\beta=\frac12$. For the general case , we obtain in this paper an explicit form of the eigenvalues, and an explicit form of the eigenvectors. The components of the eigenvectors are given by means of Hahn polynomials $Q_n(x;\alpha,\beta,m)$ [@Koekoek; @Suslov]: the even components are proportional to $Q_n(x;\alpha,\beta,m)$ and the odd components to $Q_n(x; \alpha+1,\beta-1,m)$. In order to prove our assertions about eigenvalues and eigenvectors, we need some (new) difference equations for Hahn polynomials. Section \[Hahn\] of this paper is devoted to introducing the common notation for Hahn polynomials and to proving the new difference equations. In Section \[eigen\] we obtain the main result of this paper: the explicit construction of the spectrum of $M$ for the values and the construction of its eigenvectors in terms of the Hahn polynomials. Section \[correlation\] returns to the model governed by the spin chain data . Since the spin chain is analytically solvable, we can compute the correlation function explicitly, and determine under which conditions perfect state transfer is possible. Finally, in Section \[qHahn\] we present the $q$-generalization of the results obtained (in terms of $q$-Hahn polynomials).
Although our paper is strongly inspired by the model introduced by Shi et al., it should be emphasized that our results are dealing mainly with mathematical aspects of this model. In [@Shi2005], the emphasis was on quantum state transfer. Our main result is to show that the eigenstates of the Hamiltonian in the case of Shi, , or in the extended case, , can be computed in closed form, with coefficients given as Hahn polynomial evaluations.
It should be mentioned that some completely different spin chain models related to Hahn polynomials have been considered before. The second solution of [@Albanese2004] is actually related to an interaction matrix corresponding to the Jacobi matrix of dual Hahn polynomials. In [@Chakrabarti2010], the interaction matrix corresponding to the Jacobi matrix of Hahn polynomials was studied, following some ideas of [@Regniers2009]. In that case, the matrix of eigenvectors $U$ is directly a matrix of Hahn polynomial evaluations. However, due to the complicated coefficients in the three term recurrence relations, the actual coefficients in the interaction matrix become quite involved, see e.g. [@Chakrabarti2010 Lemma 2]. In the present paper, the main innovation comes from “doubling” the matrix $U$ (hence the technique works for chains with an even length only), in a way that it contains Hahn polynomial evaluations of two different types (one with parameters $(\alpha,\beta)$ and one with $(\alpha+1,\beta-1)$), such that the interaction matrix coefficients are very simple.
Hahn polynomials and new difference equations {#Hahn}
=============================================
The Hahn polynomial $Q_n(x;\alpha, \beta, m)$ [@Koekoek; @Suslov] of degree $n$ ($n=0,1,\ldots,m$) in the variable $x$, with parameters $\alpha>-1$ and $\beta>-1$, or $\alpha<-m$ and $\beta<-m$ is defined by [@Koekoek; @Suslov]: $$\begin{gathered}
Q_n(x;\alpha,\beta,m) = {\;}_3F_2 \left( {\genfrac{}{}{0pt}{}{-n,n+\alpha+\beta+1,-x}{\alpha+1,-m}} ; 1 \right).
\label{defQ}\end{gathered}$$ Herein, the function $_3F_2$ is the generalized hypergeometric series [@Bailey; @Slater]: $$\begin{gathered}
{}_3F_2 \left( {\genfrac{}{}{0pt}{}{a,b,c}{d,e}} ; z \right)=\sum_{k=0}^\infty \frac{(a)_k(b)_k(c)_k}{(d)_k(e)_k}\frac{z^k}{k!}.
\label{defF}\end{gathered}$$ In (\[defQ\]), the series is terminating because of the appearance of the negative integer $-n$ as a numerator parameter. Note that in (\[defF\]) we use the common notation for Pochhammer symbols [@Bailey; @Slater] $(a)_k=a(a+1)\cdots(a+k-1)$ for $k=1,2,\ldots$ and $(a)_0=1$. Hahn polynomials satisfy a (discrete) orthogonality relation [@Koekoek]: $$\begin{gathered}
\sum_{x=0}^m w(x;\alpha, \beta,m) Q_l(x;\alpha, \beta, m) Q_n(x;\alpha,\beta,m) = h_n(\alpha,\beta,m) \delta_{ln},
\label{orth-Q}\end{gathered}$$ where $$\begin{gathered}
w(x;\alpha, \beta,m) = \binom{\alpha+x}{x} \binom{m+\beta-x}{m-x}, \qquad x=0,1,\ldots,m, \\
h_n (\alpha,\beta,m)= \frac{(-1)^n(n+\alpha+\beta+1)_{m+1}(\beta+1)_n n!}{(2n+\alpha+\beta+1)(\alpha+1)_n(-m)_n m!}.\end{gathered}$$ Denote the orthonormal Hahn functions as follows: $$\begin{gathered}
\tilde Q_n(x;\alpha,\beta,m) \equiv \frac{\sqrt{w(x;\alpha,\beta,m)}\, Q_n(x;\alpha,\beta,m)}{\sqrt{h_n(\alpha,\beta,m)}}.
\label{Q-tilde}\end{gathered}$$
For our construction, the essential ingredient is a set of new difference equations for Hahn polynomials. These relations involve Hahn polynomials of the same degree in variables $x$ or $x+1$, and with parameters $(\alpha,\beta)$ and $(\alpha+1,\beta-1)$; in this sense it could also be appropriate to speak of “contiguous relations” rather than difference equations.
\[proposition1\] The Hahn polynomials satisfy the following difference equations: $$\begin{gathered}
(m+\beta-x) Q_n(x;\alpha ,\beta ,m)-(m-x) Q_{n}(x+1;\alpha , \beta,m)\nonumber \\
\qquad{} =\frac{(n+\alpha +1)(n+\beta)}{\alpha+1} Q_n(x;\alpha+1,\beta-1,m), \label{Q-rec1} \\
(x+1) Q_n(x;\alpha +1,\beta -1,m)-(\alpha +x+2) Q_{n}(x+1;\alpha +1 , \beta -1,m) \nonumber \\
\qquad {}=-(\alpha+1) Q_n(x+1;\alpha,\beta,m). \label{Q-rec2}\end{gathered}$$
Both equations follow from a simple computation using the hypergeometric series expression. In the case of , the left hand side is expanded as follows: $$\begin{gathered}
(m+\beta-x) Q_n(x;\alpha ,\beta ,m)-(m-x) Q_{n}(x+1;\alpha , \beta,m) \nonumber\\
\qquad{} = \sum_{k=0}^n \frac{(-n)_k(\alpha+\beta+n+1)_k(-x)_{k-1}}{k!(\alpha+1)_k(-m)_k}
[(\beta+m-x)(k-x-1)-(m-x)(-x-1)]. \nonumber\end{gathered}$$ Rewriting the expression in square brackets as $[k(m-k+1)+(\beta+k)(k-x-1)]$, the above sum splits in two parts: $$\begin{gathered}
-\sum_{k=1}^n \frac{(-n)_k(\alpha+\beta+n+1)_k(-x)_{k-1}}{(k-1)!(\alpha+1)_k(-m)_{k-1}} +
\sum_{k=0}^n \frac{(-n)_k(\alpha+\beta+n+1)_k(-x)_{k}}{k!(\alpha+1)_k(-m)_k} (\beta+k).
\label{2parts}\end{gathered}$$ The first part can be brought in the following form: $$\begin{gathered}
-\sum_{k=1}^n \frac{(-n)_k(\alpha+\beta+n+1)_k(-x)_{k-1}}{(k-1)!(\alpha+1)_k(-m)_{k-1}}
= -\sum_{j=0}^{n-1} \frac{(-n)_{j+1}(\alpha+\beta+n+1)_{j+1}(-x)_{j}}{j!(\alpha+1)_{j+1}(-m)_{j}} \\
\qquad{} = \sum_{k=0}^n \frac{(-n)_k(\alpha+\beta+n+1)_k(-x)_k}{k!(\alpha+1)_k (-m)_k} \frac{(n-k)(\alpha+\beta+n+k+1)}{(\alpha+k+1)}.\end{gathered}$$ So becomes $$\begin{gathered}
\sum_{k=0}^n \frac{(-n)_k(\alpha+\beta+n+1)_k(-x)_k}{k!(\alpha+1)_k (-m)_k}
\Bigl[ \frac{(n-k)(\alpha+\beta+n+k+1)}{(\alpha+k+1)} + (\beta+k) \Bigr] \nonumber\\
\qquad{} =\sum_{k=0}^n \frac{(-n)_k(\alpha+\beta+n+1)_k(-x)_k}{k!(\alpha+1)_k (-m)_k}
\Bigl[ \frac{(n+\alpha+1)(n+\beta)}{(\alpha+k+1)} \Bigr] \nonumber\\
\qquad{} = \frac{(n+\alpha+1)(n+\beta)}{(\alpha+1)} \sum_{k=0}^n \frac{(-n)_k(\alpha+\beta+n+1)_k(-x)_k}{k!(\alpha+2)_k (-m)_k} \nonumber\end{gathered}$$ leading to the right hand side of .
The second equation is even simpler to prove. The left hand side of can be written as $$\begin{gathered}
(x+1)\sum_{k=0}^n A_k \frac{(-x+k-1)}{(\alpha+k+1)} - (\alpha+x+2) \sum_{k=0}^n A_k \frac{(-x-1)}{(\alpha+k+1)},
\label{prf2}\end{gathered}$$ where $$A_k= \frac{(-n)_k(\alpha+\beta+n+1)_k (-x)_{k-1}}{k!(\alpha+2)_{k-1} (-m)_k}.$$ A simple addition of the two terms in yields $$- \sum_{k=0}^n A_k (-x-1) = -(\alpha+1) \sum_{k=0}^{n} \frac{(-n)_k(\alpha+\beta+n+1)_k (-x-1)_{k}}{k!(\alpha+1)_{k} (-m)_k},$$ giving the right hand side of .
The set of difference equations will turn out to be the essential ingredient for the eigenvector construction in the next section.
Eigenvalues and eigenvectors of the interaction matrix {#eigen}
======================================================
Let $N=2m+1$ be an odd integer, and consider the $(N+1)\times(N+1)$ interaction matrix $M$ of the form with spin chain data $J_k$ determined by , i.e.$$\begin{gathered}
J_k= \begin{cases}
\sqrt{(k+1)(2m+1-k)}, & \text{if $k$ is odd,}\\
\sqrt{(k+2\alpha+2)(2m+2\beta-k)}, & \text{if $k$ is even.}
\end{cases}
\label{Jabm}\end{gathered}$$
We begin with the construction of a $(N+1)\times(N+1)$ matrix $U=(U_{ij})_{0\leq i,j \leq N}$ whose even rows are given in terms of normalized Hahn polynomials with parameters $(\alpha,\beta)$ and whose odd rows are given in terms of normalized Hahn polynomials with parameters $(\alpha+1,\beta-1)$. In order to have positive weight functions for both, we require that $\alpha>-1$ and $\beta>0$.
The $(N+1)\times(N+1)$ matrix $U$ with indices running from 0 to $N=2m+1$ is defined by $$\begin{gathered}
U_{2i,m-j} = U_{2i,m+j+1} = \frac{(-1)^i}{\sqrt{2}} \tilde Q_j(i;\alpha,\beta,m), \label{Ueven}\\
U_{2i+1,m-j} = -U_{2i+1,m+j+1} = -\frac{(-1)^i}{\sqrt{2}} \tilde Q_j(i;\alpha+1,\beta-1,m), \label{Uodd}\end{gathered}$$ where $i,j\in\{0,1,\ldots,m\}$. \[defU\]
It is easy to verify that $U$ is an orthogonal matrix. Indeed, let us compute $U^TU$: $$\begin{gathered}
(U^TU)_{jk} = \sum_{i=0}^{2m+1} U_{ij}U_{ik}
= \sum_{i=0}^{m} U_{2i,j}U_{2i,k} + \sum_{i=0}^{m} U_{2i+1,j}U_{2i+1,k}.
\label{UU}\end{gathered}$$ For $j,k\in\{0,\ldots,m\}$, gives $$\begin{gathered}
\sum_{i=0}^{m} \frac{1}{2} \tilde Q_{m-j}(i;\alpha,\beta,m) \tilde Q_{m-k}(i;\alpha,\beta,m) \\
\qquad \quad{}+ \sum_{i=0}^{m} \frac{1}{2} \tilde Q_{m-j}(i;\alpha+1,\beta-1,m) \tilde Q_{m-k}(i;\alpha+1,\beta-1,m) \\
\qquad{} = \frac{1}{2} \delta_{m-j,m-k} + \frac{1}{2} \delta_{m-j,m-k} = \delta_{jk}\end{gathered}$$ using the orthogonality of Hahn polynomials. For $j,k\in\{m+1,\ldots,2m+1\}$, the computation is essentially the same and gives again $\delta_{jk}$. For $j\in\{0,\ldots,m\}$ and $k\in\{m+1,\ldots,2m+1\}$, gives $$\begin{gathered}
\sum_{i=0}^{m} \frac{1}{2} \tilde Q_{m-j}(i;\alpha,\beta,m) \tilde Q_{k-m-1}(i;\alpha,\beta,m) \\
\qquad \quad{}- \sum_{i=0}^{m} \frac{1}{2} \tilde Q_{m-j}(i;\alpha+1,\beta-1,m) \tilde Q_{k-m-1}(i;\alpha+1,\beta-1,m) \\
\qquad{} = \frac{1}{2} \delta_{m-j,k-m-1} - \frac{1}{2} \delta_{m-j,k-m-1} = 0,\end{gathered}$$ and for $j\in\{m+1,\ldots,2m+1\}$ and $k\in\{0,\ldots,m\}$, the result is the same. So it follows that $(U^TU)_{jk}=\delta_{jk}$, or $U^TU=I$, the identity matrix. Hence $U^T$ is the inverse of $U$, so $UU^T=I$ holds as well.
Now we have the main proposition.
\[proposition2\] Let $M$ be the tridiagonal $(2m+2)\times(2m+2)$-matrix $$\begin{gathered}
M= \left( \begin{array}{ccccc}
0 & J_0 & 0 & & \\
J_0 & 0 & J_1 & \ddots & \\
0 & J_1 & 0 & \ddots & 0 \\
&\ddots & \ddots & \ddots & J_{2m} \\
& & 0 & J_{2m} & 0
\end{array} \right),\end{gathered}$$ where the $J_k$ are given in , and let $U$ be the matrix determined in Definition [\[defU\]]{}. Then $U$ is an orthogonal matrix: $$\begin{gathered}
U U^T = U^TU=I.\end{gathered}$$ Furthermore, the columns of $U$ are the eigenvectors of $M$, i.e. $$\begin{gathered}
M U = U D,
\label{MUUD}\end{gathered}$$ where $D$ is a diagonal matrix containing the eigenvalues $\epsilon_j$ of $M$: $$\begin{gathered}
D= \mathop{\rm diag}\nolimits (\epsilon_0,\epsilon_1,\ldots,\epsilon_{2m+1}), \qquad
\epsilon_{m-k}=-2\sqrt{(\alpha+k+1)(\beta+k)},\nonumber\\
\epsilon_{m+k+1}=2\sqrt{(\alpha+k+1)(\beta+k)},
\qquad k=0,1,\ldots,m.\label{epsilon}\end{gathered}$$
The orthogonality of $U$ has already been proved, so it remains to verify and . Now $$\begin{gathered}
\big(MU\big)_{ij}= \sum_{k=0}^{2m+1}M_{ik}U_{kj}=J_{i-1}U_{i-1,j}+J_{i}U_{i+1,j}.
\label{MU}\end{gathered}$$ We have to consider in four distinct cases, according to $i$ even or odd, and to $j$ belonging to $\{0,1,\ldots,m\}$ or to $\{m+1,m+2,\ldots,2m+1\}$. Let us consider the case that $i$ is odd and $j\in\{0,1,\ldots,m\}$. Then, relabelling the indices appropriately, and yield: $$\begin{gathered}
(MU)_{2i+1,m-j}=J_{2i}U_{2i,m-j}+J_{2i+1}U_{2i+2,m-j} \\
\phantom{(MU)_{2i+1,m-j}}{} =(-1)^i\sqrt{2}\sqrt{(\alpha+i+1)(m+\beta-i)}\tilde Q_j(i;\alpha,\beta,m) \\
\phantom{(MU)_{2i+1,m-j}=}{} + (-1)^{i+1}\sqrt{2}\sqrt{(i+1)(m-i)}\tilde Q_j(i+1;\alpha,\beta,m)\\
\phantom{(MU)_{2i+1,m-j}}{}= (-1)^i \sqrt{2}\sqrt{\frac{(\alpha+1)_{i+1}(\beta+1)_{m-i-1}}{i!(m-i)!h_j(\alpha,\beta,m)}}[ (\beta+m-i) Q_j(i;\alpha,\beta,m)\\
\phantom{(MU)_{2i+1,m-j}=}{}
-(m-i) Q_j(i+1;\alpha,\beta,m)].\end{gathered}$$ Applying (\[Q-rec1\]), this becomes $$\begin{gathered}
=(-1)^i \sqrt{2} \sqrt{\frac{(\alpha+1)_{i+1}(\beta+1)_{m-i-1}}{i!(m-i)!h_j(\alpha,\beta,m)}} \frac{(\alpha+j+1)(\beta+j)}{(\alpha+1)} Q_j(i;\alpha+1,\beta-1,m)\\
= -2\sqrt{(\alpha+j+1)(\beta+j)} U_{2i+1,m-j} = \epsilon_{m-j}U_{2i+1,m-j}=\big(UD\big)_{2i+1,m-j}.\end{gathered}$$ For $i$ odd and $j\in\{m+1,m+2,\ldots,2m+1\}$, the computation is essentially the same. For $i$ even (and the two cases for $j$), the computation is also similar, but now the second difference equation must be used.
Note that the spectrum of $M$ is symmetric, consisting of the values $\pm 2\sqrt{(\alpha+k+1)(\beta+k)}$ ($k=0,1,\ldots,m$). Furthermore, when $\beta=\alpha+1$, the spectrum consists of integers $\pm 2(\alpha+k+1)$. This latter case corresponds to the model of Shi et al. [@Shi2005].
Some aspects of the corresponding spin chain model {#correlation}
==================================================
Let us consider a spin chain with data determined by . The dynamics of this system is described by the unitary time evolution operator $\exp(-it\hat H)$. The transition amplitude of a single spin excitation from site $s$ to site $r$ of the spin chain is given by the time-dependent correlation function [@Bose2007; @Chakrabarti2010] $$f_{r,s}(t) = (r| \exp(-it\hat H) |s).$$ But the (orthonormal) eigenvectors of $\hat H$ in the single fermion mode are now known and given by $\varphi_j= \sum\limits_{k=0}^N U_{kj}\,|\,k)$, i.e. the columns of the matrix $U$ constructed in , , with $\hat H\varphi_j = M\varphi_j = \epsilon_j \varphi_j$. Using the orthogonality of the states $\varphi_j$, one finds [@Chakrabarti2010]: $$\begin{gathered}
f_{r,s}(t) = \sum_{j=0}^N U_{rj}U_{sj} {\rm e}^{-it\epsilon_j}.
$$
Due to the expressions , , implying $U_{r,m-j}=(-1)^r U_{r,m+j+1}$, it is appropriate to write the correlation function in the following form: $$\begin{gathered}
f_{rs}(t) =\sum_{j=0}^m\big(U_{r,m-j}U_{s,m-j} {\rm e}^{-it\epsilon_{m-j}}+
U_{r,m+j+1}U_{s,m+j+1} {\rm e}^{-it\epsilon_{m+j+1}}\big) \nonumber\\
\phantom{f_{rs}(t)}{} =\sum_{j=0}^m U_{r,m-j}U_{s,m-j} \big( \,{\rm e}^{-it\epsilon_{m-j}} +(-1)^{r+s} \,{\rm e}^{it\epsilon_{m-j}} \big). \label{CF}\end{gathered}$$ Now it is a matter of considering the different parities for $r$ and $s$. In the case they are both even, one finds $$\begin{gathered}
f_{2k,2l}(t) =(-1)^{k+l}\sqrt{w(k;\alpha,\beta,m)w(l;\alpha,\beta,m)}\nonumber\\
\phantom{f_{2k,2l}(t) =}{} \times \sum_{j=0}^{m}Q_j(k;\alpha,\beta,m) Q_j(l;\alpha,\beta,m)
\frac{\cos\big(2t\sqrt{(\alpha+j+1)(\beta+j)} \big)}{h_j(\alpha,\beta,m)}. $$ In the case the first index is odd and the second even, this becomes $$\begin{gathered}
f_{2k+1,2l}(t) =-i(-1)^{k+l}\sqrt{w(k;\alpha+1,\beta-1,m)w(l;\alpha,\beta,m)}\label{CF2}\\
\phantom{f_{2k+1,2l}(t) =}{} \times \sum_{j=0}^{m}Q_j(k;\alpha+1,\beta-1,m) Q_j(l;\alpha,\beta,m)
\frac{\sin\big(2t\sqrt{(\alpha+j+1)(\beta+j)} \big)}{\sqrt{h_j(\alpha+1,\beta-1,m)h_j(\alpha,\beta,m)}}. \nonumber\end{gathered}$$ The expressions for the case even/odd and odd/odd are similar, the main message being that due to the analytic expressions for the eigenvectors, we obtain explicit expressions for the correlation function.
Let us examine, in this context, the transition from one end of the chain ($s=0$) to the final end of the chain ($r=N=2m+1$). Expression reduces to: $$\begin{gathered}
f_{2m+1,0}(t) = -i(-1)^m\sqrt{(\beta)_{m+1}(\alpha+1)_{m+1}} \nonumber\\
\phantom{f_{2m+1,0}(t) =}{} \times \sum_{j=0}^m \frac{(2j+\alpha+\beta+1)(-m)_j}{(j+\alpha+\beta+1)_{m+1} j!}
\frac{\sin\big(2t\sqrt{(\alpha+j+1)(\beta+j)} \big)}{\sqrt{(\alpha+j+1)(\beta+j)}}.
\label{CF3}\end{gathered}$$ For general $\alpha$ and $\beta$, this expression cannot be simplified further. Let us now consider the special case that $$\beta=\alpha+1.$$ Then reduces to $$\begin{gathered}
f_{2m+1,0}(t) = -2i(-1)^m (\alpha+1)_{m+1} \sum_{j=0}^m \frac{(-m)_j}{(j+2\alpha+2)_{m+1} j!}
\sin\big(2t(\alpha+j+1)\big).
\label{CF3a}\end{gathered}$$ This can be rewritten as $$\begin{gathered}
f_{2m+1,0}(t) = -2i(-1)^m \frac{(\alpha+1)_{m+1}}{(2\alpha+2)_{m+1}} \sum_{j=0}^m \frac{(-m)_j(2\alpha+2)_j}{j!(2\alpha+m+3)_j}
\sin\big(2t(\alpha+j+1)\big).
\label{CF4}\end{gathered}$$ The last sum is of hypergeometric type ${}_2F_1$, and so it can be further simplified for special values of $t$ and/or $\alpha$. In particular, for $t=T=\pi/2$, one has $\sin\big(\pi(\alpha+j+1)\big)=-(-1)^j \sin(\pi\alpha)$. Using then Kummer’s summation formula [@Bailey; @Slater] $${}_2F_1 \left( {\genfrac{}{}{0pt}{}{-m,2\alpha+2}{2\alpha+m+3}} ; -1 \right)= \frac{(2\alpha+3)_m}{(\alpha+2)_m},$$ in the right hand side of , this expression reduces to $$\begin{gathered}
f_{N,0}\left(\frac{\pi}{2}\right)= f_{2m+1,0}\left(\frac{\pi}{2}\right) = i(-1)^m\sin(\pi\alpha).
\label{CF5}\end{gathered}$$ Note, by the way, that also for $t=2T=\pi$ one can simplify , since now $\sin\big(2\pi(\alpha+j+1)\big)=\sin(2\pi\alpha)$. Then, using Gauss’s summation formula [@Bailey; @Slater] $${}_2F_1 \left( {\genfrac{}{}{0pt}{}{-m,2\alpha+2}{2\alpha+m+3}} ; 1 \right)= \frac{(m+1)_m}{(2\alpha+m+3)_m},$$ the right hand side of reduces to $$\begin{gathered}
f_{N,0}(\pi)= f_{2m+1,0}(\pi) = -2i\sin(2\pi\alpha) (-1)^m \frac{(\alpha+1)_{m+1}(m+1)_m}{(2\alpha+2)_{2m+1}} .
$$
Note the importance of . Indeed, keeping in mind that $\alpha>-1$, we have $$\begin{gathered}
\left|f_{N,0}\left(\frac{\pi}{2}\right)\right|= \left|f_{2m+1,0}\left(\frac{\pi}{2}\right)\right| = 1 \qquad \hbox{for} \quad
\alpha=-\frac12, \frac12, \frac32, \frac52,\ldots.
\label{psf}\end{gathered}$$ So there is perfect state transfer in the chain for $\alpha$ assuming one of these values, at time $t=\pi/2$. In fact, this corresponds to the values given by Shi et al. [@Shi2005]. Note that for $\alpha=-\frac12$, the spin chain data further reduces to that of Christandl [@Christandl2004].
The case corresponds to $2\alpha+1 =2l$ with $l$ a nonnegative integer. As a matter of fact, this can still be extended slightly. Let $2\alpha+1 =\frac{2l}{2k+1}$ with both $l$ and $k$ nonnegative integers. Then for $t=T'=(2k+1)\pi/2$, the factor in becomes $\sin(2t(\alpha+j+1))=-(-1)^j \sin((2k+1)\alpha\pi)=(-1)^{j+k+l}$, and so we have the result $$\begin{gathered}
\left|f_{N,0}\left((2k+1)\frac{\pi}{2}\right)\right|= \left|f_{2m+1,0}\left((2k+1)\frac{\pi}{2}\right)\right| = 1 \qquad \hbox{for} \quad
2\alpha+1=\frac{2l}{2k+1}, \quad k,l\in{\mathbb Z}_+ .
$$ This case appears already in the paper of Qian et al. [@Qian], who use the “mirror mode concurrence” to find this extension of Shi’s result.
As far as perfect state transfer is concerned, our extension of Shi’s model by an extra parameter $\beta$ does not give rise to any new cases. In fact, just for investigating perfect state transfer, the mathematical machinery developed here is not necessary: the verification of the spectrum parity matching condition, using the entries in the interaction matrix and the spectrum itself, is sufficient. The main advantage of our analysis is the explicit computation of the correlation function. In particular, the simplicity of the expressions and , describing the transfer from one end of the chain to the other end, is striking. Apart from the model of Christandl [@Christandl2004], where the general correlation function is given in [@Chakrabarti2010 § 2], there are no other models with such an elegant and simple correlation function.
On the $\boldsymbol{q}$-generalization of the previous results {#qHahn}
==============================================================
As the classical orthogonal polynomials of hypergeometric type have a generalization in terms of basic hypergeometric series, i.e.a $q$-generalization, one may wonder whether the present construction of the tridiagonal interaction matrix $M$ can also be generalized. This is indeed the case: we can present a matrix $M_q$, whose eigenvectors are given in terms of $q$-Hahn polynomials, and whose eigenvalues are symmetric and take a simple form. In order to present these results, let us first briefly recall some notation related to $q$-series [@Gasper].
For a positive real number $q$ ($\ne 1$), the $q$-Hahn polynomial $Q_n(q^{-x};\alpha, \beta, m |q)$ of degree $n$ ($n=0,1,\ldots,m$) in $q^{-x}$ is defined by [@Koekoek; @Gasper]: $$\begin{gathered}
Q_n(q^{-x};\alpha,\beta,m|q) = {}_3\Phi_2 \left( {\genfrac{}{}{0pt}{}{q^{-n},\alpha\beta q^{n+1},q^{-x}}{\alpha q,q^{-m}}} ; q,q \right).
\label{defqQ}\end{gathered}$$ Herein, the function $_3\Phi_2$ is the basic hypergeometric series [@Bailey; @Slater; @Gasper]: $$\begin{gathered}
{\;}_3\Phi_2 \left( {\genfrac{}{}{0pt}{}{a,b,c}{d,e}} ; q, z \right)=\sum_{k=0}^\infty \frac{(a, b, c; q)_k}{(q,d, e;q)_k}
z^k.
\label{defPhi}\end{gathered}$$ Note that in (\[defPhi\]) we use the common notation for $q$-shifted factorials and their products [@Gasper]: $$\begin{gathered}
(a_1,a_2,\dots ,a_A;q)_k=(a_1;q)_k(a_2;q)_k\cdots(a_A;q)_k,\\ (a;q)_k=(1-a)(1-aq)\cdots\big(1-aq^{k-1}\big) \qquad {\rm and} \qquad (a)_0=1.
$$ In (\[defqQ\]), the series is terminating because of the appearance of $q^{-n}$ in the numerator. $q$-Hahn polynomials satisfy a (discrete) orthogonality relation [@Koekoek]: $$\begin{gathered}
\sum_{x=0}^m w(x;\alpha, \beta,m|q) Q_l\big(q^{-x};\alpha, \beta, m|q\big) Q_n\big(q^{-x};\alpha,\beta,m|q\big) = h_n(\alpha,\beta,m|q) \delta_{ln},
$$ where $$\begin{gathered}
w(x;\alpha, \beta,m|q) = \frac{(\alpha q,q^{-m};q)_x}{(q,\beta^{-1}q^{-m};q)_x} (\alpha \beta q)^{-x}, \qquad x=0,1,\ldots,m, \\
h_n (\alpha,\beta,m|q)= \frac{(\alpha\beta q^2;q)_{m}(q,\alpha\beta q^{m+2},\beta q;q)_n
(1-\alpha\beta q)(-\alpha q)^n}{(\beta q;q)_m (\alpha q)^m (\alpha q, \alpha\beta q, q^{-m};q)_n(1-\alpha\beta q^{2n+1})}
q^{({\genfrac{}{}{0pt}{}{n}{2}}) -mn}.\end{gathered}$$ We shall assume that $0<q<1$; then the weight function is positive when $0<\alpha<q^{-1}$ and $0<\beta<q^{-1}$. Denote the orthonormal $q$-Hahn functions as follows: $$\begin{gathered}
\tilde Q_n(q^{-x};\alpha,\beta,m|q) \equiv \frac{\sqrt{w(x;\alpha, \beta,m|q)} Q_n(q^{-x};\alpha,\beta,m|q)}{\sqrt{h_n(q;\alpha, \beta,m)}}.
\label{qQtilde}\end{gathered}$$
Just as in Section \[Hahn\], the main result needed is a set of difference equations for $q$-Hahn polynomials.
\[proposition3\] $q$-Hahn polynomials satisfy the following $q$-difference equations: $$\begin{gathered}
\big(1-\beta q^{m-x}\big) Q_n\big(q^{-x};\alpha ,\beta ,m|q\big) -\big(1-q^{m-x}\big) Q_{n}\big(q^{-x-1};\alpha , \beta,m |q\big)\nonumber \\
\qquad {} =\frac{(1-\alpha q^{n+1})(1-\beta q^n) q^{m-n-x}}{1-\alpha q} Q_n\big(q^{-x};\alpha q,\beta q^{-1},m|q\big), \label{qQ-rec1} \\
\big(1-q^{x+1}\big) \alpha q Q_n\big(q^{-x};\alpha q,\beta q^{-1},m|q\big)-\big(1-\alpha q^{x+2}\big) Q_{n}\big(q^{-x-1};\alpha q , \beta q^{-1},m|q\big) \nonumber \\
\qquad {} =-(1-\alpha q) Q_n\big(q^{-x-1};\alpha,\beta,m|q\big). \label{qQ-rec2}\end{gathered}$$
The proof follows the same computation as in the proof of Proposition \[proposition1\], with the replacement of Pochhammer symbols by the corresponding $q$-shifted factorials (and keeping track of the appropriate powers of $q$).
We now come to the construction of the tridiagonal matrix $M_q$ and the matrix of eigenvectors $U$. The polynomials that appear here will be $q$-Hahn polynomials with parameters $(\alpha,\beta)$ and $(\alpha q,\beta q^{-1})$. So in order to have positive weight functions for both sets of polynomials, we shall assume: $$0<q<1, \qquad 0<\alpha<q^{-1}, \qquad 0<\beta<1.$$ As before, let $N=2m+1$, and consider the $(N+1)\times(N+1)$ interaction matrix $M_q$ of the form with non-zero matrix elements given by: $$\begin{gathered}
J_{2k+1} = 2\sqrt{(1-q^{k+1})(1-q^{m-k})q^{k+1}\alpha }, \qquad
J_{2k} = 2\sqrt{(1-\alpha q^{k+1})(1-\beta q^{m-k})q^{k}} ,
\label{qJi}\end{gathered}$$ where $k=0,1,\ldots,m$. The $(N+1)\times(N+1)$ matrix $U$ with indices running from 0 to $N=2m+1$ is defined similarly as in Definition \[defU\], but in terms of $q$-Hahn polynomials : $$\begin{gathered}
U_{2i,m-j} = U_{2i,m+j+1} = \frac{(-1)^i}{\sqrt{2}} \tilde Q_j\big(q^{-i};\alpha,\beta,m|q\big), \label{qUeven}\\
U_{2i+1,m-j} = -U_{2i+1,m+j+1} = -\frac{(-1)^i}{\sqrt{2}} \tilde Q_j\big(q^{-i};\alpha q,\beta q^{-1},m|q\big), \label{qUodd}\end{gathered}$$ where $i,j\in\{0,1,\ldots,m\}$. By the same argument as in Section \[eigen\], the orthogonality of the matrix $U$ follows from the orthogonality of the $q$-Hahn polynomials and the appropriate signs.
Then the main result in the $q$-case reads:
Let $M_q$ be the tridiagonal $(2m+2)\times(2m+2)$-matrix $$\begin{gathered}
M_q= \left( \begin{array}{ccccc}
0 & J_0 & 0 & & \\
J_0 & 0 & J_1 & \ddots & \\
0 & J_1 & 0 & \ddots & 0 \\
&\ddots & \ddots & \ddots & J_{2m} \\
& & 0 & J_{2m} & 0
\end{array} \right),\end{gathered}$$ where the $J_k$ are given in , and let $U$ be the matrix determined in , . Then $U$ is an orthogonal matrix, $U U^T = U^TU=I$. Furthermore, the columns of $U$ are the eigenvectors of $M_q$, i.e. $$\begin{gathered}
M_q U = U D,
$$ where $D$ is a diagonal matrix containing the eigenvalues $\epsilon_j$ of $M_q$: $$\begin{gathered}
D= \mathop{\rm diag}\nolimits (\epsilon_0,\epsilon_1,\ldots,\epsilon_{2m+1}), \qquad
\epsilon_{m-k}=-2\sqrt{(1-\alpha q^{k+1})(1-\beta q^k)q^{m-k}}, \nonumber\\
\epsilon_{m+k+1}=2\sqrt{(1-\alpha q^{k+1})(1-\beta q^k)q^{m-k}},
\qquad k=0,1,\ldots,m. $$
The proof of this proposition is essentially the same as that of Proposition \[proposition3\], and uses the $q$-difference equations , . Note that, as in the ordinary case, the spectrum of $M_q$ is symmetric.
Once the explicit eigenvectors of the Hamiltonian are known in the $q$-generalized case , one can compute the correlation function, using the same expression . The expressions become quite involved, so we give just one example here. This is in the case of transition from one end ($s=0$) to the other end ($r=2m+1$) of the chain, and for $\beta =q \alpha$: $$\begin{gathered}
f_{2m+1,0}(t) =i(-1)^{m}q^{m/2} \alpha^{m/2}(\alpha q;q)_{m+1}\nonumber\\
\phantom{f_{2m+1,0}(t) =}{} \times \sum_{j=0}^m (-1)^j\sin\big(2t\big(1-\alpha q^{j+1}\big)q^{(m-j)/2}\big)
q^{j^2/2}
\frac{(q^{m-j+1};q)_{j}(1+\alpha q^{j+1})}{(\alpha^2 q^{j+2};q)_{m+1}(q;q)_j}.\end{gathered}$$ The $q$-generalization does not give rise to any special cases with perfect state transfer, however.
Conclusions and outlook
=======================
We have dealt with some mathematical aspects of a spin chain model of Shi et al. [@Shi2005], which is a one-parameter extension of the popular spin chain introduced by Christandl et al. [@Christandl2004]. In Christandl’s model, the single fermion eigenvalues and eigenvectors could easily be computed, the eigenvectors being related to Krawtchouk polynomials. In Shi’s model, with an extra parameter $\alpha$, there was so far no known expression of the eigenvectors. In the current paper, we have shown that these eigenvectors can be expressed in terms of Hahn polynomials. As a matter of fact, we first extend Shi’s model by introducing an extra parameter $\beta$, and then construct the eigenvectors. In this process two types of Hahn polynomials are involved, those with parameters $(\alpha,\beta)$ and those with $(\alpha+1,\beta-1)$. These Hahn polynomials are appropriately combined in a matrix $U$, yielding the eigenvectors wanted. When $\beta=\alpha+1$, the two-parameter spin chain reduces to that of Shi. And when $\alpha=-\frac12$, $\beta=\frac12$, the spin chain reduces to that of Christandl. Note, by the way, that in this last case the eigenvectors (expressed in terms of Hahn polynomials) indeed reduce to the known ones (expressed in terms of Krawtchouk polynomials). This follows from the fact that when $\alpha=-\frac12$ and $\beta=\frac12$ the Hahn polynomials, which are ${}_3F_2$ series, reduce to ${}_2F_1$ series according to $$\begin{gathered}
{}_3F_2 \left( {\genfrac{}{}{0pt}{}{-s,s+1,-i}{1/2,-m}} ; 1 \right)= (-1)^i \frac{\binom{2m+1}{2i}}{\binom{m}{i}} {\;}_2F_1 \left( {\genfrac{}{}{0pt}{}{-2i,-m-s-1}{-2m-1}} ; 2 \right),\\
{}_3F_2 \left( {\genfrac{}{}{0pt}{}{-s,s+1,-i}{3/2,-m}} ; 1 \right)= -\frac{(-1)^i}{(2s+1)} \frac{\binom{2m+1}{2i+1}}{\binom{m}{i}} {\;}_2F_1 \left( {\genfrac{}{}{0pt}{}{-2i-1,-m-s-1}{-2m-1}} ; 2 \right).\end{gathered}$$ These reductions can be obtained, e.g., from [@Atakishiyev2005 (48)]. The ${}_2F_1$ series in the right hand side correspond to (symmetric) Krawtchouk polynomials (with $p=1/2$).
Due to the explicit forms of the eigenvectors, the time-dependent correlation function $f_{r,s}(t)$ has been computed for the spin chains under consideration. In special cases, the expression of the correlation function is particularly simple, see e.g. .
In the construction of the eigenvectors, the main relations needed are two new difference equations for Hahn polynomials. We have also examined the $q$-generalization of these results. The $q$-extension of these difference equations is more or less straightforward. Also the construction of the corresponding eigenvectors in terms of $q$-Hahn polynomials has been completed.
The extension of symmetric Krawtchouk polynomials (without a parameter $\alpha$) to Hahn polynomials (with parameters $(\alpha,\alpha+1)$) may also be used in other applications. In particular, we hope to extend the finite oscillator models of [@Atakishiyev2005], by introducing such an extra parameter. It remains to be seen, in that case, how the underlying Lie algebra is deformed, and how the parameter has an influence on the finite oscillator eigenstates. This topic will be treated elsewhere.
Acknowledgements {#acknowledgements .unnumbered}
----------------
N.I. Stoilova would like to thank Professor H.D. Doebner (Clausthal, Germany) for constructive discussions. N.I. Stoilova was supported by project P6/02 of the Interuniversity Attraction Poles Programme (Belgian State – Belgian Science Policy) and by the Humboldt Foundation.
[99]{}
Shi T., Li Y., Song A., Sun C.P., Quantum-state transfer via the ferromagnetic chain in a spatially modulated field, [[*Phys. Rev. A*]{}](http://dx.doi.org/10.1103/PhysRevA.71.032309) [**71**]{} (2005), 032309, 5 pages, [quant-ph/0408152](http://arxiv.org/abs/quant-ph/0408152).
Bose S., Quantum communication through an unmodulated spin chain, [[*Phys. Rev. Lett.*]{}](http://dx.doi.org/10.1103/PhysRevLett.91.207901) [**91**]{} (2003), 207901, 4 pages, [quant-ph/0212041](http://arxiv.org/abs/quant-ph/0212041).
Bose S., Jin B.-Q., Korepin V.E., Quantum communication through a spin ring with twisted boundary conditions, [[*Phys. Rev. A*]{}](http://dx.doi.org/10.1103/PhysRevA.72.022345) [**72**]{} (2005), 022345, 4 pages, [quant-ph/0409134](http://arxiv.org/abs/quant-ph/0409134).
Bose S., Quantum communication through spin chain dynamics: an introductory overview, [[*Contemp. Phys.*]{}](http://dx.doi.org/10.1080/00107510701342313) [**48**]{} (2007), 13–30, [arXiv:0802.1224](http://arxiv.org/abs/0802.1224).
Lieb E., Wu F., Absence of Mott transition in an exact solution of short-range 1-band model in 1 dimension, [[*Phys. Rev. Lett.*]{}](http://dx.doi.org/10.1103/PhysRevLett.20.1445) [**20**]{} (1968), 1445–1448.
Jordan P., Wigner E., About the Pauli exclusion principle, [*Z. Phys.*]{} [**47**]{} (1928), 631–651.
Christandl M., Datta N., Ekert A., Landahl A.J., Perfect state transfer in quantum spin networks, [[*Phys. Rev. Lett.*]{}](http://dx.doi.org/10.1103/PhysRevLett.92.187902) [**92**]{} (2004), 187902, 4 pages, [quant-ph/0309131](http://arxiv.org/abs/quant-ph/0309131).
Albanese C., Christandl M., Datta N., Ekert A., Mirror inversion of quantum states in linear registers, [[*Phys. Rev. Lett.*]{}](http://dx.doi.org/10.1103/PhysRevLett.93.230502) [**93**]{} (2004), 230502, 4 pages, [quant-ph/0405029](http://arxiv.org/abs/quant-ph/0405029).
Christandl M., Datta N., Dorlas T.C., Ekert A., Kay A., Landahl A.J., Perfect transfer of arbitrary states in quantum spin networks, [[*Phys. Rev. A*]{}](http://dx.doi.org/10.1103/PhysRevA.71.032312) [**71**]{} (2005), 032312, 11 pages, [quant-ph/0411020](http://arxiv.org/abs/quant-ph/0411020).
Yung M.H., Bose S., Perfect state transfer, effective gates, and entanglement generation in engineered bosonic and fermionic networks, [[*Phys. Rev. A*]{}](http://dx.doi.org/10.1103/PhysRevA.71.032310) [**71**]{} (2005), 032310, 6 pages, [quant-ph/0407212](http://arxiv.org/abs/quant-ph/0407212).
Karbach P., Stolze J., Spin chains as perfect quantum state mirrors, [[*Phys. Rev. A*]{}](http://dx.doi.org/10.1103/PhysRevA.72.030301) [**72**]{} (2005), 030301, 4 pages, [quant-ph/0501007](http://arxiv.org/abs/quant-ph/0501007).
Kay A., A review of perfect state transfer and its application as a constructive tool, [[*Int. J. Quantum Inf.*]{}](http://dx.doi.org/10.1142/S0219749910006514) [**8**]{} (2010), 641–676, [arXiv:0903.4274](http://arxiv.org/abs/0903.4274).
Chakrabarti R., Van der Jeugt J., Quantum communication through a spin chain with interaction determined by a Jacobi matrix, [[*J. Phys. A: Math. Theor.*]{}](http://dx.doi.org/10.1088/1751-8113/43/8/085302) [**43**]{} (2010), 085302, 20 pages, [arXiv:0912.0837](http://arxiv.org/abs/0912.0837).
Jafarov E.I., Van der Jeugt J., Quantum state transfer in spin chains with $q$-deformed interaction terms, [[*J. Phys. A: Math. Theor.*]{}](http://dx.doi.org/10.1088/1751-8113/43/40/405301) [**43**]{} (2010), 405301, 18 pages, [arXiv:1005.2912](http://arxiv.org/abs/1005.2912).
Koekoek R., Lesky P.A., Swarttouw R.F., Hypergeometric orthogonal polynomials and their $q$-analogues, [[*Springer Monographs in Mathematics*]{}](http://dx.doi.org/10.1007/978-3-642-05014-5), Springer-Verlag, Berlin, 2010.
Nikiforov A.F., Suslov S.K., Uvarov V.B., Classical orthogonal polynomials of a discrete variable, [*Springer Series in Computational Physics*]{}, Springer-Verlag, Berlin, 1991.
Regniers G., Van der Jeugt J., Analytically solvable Hamiltonians for quantum systems with a nearest-neighbour interaction, [[*J. Phys. A: Math. Theor.*]{}](http://dx.doi.org/10.1088/1751-8113/42/12/125301) [**42**]{} (2009), 125301, 16 pages, [arXiv:0902.2308](http://arxiv.org/abs/0902.2308).
Bailey W.N., Generalized hypergeometric series, [*Cambridge Tracts in Mathematics and Mathematical Physics*]{}, no. 32, Stechert-Hafner, Inc., New York, 1964.
Slater L.J., Generalized hypergeometric functions, Cambridge University Press, Cambridge, 1966.
Qian X.-F., Li Y., Li Y., Song Z., Sun C.P., Quantum-state transfer characterized by mode entanglement, [[*Phys. Rev. A*]{}](http://dx.doi.org/10.1103/PhysRevA.72.062329) [**72**]{} (2005), 062329, 6 pages.
Gasper G., Rahman M., Basic hypergeometric series, 2nd ed., [*Encyclopedia of Mathematics and its Applications*]{}, Vol. 96, Cambridge University Press, Cambridge, 2004.
Atakishiyev N.M., Pogosyan G.S., Wolf K.B., Finite models of the oscillator, [*Phys. Part. Nuclei*]{} [**36**]{} (2005), 247–265.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Modified Chaplygin Gas has been successful in describing the cosmic history of the universe from radiation to $\Lambda$CDM in standard cosmology, while particle creation mechanism in nonlinear thermodynamics can be used to explain inflation as well as late time acceleration. The present work is an attempt to explore the possibilities of obtaining an alternative explanations to cosmic evolution when modified Chaplygin gas is used in the context of particle creation mechanism.'
author:
- Subhra Bhattacharya
- Shibaji Halder
- Subenoy Chakraborty
title: Evolving Cosmic Scenario in Modified Chaplygin Gas with Adiabatic Matter Creation
---
Introduction {#Intro}
============
The relativistic second order thermodynamic theories of Muller [@mul], Israel and Stewart [@is1]-[@his] and Pavon [*et. al.*]{} [@pav1],[@pav2] play crucial role in describing the evolution of the Universe as a sequence of dissipative processes. The theory proposes that deviations from equilibrium described by bulk stress, heat flow and shear stress can be treated as independent dynamical variables bounded by average molecular speed thereby ensuring causality. In a homogeneous and isotropic FRW universe the bulk viscous pressure is the only possible mechanism for dissipative processes. The bulk viscous pressure can be attributed to particle number changing processes in an expanding universe [@zel]-[@zim2] or it might due to coupling of the different components of the cosmic fluids [@ud]-[@sch]. Particle creation mechanism driving bulk viscous pressure has been extensively used to describe the dynamics and evolution of the early universe including early inflation and current accelerated expansion [@sc1]. Particle creation has also been related to emergent universe [@sc2].
In the present work we shall consider the cosmological implications of bulk viscous pressure due to particle creation mechanism in a universe with matter described by the Modified Chaplygin gas (MCG). Thermodynamically we shall concentrate on an isentropic universe, i.e. we shall envisage the universe as an open thermodynamic system, where entropy per particle under the mechanism of particle creation is constant [@pri; @cal], although there will be entropy creation due to particle creation driven phase space change. The viscous effects shall be described by the truncated Muller-Israel-Stewart (MIS) type theory.
The reason for considering Modified Chaplygin Gas (MCG) as the cosmic fluid is that it provides a unified Dark Matter-Dark energy manifestations in a single fluid. MCG is an exotic fluid with EoS $$p=A\rho-\frac{B}{\rho^{\alpha}},~0\leq\alpha\leq 1.\label{eos}$$ This EoS show that MCG can accommodate a radiation dominated universe for $A=1/3$ at high density to one with negative pressure at low density for the current accelerating universe [@deb1; @bena]. For $\alpha=-1,~B=1+A$ one can get the $\Lambda$CDM universe, while for $B=0,$ the EoS describes a perfect fluid, i.e. a quintessence model. Several works on MCG has established its consistency as a cosmic fluid [@wu; @bed; @costa; @deb2; @tha].
Using a FRW model of the universe with viscous effects described by the MIS theory in an isentropic universe with particle creation, we shall find expressions for the Hubble parameter in terms of particle creation rate. Corresponding relevant cosmological parameters like the scale factor, deceleration parameter and energy density are evaluated along with basic thermodynamic variables like fluid temperature $T$ and particle number density $n$ in terms of the particle creation rate. Using a single phenomenological choice of the particle creation rate we shall then show a unified cosmic evolution starting from early accelerated expansion to a late time accelerated one. Further we could successfully connect the particle creation rate, in MCG to an increasing entropy in the de-Sitter phase. Finally we shall relate the particle creation mechanism to Hawking radiation [@haw].
The paper is organised as follows: Section 2 deals with basic conditions for bulk viscous cosmology related to particle creation mechanism. Bulk viscous FRW universe with MCG has been presented in section 3. Section 4 shows a comparison of the present result for specific choices of the particle creation rate with recent observations. Section 5 shows the scalar field description corresponding to Model 3. The entropy production for the present context has been derived in section 6. Section 7 describes interrelation between particle creation process with Hawking radiation. Finally the paper ends with a brief discussion in section 8.
Bulk viscous universe with particle creation: Non equilibrium M-I-S type thermodynamic theory
==============================================================================================
The energy momentum tensor of a relativistic fluid with bulk viscosity as the only dissipative phenomenon is given by $$T^{\mu\nu}=(\rho+p+\Pi) u^{\mu}u^{\nu}+(p+\Pi)g^{\mu\nu}\label{em}$$ where $u^{\mu}$ is the 4 velocity, $\rho$ is the energy density, $p$ is the thermodynamic pressure and $\Pi$ is the bulk viscous pressure. Considering the second order non equilibrium thermodynamics, the entropy flow vector $S^{\mu}$ is characterised by the equation [@is2] $$S^{\mu}=sN^{\mu}-\frac{\tau \Pi^{2}}{2\zeta T}u^{\mu}\label{entropy}$$ where $N^{\mu}=nu^{\mu}$ is the particle flow vector with $n$ being the particle number density, $s,$ the entropy per particle, $\tau$ is the relaxation time, $T$ is the temperature of the fluid and $\zeta$ is the coefficient of bulk viscosity.
Now we consider a scenario where the non vanishing bulk viscous pressure is due to a change in fluid number density, which is characterised by the particle production rate $\Gamma=\frac{\dot{N}}{N}, ~N=na^{3}$ being the number of particles in co-moving volume $a^{3}.$ For $\Gamma>0$ we get particle creation while $\Gamma<0$ usually means particle annihilation. The varying particle number density will cause a change of phase creating a entropy production density which will be given by: $$S^{\mu}_{;\mu}=-\frac{\Pi}{T}\left[3H+\frac{\tau}{s}\dot{\Pi}+\frac{1}{2}\Pi T\left(\frac{\tau}{\zeta T}u^{\mu}\right)_{;\mu}+\varepsilon\frac{n\Gamma}{\Pi}\right]\label{eprod}$$ $\varepsilon$ being the chemical potential. For the validity of second law of thermodynamics we must have $S^{\mu}_{;\mu}=\frac{\Pi^{2}}{\zeta T}\geq 0.$ This gives the following non linear differential equation for bulk viscosity $\Pi$ [@zim2] $$\Pi^{2}\left[1+\frac{1}{2}T\left(\frac{\tau}{\zeta T}u^{\mu}\right)_{;\mu}\right]+\tau\Pi\dot{\Pi}+3H\zeta\Pi=-\zeta\varepsilon n\Gamma.\label{pi}$$ Thus any deviation from equilibrium is characterized by the bulk viscous pressure $\Pi$ in the presence of particle creation $\Gamma,$ further the above equation asserts the existence of a single causal theory even with particle creation processes taken into account.
Given the existence of particle creation $\Gamma,$ the conservation equations are modified as $$N^{\mu}_{;\mu}=n\Gamma ;~~T^{\mu\nu}_{;\nu}=0$$ which gives $$\begin{aligned}
&\dot{n}+3Hn=n\Gamma \label{consv1}\\
\text{and}~&\dot{\rho}+3H(\rho+p+\Pi)=0\label{consv2}\end{aligned}$$ with $\dot{n}=n_{,\mu}u^{\mu}.$ Comparing equations (\[consv1\]) and (\[consv2\]) with the Gibb’s relation $$Tds=d\left(\frac{\rho}{n}\right)+pd\left(\frac{1}{n}\right)\label{gibb}$$ one can get $$nT\dot{s}=-3H\Pi-(p+\rho)\Gamma.\label{entropy1}$$ Considering that the pressure $p$ and density $\rho$ are related to the thermodynamic variables $n$ and $T$ by the equations $p=p(n,T)$ and $\rho=\rho(n,T)$ and using the conservation equations (\[consv1\]) and (\[consv2\]) together with $$\frac{\partial\rho}{\partial n}=\frac{p+\rho}{n}-\frac{T}{n}\frac{dp}{dT}\label{rhovariat}$$ one can obtain the temperature evolution equation as $$\frac{\dot{T}}{T}=-3H\left[\frac{\partial p/\partial T}{\partial \rho/\partial T}+\frac{\Pi}{T\partial \rho/\partial T}\right]+\Gamma\left[\frac{\partial p/\partial T}{\partial \rho/\partial T}-\frac{p+\rho}{T\partial \rho/\partial T}\right]\label{t}$$ Alternatively using (\[entropy1\]) the above relation can be written as $$\frac{\dot{T}}{T}=-(3H-\Gamma)\frac{\partial p/\partial T}{\partial \rho/\partial T}+\frac{n\dot{s}}{\partial \rho/\partial T}.\label{tgamma}$$ Thus it is easily observed that particle production affects the temperature with an effective viscous pressure $\Pi$ together with a direct coupling.
Considering isentropic particle production characterised by constant entropy $\dot{s}=0$ the viscous pressure can be obtained directly in terms of particle production rate as $$\Pi=-\frac{\Gamma}{3H}(p+\rho).\label{pie}$$ From the above we can get a cosmic fluid characterised by changing particle number density. Also the variation of the fluid temperature is now given by $$\frac{\dot{T}}{T}=-(3H-\Gamma)\frac{\partial p}{\partial \rho}.\label{tisen}$$ Further from (\[consv1\]) for isentropic particle production the evolution of $n$ is given by $$\frac{\dot{n}}{n}=-(3H-\Gamma)\label{n}$$
Bulk viscous FRW universe with MCG as cosmic fluid {#model}
==================================================
We consider a spatially flat FRW model of the homogeneous universe as an open thermodynamic system with metric $$ds^{2}=-dt^{2}+a^{2}(t)[dr^{2}+r^{2}(d\theta^{2}+\sin^{2}\theta d\phi^{2})]\label{metric}$$ Since we shall consider the above metric in the context of non-equilibrium thermodynamics driven by particle creation mechanism, the corresponding cosmic fluid with dissipation $\Pi$ will have the field equations given by $$3H^{2}=\kappa\rho;~~~~\dot{H}=-\frac{\kappa}{2}(\rho+p+\Pi)\label{fe}$$ with $H$ the Hubble parameter and $\kappa=8\pi G$ is the Einstein’s gravitation constant. Considering MCG as the cosmic fluid with EoS given by (\[eos\]) together with (\[consv2\]) one can obtain the energy density of the fluid as $$\rho^{\alpha+1}=\frac{B}{A+1}+\frac{C}{A+1}a^{-3\mu}e^{\mu\int\Gamma dt}\label{rho}$$ with $\mu=(A+1)(\alpha+1)$ and $C$ the constant of integration. Then using (\[fe\]) above the Hubble parameter is obtained as (choosing $\kappa=1$) $$H=\frac{1}{\sqrt{3}}\left[\frac{B}{A+1}+\frac{C}{A+1}a^{-3\mu}e^{\mu\int\Gamma dt}\right]^{\frac{1}{2\alpha+2}}.\label{h}$$ Using $\dot{H}=-\frac{1}{2}\left(1-\frac{\Gamma}{3H}\right)(p+\rho)$ the deceleration parameter $q=-\left(1+\frac{\dot{H}}{H^{2}}\right)$ is obtained as $$q=-1+\frac{3}{2}\left(1-\frac{\Gamma}{3H}\right)\frac{C(A+1)}{Ba^{3\mu}e^{-\mu\int\Gamma dt}+C}.\label{dp}$$ Also using (\[tisen\]) and (\[n\]) the thermodynamic variables $T$ and $n$ can be obtained as $$T=T_{c}\rho^{-\alpha}\left(\frac{C}{A+1}\right)^{\frac{\mu-1}{\mu}}a^{-3(\mu-1)}e^{(\mu-1)\int\Gamma dt}\label{tm}$$ and $$n=n_{c}C^{\frac{1}{\mu}}a^{-3}e^{\int \Gamma dt}\label{nm}$$ where $n_{c},~T_{c}$ are constants of integration.
Equations (\[rho\])-(\[nm\]) clearly show that the relevant dynamical and thermodynamic parameters all vary as the particle creation rate $\Gamma.$ Thus we need specific choices of $\Gamma$ to trace the history of the cosmic evolution. In the absence of any known form of the particle creation parameter we make three phenomenological choices of the particle creation indicator $\Gamma.$ First two being functions of the scale factor $a,$ while the third being a function of the Hubble parameter $H.$ $$\begin{aligned}
\Gamma&=15\beta H[1-a\tanh(10-12a)]&\textbf{Model 1}\\
\Gamma& = 3\beta H(1+a^{m+1}),&\textbf{Model 2}\\
\Gamma&=3\beta H\frac{(A+1)H^{2\alpha+2}-B/3^{\alpha+1}}{(A+1)H_{0}^{2\alpha+2}-B/3^{\alpha+1}}&\textbf{Model 3}\end{aligned}$$ where $\beta$ is some positive constant and $m$ is any real number $(\neq -1).$ Model 1 and 2 are phenomenological choices based on the scale factor $a.$ (Recently a similar form as Model 1 was considered in [@nun] to describe Phantom behaviour through particle creation). In a universe dominated by bulk viscosity with perfect fluid, it is usually found that particle creation rate is proportional to the energy density [@gunzig]. Thus one can at least speculate that an universe dominated by bulk viscosity in the presence of an exotic fluid, like MCG, the particle creation rate will have some functional dependence on the energy density, hence we chose Model 3 as a function of the Hubble parameter $H.$
It is known that in a simplified model of the homogeneous and isotropic universe the dynamical variable is the scale factor $a(t),$ quantities that determine the time dependence of $a(t)$ must also determine the aspects of universe’s evolution. Taylor series expansion of $a(t)$ about present time reveals two such important terms as the Hubble parameter $H$ and the deceleration parameter $q$. The deceleration parameter is the first non linear correction of the expansion and is indicative of cosmic expansion. Essentially expansion rate of the universe is given by the Hubble parameter, such that a $H>0$ indicates an expanding universe while $q$ will indicate the time dependence of $H,$ such that an accelerating universe will have $q<0$ and decelerating universe will have $q>0.$ Hence corresponding to the above three phenomenological anasatze we shall consider the dynamic and thermal evolution of the universe by tracing the evolution of the parameters like the Hubble parameter $H,$ deceleration parameter $q,$ the energy density $\rho,$ particle number density $n$ and fluid temperature $T.$
Model 1:$\Gamma=15\beta H[1-a\tanh(10-12a)]$
---------------------------------------------
Using the expression for $\Gamma$ as given above in the equations (\[rho\])-(\[nm\]) we obtain the following expressions for $\rho,~H,~q,~n~\text{and}~T$ $$\begin{aligned}
\rho&=\left[\frac{B+Ca^{15\mu\beta-3\mu}\cosh^{\frac{5\mu\beta}{4}}(10-12a)}{A+1}\right]^{\frac{1}{\alpha+1}}\label{rho1}\\
H&=\frac{1}{\sqrt{3}}\left[\frac{B+Ca^{15\mu\beta-3\mu}\cosh^{\frac{5\mu\beta}{4}}(10-12a)}{A+1}\right]^{\frac{1}{2\alpha+2}}\label{h1}\\
q&=-1+\frac{3C(A+1)\left[ 1-5\beta(1-a\tanh(10-12a))\right]}{2Ba^{3\mu(1-5\beta)}\cosh^{-\frac{5\mu\beta}{4}}(10-12a)+2C}\label{q1}\\
n&=n_{c}C^{\frac{1}{\mu}}a^{15\beta-3}\cosh^{\frac{5\beta}{4}}(10-12a)\label{n1}\\
T&=T_{c}\left(\frac{C}{A+1}\right)^{\frac{\mu-1}{\mu}}\rho^{-\alpha}a^{(3\mu-3)(5\beta-1)}\cosh^{\frac{5\beta(\mu-1)}{4}}(10-12a)\label{t1}\end{aligned}$$
The above physical quantities are graphically evaluated using the following values of the parameters $A=\frac{1}{3},~B=3.5,~\alpha=0.5,~\beta=0.12~\text{and}~C=2$ [@sad; @costa].
-- -- --
-- -- --
Model 2: $\Gamma=3\beta H(1+a^{m+1})$
-------------------------------------
As before we can obtain the following expressions for $\rho,~H,~q,~n~\text{and}~T$ corresponding to the above $\Gamma$ as $$\begin{aligned}
\rho&=\left[\frac{B+Ca^{3\mu(\beta-1)}e^{\frac{3\mu\beta a^{m+1}}{m+1}}}{A+1}\right]^{\frac{1}{\alpha+1}}\label{rho2}\\
H&=\frac{1}{\sqrt{3}}\left[\frac{B+Ca^{3\mu(\beta-1)}e^{\frac{3\mu\beta a^{m+1}}{m+1}}}{A+1}\right]^{\frac{1}{2\alpha+2}}\label{h2}\\
q&=-1+\frac{3C(A+1)\left[1-\beta(1+a^{m+1})\right]}{2Ba^{3\mu(1-\beta)}e^{-\frac{3\mu\beta a^{m+1}}{m+1}}+2C}\label{q2}\\
n&=n_{c}C^{\frac{1}{\mu}}a^{3(\beta-1)}e^{\frac{3\beta a^{m+1}}{m+1}}\label{n2}\\
T&=T_{c}\left(\frac{C}{A+1}\right)^{\frac{\mu-1}{\mu}}\rho^{-\alpha}a^{3(\mu-1)(\beta-1)}e^{\frac{3\beta(\mu-1)a^{m+1}}{m+1}}\label{t2}\end{aligned}$$
The above physical parameters are graphically presented in Fig 2 for $m=-2$ and using the previous values for the parameters $A,~B,~\text{and}~\alpha,$ while $\beta=0.08~\text{and}~C=1.$
-- -- --
-- -- --
Model 3: $\Gamma=3\beta H\frac{(A+1)H^{2\alpha+2}-B/3^{\alpha+1}}{(A+1)H_{0}^{2\alpha+2}-B/3^{\alpha+1}}$
---------------------------------------------------------------------------------------------------------
As before we can obtain the following expressions for $\rho,~H,~q,~n~\text{and}~T$ corresponding to the above $\Gamma$ $$\begin{aligned}
\rho&=\left[\frac{\rho_{0}^{\alpha+1}}{r}+\frac{B}{A+1}\left(1-\frac{1}{r}\right)\right]^{1/(\alpha+1)}\label{rho3}\\
H&=\left(\frac{1}{A+1}\right)^{1/(2\alpha+2)}\left[\frac{y_{0}}{r}+\frac{B}{3^{\alpha+1}}\right]^{1/2\alpha+2}\label{h3}\\
q&=-1+\frac{3y_{0}(A+1)(r-\beta)}{2r(y_{0}+rB/3^{\alpha+1})}\label{q3}\\
n&=\frac{n_{0}}{r^{1/\mu}}\label{n3}\\
T&=T_{0}\left(\frac{\rho_{0}}{\rho}\right)^{\alpha}r^{-1+\frac{1}{\mu}}\label{t3}\end{aligned}$$ where $r=\beta+(1-\beta)\left(\frac{a}{a_{0}}\right)^{3\mu},~y_{0}=(A+1)H_{0}^{2\alpha+2}-B/3^{\alpha+1}$ and $H=H_{0}$ at $a=a_{0}.$
The above physical parameters are graphically presented in Fig 3 using the previous values for the parameters $A,~B,~\text{and}~\alpha,~\beta$ while $a_{0}=0.3.$
-- -- --
-- -- --
Data comparison {#data-analysis}
===============
The Hubble parameter gives the expansion rate of the universe, and its time dependence can be measured using the deceleration parameter $q(t).$ In order to understand the kinematics of the cosmological evolution we consider a Taylor series expansion of the scale factor about the present time $t_{0}$ is given by $$a(t)=a(t_{0})\left[1+H_{0}(t-t_{0})-\frac{1}{2!}q_{0}H_{0}^{2}(t-t_{0})^{2}+\frac{1}{3!}j_{0}H_{0}^{3}(t-t_{0})^{3}+\frac{1}{4!s_{0}}H_{0}^{4}(t-t_{0})^{4}+\frac{1}{5!}l_{0}H_{0}^{5}(t-t_{0})^{5}+O((t-t_{0})^{6})\right]$$ where $H_{0},~q_{0},~j_{0},~s_{0},~l_{0}$ are higher order derivatives of the scale factor considered at the present time and are more commonly known as the cosmographic Hubble, deceleration, jerk, snap and lerk functions respectively [@vis]. Considering that the scale factor $a(t)$ is related to the redshift $z$ by the relation $\frac{a(t_{0})}{a(t)}=1+z$ one can obtain the deceleration parameter $q(z)$ as a power series in $z$ as [@gui] $$q(z)=q_{0}+(-q_{0}-2q_{0}^{2}+j_{0})z+\frac{1}{2}(2q_{0}+8q_{0}^{2}+8q_{0}^{3}-7q_{0}j_{0}-4j_{0}-s_{0})z^{2}+O(z^{3})$$
[l\*[3]{}[c]{}|l\*[2]{}[c]{}]{}Data&$q_{0}$&$j_{0}$&$s_{0}$&Data&$q_{0}$&$j_{0}$\
Data1(192 SN 1a+GRB - CPL)&-0.90&3.93&-25.52&Data4(BAO data)&-0.764&1.774\
Data2(192 SN 1a+GRB - Linear)&-0.75&2.21&-12.25&Data5(Union2.1 SN1a+BAO+$H(z)$)&-0.48&0.68\
Data3(Union2 SN 1a+GRB +BAO+OHD)&-0.39&-4.925&-26.40&Data6(Union2.1 SN1a+BAO+GRB)&-0.6&0.7\
-- -- --
-- -- --
Using the above expression for $q$ we have plotted the deceleration parameter corresponding to six data sets in Fig 4. Data 1 corresponds to 192 SN 1A and 69 GRB’s with CPL parametrizations [@wang], Data 2 corresponds to the same data set with linear parametrizations [@wang] and Data 3 corresponds to Union2+BAO+OHD+GRB’s data [@xu] for low red shift range. Data 4 corresponds to most recent high red-shift BAO data [@laz], while Data 5 and Data 6 correspond to high red-shifts data from Union 2.1 compilation of SN 1a data with BAO and H(z) data, and SN 1a+GRB’s+BAO data respectively [@demia]. (It may be noted that for data sets 4, 5 and 6 only first order terms were used because $s_{0}$ values were not constrained well in the literatures cited). Parameters used in the models are $A=1/3,~B=3.4,~\alpha=0.5,~C=2,~\beta=0.1~\text{and}~m=-2.$
From the left figure of Fig. 4 we can see that our model fits well with both Data 2 and Data 3, while from the right figure of Fig. 4 it is evident that our model accurately matched the data 4, 5 and 6 in the region $.5\leq a\leq .8$ with third model providing the best match.
Field Theoretic Description {#sec-ftd}
============================
In this section we shall address the process from the field theoretic view point. With the help of Model 3 we shall show the whole dynamical process as a evolution of scalar field $\phi$ having self interacting potential $V(\phi).$ We know that in terms of the scalar field, the energy density and thermodynamic pressure of the cosmic fluid is given by: $$\rho=\frac{1}{2}\dot{\phi}^{2}+V(\phi)~~~~p_{tot}=p+\Pi=\frac{1}{2}\dot{\phi}^{2}-V(\phi).\label{phi}$$ Now using the isentropic condition (\[pie\]) together with the particle creation rate as used in Model 3 and the corresponding expression for Hubble parameter as given by (\[h3\]) we can obtain $$\phi=\phi_{s}+\frac{2}{\mu}\sqrt{\frac{\beta}{3y_{0}}}\frac{m}{\sqrt{1+m^{2}}}F\left[\sin^{-1}\sqrt{\frac{(A+1)H^{2\alpha+2}}{B/3^{\alpha+1}}}|\frac{m^{2}}{1+m^{2}}\right]$$ where $F$ is the incomplete Elliptic integral of the first kind [@abram] and $m=\sqrt{\frac{B/3^{\alpha+1}}{y_{0}/\beta}}.$ Here scalar field $\phi$ always has a value greater than $\phi_{s}.$ Consequently $$\left(\frac{a}{a_{0}}\right)^{3\mu}=\frac{y_{0}}{1-\beta}\left\lbrace\frac{B}{3^{\alpha+1}}\left[sn^{-1}\left(\frac{(\phi-\phi_{s})}{\frac{2}{\mu}\sqrt{\frac{\beta}{3y_{0}}}\frac{m}{\sqrt{1+m^{2}}}}|\frac{m^{2}}{1+m^{2}}\right)\right]^{2}-1\right\rbrace^{-1}-\frac{\beta}{1-\beta}$$ where $sn^{-1}$ is the equivalent inverse Jacobian Elliptic function. Now from (\[phi\]) we know $2V(\phi)=\rho-p-\Pi,$ substituting the values of $\rho,~p,~\Pi$ in terms of the Hubble parameter $H$ as obtained in model 3, we get: $$2V(\phi)=\nu H^{2\alpha}+\zeta H^{2}+\eta H^{2\alpha+4}.\label{vh}$$ where $$\begin{aligned}
\zeta&=\frac{B}{3^{\alpha}}\left(1-\frac{\beta}{y_{0}}\frac{B}{3^{\alpha+1}}\right)\\
\nu&=3\left(\frac{1-A}{1+A}-(2+A)\frac{\beta}{y_{0}}\frac{B}{3^{\alpha+1}}\right)\\
\eta&=\frac{3\beta}{y_{0}}(1+A)^{2}\end{aligned}$$ Assuming $\psi=\frac{B}{3^{\alpha+1}(A+1)}\left[sn^{-1}\left(\frac{(\phi-\phi_{s})}{\frac{2}{\mu}\sqrt{\frac{\beta}{3y_{0}}}\frac{m}{\sqrt{1+m^{2}}}}|\frac{m^{2}}{1+m^{2}}\right)\right]^{2}$ as the modified scalar field, the corresponding modified field potential $2V(\phi)=\mathfrak{V}(\psi)$ is given by: $$\mathfrak{V}(\psi)=\zeta\psi^{1-\delta}+\nu\psi^{\delta}+\eta\psi^{1+\delta}$$ where $\delta=\frac{1}{1+\alpha}.$ Thus, the potential is essentially a linear combination of different powers of the modified scalar field $\psi.$ Choosing $\alpha=0.5$ one can obtain $$\mathfrak{V}(\psi)=\zeta\psi^{1/3}+\nu\psi^{2/3}+\eta\psi^{5/3}.$$ From the above relations the particle creation rate $\Gamma$ takes the form $$\Gamma=\frac{3\beta}{y_{0}}\left[(A+1)\psi^{1+\frac{1}{2\alpha+2}}-\frac{B}{3^{\alpha+1}}\psi^{\frac{1}{2\alpha+2}}\right].$$ Hence we obtain a continuous description of the scalar field evolution and corresponding potential for different cosmological epochs, establishing the correlation between scalar field and fluid evolution.
For models 1 and 2 however, it is difficult to obtain explicit analytic expressions for $\phi$ and $V(\phi).$ They are thus solved by numerical methods and the corresponding scalar field evolution has been represented graphically in Figure 5. The leftmost figure corresponds to model 1, the middle to model 2 and the right most to model 3.
-- -- --
-- -- --
Entropy production {#sec-ent}
===================
Eliminating $\Pi$ from (\[entropy1\]) and using (\[consv1\]) and (\[pie\]) we obtain the rate of change of entropy per particle as $$nT\dot{s}=-\frac{\dot{n}}{n}(p+\rho)+\dot{\rho},$$ after some algebraic manipulations which gives $$TX^{\frac{\alpha}{\alpha+1}}\dot{s}=\frac{A+1}{\alpha+1}\left[\frac{dX}{dt}+n^{-1-\alpha}\frac{dY}{dt}\right]\label{sdot1}$$ where $$X=\left(\frac{\rho}{n}\right)^{\alpha+1}~\text{and}~Y=\frac{B(\alpha+1)}{A+1}\log n-\frac{A}{A+1}\rho^{\alpha+1}.\label{substs}$$ Also from (\[t\]) using (\[consv1\]) together with $\Pi=-\rho\left(\gamma+\frac{2}{3}\frac{\dot{H}}{H^{2}}\right)$ one can obtain $$T=T_{i}e^{[\int_{n_{i}}^{n}\frac{v-u\gamma}{n}dn+\int_{\rho_{i}}^{\rho}\frac{u}{\rho}d\rho]}\label{tval}$$ where $u$ and $v$ are the ratios, given by $u=\frac{\rho}{T\partial\rho/\partial T},~v=\frac{\partial p/\partial T}{\partial\rho/\partial T}$ and $\gamma=\frac{p+\rho}{\rho}.$ Subscript $i$ here refers to some initial time. Using this value of $T$ in (\[sdot1\]) one obtains $$\dot{s}=\left(\frac{n}{\rho}\right)^{\alpha}\frac{1}{T_{i}}e^{-[\int_{n_{i}}^{n}\frac{v-u\gamma}{n}dn+\int_{\rho_{i}}^{\rho}\frac{u}{\rho}d\rho]}\frac{A+1}{\alpha+1}\left[\frac{dX}{dt}+n^{-1-\alpha}\frac{dY}{dt}\right].\label{sdot2}$$ For the inflationary phase with $\rho=\rho_{i}$ and $H=H_{i}$ the above expression for $\dot{s}$ simplifies to $$\dot{s}=\frac{\gamma_{i}\rho_{i}}{nT_{i}}\left(\frac{n}{n_{i}}\right)^{-v_{i}+u_{i}\gamma_{i}}(3H_{i}-\Gamma)\label{sdot3}$$ where $\gamma_{i}=A+1-\frac{B}{\rho_{i}^{\alpha+1}},~u_{i}=u|_{\rho_{i}},~v_{i}=v|_{\rho_{i}}.$ Corresponding to the radiation era with high density, we choose $\gamma_{i}=\frac{4}{3}$ and $u_{i}=1,~v_{i}=\frac{1}{3}.$ Using these in (\[sdot3\]) one obtains $$\dot{s}=\frac{4\rho_{i}}{3n_{i}T_{i}}(3H_{i}-\Gamma).\label{sdot4}$$ Integrating the above we get $$s=\frac{4\rho_{i}}{n_{i}T_{i}}\left[(t-t_{i})H_{i}-\frac{1}{3}\int_{t_{i}}^{t}\Gamma dt\right]+s(t_{i}).\label{s}$$ From the above we find that $s$ has an additional dependence on the particle creation rate $\Gamma.$ Further one can find that the change of entropy in comoving volume given by $\mathcal{E}=sna^{3}$ using MCG as matter is similar to that obtained in [@zim2] with normal matter with bulk viscosity driven by particle creation. Using above result we can write an explicit expression for $\mathcal{E}$ as $$\mathcal{E}=\left\lbrace \frac{4\rho_{i}}{n_{i}T_{i}}\left[(t-t_{i})H_{i}-\frac{1}{3}\int_{t_{i}}^{t}\Gamma dt\right]+s(t_{i})\right\rbrace N_{i}e^{\int_{t_{i}}^{t}\Gamma dt}\label{v}$$ which clearly shows that for $\Gamma\neq 0$ there is an exponential increase of comoving entropy in MCG in the de Sitter phase. Thus we can safely commit that with MCG as cosmic fluid, the results above are in complete agreement to the results obtained in [@zim2] corresponding to the normal fluid, and further we assert that the viscous pressure $\Pi$ is connected to an increasing particle number rather than with changing entropy per particle.
Interrelation between particle creation with MCG and Hawking Radiation {#section-hwr}
======================================================================
In the previous sections it is found that the dissipative phenomenon in the cosmic substratum leads to non-equilibrium thermodynamics with particle creation mechanism. Further the dissipative effect is only in the form of effective bulk viscous pressure due to homogeneity and isotropic nature of space-time. The present section is an attempt to show some inner relationship between particle creation rate and Hawking temperature.
In the context of universal thermodynamics, the process of Hawking radiation is just the opposite to black hole (BH) evaporation. The particles created just outside the event horizon in the BH evaporation escape outside towards asymptotic infinity, but in universal thermodynamics the created particles near the trapping horizon move inside the horizon. Also this flow of particles will be uniform in all directions due to the isotropic nature of space-time. Further at the beginning, the BH evaporation process is very slow, subsequently with the decrease of the BH size, the process becomes faster and faster, so that the temperature becomes larger and larger until quantum gravity effects become important due to the plank size of the BH. In universal thermodynamics, on the other hand, the universe at the beginning is of the Planck size and quantum gravity effects are important, then gradually with the expansion of the universe Hawking radiation comes into effect and the temperature gradually decreases.
There is another basic difference between the two processes namely, due to Hawking radiation in the context of BH there is a loss of energy and hence it is termed as BH evaporation, while due to Hawking radiation in universal thermodynamics, the universe gains energy and hence it should be referred as Hawking condensation. It is reasonable to assume that the Hawking radiation follows the Stephen-Boltzman radiation law [@haw; @mod]. $$P=\frac{dQ}{dt}=\sigma A_{\tau}T^{4}\label{st}$$ where $\sigma=\frac{\pi^{2}\kappa_{B}^{2}}{60\hbar^{3}c^{2}}$ is Stephen-Boltzman constant, $A_{\tau}$ is the area of the bounding trapping horizon and $T=\frac{\hbar H}{2\pi\kappa_{B}}$ is the Hawking temperature. Using this heat variation in the first law of thermodynamics $$\dfrac{dQ}{dT}=\dfrac{d}{dt}(\rho V_{H})+p\dfrac{dV_{H}}{dt}=\sigma A_{H}T^{4} \label{flt}$$ one gets $$\dot{\rho}+3H(\rho+p)=3\sigma T^{4}\label{ec}$$ where $A_{H}$ and $V_{H}$ are respectively the area and volume bounded by the horizon. The non-zero r.h.s. of the above equation gives some dissipation due to Hawking radiation. Thus comparing with energy-momentum conservation relation (\[consv2\]), the effective bulk viscous pressure is given by $$\Pi=-\frac{\sigma T^{4}}{H}.$$ Further using the relation(\[pie\]) for isentropic process the particle creation rate is given by $$\Gamma=\frac{\sigma T^{4}}{H(1+A-B(3H^{2})^{-(\alpha+1)})}$$ Now using the value of $T,$ the Hawking temperature, as given above, we have $$\Gamma=\frac{\sigma\hbar^{4} H^{2\alpha+5}}{(2\pi\kappa_{B})^{4}\lbrace(1+A)H^{2\alpha+2}-\frac{B}{3^{\alpha+1}}\rbrace}.$$ Hence for large $H,~\Gamma\propto H^{3}.$ Since in the early phase of the evolution of the universe the MCG behaves as perfect fluid (with constant equation of state $p=A\rho$) we have $$H^{-2}=H_{0}^{-2}+\left(\frac{a}{a_{0}}\right)^{3(A+1)}$$ i.e. $H\sim H_{0}$ for $a\ll a_{0}$ while $H\sim a^{-\frac{3(1+A)}{2}}$ for $a\gg a_{0}.$ Thus one obtains the usual exponential expansion at the early phase and subsequently, the evolution follows the standard cosmology. On the other hand when $H$ is small (i.e. at the late phase of evolution) then instead of particle creation there will be particle annihilation and $\Gamma\propto H^{2\alpha+5}.$ This is not a physically realistic situation and hence Hawking type radiation is not possible at later phase of evolution.
Discussions {#discu}
===========
In the above sections MCG model is considered as a candidate for describing cosmic evolution with dissipation in the form of bulk viscosity due to the mechanism of particle creation. For three different choices for the particle creation rate (as a function of Hubble parameter and scale factor) it is possible to show a complete cosmic scenario from inflation to present late time acceleration. Also the thermodynamic parameters namely density, temperature and particle number density are presented both analytically and graphically. The deceleration for the three models are compared with different observed results and it is found that the present models match with observed results for different ranges of the red shift parameter (or the scale factor). One of the three models (Model 3) has been shown equivalent to a a scalar field, with self interacting potential description. Analytic expression for the scalar field and the potential function has also been evaluated. Considering the particle creation mechanism in the context of non equilibrium thermodynamic prescription, the entropy production in the cosmic volume has been evaluated for the MCG model. We could show that in the radiation era, corresponding to high density the comoving entropy will have an exponential increase that is proportional to the particle creation rate $\Gamma.$ Finally an attempt has been made to find an analogy of the present model in Hawking radiation. Thus we can conclude that with MCG as cosmic fluid it is possible to obtain a continuous cosmic scenario in the presence of bulk viscosity due to particle creation. Unlike the cosmic scenario in the presence of a perfect fluid, that requires different particle creation parameters for different cosmic era [@sc1], a cosmic scenario in the presence of MGC could be described by a single continuous choice of the particle create rate $\Gamma.$ It is only natural if one considers that MCG can as such accommodate an universe starting with inflationary scenario to the current accelerating one, while particle creation effects can explain both inflationary and late time acceleration of the universe. We also note that using equations (\[eos\]) and (\[rho\]) it is possible to evaluate the effective equation of state parameter $\omega_{eff}=\frac{\rho+p+\Pi}{\rho}$ for the models 1, 2 and 3 in terms of the scale factor $a.$ Defining constraints between the free parameters it is possible to obtain a qualitative behaviour of of $\omega_{eff}$ w.r.t. the scale factor. Figure 6 is the corresponding graphical representation. The figures show that at a very early time, our models can have an effective equation of state parameter $\omega_{eff}=-1$ thus describing an accelerated expansion, while at the present time one can obtain an effective equation of state having phantom attributes.
In conclusion we state that the present work successfully describes the cosmic evolution using a single continuous particle creation model in Modified Chaplygin gas, incorporating the early and late time accelerated expansion of the universe. It remains to be seen whether such continuous models can describe in details the inflationary cosmology successfully, and can be considered as a future work.
-- -- --
-- -- --
Acknowledgments
===============
SB acknowledges UGC’s Faculty Recharge Programme. SC thanks IUCAA, Pune, India, for their warm hospitality while working on this project.
I. Muller, [*Z. Physik*]{} [**198**]{}, 329 (1967).
W. Israel, [*Ann. Phys.*]{} [**100**]{}, 310 (1976).
W. Israel and J. M. Stewart [*Proc. R. Soc. Lond. A*]{} [**365**]{}, 43 (1979); [*Ann. Phys.*]{} [**118**]{}, 341 (1979)
W. A. Hiscock and L. Lindblom, [*Ann. Phys.*]{} [**151**]{}, 466 (1983).
D. Pavon, D. Jou and J. Casas-Vazquesz, [*Ann. Inst. H. Poincare Sec. A*]{} [**36**]{}, 79 (1982).
D. Pavon, J. Bafaluy and D. Jou, [*Class. Quant. Grav.*]{} [**8**]{}, 347 (1991).
Ya. B. Zel’dovich, [*Sov. Phys. JETP Lett.*]{} [**12**]{}, 307 (1970).
G. L. Murphy [*Phys. Rev. D*]{} [**8**]{}, 4231 (1973).
B. L. Hu, [*Phys. Lett.* ]{} [**A90**]{}, 375 (1982).
J. D. Barrow, [*Phys. Lett. B*]{} [**180**]{}, 335 (1986).
I. Progogine, J. Geheniau, E. Gunzig and P. Nardone, [*Proc. Natl. Acad. Sci. USA*]{} [**85**]{}, 7428 (1988); [*Gen. Relt. Grav.* ]{} [**21**]{}, 767 (1989).
M. O. Calvao, J. A. S. Lima, I. Waga, [*Phys. Lett.* ]{} [**A162**]{}, 223 (1992).
W. Zimdahl, [*Mon. Not. R. Astron. Soc*]{} [**280**]{}, 1239 (1996).
W. Zimdahl, D. Pavon, [*Phys. Lett.* ]{} [**A176**]{}, 67 (1993); [*Mon. Not. R. Astron. Soc*]{} [**266**]{}, 872 (1994); [*Gen. Relt. Grav.* ]{} [**26**]{}, 1259 (1994).
N. Udey and W. Israel, [*Mon. Not. R. Astron. Soc*]{} [**199**]{}, 1137 (1982).
W. Zimdahl, [*Mon. Not. R. Astron. Soc*]{} [**280**]{}, 1239 (1996).
S. Weinberg, [*ApJ*]{} [**168**]{}, 175 (1971).
N. Staumann, [*Helv. Phys. Acta*]{} [**49**]{}, 269 (1976).
M. A. Schweizer, [*ApJ.*]{} [**258**]{}, 798 (1982).
S. Chakraborty, S. Saha, [*Phys. Rev. D*]{} [**90**]{}, 123505 (2014).
S. Chakraborty, [*Phys. Lett. B*]{} [**732**]{}, 81 (2014).
U. Debnath, A. Banerjee, S. Chakraborty, [*Class. Quant. Grav.*]{} [**21**]{}, 5609 (2004).
H. B. Benaoum, arXiv: hep-th/0205140.
Y. Wu, S. Li, J. Lu, X. Yang, [*Mod. Phys. Lett. A* ]{} [**22**]{}, 783 (2007).
M. L. Bedran, V. Soares, M. E. Araujo, [*Phys. Lett. B*]{} [**659**]{}, 462 (2008).
S. S. e Costa, M. Ujevic, A. F. dos Santos, [*Gen. Relt. Grav.* ]{} [**40**]{}, 1683 (2008).
U. Debnath, S. Chakraborty, [*Int. J. Theor. Phys.*]{} [**47**]{}, 2663 (2008).
P. Thakur, S. Ghose, B. C. Paul, [*Mon. Not. R. Astron. Soc*]{} [**397**]{}, 1935 (2009).
R. C. Nunes, D. Pavon, [*Phys. Rev. D*]{} [**91**]{}, 063526 (2015).
E. Gunzig, R. Maartens, A. V. Nesteruk, [*Class. Quant. Grav.*]{} [**15**]{}, 923 (1998)
H. Saadat, B. Pourhassan, [*Astrophys. Space Sci.*]{} [**343**]{}, 783 (2013).
M. Visser, [*Class. Quant. Grav.*]{} [**21**]{}, 2603 (2004).
A. C. C. Guimaraes, J. A. S. Lima, [*Class. Quant. Grav.*]{} [**28**]{}, 125026 (2011).
F. Y. Wang, Z. G. Dai, S. Qi, [*Astron. Astrophys.*]{} [**507**]{}, 53 (2009).
L. Xu, Y. Wang, [*Phys. Lett. B*]{} [**702**]{}, 114 (2011).
R. Lazkoz, J. Alconiz, C. Escamilla-Rivera, V. Salzano, I. Sendra, [*J. Cosmol. Astropart. Phys.*]{} [**12**]{}, 005 (2013).
M. Demianski, E. Piedipalumbo, D. Sawant, L. Amati, [*Astron. Astrophys.*]{} [**598**]{}, A113 (2017).
S. W. Hawking, [*Commun. Math. Phys.*]{} [**43**]{}, 199 (1975).
S. K. Modak, D. Singleton, [*Phys. Rev. D*]{} [**86**]{}, 123515 (2012).
M. Abramowitz, I. A. Stegun, [*Handbook of Mathematical Functions with Formulas, Graphs and Mathematical Tables*]{} (National Bureau of Standards, Applied Mathematics Series, 1972).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Let $G$ be a reductive complex Lie group acting holomorphically on normal Stein spaces $X$ and $Y$, which are locally $G$-biholomorphic over a common categorical quotient $Q$. When is there a global $G$-biholomorphism $X\to Y$?
If the actions of $G$ on $X$ and $Y$ are what we, with justification, call generic, we prove that the obstruction to solving this local-to-global problem is topological and provide sufficient conditions for it to vanish. Our main tool is the equivariant version of Grauert’s Oka principle due to Heinzner and Kutzschebauch.
We prove that $X$ and $Y$ are $G$-biholomorphic if $X$ is $K$-contractible, where $K$ is a maximal compact subgroup of $G$, or if $X$ and $Y$ are smooth and there is a $G$-diffeomorphism $\psi:X\to Y$ over $Q$, which is holomorphic when restricted to each fibre of the quotient map $X\to Q$. We prove a similar theorem when $\psi$ is only a $G$-homeomorphism, but with an assumption about its action on $G$-finite functions. When $G$ is abelian, we obtain stronger theorems. Our results can be interpreted as instances of the Oka principle for sections of the sheaf of $G$-biholomorphisms from $X$ to $Y$ over $Q$. This sheaf can be badly singular, even for a low-dimensional representation of $\mathrm{SL}_2({{\mathbb C}})$.
Our work is in part motivated by the linearisation problem for actions on ${{\mathbb C}}^n$. It follows from one of our main results that a holomorphic $G$-action on ${{\mathbb C}}^n$, which is locally $G$-biholomorphic over a common quotient to a generic linear action, is linearisable.
address:
- 'Frank Kutzschebauch, Institute of Mathematics, University of Bern, Sidlerstrasse 5, CH-3012 Bern, Switzerland'
- 'Finnur Lárusson, School of Mathematical Sciences, University of Adelaide, Adelaide SA 5005, Australia'
- 'Gerald W. Schwarz, Department of Mathematics, Brandeis University, Waltham MA 02454-9110, USA'
author:
- 'Frank Kutzschebauch, Finnur Lárusson, Gerald W. Schwarz'
date: '20 March 2013. Most recent minor changes 14 August 2013'
title: An Oka principle for equivariant isomorphisms
---
Introduction {#s:introduction}
============
In renowned work from the late 1950s, Grauert showed that a holomorphic principal $G$-bundle over a Stein space, where $G$ is a complex Lie group, has a holomorphic section if it has a continuous section [@Grauert]. In fact, every continuous section can be deformed to a holomorphic section. This is an instance of the Oka principle, a metatheorem supported by many actual theorems, saying that on Stein spaces, there are only topological obstructions to solving holomorphic problems that can be cohomologically formulated. In a seminal paper of 1989, Gromov showed that the structure group is immaterial, so Grauert’s theorem holds for any holomorphic fibre bundle whose fibre is a complex Lie group [@Gromov]. And recently, Forstnerič has shown that Grauert’s theorem holds even more generally for sections of any holomorphic submersion over a Stein space with the structure of a stratified holomorphic fibre bundle with complex Lie groups as fibres [@Forstneric2010]. (We should say that we have not stated the theorems of Grauert, Gromov, and Forstnerič in their full strength.) For more information on the Oka principle, see the monograph [@Forstneric2011] and the survey [@Forstneric-Larusson].
In this paper, we prove Oka principles for sections of sheaves of groups that arise naturally in geometric invariant theory and that can be so singular, even for a low-dimensional representation of $\mathrm{SL}_2({{\mathbb C}})$ (Example \[e:canonical-bad-example\]), that they are not represented by a complex space over the base. The motivation for our study is to extend the Oka principle to singular bundles that lie beyond the reach of the theorems mentioned above, and at the same time to explore applications of the Oka principle in geometric invariant theory. Also, there is more specific motivation coming from the so-called linearisation problem, which we describe at the end of this section.
For more details on the following, see the next section. Let $G$ be a reductive complex Lie group. Let $X$ and $Y$ be normal Stein spaces (always taken to be connected) on which $G$ acts holomorphically. The categorical quotients $X{{\operatorname{/\!\!/}}}G$ and $Y{{\operatorname{/\!\!/}}}G$ are normal Stein spaces. Assume that there is a biholomorphism $\tau:X{{\operatorname{/\!\!/}}}G\to Y{{\operatorname{/\!\!/}}}G$ that locally lifts to $G$-equivariant biholomorphisms between $G$-saturated open subsets of $X$ and $Y$. We use $\tau$ to identify the quotients, and call the common quotient $Q$ with quotient maps $p:X\to Q$ and $r:Y\to Q$. Our assumption, then, is that there is an open cover $(U_i)_{i\in I}$ of $Q$ and $G$-equivariant biholomorphisms $\phi_i:p^{-1}(U_i)\to r^{-1}(U_i)$ over $U_i$ (meaning that $\phi_i$ descends to the identity map of $U_i$). We express the assumption by saying that $X$ and $Y$ are *locally $G$-biholomorphic over a common quotient*.
We want to conclude that there is a *global* $G$-biholomorphism $X\to Y$. If $X$ and $Y$ are what we, with justification, call generic (see below), we prove that the obstruction to solving our local-to-global problem is topological and provide sufficient conditions for it to vanish.
If $X$ is smooth, the common quotient $Q$ has a natural stratification, the *Luna stratification*. We call the corresponding stratified space the *Luna quotient*. There is a unique open stratum $Q_\mathrm{pr}$, the *principal stratum*, and we set $X_\mathrm{pr}=p^{-1}(Q_\mathrm{pr})$. If $X$ is only normal, then we still have a stratification by *isotropy type*. There is a unique open stratum $Q_\mathrm{pr}$ in this case also, and the two definitions coincide when $X$ is smooth. We say that $X$ (or the $G$-action on $X$) is *$2$-principal* if $X\setminus X_\mathrm{pr}$ has codimension at least 2 in $X$. If, in addition, $X_\mathrm{pr}$ consists of closed orbits with trivial stabiliser, we say that $X$ (or the $G$-action on $X$) is *generic*. In this case, the quotient map $X_\mathrm{pr}\to Q_\mathrm{pr}$ is a principal $G$-bundle [@Snow Cor. 5.5]. We justify the term generic in Remark \[r:justification\] below. Our assumptions on $X$ and $Y$ show that $X$ is generic if and only if $Y$ is generic.
\[r:assumptions\] If $X$ and $Y$ are smooth and locally $G$-biholomorphic over a common quotient (in particular, if $X$ and $Y$ are $G$-biholomorphic), then $X$ and $Y$ have isomorphic Luna quotients. In [@Schwarz2013b Ex. 4.4], there is an example of an automorphism of a Luna quotient of a generic $G$-module which does not lift over any neighbourhood of the image of the origin. Thus an isomorphism of Luna quotients need not lift to a $G$-biholomorphism, even locally. By slice theory, if the Luna quotients are isomorphic, then there are $G$-biholomorphisms $\phi_i:p^{-1}(U_i)\to r^{-1}(U_i)$ as above, except that $\phi_i$ need not descend to the identity map of $U_i$. *We do not know whether $X$ and $Y$ having isomorphic Luna quotients implies that they are locally $G$-biholomorphic over a common quotient.*
Let $\psi_{ij}$ be the $G$-biholomorphism defined as $\phi_i^{-1}\circ\phi_j$ on $p^{-1}(U_i\cap U_j)$. Then $(\psi_{ij})$ is a cocycle with respect to the open cover $(U_i)$ of $Q$ with coefficients in the sheaf ${{\mathscr A}}$ of groups of $G$-biholomorphisms of $X$ over $Q$. There is a $G$-biholomorphism $X\to Y$ over $Q$ if and only if the cocycle splits, so the obstruction to $X$ and $Y$ being $G$-biholomorphic over $Q$ is an element of the cohomology set $H^1(Q, {{\mathscr A}})$.
The important consequence of genericity is that the cocycle $(\psi_{ij})$ may be viewed as consisting of $G$-equivariant holomorphic maps $p^{-1}(U_i\cap U_j)\to G$, where $G$ acts on the target $G$ by conjugation (Lemma \[l:g-biholomorphisms\]). The cocycle thus defines a holomorphic principal bundle $E$ over $X$ with two commuting actions of $G$, one being part of the principal bundle structure, and the other making the projection $E\to X$ equivariant. The bundle $E$ with the two $G$-actions is holomorphically trivial if and only if the cocycle splits.
We are now able to apply our fundamental tool, the equivariant version of Grauert’s Oka principle due to Heinzner and Kutzschebauch [@Heinzner-Kutzschebauch], which implies that $E$ is holomorphically trivial if it is topologically trivial. We conclude that in the generic case, the obstruction to $X$ and $Y$ being $G$-biholomorphic is topological (Theorem \[t:topological-obstruction\]).
Using, among other things, the equivariant version of the theory of universal bundles over classifying spaces, we go on to prove the first main result of the paper (Theorems \[t:torus\] and \[t:K-contractible\]). Let $K$ be a maximal compact subgroup of $G$. We say that $X$ is $K$-contractible if the identity map of $X$ can be joined to a constant map by a continuous path of $K$-equivariant continuous maps $X\to X$. The value of the constant map is then a fixed point of the $K$-action on $X$, and hence of the $G$-action as well.
\[t:first-main\] Let $G$ be a reductive complex Lie group, and let $K$ be a maximal compact subgroup of $G$. Let $X$ and $Y$ be normal Stein spaces on which $G$ acts holomorphically and generically, such that $X$ and $Y$ are locally $G$-biholomorphic over a common quotient. If $X$ is $K$-contractible, then $X$ and $Y$ are $G$-biholomorphic. If $G$ is abelian and $X$ is smooth, it suffices that $X$ be $\mathbb Z$-acyclic.
Our second main result is the following (Corollary \[c:strict-diffeomorphisms\]).
\[t:second-main\] Let $G$ be a reductive complex Lie group acting holomorphically and generically on Stein manifolds $X$ and $Y$, which are locally $G$-biholomorphic over a common quotient $Q$. Suppose there is a $G$-equivariant diffeomorphism $X\to Y$ over $Q$, which is holomorphic when restricted to each fibre of the quotient map. Then $X$ and $Y$ are $G$-biholomorphic over $Q$.
In future work we hope to remove the genericity assumption, but this will require different methods.
We would like to interpret Theorems \[t:first-main\] and \[t:second-main\] as follows. Let $\mathscr I$ be the sheaf of sets of $G$-biholomorphisms from $X$ to $Y$ over $Q$. It is locally isomorphic to the sheaf ${{\mathscr A}}$, and is an ${{\mathscr A}}$-torsor, meaning that each stalk of ${{\mathscr A}}$ acts freely and transitively on the corresponding stalk of $\mathscr I$. Theorems \[t:first-main\] and \[t:second-main\] each provide a sufficient condition for $\mathscr I$ to have a holomorphic section. Theorem \[t:first-main\] says that it does if $X$ is $K$-contractible. Theorem \[t:second-main\] says that it does if it has a smooth section.
We conjecture that $\mathscr I$ has a holomorphic section if it has a continuous section, but this we cannot prove in general. We offer two partial results in this direction. The first (Theorem \[t:codim-3\]) says that if $X$ is smooth and $G$ is abelian and the complement of the principal stratum $Q_\mathrm{pr}$ has codimension at least 3 in $Q$, then $\mathscr I$ has a holomorphic section over $Q$ if it has a continuous section merely over $Q_\mathrm{pr}$.
The second result (Theorem \[t:strongly\]) replaces the diffeomorphism in Theorem \[t:second-main\] by a homeomorphism satisfying a technical condition described at the beginning of Section \[s:continuous\]. Thus, if $\mathscr I$ has a continuous section satisfying this condition, then $\mathscr I$ has a holomorphic section.
We remark that we do not prove Theorem \[t:second-main\] and Theorem \[t:strongly\] by deforming the diffeomorphism or homeomorphism in question to a biholomorphism. Rather, we deform it to a diffeomorphism or homeomorphism of a kind that we call *special*, the existence of which implies the existence of a biholomorphism by the equivariant Oka principle of Heinzner and Kutzschebauch.
We conclude this section with a few words about the linearisation problem. The groups of holomorphic and algebraic automorphisms of ${{\mathbb C}}^n$ are infinite-dimensional and quite mysterious for $n\geq 2$ in the holomorphic case and $n\geq 3$ in the algebraic case. It is of interest to study complex Lie subgroups of these groups, up to conjugacy. The problem of linearising actions of reductive groups on ${{\mathbb C}}^n$ has attracted much attention both in the algebraic and holomorphic settings ([@Huckleberry], [@Kraft1996]).
**Linearisation problem.** Let a reductive complex Lie group $G$ act algebraically (resp. holomorphically) on ${{\mathbb C}}^n$. Is it conjugate inside the group of algebraic (resp. holomorphic) automorphisms of ${{\mathbb C}}^n$ to a subgroup of $\mathrm{GL}_n({{\mathbb C}})$? In other words, is there an algebraic (resp. holomorphic) change of coordinates on ${{\mathbb C}}^n$ that makes $G$ act by linear transformations?
The first counterexamples for the algebraic linearisation problem were constructed by Schwarz [@Schwarz1989] for $n\geq 4$. His examples are holomorphically linearisable. The holomorphic linearisation problem for ${{\mathbb C}}^*$-actions on ${{\mathbb C}}^n$, $n\geq 4$, was solved in the negative by Derksen and Kutzschebauch [@Derksen-Kutzschebauch], who constructed holomorphic ${{\mathbb C}}^*$-actions on ${{\mathbb C}}^n$ whose stratified categorical quotients are not isomorphic to the stratified quotient of any linear action. For $n\geq 5$, there are even holomorphically parametrised families of mutually holomorphically inequivalent holomorphic ${{\mathbb C}}^*$-actions on ${{\mathbb C}}^n$ [@Kutzschebauch-Lodin].
Our contribution to the linearisation problem is the following consequence of Theorem \[t:first-main\] (Corollaries \[c:first-corollary\] and \[c:Franks-dream\]).
Let $X$ be a Stein manifold on which a reductive complex Lie group $G$ acts holomorphically. If $X$ is locally $G$-biholomorphic over a common quotient to a generic $G$-module $V$, then $X$ is $G$-biholomorphic to $V$.
Note that the theorem gives a sufficient condition for a Stein manifold to be biholomorphic to ${{\mathbb C}}^n$.
*Acknowledgement.* We thank G. Tomassini for pointing us to the reference [@Banica-Stanasila], which helped us complete the proof of Theorem \[t:codim-3\]. We also thank the referees for valuable comments that helped us improve the exposition.
Generic actions {#s:generic}
===============
We start with some background. For more information, see [@Luna] and [@Snow Sec. 6]. Let $X$ be a normal Stein space with a holomorphic action of a reductive complex Lie group $G$. The categorical quotient $Q=X{{\operatorname{/\!\!/}}}G$ of $X$ by the action of $G$ is the set of closed orbits in $X$ with a reduced Stein structure that makes the quotient map $p:X\to Q$ the universal $G$-invariant holomorphic map from $X$ to a Stein space. Since $X$ is normal, $Q$ is normal. If $U$ is an open subset of $Q$, then ${{\mathscr O}}_X(p^{-1}(U))^G \cong {{\mathscr O}}_Q(U)$. If $X$ is a $G$-module, then $Q$ is just the complex space corresponding to the affine algebraic variety with coordinate ring ${{\mathscr O}}_\mathrm{alg}(X)^G$. We say that a subset of $X$ is *$G$-saturated* if it is a union of fibres of $p$. If $Gx$ is a closed orbit, then the stabiliser (or isotropy group) $G_x$ is reductive. We say that closed orbits $Gx$ and $G{x'}$ have the same *isotropy type* if $G_x$ is $G$-conjugate to $G_{x'}$. Thus we get the *isotropy type stratification* of $Q$ with strata whose labels are conjugacy classes of reductive subgroups of $G$.
Let $H$ be a reductive subgroup of $G$ and let $Z$ be a normal Stein $H$-space. Define $G\times^H Z$ to be the orbit space of the free $H$-action on $G \times Z$ given by $h\cdot (g,z) = (g h^{-1} , h z)$. We denote the $H$-orbit of $(g,z)$ by $[g,z]$. Since $G\times Z$ is Stein and normal, so is $G\times^H Z$. Note that $G$ acts naturally on $G\times^HZ$ on the left.
Let $q\in Q$ and take a point $x$ in the unique closed $G$-orbit in $p^{-1}(q)$. Set $H=G_x$. The *slice theorem* states that there is a locally closed $H$-saturated Stein subvariety $S$ of $X$ (the *slice*) containing $x$ such that $G\times^H S\to X$, $[g,s]\mapsto gs$, is a $G$-biholomorphism onto a $G$-saturated neighbourhood of $x$ in $X$. If $x$ is a smooth point of $X$, the $H$-module $T_x X / T_x(Gx)$ is called the *slice representation* at $x$. In this case, we may assume that the slice $S$ is $H$-biholomorphic to an $H$-saturated neighbourhood of the origin $0$ in $T_x X / T_x(Gx)$, with $0$ corresponding to $x$. Since $H$ is reductive, we can identify $T_x X / T_x(Gx)$ with an $H$-stable complement $W$ to $T_x(Gx)$ in $T_x X$. Write the $H$-module $W$ as $W^H \oplus W'$. We may choose the slice $S$ to be $H$-biholomorphic to $B_1\times B_2$, where $0\in B_1\subset W^H$ and $0\in B_2\subset W'$. Then $B_1$ maps biholomorphically onto a neighbourhood $U$ of $p(x)$ in the stratum through $p(x)$. Thus $x$ has a $G$-saturated neighbourhood biholomorphic to $U\times(G\times^H B_2)$, and $p:p^{-1}(U)\to U$ is a trivial bundle with fibre $G\times^H {{\mathscr N}}(W')$, where ${{\mathscr N}}(W')$ is the null cone of $W'$, that is, the union of the $H$-orbits whose closure contains $0$.
Assume that $X$ is smooth and that $G_x=gG_{x'}g^{-1}=G_{gx'}$, where $Gx$ and $Gx'$ are closed. We say that $Gx$ and $Gx'$ have the same *slice type* if the slice representations of $G_x$ at $x$ and $gx'$ are isomorphic. Declaring points in $Q$ to be equivalent if the closed orbits above them have the same slice type defines a holomorphic stratification of $Q$, called the *Luna stratification*. The strata are smooth locally closed subvarieties of $Q$. We call $Q$, viewed as a stratified complex space with each stratum labelled by the isomorphism class of the corresponding slice representation, the *Luna quotient* of $X$ by the action of $G$. An isomorphism of Luna quotients is a biholomorphism respecting the additional structure. Precisely one Luna stratum is open in $Q$: this is the *principal stratum* $Q_\mathrm{pr}$. The corresponding stabiliser, well defined up to conjugation in $G$, is called the *principal stabiliser*. The closed orbits above the principal stratum are called *principal orbits*. The stratification of $Q$ by isotropy type is coarser than the Luna stratification, but the connected components of the strata in the two stratifications are the same. Both stratifications are locally finite, meaning that each point of $Q$ has a neighbourhood that intersects only finitely many strata. The isotropy type stratification exists even if $X$ is not smooth, and then there is still an open and dense principal stratum $Q_\mathrm{pr}$.
Let $X$ and $Y$ be normal Stein spaces (always assumed connected) on which a reductive complex Lie group $G$ acts holomorphically, such that $X$ and $Y$ are locally $G$-biholomorphic over a common quotient $Q$. Call the quotient maps $p:X\to Q$ and $r:Y\to Q$. As in the introduction, there is an open cover $(U_i)$ of $Q$ and a cocycle $(\psi_{ij})$ with coefficients in the sheaf ${{\mathscr A}}$ of groups of $G$-biholomorphisms of $X$ over $Q$. The obstruction to $X$ and $Y$ being $G$-biholomorphic is the corresponding element in $H^1(Q,{{\mathscr A}})$.
\[e:obstruction\] Let $Q$ be a Stein manifold with $H^2(Q,{{\mathbb Z}})\neq 0$, and let $L$ and $M$ be non-isomorphic holomorphic line bundles on $Q$. Let $X$ and $Y$ be the manifolds obtained from $L$ and $M$, respectively, by removing the zero sections. Vector bundles over Stein manifolds are Stein, and the complement of a hypersurface in a Stein manifold is Stein, so $X$ and $Y$ are Stein. The actions of ${{\mathbb C}}^*$ on $L$ and $M$ by scalar multiplication in each fibre restrict to actions on $X$ and $Y$. The actions are obviously free and generic. The categorical quotient of both $X$ and $Y$ is $Q$ with a trivial stratification. Clearly, $X$ and $Y$ are locally $G$-biholomorphic over $Q$. It is easily seen that a ${{\mathbb C}}^*$-equivariant biholomorphism $X\to Y$ over $Q$ would extend to an isomorphism $L\to M$. Conversely, an isomorphism $L\to M$ restricts to a ${{\mathbb C}}^*$-equivariant biholomorphism $X\to Y$ over $Q$. Here, the sheaf ${{\mathscr A}}$ of $G$-biholomorphisms of $X$ over $Q$ is simply the sheaf ${{\mathscr O}}^*$ of nowhere-vanishing holomorphic functions. Indeed, the obstruction to $X$ and $Y$ being $G$-biholomorphic over $Q$ is given by the element of $H^1(Q,{{\mathscr A}})\cong H^2(Q,{{\mathbb Z}})$ represented by $L\otimes M^*$.
So as not to break the flow of this section, we have postponed to the next section an example showing that the sheaf ${{\mathscr A}}$ can be badly behaved in that it need not be represented by a complex space.
In the introduction, we defined what it means for the action of $G$ on a normal Stein space $X$ to be generic and 2-principal. The action of $G$ on $X$ is said to be *stable* if there is a nonempty open subset of $X$ consisting of closed orbits; equivalently, $X_\mathrm{pr}=p^{-1}(Q_\mathrm{pr})$ consists of closed orbits. We can reduce the stable 2-principal case to the generic case as follows.
Let $X$ be stable and 2-principal and let $H$ be a principal stabiliser. Let $X_H$ be the union of the irreducible components of the $H$-fixed point set $X^H$ which intersect $X_\mathrm{pr}$. Then there is a $G$-equivariant biholomorphism $$\phi:G\times^{N_G(H)} X_H \to X,\quad [g,x]\mapsto gx.$$ Moreover, $X_H$ is normal and the action of $N_G(H)/H$ on $X_H$ is generic.
Here, $N_G(H)$ denotes the normaliser of $H$ in $G$. Without the assumption that $X$ is stable and 2-principal, $X{{\operatorname{/\!\!/}}}G$ is biholomorphic to $X_H{{\operatorname{/\!\!/}}}N_G(H)$ [@Luna-Richardson]. See also [@Schwarz1995 Thm. 7.5].
Clearly $X_H$ is $N_G(H)$-stable. Set $X'=G\times^{N_G(H)} X_H$. Suppose that $x\in X_H\cap X_\mathrm{pr}$ and $gx \in X_H$. Then the stabiliser $G_{gx}$ contains $H$ and is conjugate to $H$, so $G_{gx}=H$. But $G_{gx}=gG_xg^{-1}=gHg^{-1}$, so $g\in N_G(H)$. Hence every principal orbit intersects $X_H$ in an $N_G(H)$-orbit with stabiliser $H$, and $$Gx= G\times^{N_G(H)}(Gx\cap X_H).$$ By construction then, $\phi$ induces a biholomorphism $\phi_\mathrm{pr}: X'_\mathrm{pr} \to X_\mathrm{pr}$, and $X'_\mathrm{pr}$ is open and dense in $X'$ since $X_\mathrm{pr}\cap X_H$ is dense in $X_H$. Since ${{\operatorname{codim}\,}}X\setminus X_\mathrm{pr}\geq 2$, the inverse of $\phi_\mathrm{pr}$ extends to be holomorphic on $X$. Hence $\phi$ is a biholomorphism. Observe that the Luna quotient of $X_H$ by $N_G(H)/H$ is determined by that of $X$ by $G$, and vice versa. Since $X'$ is a bundle over $G/N_G(H)$, if ${{\operatorname{codim}\,}}X_H\setminus (X_H)_\mathrm{pr}<2$, then ${{\operatorname{codim}\,}}X'\setminus (X')_\mathrm{pr}<2$: a contradiction. Hence ${{\operatorname{codim}\,}}X_H\setminus (X_H)_\mathrm{pr} \geq 2$. Finally, $X_H$ is normal since $X'\cong X$ is normal.
Let $X$ and $Y$ be locally $G$-biholomorphic over a common quotient and let $H$ be a principal stabiliser. Suppose that $X$ is stable and 2-principal. Then $Y$ is stable and 2-principal. Let $X'$, $Y'$, $X_H$, and $Y_H$ be as above. Then $X$ is $G$-biholomorphic to $Y$ if and only if $X'$ is $G$-biholomorphic to $Y'$ if and only if $X_H$ is $N_G(H)/H$-biholomorphic to $Y_H$.
Using the results above, one can prove versions of our main theorems with the hypothesis of genericity replaced by the assumption that the actions are 2-principal and stable. We leave this as an exercise for the reader.
\[r:justification\] Our use of the term *generic* is rigorously justified in the case of $G$-modules. If $G$ is simple, then, up to isomorphism, all but finitely many $G$-modules $V$ with $V^G=0$ are 2-principal and stable [@Schwarz1995 Cor. 11.6 (1)]. There is a similar result for semisimple groups [@Schwarz1995 Cor. 11.6 (2)]. If a $G$-module $V$ is 2-principal and stable, then the principal stabiliser $H$ is the kernel of the action of $G$ on $V$, so by replacing $G$ by $G/H$, we may assume that the principal stabiliser is trivial [@Schwarz2013 Rem. 2.6]. A random ${{\mathbb C}}^*$-module is generic, although infinite families of counterexamples exist. More precisely, a faithful $n$-dimensional ${{\mathbb C}}^*$-representation without zero weights is generic if and only if it has at least two positive weights and at least two negative weights and any $n-1$ weights are coprime. As justification for our use of the term *generic* for arbitrary Stein $G$-manifolds $X$, we note that $X$ is generic if and only if each of its slice representations is generic.
For us, the important consequence of $X$ being generic is that sections of ${{\mathscr A}}$ over an open subset $U$ of $Q$ may be identified with $G$-equivariant holomorphic maps $p^{-1}(U)\to G$, where $G$ acts on the target $G$ by conjugation: $h\cdot g=hgh^{-1}$.
\[l:g-biholomorphisms\] Let $X$ be a normal Stein space with a generic holomorphic action of a reductive complex Lie group $G$. Let $p:X\to Q$ be the categorical quotient map. Let $U$ be an open subset of $Q$. There is a bijective correspondence between $G$-equivariant biholomorphisms $\psi$ of $p^{-1}(U)$ over $U$ and $G$-equivariant holomorphic maps $\gamma:p^{-1}(U)\to G$, where $G$ acts on itself by conjugation, given by $\psi(x)=\gamma(x)x$.
Clearly, every $G$-equivariant holomorphic map $\gamma:p^{-1}(U)\to G$ induces a $G$-biholomorphism of $p^{-1}(U)$ over $U$ by the formula $x\mapsto \gamma(x) x$. We need to show that every $G$-biholomorphism is induced by a unique such map.
The assumption that the action is generic implies that holomorphic functions, and therefore also holomorphic maps into Stein spaces, extend uniquely from $X_\mathrm{pr}$ to $X$. Hence we may assume that $Q_\mathrm{pr}=Q$, so $p:X\to Q$ is a principal $G$-bundle.
Now if $\psi:p^{-1}(U)\to p^{-1}(U)$ is a $G$-biholomorphism over $U$, then, since the action on each fibre is free and transitive, there is a unique map $\gamma:p^{-1}(U)\to G$ with $\psi(x)=\gamma(x)x$ for all $x\in p^{-1}(U)$, and, since the action is free, $$(\gamma(gx)g)x=\psi(gx)= g\psi(x)=g\gamma(x)x$$ implies that $\gamma$ is $G$-equivariant. To show that $\gamma$ is holomorphic, we holomorphically and $G$-equivariantly trivialise $p$ over a small open subset $V$ of $U$, making $p^{-1}(V)$ isomorphic to $G\times V$ such that $G$ acts by left multiplication in the first component. Then $\gamma(g,y)=(\mathrm{pr}_1\circ\psi(g,y))g^{-1}$, so $\gamma$ is holomorphic.
We may now view the cocycle $(\psi_{ij})$ as consisting of $G$-equivariant holomorphic maps $p^{-1}(U_{ij})\to G$. Ignoring the $G$-equivariance for the moment, the cocycle defines a holomorphic principal bundle $E$ over $X$. The total space of $E$ is obtained from the disjoint union of $p^{-1}(U_i)\times G$, $i\in I$, by identifying $(x,g)\in p^{-1}(U_i)\times G$ with $(x,\psi_{ji}(x)g)\in p^{-1}(U_j)\times G$ when $x\in p^{-1}(U_{ij})$. The action of $G$ on the fibres is by right multiplication.
We define a holomorphic $G$-action on $E$ by setting $h\cdot(x,g)=(hx, hg)$ for $(x,g)\in p^{-1}(U_i)\times G$ and $h\in G$. The action is well defined since $$(hx,h\psi_{ji}(x)g) = (hx, \psi_{ji}(hx)hg)$$ is identified with $(hx,hg)$ by the $G$-equivariance of the cocycle. The action commutes with the action of $G$ on the fibres (left and right multiplications commute), and the projection $E\to X$ is $G$-equivariant. Thus, in the language of Lashof [@Lashof], which we shall use below, $E$ is a holomorphic principal $G$-$G$-bundle over $X$, where the first $G$ acts on $E$ by a lift of its action on $X$, and the second $G$ acts on each fibre of $E$. (In the language of [@Heinzner-Kutzschebauch], $E$ is a holomorphic $G$-principal $G$-bundle, where the first $G$ acts on each fibre and the second by a lift of the action on the base.)
Recall that $X$ and $Y$ are $G$-biholomorphic over $Q$ if and only if the cocycle $(\psi_{ij})$ splits (as a cocycle of $G$-equivariant holomorphic maps to $G$). Equivalently, $E$ is trivial as a holomorphic principal $G$-$G$-bundle, meaning that it is isomorphic as a holomorphic $G$-$G$-bundle to the trivial principal $G$-$G$-bundle $X\times G$ with the action $(h,h')\cdot(x,g)=(h x, h g h')$. By the equivariant Oka-Grauert principle of Heinzner and Kutzschebauch [@Heinzner-Kutzschebauch p. 341], this holds if (and of course only if) there is a topological $K$-equivariant isomorphism from $E$ to the trivial principal $G$-$G$-bundle, that is, if $E$ is trivial as a topological principal $K$-$G$-bundle. Here, $K$ denotes a maximal compact subgroup of $G$. (In our situation, the complexification $X^{{\mathbb C}}$ discussed in [@Heinzner-Kutzschebauch] is $X$ itself.)
Equivalently, the cocycle $(\psi_{ij})$ splits as a cocycle of $K$-equivariant continuous maps to $G$. In other words, there is a $K$-equivariant homeomorphism $\sigma:X\to Y$ of a certain form. Namely, on $p^{-1}(U_i)$, we have $\sigma(x)=\phi_i(\gamma_i(x)x)$, where $\gamma_i:p^{-1}(U_i)\to G$ is continuous and $K$-equivariant.
In particular, we have proved the following result.
\[t:topological-obstruction\] Let $G$ be a reductive complex Lie group acting holomorphically and generically on normal Stein spaces $X$ and $Y$, which are locally $G$-biholomorphic over a common quotient. The obstruction to $X$ and $Y$ being $G$-biholomorphic is topological. Namely, there is a bundle naturally arising from the given data whose topological triviality is equivalent to $X$ and $Y$ being $G$-biholomorphic.
In the following sections, we will provide sufficient conditions for the obstruction to vanish, starting with the case when $G$ is abelian.
To close this section, we remark that to conclude that $E$ is trivial as a holomorphic principal $G$-$G$-bundle over $X$, it suffices to know that $E$ is trivial as a topological principal $K$-$G$-bundle over a Kempf-Ness set $R$ in $X$ [@Heinzner-Kutzschebauch p. 341]. In other words, it suffices to split the cocycle $(\psi_{ij})$ on $R$.
We also remark, in the context of $G$ versus $K$, that if $X={{\mathbb C}}^n$ or, more generally, $X$ has no nonconstant plurisubharmonic functions that are bounded above, then a holomorphic action of $G$ on $X$ is the same thing as an action (continuous or, equivalently, real analytic) of $K$ on $X$ by biholomorphisms [@Kutzschebauch].
An example {#s:example}
==========
Kraft and Schwarz have shown that if $G$ is a reductive complex Lie group and $X$ is an affine $G$-variety such the categorical quotient map $X\to X{{\operatorname{/\!\!/}}}G$ is flat (this is a stringent assumption), then the functor associating to a morphism $Z\to X{{\operatorname{/\!\!/}}}G$ the group of $G$-automorphisms of the pullback $Z\times_{X{{\operatorname{/\!\!/}}}G} X$ over $Z$ is represented by an affine group scheme over the quotient [@Kraft-Schwarz Prop. III.2.2]. The following example shows that this may fail when the quotient map is not flat, even for a low-dimensional representation of $\mathrm{SL}_2({{\mathbb C}})$.
\[e:canonical-bad-example\] We let $G=\mathrm{SL}_2({{\mathbb C}})$ and consider the $G$-module $V={{\mathbb C}}^2\oplus{{\mathbb C}}^2\oplus{{\mathbb C}}^2\cong{{\mathbb C}}^6$ with the action $g\cdot(v_1,v_2,v_3)=(gv_1,gv_2,gv_3)$. This is a well-studied action: see e.g. [@Weyl], [@DeConcini-Procesi], and [@Kraft1984 Sec. I.4].
The categorical quotient map is $\pi:V\to Q={{\mathbb C}}^3$, $(v_1,v_2,v_3)\mapsto (f_3,f_2,f_1)$, where $f_1=\det [v_2\ v_3]$, $f_2=\det[v_1\ v_3]$, $f_3=\det[v_1\ v_2]$ [@DeConcini-Procesi Thm. 6.6]. If $\pi(v_1,v_2,v_3)\neq 0$, that is, the three vectors span ${{\mathbb C}}^2$, then the $\pi$-fibre through $(v_1,v_2,v_3)$ is simply the orbit through $(v_1,v_2,v_3)$, and is isomorphic to $G$ with trivial stabiliser. In particular, the action is generic with principal stratum $Q^*=Q\setminus\{0\}$. The null cone $N=\pi^{-1}(0,0,0)$, which consists of triples of vectors $(v_1,v_2,v_3)$ that span a line or are all zero, is a vector bundle of rank $2$ over $\mathbb P^2$ with the zero section blown down to a point [@Kraft1984 p. 28]. The point corresponds to the triple $(0,0,0)$ and is the unique closed orbit in $N$. The non-closed orbits in $N$ are the fibres of the vector bundle with zero removed.
It is clear that the group of $G$-automorphisms of each principal fibre is $G$ itself. In fact, over the principal stratum $Q^*$, since $G$ acts freely, $\pi$ is a principal $G$-bundle ([@Luna Cor. 5], [@Snow Cor. 5.5]). Thus, over $Q^*$, the sheaf ${{\mathscr A}}$ of $G$-biholomorphisms of $V$ over $Q$ is the sheaf of sections of a holomorphic principal $G$-bundle.
There is another useful action on $V$. Viewing $V$ as ${{\mathbb C}}^2\otimes{{\mathbb C}}^3$, we see that there is an action of $H=\mathrm{GL}_3({{\mathbb C}})$ on $V$, commuting with the $G$-action. Since the actions commute, the $H$-action descends to $Q$, and $\pi:V\to Q$ is $H$-equivariant. Also, $H$ acts by conjugation on the $G$-automorphisms of $V$ over $Q$. Clearly, $H$ preserves the null cone $N$, so each element of $H$ induces a $G$-automorphism of $N$.
It is easily seen that $H$ injects into ${{\operatorname{Aut}}}^G(N)$. In fact, ${{\operatorname{Aut}}}^G(N)=H$ (this is not required for our arguments below). Namely, the occurrences of the representation ${{\mathbb C}}^2$ in ${{\mathscr O}}_\mathrm{alg}(N)$ are spanned by the three copies of ${{\mathbb C}}^2$ in degree 1, and these copies of ${{\mathbb C}}^2$ generate ${{\mathscr O}}_\mathrm{alg}(N)$. Thus an element of ${{\operatorname{Aut}}}^G(N)$ permutes the copies of ${{\mathbb C}}^2$ linearly and corresponds to an element of $H$.
We will show that the sheaf ${{\mathscr A}}$ is not representable, in the sense that there is no group object $\alpha:A\to Q$ in the category of complex spaces over $Q$ representing the functor that takes a holomorphic map $f:Y\to Q$ of reduced complex spaces to the group of $G$-automorphisms of $Y\times_Q V$ over $Y$, meaning that there is a natural group isomorphism between the group ${{\operatorname{Aut}}}_Y^G(Y\times_Q V)$ and the group ${{\operatorname{Hom}}}(f,\alpha)$ of all holomorphic maps $g:Y\to A$ with $\alpha\circ g=f$, which is naturally identified with the group of holomorphic sections of $Y\times_Q A$ over $Y$. Informally speaking, ${{\mathscr A}}$ has a bad singularity over $0$. Since $Y$ is reduced, ${{\operatorname{Hom}}}(f,\alpha)$ is naturally identified with ${{\operatorname{Hom}}}(f,\tilde\alpha)$, where $\tilde\alpha:\tilde A\to Q$ is the reduction of $\alpha$. Thus we may assume that $A$ is reduced.
Suppose such a representing $\alpha$ exists. We will derive a contradiction. First, letting $Y$ run through the points of $Q$, we see that the fibres of $\alpha$ over $Q^*$ are $G$. The fibre over $0$ is ${{\operatorname{Aut}}}^G(N)\supset H$. Over $Q^*$, $A$ is connected (since $G$ is connected) and 6-dimensional, so $\alpha^{-1}(Q^*)$ lies in a 6-dimensional irreducible component $C$ of $A$. Clearly, $H\cap C\neq C$, so $H\cap C$ is a closed subgroup of $H$ of dimension at most 5.
We want to know that $H\cap C$ is a *normal* subgroup of $H$, because this narrows it down drastically. As a consequence of the Lie algebra $\mathfrak{sl}_3$ being simple, a normal subgroup of $H$ either consists of scalar matrices or contains $\mathrm{SL}_3({{\mathbb C}})$. If $H\cap C$ is normal, since it is at most 5-dimensional, it must consist of scalar matrices, which is easily contradicted. Indeed, take $f$ in the universal property to be the inclusion of the line $Y=\{(t,0,0):t\in{{\mathbb C}}\}$ into $Q$. For $t\neq 0$, $$\pi^{-1}(t,0,0)=\{(v_1, v_2, 0)\in V : \det[v_1\ v_2]=t\} \cong G,$$ and of course $\pi^{-1}(0,0,0)=N$. These are the fibres of $Y\times_Q V\to Y$. There is a $G$-automorphism of $Y\times_Q V$ over $Y$ given by $(v_1,v_2,v_3)\mapsto (2v_1, \tfrac 1 2 v_2, v_3)$. Its restriction to $N$ is given by a non-scalar element of $H$. (Alternatively, if we know that ${{\operatorname{Aut}}}^G(N)=H$, we can simply observe that the fibre dimension of $\alpha\vert C$ cannot drop from 3 down to 0 or 1 over $0\in Q$.)
To see that $H\cap C$ is normal, we verify that the action of $H$ by conjugation is built into $A$ via its universal property. Let $h\in H$ and take $f$ in the universal property above to be $A \xrightarrow{\alpha} Q \xrightarrow{h} Q$. Then we have a natural bijection between ${{\operatorname{Hom}}}(f,\alpha)$, which is the set of liftings of $h$ to $A$, and ${{\operatorname{Aut}}}_A^G(A\times_Q V)$. The desired action of $h$ on $A$ is the lifting corresponding to the $G$-automorphism $$(\phi,v)\mapsto(\phi, (h\phi h^{-1})(v))$$ of $A\times_Q V$ over $A$.
Abelian reductive groups {#s:torus}
========================
We use the notation established at the beginning of Section \[s:generic\]. In this section, we take $G$ to be abelian. Then the conjugation action of $G$ on itself is trivial, so a $G$-equivariant map $p^{-1}(U)\to G$, where $U\subset Q$ is open, is simply a $G$-invariant map. Thus, by the universal property of the categorical quotient, ${{\mathscr A}}$ may be identified with the sheaf of holomorphic maps from open subsets of $Q$ into $G$.
If $G$ is a torus $({{\mathbb C}}^*)^k$, $k\geq 1$, then $$H^1(Q,{{\mathscr A}})\cong H^1(Q,{{\mathscr O}}^*)^k\cong H^2(Q,{{\mathbb Z}})^k$$ since $Q$ is Stein. The second isomorphism is Oka’s original Oka principle! The following proposition shows that $H^2(Q,{{\mathbb Z}})$ vanishes if $X$ is smooth and ${{\mathbb Z}}$-acyclic.
\[p:acyclic\] Let $G$ be a reductive complex Lie group acting holomorphically on a $\mathbb Z$-acyclic Stein manifold $X$. Then $X{{\operatorname{/\!\!/}}}G$ is $\mathbb Z$-acyclic.
First, $X$ has a real-analytic $K$-invariant strictly plurisubharmonic exhaustion, and the corresponding Kempf-Ness set $R$ is a $K$-equivariant strong deformation retract of $X$ [@Heinzner-Huckleberry p. 23]. Hence, $R/K$ is a strong deformation retract of $X/K$. Also, $X{{\operatorname{/\!\!/}}}K=X{{\operatorname{/\!\!/}}}G$ is homeomorphic to $R/K$ [@Heinzner-Huckleberry p. 22]. Therefore it suffices to show that the orbit space $X/K$ is $\mathbb Z$-acyclic.
By a theorem of Oliver [@Oliver], to conclude that $X/K$ is $\mathbb Z$-acyclic, we need to know that $X$ is paracompact of finite cohomological dimension and with finitely many $K$-orbit types (finitely many conjugacy classes of stabilisers). The first two conditions are evident.
To verify the third condition, we use a theorem of Mann [@Mann], which states that a compact Lie group acting on an orientable cohomology manifold over $\mathbb Z$ with finitely generated integral cohomology has only finitely many orbit types. We conclude that the action of $K$ on $X$ has only finitely many orbit types.
The following theorem is now immediate for tori, and with a little more work we can prove it for abelian groups in general.
\[t:torus\] Let $X$ and $Y$ be Stein manifolds on which a reductive complex Lie group $G$ acts holomorphically and generically, such that $X$ and $Y$ are locally $G$-biholomorphic over a common quotient. If $G$ is abelian and $X$ is $\mathbb Z$-acyclic, then $X$ and $Y$ are $G$-biholomorphic.
Since $G$ is abelian, it is an extension of a torus $T=({{\mathbb C}}^*)^k$ by a finite abelian group $F$. The short exact sequence $0\to T\to G\to F\to 0$ induces a short exact sequence $$0 \to {{\mathscr O}}(\cdot,T) \to {{\mathscr O}}(\cdot,G)\to {{\mathscr O}}(\cdot,F)\to 0$$ of sheaves of abelian groups on $Q=X{{\operatorname{/\!\!/}}}G$. Clearly, ${{\mathscr O}}(\cdot, T)=({{\mathscr O}}^*)^k$, ${{\mathscr O}}(\cdot, G)={{\mathscr A}}$, and ${{\mathscr O}}(\cdot,F)$ is simply the sheaf of locally constant functions with values in $F$. Consider the long exact sequence $$\cdots \to H^1(Q, {{\mathscr O}}^*)^k \to H^1(Q, {{\mathscr A}}) \to H^1(Q, F) \to \cdots.$$ By Proposition \[p:acyclic\], $H^1(Q,{{\mathscr O}}^*)\cong H^2(Q,{{\mathbb Z}})=0$. By Proposition \[p:acyclic\] and universal coefficients, $H^1(Q,F)=0$ (for this we require both $H^1(Q,{{\mathbb Z}})$ and $H^2(Q,{{\mathbb Z}})$ to vanish). Hence $H^1(Q,{{\mathscr A}})=0$, so the obstruction to $X$ and $Y$ being $G$-biholomorphic over $Q$ vanishes.
Although we do not need it, we remark that arguments of Kraft, Petrie, and Randall [@Kraft-Petrie-Randall] in the algebraic case carry over to the analytic case and, combined with Proposition \[p:acyclic\], show that if $G$ is a reductive complex Lie group acting holomorphically on a contractible Stein manifold $X$, then $X{{\operatorname{/\!\!/}}}G$ is contractible. If there is a connected $G$-orbit, for instance if $G$ is connected or has a fixed point in $X$, then the result follows easily from Proposition \[p:acyclic\] and [@Bredon Cor. II.6.3].
A deep theorem of Hochster and Roberts [@Hochster-Roberts] states that the categorical quotient of a smooth affine variety by the action of a reductive group is Cohen-Macaulay. Slice theory allows us to easily extend the theorem to the holomorphic setting. We note that Cohen-Macaulay in the algebraic sense is equivalent to Cohen-Macaulay in the holomorphic sense. Namely, by GAGA [@Serre Prop. 3], each stalk of the algebraic structure sheaf has the same completion as the corresponding stalk of the holomorphic structure sheaf, so one is Cohen-Macaulay if and only if the other one is [@Eisenbud Prop. 18.8].
\[p:cohen-macaulay\] Let a reductive complex Lie group $G$ act holomorphically on a Stein manifold $X$. The categorical quotient $X{{\operatorname{/\!\!/}}}G$ is Cohen-Macaulay.
Let $q$ be a point in $X{{\operatorname{/\!\!/}}}G$, and let $x$ be a point in the closed orbit over $q$ with stabiliser $H$. By the slice theorem, a neighbourhood of $q$ in $X{{\operatorname{/\!\!/}}}G$ is biholomorphic to an open subset of the quotient of $W$ by the reductive group $H$, where $W$ is the slice representation of $H$. The action of $H$ on $W$ is linear algebraic. By [@Hochster-Roberts], $W{{\operatorname{/\!\!/}}}H$ is Cohen-Macaulay, so the stalk ${{\mathscr O}}_{X{{\operatorname{/\!\!/}}}G,q}$ is Cohen-Macaulay. Thus $X{{\operatorname{/\!\!/}}}G$ is Cohen-Macaulay
\[t:codim-3\] Let $X$ and $Y$ be Stein manifolds on which a reductive complex Lie group $G$ acts holomorphically and generically, such that $X$ and $Y$ are locally $G$-biholomorphic over a common quotient $Q$. Assume moreover that $G$ is abelian and that the complement of the principal stratum $Q_\mathrm{pr}$ has codimension at least 3 in $Q$. If there is an equivariant homeomorphism $X_\mathrm{pr}\to Y_\mathrm{pr}$ over $Q_\mathrm{pr}$, then $X$ and $Y$ are $G$-biholomorphic over $Q$.
We use the same notation as in the proof of Theorem \[t:torus\]. Consider the commuting diagram $$\xymatrix{
H^0(Q_\mathrm{pr}, F) \ar[r] \ar@{=}[d] & H^1(Q_\mathrm{pr}, {{\mathscr O}}^*)^k \ar[r] \ar[d]^\alpha & H^1(Q_\mathrm{pr}, {{\mathscr A}}) \ar[r] \ar[d]^\beta & H^1(Q_\mathrm{pr}, F) \ar@{=}[d] \\
H^0(Q_\mathrm{pr}, F) \ar[r] & H^1(Q_\mathrm{pr}, \mathscr C^*)^k \ar[r] & H^1(Q_\mathrm{pr}, \mathscr C(\cdot, G)) \ar[r] & H^1(Q_\mathrm{pr}, F) }$$ with exact rows (here, $\mathscr C$ refers to continuous functions). Let $\omega\in H^1(Q_\mathrm{pr}, {{\mathscr A}})$ be the obstruction to $X_\mathrm{pr}$ and $Y_\mathrm{pr}$ being $G$-biholomorphic over $Q_\mathrm{pr}$. By assumption, $\beta(\omega)=0$. We will show that $\alpha$ is injective. It follows that $\beta$ is injective, so $\omega=0$. Since $X\setminus X_\mathrm{pr}$ has codimension at least 2 in $X$ by the genericity assumption, any $G$-biholomorphism $X_\mathrm{pr}\to Y_\mathrm{pr}$ over $Q_\mathrm{pr}$ extends to a $G$-biholomorphism $X\to Y$ over $Q$.
Now consider the commuting diagram $$\xymatrix{
H^1(Q_\mathrm{pr}, {{\mathbb Z}}) \ar[r] \ar@{=}[d] & H^1(Q_\mathrm{pr}, {{\mathscr O}}) \ar[r] \ar[d] & H^1(Q_\mathrm{pr}, {{\mathscr O}}^*) \ar[r] \ar[d] & H^2(Q_\mathrm{pr}, {{\mathbb Z}}) \ar@{=}[d] \\
H^1(Q_\mathrm{pr}, {{\mathbb Z}}) \ar[r] & H^1(Q_\mathrm{pr}, \mathscr C)=0 \ar[r] & H^1(Q_\mathrm{pr}, \mathscr C^*) \ar[r] & H^2(Q_\mathrm{pr}, {{\mathbb Z}}) }$$ with exact rows. By Proposition \[p:cohen-macaulay\], $Q$ is Cohen-Macaulay, and since $Q\setminus Q_\mathrm{pr}$ has codimension at least 3, the vanishing theorem for local cohomology of Scheja and Trautmann ([@Scheja], [@Trautmann], [@Siu-Trautmann Thm. 1.14], [@Banica-Stanasila Thm. II.3.6]) implies that $H^1(Q_\mathrm{pr},{{\mathscr O}})\cong H^1(Q,{{\mathscr O}})=0$. Hence $\alpha$ is injective.
General reductive groups
========================
We continue to use the notation established at the beginning of Section \[s:generic\]. We refer to Lashof’s foundational paper for the results we need about the classification of topological $K$-$G$-bundles over $X$ (see also [@tomDieck]). Since $G$ is isomorphic to a closed subgroup of $\mathrm{GL}_n({{\mathbb C}})$ for some $n$, every $K$-$G$-bundle over $X$ is numerable [@Lashof Prop. 1.11 and Cor. 1.13]. (Numerability is a technical condition whose definition we omit.)
We say that $X$ is $K$-contractible if the identity map of $X$ can be joined to a constant map by a continuous path of $K$-equivariant continuous maps $X\to X$. The value of the constant map must then be a fixed point $x_0$ of the $K$-action. A $K$-module is obviously $K$-contractible.
By [@Lashof Cor. 2.11], if $X$ is $K$-contractible, then every numerable $K$-$G$-bundle over $X$ is isomorphic as a $K$-$G$-bundle to its own pullback by the map that takes all of $X$ to $x_0$. Thus, by the discussion preceding Theorem \[t:topological-obstruction\], we have the following theorem and corollaries.
\[t:K-contractible\] Let $G$ be a reductive complex Lie group with maximal compact subgroup $K$. Let $X$ and $Y$ be normal Stein spaces on which $G$ acts holomorphically and generically, such that $X$ and $Y$ are locally $G$-biholomorphic over a common quotient. If $X$ is $K$-contractible, then $X$ and $Y$ are $G$-biholomorphic.
\[c:first-corollary\] Let $X$ be a Stein manifold on which $G$ acts holomorphically and let $V$ be a $G$-module, such that $X$ and $V$ are locally $G$-biholomorphic over a common quotient. If $X$ and $V$ are generic, then $X$ and $V$ are $G$-biholomorphic.
\[c:Franks-dream\] A holomorphic $G$-action on ${{\mathbb C}}^n$, which is locally $G$-biholomorphic over a common quotient to a generic linear action, is linearisable.
Strict equivariant diffeomorphisms {#s:smooth}
==================================
Let $X$ and $Y$ be Stein manifolds on which a reductive complex Lie group $G$ acts holomorphically, such that $X$ and $Y$ are locally $G$-biholomorphic over a common quotient $Q$ with quotient maps $p:X\to Q$ and $r:Y\to Q$. A $G$-equivariant diffeomorphism $X\to Y$ which induces the identity on $Q$ and is biholomorphic from $p^{-1}(q)$ onto $r^{-1}(q)$ for every $q\in Q$ will be called a *strict $G$-diffeomorphism* from $X$ to $Y$. If a strict $G$-diffeomorphism $X\to Y$ exists, then we call $X$ and $Y$ *strictly $G$-diffeomorphic*.
A $G$-equivariant diffeomorphism $\psi$ of $X$ is called *special* if it is of the form $\psi (x) = \gamma (x) \cdot x$ for some smooth $G$-equivariant map $\gamma: X \to G$, where $G$ acts on the target $G$ by conjugation. A $G$-equivariant diffeomorphism $\psi:X\to Y$ is called *special* if for some open cover $(U_i)$ of $Q$ and $G$-biholomorphisms $\phi_i:p^{-1}(U_i)\to r^{-1}(U_i)$ over $U_i$, the diffeomorphisms $\phi_i^{-1} \circ \psi$ of $p^{-1}(U_i)$ are special.
If the action on $X$ (and thus on $Y$) is generic, then every $G$-biholomorphism over the quotient is a special (and obviously a strict) $G$-diffeomorphism (Lemma \[l:g-biholomorphisms\]). Over the principal stratum of a generic action, the notions of a special and a strict $G$-diffeomorphism coincide. Clearly, the special $G$-diffeomorphisms of $X$ form a group, and so do the strict ones.
If the action is generic, then the definition of a special $G$-diffeomorphism $X\to Y$ does not depend on the choice of the cover $(U_i)$ of $Q$ and the $G$-biholomorphisms $\phi_i:p^{-1}(U_i)\to r^{-1}(U_i)$ over $U_i$, since any $G$-biholomorphism over the quotient is special, and the special $G$-diffeomorphisms form a group.
\[t:strict-diffeomorphisms\] Let $G$ be a reductive complex Lie group acting holomorphically and generically on Stein manifolds $X$ and $Y$, which are locally $G$-biholomorphic over a common quotient. Every strict $G$-diffeomorphism $X\to Y$ is homotopic, via a continuous path of strict $G$-diffeomorphisms, to a special (and strict) $G$-diffeomorphism.
By the discussion preceding Theorem \[t:topological-obstruction\], the obstruction to $X$ and $Y$ being $G$-biholomorphic over $Q$ vanishes if there is a special $G$-diffeomorphism $X\to Y$. Hence the following corollary is immediate.
\[c:strict-diffeomorphisms\] Let $G$ be a reductive complex Lie group acting holomorphically and generically on Stein manifolds $X$ and $Y$, which are locally $G$-biholomorphic over a common quotient. If $X$ and $Y$ are strictly $G$-diffeomorphic, then they are $G$-biholomorphic.
The remainder of this section is devoted to the proof of the theorem. We start by constructing the desired homotopy in a particular local setting. Let $H$ be a reductive subgroup of $G$, and let $W$ be an $H$-module, not necessarily generic, such that $W^H=0$. Let $p_1,\ldots,p_k$ be homogeneous generators of ${{\mathscr O}}_\mathrm{alg}(W)^H$ of degrees $d_1, \ldots, d_k$ respectively, and let $B=\{w\in W : \lvert p_i\rvert<a_i\}$ for some $a_1,\ldots,a_k>0$. Let $T_W$ (resp. $T_B$) be the tube $G\times^HW$ (resp. $G\times^HB$). Then $T_W$ and $T_B$ are bundles over $G/H$ with fibres $W$ and $B$, respectively. Below, when talking about derivatives in the fibre-direction, we mean the fibres of these bundles. Note that the $G$-action on $T_W$ is generic if and only if the $H$-action on $W$ is generic.
The null fibre of the quotient map $T_B \to T_B{{\operatorname{/\!\!/}}}G\cong B{{\operatorname{/\!\!/}}}H$ is $G\times^H{{\mathscr N}}(W)$, where ${{\mathscr N}}(W)$ denotes the null cone of the $H$-representation on $W$. The unique closed orbit in the null fibre is the zero section $Z$ of $T_W$. Now let $\phi:T_B\to T_B$ be a strict $G$-diffeomorphism. Then $\phi$ must preserve $Z$. Let $\delta\phi:T_W\to T_W$ denote the derivative of $\phi$ in the fibre-direction along $Z$.
For $t\in {{\mathbb C}}^*$, denote by $\alpha_t : T_W\to T_W$ the $G$-biholomorphism defined by $\alpha_t ([g, w])= [g,tw]$. Note that $p_1,\ldots,p_k$ correspond to generators $F_1,\ldots,F_k$ of ${{\mathscr O}}_\mathrm{alg}(T_W)^G$ with $F_i ([g, w])= p_i (w)$, where $F_i \circ \alpha_t = t^{d_i} F_i$. Let $\phi_t = \alpha_t^{-1} \circ \phi \circ \alpha_t$ for $ t>0$, and let $\phi_0=\delta\phi$.
The family $\phi_t$, $t\in[0,1]$, is a homotopy of strict $G$-diffeomorphisms of $T_B$ joining $\phi$ to $\delta \phi$. If the $H$-action on $W$ is generic, then $\phi_0$ is a special (and strict) $G$-diffeomorphism.
Since $\alpha_t$ is holomorphic, the maps $\phi_t : T_W\to T_W$, $t\in(0,1]$, are strict $G$-diffeomorphisms. Moreover, they induce the identity map on the quotient since they preserve the invariants: $$\begin{aligned}
F_i ( \alpha_t^{-1} \circ \phi \circ \alpha_t ([g, w])) &= t^{-d_i} F_i (\phi \circ \alpha_t ([g, w])) = t^{-d_i} F_i ( \alpha_t ([g, w])) \\ &= t^{-d_i} t^{d_i}F_i ([g, w]) = F_i ([g, w]).\end{aligned}$$ This shows, in particular, that for every $t\in (0,1]$, $\phi_t$ is a strict $G$-diffeomorphism of $T_B$ and that $\phi_0$ preserves the invariants. By differentiability of $\phi$, $\lim\limits_{t\to 0} \phi_t$ exists locally uniformly on $T_B$ and equals $\phi_0=\delta\phi $. Therefore, $\phi_0$ is a $G$-diffeomorphism over the quotient.
By Lemma \[l:g-biholomorphisms\], to prove that $\phi_0$ is special in the generic case, it suffices to show that $\phi_0$ is holomorphic, that is, that the the derivative of $\phi$ in the fibre-direction along $Z$ is complex-linear and not merely real-linear. By assumption, $\phi$ restricts to a $G$-biholomorphism of $G\times^H {{\mathscr N}}(W)$, so it has a complex-linear derivative in the fibre-direction along the Zariski tangent space of ${{\mathscr N}}(W)$. Since $W^H=0$ by assumption, the Zariski tangent space at $0\in W$ to ${{\mathscr N}}(W)$ is all of $W$. This shows that $\phi_0$ is complex-linear.
We can explicitly describe the form of $\phi_0$. We have $\phi_0([g,w])=\gamma([g,w])\cdot ([g,w])$ with $\gamma:G\times^HW\to G$ equivariant and algebraic. Then $\phi_0([e,hw])=h\delta\phi[(e,w])h{{^{-1}}}$, so that $\gamma([e,w]):W\to G$ is $H$-equivariant. But for $t\in{{\mathbb R}}$, $\phi_0([e,tw])=t\delta\phi([e,w])$, so $$t[\gamma([e,tw]),w]=[\gamma([e,tw]),tw]=t[\gamma([e,w]),w].$$ Hence $\gamma([e,w])\cdot [e,w]=\gamma([e,0])\cdot [e,w]$, where $\gamma([e,0])$ is an element $g_0$ of $G$ centralising $H$. Thus $\gamma([g,w])([g,w])=([gg_0,w])$ and $\gamma([g,w])=gg_0g{{^{-1}}}$. The element $g_0$ is unique, since for $w$ a principal point of $W$, $\gamma([e,w])$ is unique (and equal to $g_0$). It is easy to see that any choice of $g_0$ in the centraliser of $H$ in $G$ gives an equivariant $\gamma:G\times^H W \to G$.
Next we describe the homotopy in the general local setting. In addition to the notation above, let $C$ be a Stein manifold (with a trivial $G$-action), and let $ \phi:C \times T_B\to C\times T_B$ be a strict $G$-diffeomorphism. Since $\phi$ induces the identity on the quotient, it is of the form $\phi (c, [g, w]) = (c, \tilde \phi (c, [g,w]))$, where $\tilde \phi : C \times T_B \to T_B$ may be viewed as a smooth family of strict $G$-diffeomorphisms of $T_B$ parametrised by $C$. Let $\delta \tilde \phi : C \times T_W\to T_W$ be the derivative of $\tilde \phi$ in the fibre-direction along the zero section of $T_W$, and set $\delta\phi (c, [g,w]) = (c, \delta \tilde \phi (c, [g,w]))$.
Applying the above homotopy in this parametrised setting, we get the following result. We let $\alpha_t$ act on $C\times T_B$ as the identity on the first factor and as defined above on the second factor.
\[l:parameters\] The family $\phi_t : C \times T_B \to C\times T_B$, $t\in[0,1]$, with $$\phi_t (c, [g, w])=(c, \alpha_t^{-1} \circ \tilde \phi \circ \alpha_t (c, [g,w]))$$ for $t>0$ and $$\phi_0 (c, [g, w])=(c, \delta \tilde \phi (c, [g, w])),$$ is a homotopy of strict $G$-diffeomorphisms joining $\phi$ to $\delta \phi $. If the $G$-action on $T_B$ (equivalently, on $C \times T_B$) is generic, then $\phi_0$ is a special (and strict) $G$-diffeomorphism.
We now want to change our homotopy so that it is the identity away from a neighbourhood of a given point. Let $c_0\in C$ and let $\rho: C\times T_B{{\operatorname{/\!\!/}}}G\to[0,1]$ be a smooth function which is 1 on a neighbourhood $U$ of $p(\{c_0\}\times Z)$ and has compact support $F$. Let $\tau(t,z)=1+(t-1)\rho(z)$ for $t\in[0,1]$ and $z\in C\times T_B{{\operatorname{/\!\!/}}}G$. Then $\tau(t,z)=1$ for $z\notin F$, and $\tau(t,z)=t$ for $z\in U$. Let $\tau(t,z)\cdot (c,[g,w])$ denote $(c,[g,\tau(t,z)w])$ for $(c,[g,w])\in C\times T_B$, where $z=p(c,[g,w])$.
\[c:phi’\] Let $\rho$ be as above and let $\phi:C\times T_B \to C\times T_B$ be a strict $G$-diffeomorphism. Let $\phi_t(x)=\tau(t,z){{^{-1}}}\cdot \phi(\tau(t,z)\cdot x)$ for $x\in C\times T_B$, $z=p(x)$, and $t\in(0,1]$. Set $\phi_0=\lim\limits_{t\to 0}\phi_t$. The family $\phi_t$, $t\in[0,1]$, is a homotopy of strict $G$-diffeomorphisms joining $\phi$ to $\phi_0$. Moreover, for each $t\in[0,1]$, $\phi_t$ equals $\phi$ over the complement of $F$, and $\phi_0$ equals $\delta\phi$ over $U$.
\[l:special-remains\] If in the above situation there is a $G$-saturated open subset $\Omega$ of $C\times T_B$ which is invariant under $x\mapsto \tau(t,z)\cdot x$ such that the restriction of $\phi$ to $\Omega$ is special, then all the strict $G$-diffeomorphisms $\phi_t$ are special when restricted to $\Omega$.
The assumption is that for $(c, [g, w]) \in \Omega$, we have $\phi (c, [g, w]) = (c, \gamma(c, [g, w]) \cdot [g, w])$, where $\gamma : \Omega \to G$ is smooth and $G$-equivariant. It follows from the definition of $\phi_t$ that $\phi_t (c, [g, w]) = (c, \gamma_t (c, [g, w])\cdot [g, w])$, where $ \gamma_t (x) = \gamma (\tau(t,z)\cdot x)$, which is well defined by hypothesis. The map $\gamma_t$ is clearly smooth and $G$-equivariant.
Let $\psi:X\to Y$ be a strict $G$-diffeomorphism. Consider the stratification of $Q$ by the connected components of the Luna strata. Let $S_k$ denote the union of the strata of dimension $k$. We will inductively find homotopies of $\psi$ such that it becomes special over an open neighbourhood $\Omega$ of $S_0 \cup \cdots \cup S_k$. Since over the principal stratum any strict $G$-diffeomorphism is special, we are done once we reach the case $k=\dim Q-1$. Each step of the finite induction will be done by a countable induction.
Let $S$ be a stratum of minimal dimension. Let $H$ be a corresponding stabiliser and let $(W,H)$ be the nontrivial part of the slice representation. Cover $S$ by a locally finite collection of compact sets $K_i$, $i\in\mathbb N$, such that each $K_i$ lies in an open subset $U_i$ of $Q$ with $p{{^{-1}}}(U_i)\cong C_i\times (G\times^HB_i)$, where $C_i$ is a Stein open subset of $S$, and $B_i\subset W$ is as before. We may assume that we have the same decomposition of $r{{^{-1}}}(U_i)$, so we may view $\psi$ over $U_i$ as a map of $C_i\times(G\times^HB_i)$ to itself. We may also assume that any $K_j$ which intersects $K_i$ is contained in $U_i$. By induction we suppose that there is a neighbourhood $\Omega_{n-1}$ of $K_1\cup\dots\cup K_{n-1}$ such that $\psi$ is strict and special on $p{{^{-1}}}(\Omega_{n-1})$. Let $\rho_n(z)$ be smooth, $0\leq\rho_n\leq 1$, such that $\rho_n=1$ on a neighbourhood of $K_n$ and $\rho_n$ has compact support in $U_n$. Then Corollary \[c:phi’\] provides a homotopy from $\psi$ to $\psi'$, where $\psi'$ is special on a neighbourhood of $K_n$. Now let $\beta_t(x)$ be the endomorphism of $U_n$ which is induced by the endomorphism of $p^{-1}(U_n)$ which sends $(c,[g,w])$ to $(c,[g,\tau(t,z)w])$, where $z$ is the image of $(c,[g,w])$ in $U_n$ and $\tau(t,z)=1+(t-1)\rho_n(z)$ as before. Suppose that $K_j$, $j<n$, does not intersect $K_n$. Then we may assume that $\rho_n$ vanishes on $(C_n\cap K_j)\times(G\times^HB_n)$, where we identify $C_n$ with its image in $S$.
Note that $\beta_t$ smoothly extends to be the identity outside of $U_n$. It is also the identity on $U_{n}\cap S$ and on $K_1\cup\cdots\cup K_{n-1}$. Thus there is a neighbourhood $\Omega_{n-1}'\subset\Omega_{n-1}$ of $K_1\cup\dots\cup K_{n-1}$ such that $\beta_t(z)$ maps $\Omega_{n-1}'$ into $\Omega_{n-1}$. By Lemma \[l:special-remains\], $\phi'$ is special over $\Omega_{n-1}'\cup U_n'$, where $U_n'$ is a neighbourhood of $K_n$. This gives us $\Omega_n$ and we continue inductively. We could run into a problem if we are continually shrinking the neighbourhoods $\Omega_n$ near a point $z\in S$. But there is a neighbourhood of $z$ which does not intersect the support of any $\rho_n$ for $n$ sufficiently large. Thus for $n$ sufficiently large, each $\Omega_n$ contains a fixed neighbourhood of $z$. Hence there is a neighbourhood $\Omega$ of $S$ and a homotopy of $\phi$ to $\phi'$, such that $\phi'$ is strict and special over $\Omega$. Since the strata of minimal dimension are closed and disjoint, we may find $\Omega$ as desired for all the strata of minimal dimension.
We may assume by induction that $\phi$ is special on a neighbourhood $\Omega$ of the union of the strata of dimension less than $m\geq 0$ (the union is a closed set). Let $\Omega_1$ and $\Omega_2$ be smaller neighbourhoods such that $\overline{\Omega}_1 \subset \Omega_2\subset \overline{\Omega}_2\subset\Omega$. Let $S$ be a stratum of dimension $m$. Choose a locally finite covering of $S$ by compact sets $K_i$ as above. We may suppose that each $K_i$ either lies entirely inside $\Omega$, lies entirely outside $\overline{\Omega}$, or does not intersect $\overline{\Omega}_2$. We can now apply the above process to the $K_i$ that lie outside $\overline{\Omega}$, and to the $K_i$ that do not intersect $\overline{\Omega}_2$ to find a homotopy of $\psi$ such that it remains the same on $\Omega_1$ and becomes special on a neighbourhood of $S$. We can do this for all the strata of dimension $m$, completing our induction.
Strong equivariant homeomorphisms {#s:continuous}
=================================
As before, we let $ G$ be a reductive complex Lie group, and $X$ and $Y$ be Stein manifolds on which $G$ acts holomorphically, such that $X$ and $Y$ are locally $G$-biholomorphic over a common quotient $Q$. Let $p:X \to Q$ and $r:Y\to Q$ be the quotient maps. In the previous section we showed that in the generic case we can deform a strict $G$-diffeomorphism $X\to Y$ to one which is also special. In this section we define the notion of a *strong $G$-homeomorphism* $X\to Y$, and show, in the generic case, that it can be deformed to a special $G$-homeomorphism, that is, one locally of the form $x\mapsto\gamma(x)\cdot x$, where $\gamma: X\to G$ is continuous and $G$-equivariant, with $G$ acting on the target $G$ by conjugation. The proof is largely the same as in the previous section, aside from some technical lemmas.
Let $H$ be a reductive subgroup of $G$ and let $W$ be an $H$-module. Let $B$ be an $H$-saturated neighbourhood of $0\in W$. Let $T_W=G\times^HW$ and $T_B=G\times^HB$. There is a $G$-stable finite-dimensional subspace $V\subset {{\mathscr O}}_\mathrm{alg}(T_W)$ which generates ${{\mathscr O}}_\mathrm{alg}(T_W)$, and also generates the $G$-finite elements of ${{\mathscr O}}(T_B)$ as a module over ${{\mathscr O}}(T_B)^G\cong {{\mathscr O}}(B)^H$. We may think of the elements of $V$ as $G$-equivariant maps from $T_W$ to $V^*$ sending $[g,w]\in T_W$ to the element of $V^*$ whose value at $F\in V$ is $F([g,w])$. Let $\{F_i\}$ be a basis of $V$. We say that a $G$-equivariant homeomorphism $\psi:T_B\to T_B$ is *strong* if it lies over the identity of $T_B{{\operatorname{/\!\!/}}}G$ and sends $F_i$ to $\sum_{j} a_{ij}F_j$, where the $a_{ij}$ are continuous functions on $T_B{{\operatorname{/\!\!/}}}G$. It is easy to see that the definition does not depend on our choice of $V$ and the generators $F_i$. Since the $F_i$ generate the coordinate ring of every fibre of $p:T_B\to Q_B=T_B{{\operatorname{/\!\!/}}}G$, $\psi$ restricts to an algebraic isomorphism of each fibre of $p$.
Now let $\psi:X\to Y$ be a $G$-homeomorphism over $Q$. Let $(U_i)$ be an open cover of $Q$ such that there are $G$-biholomorphisms $\phi_i: p{{^{-1}}}(U_i)\to r{{^{-1}}}(U_i)\cong G\times^{H_i}B_i$, where $B_i$ is an $H_i$-saturated neighbourhood of $0\in W_i$, $H_i$ is a reductive subgroup of $G$, and $W_i$ is an $H_i$-module. We say that $\psi$ is *strong* if each $\phi_i{{^{-1}}}\circ\psi:G\times^{H_i}B_i\to G\times^{H_i}B_i$ is strong. Again, the definition does not depend on the choices made.
We may think of a strong $G$-homeomorphism $X\to X$ as a particular type of continuous family of algebraic (equivalently, holomorphic) $G$-isomorphisms of the fibres of $p$. If the group scheme corresponding to $X$, as constructed in [@Kraft-Schwarz], existed, then the strong $G$-homeomorphisms $X\to X$ would be the global continuous sections of the group scheme. Note that a strong $G$-diffeomorphism is also strict.
Our main result in this section is the following counterpart to Theorem \[t:strict-diffeomorphisms\] and Corollary \[c:strict-diffeomorphisms\].
\[t:strongly\] Let $G$ be a reductive complex Lie group acting holomorphically and generically on Stein manifolds $X$ and $Y$, which are locally $G$-biholomorphic over a common quotient. Every strong $G$-homeomorphism $X\to Y$ is homotopic, via a continuous path of strong $G$-homeomorphisms, to a special (and strong) $G$-homeomorphism.
Hence, if $X$ and $Y$ are strongly $G$-homeomorphic, then they are $G$-biholomorphic.
The key to the theorem is Lemma \[l:fundamental\] below which says that strongly continuous maps are somewhat differentiable.
Let $H$, $W$, $B$, and the $F_i$ be as above. Assume that $W^H=0$. Let $p_1,\ldots,p_m$ be homogeneous generators of ${{\mathscr O}}_\mathrm{alg}(W)^H$. We have a grading on ${{\mathscr O}}_\mathrm{alg}(T_W)=\bigoplus{{\mathscr O}}_\mathrm{alg}(T_W)_r$ corresponding to the ${{\mathbb C}}^*$-action on $W$. Thus $F_i$ has degree $r$ if $F_i([g,tw])=t^rF_i([g,w])$. The elements of degree zero are the pullbacks to ${{\mathscr O}}_\mathrm{alg}(T_W)$ of the elements of ${{\mathscr O}}_\mathrm{alg}(G/H)$. Now ${{\mathscr O}}_\mathrm{alg}(T_W)$ is the direct sum of covariants of the various degrees $r\geq 0$. Let $d_j$ be the degree of $F_j$. We ignore the covariants of degree zero, that is, we do not put them in our list. We also ignore the invariants, that is, we only consider $F_j$ which transform by a nontrivial representation $V_i$ of $G$. We may assume that $d_j=1$ if and only if $j\leq k$.
\[l:gens\] ${{\mathscr O}}_\mathrm{alg}(T_W)$ is generated by ${{\mathscr O}}_\mathrm{alg}(T_W)_1$ as an ${{\mathscr O}}_\mathrm{alg}(G/H)$-algebra.
Let $W_g=\{[g,w]:w\in W\}$. The restrictions of the $F_i$, $i\leq k$, must generate $W_g^*$, since all the covariants $F_j$ restricted to $W_g$ generate ${{\mathscr O}}_\mathrm{alg}(W_g)$ and the covariants of higher degree obviously cannot help. We are also using the assumption that $W^H=0$. Thus the map $T_W\to G/H\times{{\mathbb C}}^k$ is an embedding, where the map to ${{\mathbb C}}^k$ comes from the $F_i$, $i\leq k$. Thus ${{\mathscr O}}_\mathrm{alg}(T_W)$ is generated by ${{\mathscr O}}_\mathrm{alg}(T_W)_1$ over ${{\mathscr O}}_\mathrm{alg}(G/H)$.
Let $\psi:T_B\to T_B$ be a strong $G$-homeomorphism. We assume that $B$ is stable under multiplication by $t\in[0,1]$. Let $(a_{ij}(z))$ be the $n\times n$ matrix of continuous functions such that $\psi^*F_i=\sum a_{ij}(z)F_j$, $z\in Q_B=T_B{{\operatorname{/\!\!/}}}G$. Note that $\psi^*$ has to send the covariants corresponding to the $G$-module $V_i$ to the covariants of the same type.
For $x\in T_B$, let $z$ denote $p(x)\in Q_B$, and let $x\mapsto t\cdot x$ denote the action of $t\in [0,1]$, that is, $t\cdot [g,w]=[g,tw]$. Let $t\cdot z$ denote $p(t\cdot x)$. Let $\psi_t(x)=t{{^{-1}}}\cdot \psi(t\cdot x)$, $x\in X$, $t\in[0,1]$. As observed before, $\psi$ preserves the zero section $Z$ of $T_B$. Thus $d\psi$ restricted to the Zariski tangent space at the origin of ${{\mathscr N}}(W_e)$ (which is $W_e$) gives us a complex-linear map of $W_e$ to $W_g$, where $\psi$ sends $[e,0]$ to $[g,0]$. We calculate this derivative in the usual way below.
\[l:fundamental\] In the above setting, the following hold.
1. $(\psi_t^*F_i)(x)=\sum_j t^{d_j-d_i}a_{ij}(t\cdot z) F_j(x)$, $x\in T_B$.
2. The limit as $t\to 0$ of $\psi_t$ acts on the $F_i$, $i\leq k$, by the matrix $L$ with entries $a_{ij}(0)$, $i, j\leq k$.
3. $\psi$ has a normal derivative $\delta\psi $ along the zero section of $T_B$, and $\delta\psi$ is a complex-linear $G$-vector bundle isomorphism of $T_W$ which preserves the invariants.
4. There are continuous $b_{ij}(t,z)$ such that $(\psi_t^*F_i)(x)=\sum_j b_{ij}(t,z)F_j(x)$ for all $i$.
5. If $W$ is generic, then $\psi_0=\delta\psi$ is special.
As seen before, the $\psi_t$ lie over the identity of $Q_B$. Now $$\begin{aligned}
(\psi_t^*F_i)(x)&=F_i(t{{^{-1}}}\psi(t\cdot x))=t^{-d_i}(\psi^*F_i)(t\cdot x)=t^{-d_i}\sum_j a_{ij}(t\cdot z)F_j(t\cdot x) \\ &=\sum_j t^{d_j-d_i}a_{ij}(t\cdot z)F_j(x) \end{aligned}$$ and we have (1). Now let $i\leq k$, so that $d_i=1$. Then $$\lim_{t\to 0}\psi_t^*F_i(x)=\lim_{t\to 0}\sum_j t^{d_j-1}a_{ij}(t\cdot z) F_j(x).$$ Since $d_j\geq 1$, and $d_j>1$ for $j>k$, the limit is $\sum\limits_{j\leq k}a_{ij}(0)F_j(x)$. Thus the directional derivative of $\psi$ in the direction of $x$ exists and the derivative acting on the $F_i$ for $i\leq k$ is given by the matrix $L$.
Next we prove differentiability. Let $\delta\psi$ be the bundle map of $T_B$ given by the directional derivatives. Choose a norm on the vector bundle $T_W$ and consider $x\in T_B$ of norm at most $\epsilon$ for some $\epsilon>0$. We must show that $\lim\limits_{t\to 0} t{{^{-1}}}(\psi(tx)-\delta\psi(tx))=0$ locally uniformly for such $x$. For $i\leq k$, we have $$\begin{aligned}
\lim_{t\to 0}F_i(t{{^{-1}}}\psi(tx)-\delta\psi(tx))&=\lim_{t\to 0}(\psi_t^*F_i-\delta\psi^*F_i)(x) \\ &=\lim_{t\to 0}\sum_j t^{d_j-1}(a_{ij}(t\cdot z)-a_{ij}(0)) F_j(x).\end{aligned}$$ The last limit vanishes since $a_{ij}$ is continuous at $z=0$, and the vanishing is locally uniform in $x$. Hence $\psi$ is differentiable in the normal direction and $(\delta\psi)^*$ acts on those $F_i$ of degree 1 by the matrix $L$. Thus $\delta\psi$ coincides with $d\psi$ calculated on the Zariski tangent spaces of the null cones; hence it is complex-linear. Since $\psi$ is $G$-equivariant, so is $\delta\psi$, and since the $\psi_t$ preserve the invariants, so does $\delta\psi$.
In order to prove (4), we note that by Lemma \[l:gens\], for $i>k$, $F_i=\sum_s f_s Q_s(F_1,\dots,F_k)$, where each $Q_s$ is homogeneous of degree $d_i$ and $f_s\in{{\mathscr O}}_\mathrm{alg}(G/H)$. When we expand any $Q_s$, we get an expression in the invariants times the generators $F_j$. If $F_j$ occurs, then since we have an expression starting in degree at least $d_i$, either $d_j>d_i$, or the coefficient of $F_j$ has to be an invariant with lowest-degree term of degree at least $d_i-d_j$.
Expanding $\psi_t^*F_i$ using the expression above for $F_i$, we obtain $$\begin{aligned}
(\psi_t^*F_i)(x)&=\sum_s f_sQ_s(\psi_t^*F_1,\dots,\psi_t^*F_k)(x) \\ &=\sum_s t^{-d_i}f_s Q_s\left(\sum_j a_{1j}(t\cdot z)F_j,\dots,\sum_j a_{kj}(t\cdot z) F_j\right)(t\cdot x).\end{aligned}$$ Hence $$(\psi_t^*F_i)(x)=\sum_s\sum_jf_{sj} t^{d_j-d_i}a'_{ijs}(t\cdot z) P_{ijs}(t\cdot x) F_j(x),$$ where $a'_{ijs}$ is a polynomial in the $a_{pq}$, $P_{ijs}$ is an invariant, and $f_{sj}\in{{\mathscr O}}_\mathrm{alg}(G/H)$. If $d_j\geq d_i$, then the coefficient of $F_j(x)$ has the desired form. If $d_j<d_i$, then the lowest-degree monomial of $P_{ijs}$ has degree at least $d_i-d_j$. Set $P'_{ijs}(t,z)=t^{d_i-d_j}P_{ijs}(t\cdot x)$. Then $P'_{ijs}$ is continuous on ${{\mathbb R}}\times Z$ and the coefficient of $F_j(x)$ in $(\psi_t^*F_i)(x)$ is $\sum_s f_{sj}a'_{ijs}(t\cdot z)P'_{ijs}(t, z)$ giving (4). Part (5) is clear.
Using the arguments of the previous section, we prove the analogue of Lemma \[l:parameters\] for continuous parameters. Using (4) above, we establish the analogue of Corollary \[c:phi’\], where the homotopy is through strong $G$-homeomorphisms, and the continuous analogue of Lemma \[l:special-remains\] is straightforward. Now we can simply repeat the proof of Theorem \[t:strict-diffeomorphisms\] using the analogues of the lemmas and corollary.
[99]{}
Bănică, C. and O. Stănăşilă. *Algebraic methods in the global theory of complex spaces.* Editura Academiei and John Wiley and Sons, 1976.
Bredon, G. E. *Introduction to compact transformation groups.* Pure and Applied Mathematics, Vol. 46. Academic Press, 1972.
De Concini, C. and C. Procesi. *A characteristic free approach to invariant theory.* Advances in Math. **21** (1976) 330–354.
Derksen, H. and F. Kutzschebauch. *Nonlinearizable holomorphic group actions.* Math. Ann. **311** (1998) 41–53.
Eisenbud, D. *Commutative algebra. With a view toward algebraic geometry.* Graduate Texts in Mathematics, 150. Springer-Verlag, 1995
Forstnerič, F. *Oka maps.* C. R. Acad. Sci. Paris, Ser. I **348** (2010) 145–148.
Forstnerič, F. *Stein manifolds and holomorphic mappings.* Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge, 56. Springer-Verlag, 2011.
Forstnerič, F. and F. Lárusson. *Survey of Oka theory.* New York J. Math. **17a** (2011) 11–38.
Grauert, H. *Analytische Faserungen über holomorph-vollständigen Räumen.* Math. Ann. **135** (1958) 263–273.
Gromov, M. *Oka’s principle for holomorphic sections of elliptic bundles.* J. Amer. Math. Soc. **2** (1989) 851–897.
Heinzner, P. and A. Huckleberry. *Invariant plurisubharmonic exhaustions and retractions.* Manuscripta Math. **83** (1994) 19–29.
Heinzner, P. and F. Kutzschebauch. *An equivariant version of Grauert’s Oka principle.* Invent. Math. **119** (1995) 317–346.
Hochster, M. and J. L. Roberts. *Rings of invariants of reductive groups acting on regular rings are Cohen-Macaulay.* Advances in Math. **13** (1974) 115–175.
Huckleberry, A. T. *Actions of groups of holomorphic transformations.* Several complex variables, VI, 143–196, Encyclopaedia Math. Sci., 69. Springer-Verlag, 1990.
Kraft, H. *Geometrische Methoden in der Invariantentheorie.* Aspects of Mathematics, D1. Vieweg, 1984.
Kraft, H. *Challenging problems on affine $n$-space.* Séminaire Bourbaki, Vol. 1994/95. Astérisque No. 237 (1996), Exp. No. 802, 5, 295–317.
Kraft, H., T. Petrie, and J. D. Randall. *Quotient varieties.* Adv. Math. **74** (1989) 145–162.
Kraft, H. and G. W. Schwarz. *Reductive group actions with one-dimensional quotient.* Inst. Hautes Études Sci. Publ. Math. **76** (1992) 1–97.
Kutzschebauch, F. *Compact and reductive subgroups of the group of holomorphic automorphisms of ${{\mathbb C}}^n$.* Singularities and complex analytic geometry (Kyoto, 1997). S= urikaisekikenky= usho K= oky= uroku no. 1033 (1998) 81–93.
Kutzschebauch, F. and S. Lodin. *Holomorphic families of non-equivalent embeddings and of holomorphic group actions on affine space.* Duke Math. J. **162** (2013) 49–94.
Lashof, R. K. *Equivariant bundles.* Illinois J. Math. **26** (1982) 257–271.
Luna, D. *Slices étales.* Sur les groupes algébriques, 81–105. Bull. Soc. Math. France, Memoire 33. Soc. Math. France, 1973.
Luna, D. and R. W. Richardson. *A generalization of the Chevalley restriction theorem.* Duke Math. J. **46** (1979) 487–496.
Mann, L. N. *Finite orbit structure on locally compact manifolds.* Michigan Math. J. **9** (1962) 87–92.
Oliver, R. *A proof of the Conner conjecture.* Ann. of Math. (2) **103** (1976) 637–644.
Scheja, G. *Fortsetzungssätze der komplex-analytischen Cohomologie und ihre algebraische Charakterisierung.* Math. Ann. **157** (1964) 75–94.
Schwarz, G. W. *Exotic algebraic group actions.* C. R. Acad. Sci. Paris Sér. I Math. **309** (1989) 89–94.
Schwarz, G. W. *Lifting differential operators from orbit spaces.* Ann. Sci. École Norm. Sup. (4) **28** (1995) 253–305.
Schwarz, G. W. *Vector fields and Luna strata.* J. Pure Appl. Algebra **217** (2013) 54–58.
Schwarz, G. W. *Quotients, automorphisms and differential operators.* J. London Math. Soc., to appear, [arXiv:1201.6369]{}.
Serre, J.-P. *Géométrie algébrique et géométrie analytique.* Ann. Inst. Fourier, Grenoble **6** (1956) 1–42.
Siu, Y. T. and G. Trautmann. *Gap-sheaves and extension of coherent analytic subsheaves.* Lecture Notes in Mathematics, Vol. 172. Springer-Verlag, 1971.
Snow, D. M. *Reductive group actions on Stein spaces.* Math. Ann. **259** (1982) 79–97.
tom Dieck, T. *Faserbündel mit Gruppenoperation.* Arch. Math. **20** (1969) 136–143.
Trautmann, G. *Ein Kontinuitätssatz für die Fortsetzung kohärenter analytischer Garben.* Arch. Math. (Basel) **18** (1967) 188–196.
Weyl, H. *The Classical Groups. Their Invariants and Representations.* Princeton University Press, 1939.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'D. Borisov$^a$ and P. Exner$^{b,c}$'
title: '**Exponential splitting of bound states in a waveguide with a pair of distant windows**'
---
> [\
> `[email protected]`, `[email protected]`]{}
> [We consider Laplacian in a straight planar strip with Dirichlet boundary which has two Neumann “windows” of the same length the centers of which are $2l$ apart, and study the asymptotic behaviour of the discrete spectrum as $l\to\infty$. It is shown that there are pairs of eigenvalues around each isolated eigenvalue of a single-window strip and their distances vanish exponentially in the limit $l\to\infty$. We derive an asymptotic expansion also in the case where a single window gives rise to a threshold resonance which the presence of the other window turns into a single isolated eigenvalue.]{}
Introduction
============
Geometrically induced bound states in waveguide systems have attracted a lot of attention recently. The main reason is that they represent an interesting physical effect with important applications in nanophysical devices, but also in flat electromagnetic waveguides – cf. [@LCM] and references therein. At the same time, such a discrete spectrum poses many interesting mathematical questions.
One of the simplest systems of this kind is a straight hard-wall strip in the plane with a “window” or several “windows” in its boundary modeled by switching the Dirichlet boundary condition to Neumann in the Laplace operator which will be the Hamiltonian of our system. By an easy symmetry argument it represents the nontrivial part of the problem for a pair of adjacent parallel waveguides coupled by a window or several windows in the common boundary [@ESTV]; this explains the name we use for the Neumann segments.
The discrete spectrum of such a system is nonempty once a Neumann window is present. Various properties of these bound states were analyzed including their number and behaviour with respect to parameters. Recently we discussed the way in which the eigenvalues emerge from the continuous spectrum as the window width is increasing [@BEG]; we refer to this paper for references to earlier work. Here we address a different question: we consider a strip with a pair of identical Neumann windows at the same side of the boundary and ask about the behaviour of the discrete spectrum as the distance between them grows.
There is a natural analogy with the multiple-well problem in the usual Schrödinger operator theory – see [@BCD] and references therein or [@Da Sec. 8.6] – even if the nature of the effect is different. Recall that in waveguides of the considered type there are no classically closed trajectories apart of the trivial set of measure zero, and likewise, there are no classically forbidden regions. Hence the semiclassical analysis does not apply here, in particular, there is no Agmon metric to gauge the distance of the windows which replace potential wells in our situation.
Nevertheless, the picture we obtain is similar to double-well Schrödinger operators. If the half-distance $l$ between the windows is large, there is pair of eigenvalues, above and below each isolated eigenvalue of the corresponding single-window strip. We will derive an asymptotic expansion which shows that the pair splitting vanishes exponentially as $l\to\infty$ together with the appropriate expansion for the eigenfunctions. On the other hand, the analogy a double-well Schrödinger operator can be misleading. This is illustrated by the case when the single-window strip has a threshold resonance, which turns into a (single) isolated eigenvalue under influence of the other window. We derive the asymptotic expansion as $l\to\infty$ for this case too; it appears that it is exponential again with the power determined by the term coming from the second transverse mode present in the expansion of the resonance wavefunction.
Let us describe briefly the contents of the paper. In the next section we formulate the problem precisely and state two theorems which express our main results. In Section 3 we collect general properties of the involved operator. Before coming to the proper proofs, we analyze in Sections 4 and 5 strip with a single window, in particular, we show how the original question stated in PDE terms can be reformulated as a pair Fredholm problem, the second being obtained from the first one as a perturbation. Finally, in Sections 6 and 7 we prove Thms. \[th1\] and \[th3\].
Formulation of the problem and the\
main results
===================================
Let $x=(x_1,x_2)$ be Cartesian coordinates and suppose that $\Pi$ is a horizontal strip of a width $d$, i.e. $\Pi:=\{x: 0<x_2<d\}$. In the lower boundary of the strip we select two segments of the same length $2a$. The distance between these segments, denoted as $2l$, will be large playing the role of parameter in our asymptotic expansions. We will employ the symbol $\g_l(a)$ for the union of these segments, $\g_l(a):=\g_l^+(a)\cup\g_l^-(a)$, where $\g_l^\pm(a)=\{x: |x_1\mp l|<a,\, x_2=0\}$. The remaining part of the boundary of $\Pi$ will be indicated by $\G_l(a)$ (cf. Figure \[twowindows\]). The main object of our interest are discrete eigenvalues of the Laplacian in $\Pi$ with Dirichlet boundary condition on $\G_l(a)$ and Neumann one on $\g_l$. We denote such an operator by $H_l(a)$ and look what happens if $l\to\infty$.
In order to formulate main results of this paper we need some additional notations and preliminary results concerning a single-window strip. Denote $\g(a):=\{x: |x_1|<a,\, x_2=0\}$, where $\G(a):=\p\Pi\setminus\g(a)$. It was proven in [@ESTV] that the Laplacian in $\Pi$ with Dirichlet condition on $\G(a)$ and Neumann one on $\g(a)$ has (simple) eigenvalues below the threshold of the continuous spectrum for any $a>0$; their number is finite and depends on $a$. We will indicate the operator in question and its eigenvalues by $H(a)$ and $\l_j(a)$, $j=1,\ldots,n$, respectively, with the natural ordering, $\l_1(a)<\l_2(a)<\ldots<\l_n(a)<\frac{\pi^2}{d^2}$, supposing that the corresponding eigenfunctions $\psi_j$ are normalized in $L^2(\Pi)$. Furthermore, it was shown in [@ESTV] that there are critical values of size of Neumann segment, $0=a_0<a_1<a_2<\cdots$, for which the system has in addition a threshold resonance, i.e. the equation $(H(a_n)+1)\psi=0$ has a nontrivial solution $\psi^n(x)$ unique up to a multiplicative constant. This solution and eigenfunction $\psi_j$ mentioned above have a definite parity with respect to $x_1$ and behave in the limit $x_1\to+\infty$ as $$\begin{aligned}
\label{1.1}
&\psi^n(x)=\sqrt{\frac{2}{d}}\,\sin\left(\frac{\pi x_2}{d}\right)
+ \b_n\, \Ex^{-\frac{\pi\sqrt{3}}{d}\,x_1}\sin\left(\frac{2\pi
x_2}{d}\right)+
\mathcal{O}\left(\Ex^{-\frac{\pi\sqrt{8}}{d}\,x_1}\right),
\\
&
\psi_j(x)=\a_j\,\Ex^{-\sqrt{\frac{\pi^2}{d^2}-\l_0}\,x_1}\sin\left(\frac{2\pi
x_2}{d}\right)+
\mathcal{O}\left(\Ex^{-\sqrt{\frac{4\pi^2}{d^2}-\l_0}\,x_1}\right),
\label{1.5}\end{aligned}$$ with some constants $\a_j$, $\b_n$; it is clear that $\a_j=\a_j(a)$. While normalization of $\psi_j$ is natural, the normalization of $\psi^n$ can be arbitrary, of course; we choose it in such a way that asymptotically the function coincides with the first normalized transverse mode. Needless to say, when the window size is made larger than the critical value, the threshold resonance turns into a true eigenvalue.
Now we are ready to formulate the main results.
\[th1\] Let the window length be non-critical, i.e. $a\in(a_{n-1},a_n)$ for some $n\in\mathbb{N}$. Then the operator $H_l(a)$ has for any $l$ large enough exactly $2n$ eigenvalues $\l_j^\pm(l,a)$, $j=1,\ldots,n$, situated in the interval $(\frac{\pi^2}{4d^2},
\frac{\pi^2}{d^2})$. Each of them is simple and has the asymptotic expansions $$\label{asm}
\l_j^\pm(l,a)=\l_j(a)\mp\mu_j(a)\,
\mathrm{e}^{-2l\sqrt{\frac{\pi^2}{d^2}-\l_j(a)}}+
\mathcal{O}\left(\mathrm{e}^{-(4\sqrt{\frac{\pi^2}{d^2}
-\l_j(a)}-\si)l}\right)\,,$$ as $l\to\infty$ for $j=1,\dots,n$, where $\si$ is an arbitrary fixed positive number. The coefficient $\mu_j$ is given by $$\mu_j(a):=\a_j(a)^2d\,\sqrt{\frac{\pi^2}{d^2}-\l_0}\,,\label{ltd1}$$ or alternatively by $$\mu_j(a):=\frac{\pi^2}{d^3\sqrt{\frac{\pi^2}{d^2}-\l_j(a)}}
\left(\int\limits_{\g(a)}\psi_j(x)\,
\mathrm{e}^{\sqrt{\frac{\pi^2}{d^2}-\l_j(a)}\,x_1}\,
dx_1\right)^{\!\!2}.\label{ldt}$$ The eigenfunctions $\psi_j^\pm(x)$ associated with eigenvalues $\l_j^\pm(l,a)$, $j=1,\ldots,n$, have a definite parity being even for $\l_j^+(l,a)$ and odd for $\l_j^-(l,a)$. Furthermore, in the halfstrips $\Pi^\pm:=\{x: \pm x_1>0,\: 0<x_2<d\}$ they can be approximated by $$\begin{aligned}
&\psi_j^+(x)=\psi_j(x_1\mp l,x_2)+\mathcal{O}\left(
\mathrm{e}^{-(2\sqrt{\frac{\pi^2}{d^2} -\l_j(a)}-\si)l}\right),
\\
&\psi_j^-(x)=\pm\psi_j(x_1\mp l,x_2)+\mathcal{O}\left(
\mathrm{e}^{-(2\sqrt{\frac{\pi^2}{d^2} -\l_j(a)}-\si)l}\right),\end{aligned}$$ in $\SS^1(\Pi^\pm)$ as $l\to\infty$.
\[th3\] Let the Neumann segment have a critical size, $a=a_n$. Then the operator $H_l(a)$ has $2n+1$ eigenvalues in $(\frac{\pi^2}{4d^2},
\frac{\pi^2}{d^2})$ for $l$ large enough. The first $2n$ of them together with the associated eigenfunctions behave according to Theorem \[th1\], while the last one, $\l^+_{n+1}(l,a_n)$, exhibits the asymptotics $$\label{1.3}
\l^+_{n+1}(l,a_n) =\frac{\pi^2}{d^2}-
\mu\,\Ex^{-\frac{4\sqrt{3}\pi}{d}\,l}
+\mathcal{O}\left(\Ex^{-\frac{2(\sqrt{8}+\sqrt{3})\pi}{d}\,l}\right)\,,$$ where $$\mu:=3\b_n^4 d^2\,,\label{1.4}$$ or alternatively, $$\label{1.6}
\mu:=\frac{16}{3d^2}\left(\int\limits_{\g(a_n)}\psi^n(x)\,
\Ex^{\frac{\pi\sqrt{3}}{d}\,x_1}\,dx_1\right)^4.$$ The associated eigenfunction $\psi^+_{n+1}$ is even w.r.t. $x_1$ and for any $R$ in the rectangles $\{x: |x_1\mp l|<R\}\cap\Pi$ it can be approximated for large values of $l$ as $$\psi_{n+1}^+(x)= \psi^n(x_1\mp l,x_2)+ \mathcal{O}\left(
\Ex^{-\frac{2\sqrt{3}\pi}{d}\,l}\right)\,$$ in $\SS^1$-norm. In addition it behaves in the limits $x_1\to\pm\infty$ as $$\begin{aligned}
\psi^+_{n+1}(x) &=\sqrt{\frac{2}{d}}\, \Ex^{-\varkappa|x_1|}\sin
{\frac{\pi x_2}{d}}+
\mathcal{O}\left(\Ex^{-\frac{\pi\sqrt{3}}{d}|x_1|}\right)\,,
\\ \varkappa & :=\sqrt{\frac{\pi^2}{d^2}-\l_{n+1}}
=\sqrt{\mu}\,\Ex^{-2\frac{\sqrt{3}\pi}{d}l}+
\mathcal{O}\left(\Ex^{-\frac{2\sqrt{8}\pi}{d}l}\right)\,.\end{aligned}$$
Before proceeding further, let us recall what we have said in the introduction about the analogy with the multi-well problem for Schrödinger operators. As in that case a (simple) eigenvalue of the single window problem gives rise to a pair of eigenvalues (corresponding to eigenfunctions of different parities) which are exponentially close to each other with respect to the window distance, and moreover, in the generic case the splitting is determined by the eigenvalue distance from the threshold. At a glance the multiplicity is doubled by the perturbation, however, in reality the problem decomposes due to mirror symmetry into a pair of problems with definite parities whose eigenvalues tend to the same limit (see below and Sec. 5). On the other hand, the asymptotics (\[1.3\]) in the critical case differs from what the Schrödinger operator analogy would suggest being determined by the distance from the second transverse eigenvalue.
Let us now describe our way to prove Theorems \[th1\] and \[th3\]. The main idea is to reduce the eigenvalue problem at hand to a Fredholm operator equation of the second kind with a regular perturbation. Investigating this problem we will get the result both in the generic situation described in Theorem \[th1\] and for perturbation of a threshold resonance, just the analysis in the latter case is more subtle.
Our task can be simplified by taking into account the symmetry of the problem with respect to reflections, $x_1\to -x_1$, which means that the operator decomposes into orthogonal sum of parts of a definite parity which can be considered separately. This allows us to cut the strip $\Pi$ into a pair of halfstrips $\Pi^\pm$ and to consider the Laplacian in $\Pi^+$ with Dirichlet condition everywhere at the horizontal boundaries of the halfstrip except for $\g^+_l(a)$, where the boundary condition is Neumann. According to the chosen parity of an eigenfunction $\psi$ we impose at that Dirichlet condition for odd eigenfunctions of the original problem at the vertical part of the boundary, $x_1=0$, or Neumann for the even ones. Moreover, it is convenient to shift the halfstrip by $x_1\to x_1-l$ in order to fix position of the Neumann segment of the boundary. As a result, we arrive at the following pair of eigenvalues problems, $$\label{1.2}
\begin{aligned}
{}&-\D\psi=\l\psi\,,\quad x\in\Pi^l\,, \\ \psi=0\,,\;\;
x\in\G(a)\,,\quad\; {}&\frac{\p\psi}{\p x_2}=0\,,\;\;
x\in\g(a)\,,\quad\; hu=0\,,\;\; x_1=-l\,.
\end{aligned}$$ Here $\Pi^l:=\{x\in\Pi:\: x_1>-l\}$ is the shifted halfstrip and $h$ is the boundary operator which acts as $h u=u$ or $h
u=\frac{\p u}{\p x_1}$ in the odd and even case, respectively. Eigenvalues of (\[1.2\]) obviously coincide with those of $H_l(a)$ and by the even/odd extension one gets the eigenfunctions of the original problem.
Finally, we remark that the problem has a simple behaviour with respect to scaling transformations which allows us to perform the proofs for $d=\pi$ only.
Preliminaries
=============
Let us collect first some general properties of the spectrum of our operators.
\[lmpr\] The discrete spectrum of the operator $H_l(a)$ is non-empty for any $l>a>0$. It consists of a finite number of simple eigenvalues contained in the interval $\left(\frac{1}{4},1 \right)$ for $d=\pi$ which depend continuously on $l$ and $a$; for a fixed $a$ those corresponding to even and odd eigenfunctions are increasing and decreasing, respectively, as functions of the window separation parameter $l$. All eigenvalues of $H_l(a)$ which remain separated from the continuum converge to those of $H(a)$ as $l\to+\infty$, and to each eigenvalue of $H(a)$ there exists a pair of eigenvalues of $H_l(a)$ associated with eigenfunctions of opposite parities converging to that eigenvalue of $H_l(a)$. If the Neumann segment has a critical width, $a=a_n$, then there is a unique eigenvalue (corresponding to an even eigenfunction) which tends to one as $l\to+\infty$.
[**Proof.**]{} By the minimax principle and an elementary bracketing estimate the eigenvalues of $H_l(a)$ can be squeezed between those of $H(l+a)$ and $H(a)$. The essential spectrum of all the three operators is the same being equal to $[1,\infty)$; this fact in combination with the results of [@ESTV] shows that $\sigma_\mathrm{disc} (H_l(a))$ is non-empty, finite, and contained in $\left(\frac{1}{4},1 \right)$. A similar bracketing argument shows that the eigenvalues $\l_j^\pm(l,a)$ of the problem (\[1.2\]) with Neumann and Dirichlet boundary condition at $x_1=-l$, respectively, satisfy $\l_j^+(l,a) \le \l_j(a) \le
\l_j^-(l,a)$ for $j=1,\dots,n$, where the upper bound is replaced by one if the Dirichlet problem has less than $j$ eigenvalues. In fact, bracketing implies also the stated monotonous behaviour with respect to $l$, i.e. $$\label{DNest}
\l_j^+(l',a) \le \l_j^+(l,a) \le \l_j(a) \le \l_j^-(l,a) \le
\l_j^-(l',a)$$ for $l'\ge l$ with the same convention as above; it is sufficient to write $\Pi_{l'}$ as a union of $\Pi_l$ and a rectangle separated by an additional Neumann or Dirichlet boundary condition and to realize that in neither of these cases the rectangle can contribute to the spectrum below the continuum threshold, because it has Dirichlet condition at the horizontal part of the boundary. In addition, the standard domain-changing argument [@Kt Sec. VII.6.5] shows that the functions $\l_j^\pm(\cdot,a)$ are continuous. In view of the monotonicity mentioned above their limits as $l\to\infty$ exist; it remains to check that $\l_j^\pm(l,a)\to \l_j(a)$.
Take $\psi\in D(H(a))$ and a function $g\in C_0^\infty$ such that $g(x)=0$ for $x\le 0$ and $g(x)=1$ for $x\ge 1$. Denoting $h_l(x,y):= g(2(x+l)/l)$ we can construct a family $\{\psi_l\}$ by $\psi_l(x,y):= \psi(x,y)h_l(x,y)$; by construction the function $\psi_l$ belongs to the domain of $H^-_l(a)$ which is the Laplacian with the boundary condition as in (\[1.2\]) for $hu=u$. Using the fact that $\|\nabla h_l\|^2= 2\|g'\|^2 l^{-1}$ and $\|\Delta h_l\|^2= 8\|g''\|^2 l^{-3}$ one can check easily that $\psi_l\to \psi$ and $H^-_l(a)\psi_l \to H(a)\psi$ as $l\to
\infty$, so $H^-_l(a)\to H(a)$ in the strong-graph sense. By [@RS Thm VIII.26] this is equivalent to the strong resolvent convergence, hence to each $\l_j(a)$ there is a family of $\l^-_j(a)$ converging to that value. Since the spectrum of $H_l(a)$ in $(\frac{1}{4},1)$ is discrete, simple, finite, and depends monotonously on $l$, we get the desired result. In a similar way one can check that $\l_j^+(l,a)\to \l_j(a)$ as $l\to\infty$.
The continuity w.r.t. $a$ is proved as the $a$-continuity in case of a single Neumann window. We expand the solution inside and outside the window regions with respect to the appropriate transverse bases and match the Ansätze smoothly at the window edges. This yields an infinite family of linear equations for the coefficients of the expansions, which can be regarded as a search for the kernel of a certain operator in the $\ell^2$ space of the coefficients with a properly chosen weight. One has to check that this operator is Hilbert-Schmidt and continuous with respect to the parameters in the Hilbert-Schmidt norm. The argument is analogous to that from the proof of Proposition 2.1 of [@BEG], so we skip the details; the only difference is that due to the lack of symmetry the matching has to be performed at each window separately and the coefficient space is “twice as large”.
It remains to check the last claim. Using bracketing once more we see that if the presence of the other window turns a threshold resonance into an eigenvalue, the corresponding eigenfunction must be symmetric; in view of the proved monotonicity it is sufficient to show that this happens for $l$ large enough. Since this part of the proposition is not used in the proof of the claim of Theorem \[th3\] concerning the first $2n$ eigenvalues $\l_j^\pm$, we may assume that it is already proven and that we thus know that for large $l$ the operator $H_l(a_n)$ posesses $2n$ eigenvalues $\l_j^\pm \in \left(\frac{1}{4},1 \right)$ corresponding to eigenfunctions $\psi_j^\pm$. We seek a $(2n+1)$-dimensional subspace such that for any $\psi$ from it we have $(\psi,H_l(a_n)\psi)- \|\psi\|^2<0$. To this aim we employ a Goldstone-Jaffe-type argument inspired by [@ESTV] and choose $$\psi= c_0(\chi_{L,\varsigma}\psi^n + \varepsilon p)+ \sum_{j=1}^n
c_j^\pm \psi_j^\pm\,,$$ where $\psi^n$ is the resonance function (\[1.1\]), $\chi_{L,\varsigma}:\mathbb{R}\to (0,1]$ equals one in $(-L,L)$ for some $L>l+a$ and $\chi_{L,\varsigma}= \exp(-\varsigma(|x|-L))$ otherwise, and $p$ is a $C_0^\infty$ function supported in the other window region. In view of the asymptotic behaviour (\[1.1\]) such functions span a subspace of the needed dimension. Evaluating the energy form $(\psi,H_l(a_n)\psi)-
\|\psi\|^2$ we see that if some of the coefficients $c_j^\pm$ is nonzero, it is negative even with $\varepsilon=0$. In the opposite case we use the fact that in the leading term we have, as in [@ESTV], two competing terms, one linear in $\varepsilon$ and the other positive coming from the tails of $\psi$ controlled by the parameters $L$ and $\varsigma$; we can choose them in such a way that the form is negative again.
Hence $H_l(a_n)$ has for $l$ sufficiently large at least $2n+1$ eigenvalues. In fact, it has exactly this number, because its symmetric and antisymmetric parts have for large $l$ enough $l+1$ and $l$ eigenvalues, respectively, otherwise we would have an contradiction with the monotonicity and continuity properties stated above. In particular, the largest eigenvalue is increasing w.r.t. $l$ since it corresponds to an even eigenfunction. In view of (\[DNest\]) and the fact that $a_n$ is the critical width, we conclude that $\l_{n+1}^+\to 1-$ as $l\to\infty$.
Analysis of the limiting operator
=================================
After these preliminaries let us pass to the proper subject of the paper. First we are going to discuss the limiting, i.e. one-window operator which means to analyze the following boundary value problem, $$\label{2.1}
-(\D+\l) u=f\,,\;\; x\in\Pi\,,\qquad u=0\,,\;\, x\in\G(a)\,,
\quad\; \frac{\p u}{\p x_2}=0\,,\;\, x\in\g(a)\,.$$ The right hand side $f$ is here assumed to be finite and to belong to $L^2(\Pi)$; our aim is to discuss the existence and uniqueness of the solution to (\[2.1\]) as well as its dependence on $\l$. The method we use is to reduce (\[2.1\]) to a Fredholm operator equation. Then our task will be reduced to analysis of operator families, in particular their holomorphic dependence on the spectral parameter $\l$ (one need not specify at that the topology – cf. [@RS Sec. VI.3]). The reduction will follow a general scheme proposed by Sanchez-Palencia [@SP] and it will be analogous to the treatment of a similar problem in [@BEG Sec. 3.1].
We will use the symbol $\mathcal{D}_\d$ to indicate the open subset $\{\l: \RE\l<\d\}$ of the complex plane. The structure of the solution to the problem (\[2.1\]) for $\l$ close to $1$ and for $\l$ separated from $1$ is different. This is the reason why we will consider these two cases separately. We suppose first that $\l\in\mathcal{D}_\d$, where $\l_n(a)<\d<1$. In this situation it is sufficient to consider solutions of the problem (\[2.1\]) in the class of functions which behave as $\mathcal{O}
(\Ex^{-\sqrt{1-\l}|x_1|})$ in the limit $|x_1|\to\infty$.
Since the function $f$ is finite by assumption, its support lies inside the rectangle $\Pi_b:=\Pi\cap\{x: |x_1|<b\}$ for some $b>0$. Consider two boundary value problems, $$-(\D+\l) v^\pm=g\,,\quad x\in\Pi^\pm\,,\qquad v^\pm=0\,,\quad
x\in\p\Pi^\pm\,, \label{2.2}$$ where $g$ is an arbitrary function from $L^2(\Pi)$ with the support contained in $\Pi_A$ for some $A\ge\max\{a,b-1\}$. This choice is given by the requirement that $\Pi_A$ contains both the window and the support of $f$ in such a way which will make the smooth interpolation (\[2.5\]) used below possible. The problems (\[2.2\]) can be easily solved by separation of the variables; using the explicit form of Green’s function of Laplace-Dirichlet problem on a halfline we get $$\begin{gathered}
v^\pm(x)=\int\limits_{\Pi^\pm} G^\pm(x,t,\l)\,g(t)\,
d^2t\,,\label{2.3} \\
G^\pm(x,t,\l)=\sum\limits_{j=1}^\infty\frac{1}{\pi
\kappa_j(\l)}\left(\Ex^{-\kappa_j(\l)|x_1-t_1|}-\Ex^{\mp
\kappa_j(\l)(x_1+t_1)}\right)\, \sin jx_2\, \sin jt_2\,,
\label{2.12}\end{gathered}$$ where $\kappa_j(\l):=\sqrt{j^2-\l}\,$. In the following we will also employ the “glued” function $v$ equal to $v^+$ if $x_1\ge
0$ and to $v^-$ if $x_1<0$. The functions $v^\pm$ can be naturally regarded as results of action of the bounded linear operators $T_1^\pm(\l)$, i.e. we have $v^\pm=T_1^\pm(\l)g$, where $T^\pm_1:
L^2(\Pi^\pm_A)\to \SS^2(\Pi^\pm)$ with the “halved” rectangles $\Pi^\pm_A=\Pi\cap\{x: 0<\pm x_1<A\}$. It is easy to check that the operator families $T^\pm_1$ are holomorphic in $\l\in
\mathcal{D}_\d$. In the next step we consider the problem $$\label{2.4}
\D w=\D v\,,\;\; x\in\Pi_A\,,\quad\; \frac{\p w}{\p x_2}=0\,,\;\;
x\in\g(a)\,,\quad w=v\,,\;\; x\in\p\Pi_A\setminus\g(a)\,.$$ The function $v$ may have according to its definition given above a weak discontinuity, i.e. a jump of the first derivatives. Thus we have to say what we mean by $\D v$ in (\[2.4\]): it is the function from $L^2(\Pi)$ which coincides with $\D v^+$ if $x_1>0$ and with $\D v^-$ if $x_1<0$. With the problem (\[2.2\]) in mind we can also write $\D v=-(\l v+g)$. The problem (\[2.4\]) is posed in a bounded domain, hence the standard theory of elliptic boundary value problems is applicable. In particular, we can infer using [@Ld] that the function $w$ exists, it is unique and belongs to $\SS^1(\Pi_A)$. We will also consider its restriction avoiding the points where the boundary condition changes, regarded as an element of $\SS^1(\Pi_A)\cap\SS^2(\Pi_{A}\setminus S_r)$ for any $r>0$, where $S_r=\{x: (x_1\pm a)^2+x^2_2<r^2\}$. In this way we introduce a linear bounded operator $T_2: L^2(\Pi_A)\to
\SS^1(\Pi_A)\cap\SS^2(\Pi_{A}\setminus S_r)$ (for any $r$) such that $w=T_2g$.
Next we employ a smooth interpolation. Let $\chi$ be an infinitely differentiable mollifier function such that $\chi(\tau)=1$ if $|\tau|<A-1$ while for $|\tau|>A$ it vanishes. We will construct a solution to the problem (\[2.1\]) interpolating between the functions $v$ and $w$, specifically $$\label{2.5}
u(x)=\chi(x_1)w(x)+(1-\chi(x_1))v(x).$$ Since $w=T_2 g$ and $v^\pm=T_1^\pm(\l) g$, we can also regard $u$ as the result of an action of some linear operator $T_3(\l)$ which maps $L^2(\Pi_A)$ into $\SS^1(\Pi)\cap\SS^2(\Pi\setminus
S_r)$ for a fixed $r>0$. Such an operator $T_3$ is linear and bounded, and as an operator family with respect to $\l$ it is again holomorphic.
Owing to the definition of $w$ and $v$ the function $u$ satisfies all the boundary conditions involved in (\[2.1\]), and consequently, it represents a solution to (\[2.1\]) if and only if it satisfies the differential equation in question. Substituting (\[2.5\]) into the latter and taking into account (\[2.2\]), (\[2.4\]), we arrive at the equation $$\label{2.6}
g+T_4(\l)g=f\,,$$ where $T_4: L^2(\Pi_A)\to L^2(\Pi_A)$ is a linear bounded operator defined by $$\label{2.7}
T_4(\l)g:=-2\nabla_x \chi \cdot\nabla_x (w-v) -(w-v)(\D+\l)\chi\,,$$ where the dot in the first term denotes the inner product in $\mathbb{R}^2$. The relation (\[2.6\]) is the sought Fredholm equation, considered in the space $L^2(\Pi_A)$. Naturally the first thing to do here is to check the compactness of the operator $T_4$. It can be done as follows. The function $w-v$ belongs to $\SS^1(\Pi_A)$, thus the operator mapping $g$ into $w-v$ is bounded as an operator from $L^2(\Pi_A)$ into $\SS^1(\Pi_A)$, and consequently, it is compact as an operator in the space $L^2(\Pi_A)$; this solves the question for the second term at the right-hand side of (\[2.7\]). Furthermore, due to the definition of the mollifier $\chi$ the support of $\nabla_x \chi$ lies within $\overline{\Pi}_A\setminus\Pi_{A-1}$. This domain does not contain the endpoints of the segment $\g$. Hence $w-v\in\SS^2(\supp{
\nabla_x \chi})$, and therefore $\nabla_x (w-v)$ considered as an element of $L^2(\supp{\nabla_x \chi})$ results from action of a compact operator mapping $L^2(\Pi_A)$ into $L^2(\supp{\nabla_x
\chi})$; this concludes the proof of compactness $T_4(\l)$ considered as an operator in the space $L^2(\Pi_A)$. In a similar way one can check that $T_4(\l)$ is a holomorphic operator family w.r.t. $\l$.
This conclusion allows us to apply to (\[2.6\]) the standard Fredholm technique; we will see that solution to (\[2.6\]) exists and is unique for almost all $\l$ except for points where a nontrivial solution for (\[2.6\]) with zero right-hand side exists. This will yield a solution to our original problem because the two are equivalent; this is the contents of the following lemma the proof of which we skip because it is completely analogous to that of Proposition 3.2 in [@BEG].
\[lm2.1\] To any solution $g$ of (\[2.6\]) there is a unique solution $u=T_3(\l)g$ of (\[2.1\]), and vice versa, for each solution of (\[2.1\]) there exists a unique $g$ solving (\[2.6\]) such that $u=T_3(\l)g$. The equivalence holds for any $\l\in
\mathcal{D}_\d$.
Thus the equation (\[2.6\]) says how to find a bounded solution to (\[2.1\]): one should solve the equation (\[2.7\]) and then to construct the solution of (\[2.1\]) by the procedure described above, i.e. by putting $u=T_3(\l)g$.
Since the operator family $T_4(\l)$ is holomorphic, the corresponding resolvent family $(I+T_4(\l))^{-1}$ is meromorphic and its only poles are exactly the eigenvalues of $H(a)$ – cf. [@SP Chap. 16, Th. 7.1]. In order to prove Theorem \[th1\], we need to know more about the behavior of $(I+T_4(\l))^{-1}$ in the vicinity of these poles.
\[lm2.2\] Let $\l_0<1$ be an eigenvalue of $H(a)$. Then for any $\l$ close enough to $\l_0$ the following representation is valid, $$\label{2.8}
(I+T_4(\l))^{-1}=\frac{\phi}{\l-\l_0}T_5+T_6(\l)\,,$$ where $T_5 f:=-(f,\psi)_{L^2(\Pi)}$ and $T_6: L^2(\Pi_A)\to
L^2(\Pi_A)$ is a bounded linear operator which is holomorphic in $\l$. Furthermore, $\phi$ is such that $\psi=T_3(\l_0)\phi$, where $\psi$ is an eigenfunction of $H(a)$ associated with $\l_0$ and normalized in $L^2(\Pi)$.
[**Proof.**]{} We assume throughout that $\l\in\mathcal{D}_\d$ lies in a small neighborhood of $\l_0$ containing no other eigenvalues of $H(a)$. As we have already mentioned, the operator family $(I+T_4(\l))^{-1}$ has a pole at $\l_0$. It means that the vector-valued function $g:\lambda\mapsto
(I+T_4(\l))^{-1}f$ satisfies $$\label{2.9}
g(\lambda) =\frac{g_{-q}}{(\l-\l_0)^q}+\frac{\t
g(\l)}{(\l-\l_0)^{q-1}}\,,$$ where $q$ is a positive integer and $\t g$ is holomorphic in $\l$. Substituting this representation into (\[2.6\]) and calculating the coefficients of $(\l-\l_0)^{-q}$ we see that $g_{-q}$ must satisfy the equation $g_{-q}+T(\l_0)g_{-q}=0$, in other words $g_{-q}=\phi T_5 f$, where $T_5 f$ is a number depending on $f$. Together with (\[2.9\]) this means that the solution to (\[2.1\]) associated with $g$, i.e. $u=T_3(\l)g$, can be written as $$\label{2.10}
u(x,\l)=\frac{T_5 f}{(\l-\l_0)^q}\psi(x)+\frac{\t
u(x,\l)}{(\l-\l_0)^{q-1}}\,,$$ where is $\t u$ is holomorphic in $\l$. Due to the definition of $T_3(\l)$ this formula is valid in the sense of $\SS^1(\Pi)$-norm as well as in $\SS^2(\Pi\setminus\Pi_A)$. Taking the inner product of (\[2.1\]) with $\psi$, using the fact that the latter is an eigenfunction of $H(a)$, and performing an integration by parts in $\Pi_R$ with $R$ large enough we find $$\label{2.11}
-\int\limits_{\p\Pi_R}\left(\psi\frac{\p u}{\p\nu}-u\frac{\p
\psi}{\p\nu}\right)+\l_0(u,\psi)_{L^2(\Pi_R)}=(f,\psi)_{L^2(\Pi_R)}+
\l(u,\psi)_{L^2(\Pi_R)}\,.$$ The functions $u$ and $\psi$ behave at infinity as $\mathcal{O}(\Ex^{-|x_1|\sqrt{1-\l}})$ and $\mathcal{O}(\Ex^{-|x_1|\sqrt{1-\l_0}})$, respectively. With this fact in mind we can pass to the limit $R\to\infty$ in (\[2.11\]) for each fixed value of $\l$; this implies the identity $$\l_0(u,\psi)_{L^2(\Pi)}=(f,\psi)_{L^2(\Pi)}+
\l(u,\psi)_{L^2(\Pi)}\,.$$ Substituting to it from (\[2.10\]) and computing the coefficients at the same powers of $\l\!-\!\l_0$, we see first that $q=1$, and furthermore, that $T_5 f=-(f,\psi)_{L^2(\Pi)}$. This completes the proof.
We will also need to know the behavior of the inverse $(I+T_4(\l))^{-1}$ as $\l\to 1$. For the right hand side $f$ in (\[2.1\]) with a definite parity w.r.t. $x_1=0$ it was done in [@BEG], here we have just to show how to extend this result to our case. We will assume that $\l$ lies in a small neighborhood of one and that this neighborhood contains no eigenvalues of $H(a)$. First of all, however, we should characterize the class of functions in which we will seek the solution of (\[2.1\]) in this case. Instead of $\l$ we introduce another parameter by setting $\l=1-\kappa^2$, where $\kappa$ lies in a small neighborhood of zero. The only restriction to the size of this neighborhood correction is that the associated values of $\l$ should not coincide with eigenvalues of the operator $H(a)$. If $\kappa$ is real and a solution to the problem (\[2.1\]) exists, it is unique and holomorphic in $\kappa$. This fact follows from the arguments given above, because for such a $\l=1-\kappa^2$ the equation (\[2.6\]) is uniquely solvable. The said solution can be extended to all values of $\kappa$ in the vicinity of zero so that this extension will be an analytic function of $\kappa$. The existence of such an extension is guaranteed by the definition of the functions $v^\pm$ in (\[2.3\]) where $\kappa_1(\l)$ is nothing else than $\kappa$ introduced above. We see that the formulae (\[2.3\]) are valid not only for real $\kappa$ but also in a complex neighborhood including $\kappa=0$, because the kernels (\[2.12\]) have finite limits as $\kappa=0$, namely $$\label{2.13}
\begin{aligned}
G^\pm(x,t,1)={}&-\frac{1}{\pi}(|x_1-t_1|\mp(x_1+t_1))\sin x_2 \sin
t_2 \\
{}& +\sum\limits_{j=2}^\infty\frac{1}{\pi
\kappa_j(1)}\left(\Ex^{-\kappa_j(1)|x_1-t_1|}-\Ex^{\mp
\kappa_j(1)(x_1+t_1)}\right)\sin jx_2 \sin jt_2\,.
\end{aligned}$$ This is why we are able to extend the solution of the problem (\[2.1\]) analytically to all values of $\kappa$ in the vicinity of zero. We should also stress that the function $u$ given by (\[2.5\]) decays exponentially at infinity if $\RE \kappa>0$, it is bounded for $\RE \kappa=0$ and increases exponentially provided $\RE \kappa<0$.
In this approach all the operators introduced above preserve their properties when we vary the range of the variables passing from unbounded domains the cut-off ones treating, for instance, $T_1^\pm$ as operators mapping $L^2(\Pi_A^\pm)$ into $\SS^2(\Pi_R^\pm)$ for any $R$. As another notational simplification we will not introduce an extra symbol for the composed mapping $\kappa\mapsto T_i(1\!-\!\kappa^2)$ and write instead just $T_i(\kappa)$.
Mimicking the argument used in the proof of [@BEG Thm 3.4], one can check the following claim:
\[lm2.3\] If the Neumann segment of the boundary does not have a critical size, the operator $(I+T_4(\kappa))^{-1}$ exists and is uniformly bounded in $\kappa$ in the vicinity of zero. In the opposite case, i.e. $a=a_n$, we have in a punctured neighborhood of zero the following representation, $$\label{2.14}
(I+T_4(\kappa))^{-1}=\frac{\phi^n}{\kappa}T_7+T_8(\kappa)\,,$$ where $T_7 f:=\frac{1}{2}(f,\psi^n)_{L^2(\Pi)}$ and $T_8:
L^2(\Pi_A)\to L^2(\Pi_A)$ is a bounded linear operator which is holomorphic in $\kappa$. Furthermore, $\phi^n$ is such that $\psi^n=T_3(\kappa=0)\phi^n$, where $\psi^n$ solves the equation $(H(a_n)+1)\psi^n=0$ and behaves at infinity in accordance with (\[1.1\]).
Analysis of perturbed operator
==============================
The main purpose of this section is to reduce the problem $$\label{3.1}
\begin{aligned}
{}&-(\D+\l)u=f\,,\quad x\in\Pi^l\,, \\
\psi=0\,,\;\; x\in\G(a)\,,\quad\; {}&\frac{\p\psi}{\p x_1}=0\,,
\;\; x\in\g(a)\,,\quad\; hu=0\,,\;\; x_1=0\,, \phantom{AAAA}
\end{aligned}$$ to an operator equation similar to (\[2.6\]). We will show that the problem (\[3.1\]) can be reduced to solution of a Fredholm equation which is a regular perturbation of the equation (\[2.6\]). We will start from the case of Dirichlet condition at the cut $x_1=0$, i.e., $hu=u$. We are going to employ the same scheme as in previous section and use the same notations unless stated otherwise.
First we will treat the case $\l\in \mathcal{D}_\d$. In analogy with (\[2.2\]) we consider two problems, $$\begin{aligned}
{}&-(\D+\l) v^+_l=g\,,\;\; x\in\Pi^+\,,\quad\; v^+_l=0\,,\;\;
x\in\p\Pi^+\,,\label{3.2}
\\
{}&-(\D+\l) v^-_l=g\,,\;\; x\in\Pi^-_l\,,\quad\; v^-_l=0\,,\;\;
x\in\p\Pi^-_l\,.\label{3.3}\end{aligned}$$ The first one coincides with (\[2.2\]) for $v^+_l$, while in (\[3.3\]) we take into account the perturbation. Consequently, we have $v^+_l:= v^+$, where $v^+$ is the function from (\[2.3\]). The problem (\[3.3\]) differs from (\[2.2\]) but it can be solved again by separation of variables. It is convenient to write its solution $v_l^-$ in the following form, $$\begin{gathered}
v^-_l(x)=v^-(x)+\int\limits_{\Pi^-}
G^-_l(x,t,\l)\,g(t)\,d^2t\,,\label{3.4}
\\
G^-_l(x,t,\l)=-\sum\limits_{j=1}^\infty\frac{2\,\Ex^{-\kappa_j(\l)
l}}{\pi \kappa_j(\l)\sinh \kappa_j(\l) l}\,\sinh \kappa_j(\l)
x_1\,\sinh \kappa_j(\l) t_1 \sin j x_2\,\sin j t_2\,,\label{3.7}\end{gathered}$$ where $v^-$ is given by (\[2.3\]); we keep in mind here that $g$ is finite, and therefore its support lies inside $\Pi_l$ for all $l$ large enough. As in previous section we can introduce a linear bounded operator $T_{9}(\l): L^2(\Pi_A^-)\to \SS^2(\Pi^-_l)$ such that $v^-_l=T_{9}(\l,l) g$. This operator can be represented as the sum $T_{9}(\l,l)=T_1^-(\l)+T_{10}(\l,l)$, where $T_{10}(\l,l):
L^2(\Pi_A^-)\to \SS^2(\Pi^-_l)$ is holomorphic in $\l$, jointly continuous with respect to $(\l,l)$ provided $\l\in\mathcal{D}_\d$, $l\in[l_0,+\infty]$, and $l_0$ is a fixed number large enough. The norm of the operator $T_{10}$ is of order $\mathcal{O}(\Ex^{-l\sqrt{1-\l}})$ as $l\to+\infty$, hence for $\l\in\mathcal{D}_\d$ we may consider this operator as an exponentially small perturbation.
The analogue of the function $w$ (denoted here by $w_l$) is defined as above without any changes, i.e. as a solution of the problem (\[2.4\]) with $v$ replaced by $$v_l:=\left\{
\begin{aligned}
{}&v^+_l\,,\quad x_1>0\,,
\\
{}& v^-_l\,,\quad x_1<0\,.
\end{aligned}\right.$$ The solution of (\[3.1\]) is then constructed as an interpolation (\[2.5\]) with $v$ and $w$ replaced by $v_l$ and $w_l$; this leads us to the desired operator equation, $$g+T_4(\l)g+T_{11}(\l,l)=f\,.\label{3.5}$$ Here $T_4$ is the operator appearing in (\[2.6\]) and $T_{11}(\l,l):L^2(\Pi_A)\to L^2(\Pi_A)$ is a compact linear operator which is holomorphic in $\l$ and jointly continuous w.r.t. $(\l,l)$ provided $\l\in\mathcal{D}_\d$, and $l\in[l_0,+\infty]$. The norm of the last named operator is exponentially small as $l\to+\infty$ uniformly in $\l\in\mathcal{D}_\d$: $$\label{3.10}
\|T_{11}\|=\mathcal{O}(\Ex^{-2l\sqrt{1-\l}})\,.$$ The solution to the problem (\[3.1\]) can be reconstructed from the function $g$ by $u=T_3(\l)g+T_{12}(\l,l)g$, where $T_{12}:
L^2(\Pi_A)\to\SS^1(\Pi^l)$ is a linear bounded operator the norm of which satisfies $$\label{3.8}
\|T_{12}\|=\mathcal{O}(\Ex^{-l\sqrt{1-\l}})\,.$$ This operator is also holomorphic in $\l$ and jointly continuous with respect to $(\l,l)\in \mathcal{D}_\d\times[\l_0,+\infty]$. The equation (\[3.5\]) is a second-kind Fredholm operator equation and it is equivalent to the problem (\[3.1\]); this claim can be checked in the same way as we did it for (\[2.6\]) in the previous section.
The case of $hu=\frac{\p u}{\p x_1}$ is treated in full analogy. The only difference due to another boundary condition at $x_1=0$ is the definition of the operator $T_{10}$ which is now described by the kernel $$\label{3.6}
G^-_l(x,t,\l)=\sum\limits_{j=1}^\infty\frac{2\Ex^{-\kappa_j(\l)
l}}{\pi \kappa_j(\l)\cosh \kappa_j(\l) l}\, \sinh \kappa_j(\l)
x_1\, \sinh \kappa_j(\l) t_1\,\sin j x_2\,\sin j t_2\,.$$ All the arguments used above remain valid.
On the other hand, for $\l$ in the vicinity of one almost all the above arguments remain valid provided we replace $\l$ by $(1-\kappa^2)$. In analogy with the previous section the operators introduced here may be considered on cut-off strips, i.e. as $T_9(\kappa,l):L^2(\Pi_A^-)\to \SS^2(\Pi^-_R)$, $T_{10}(\kappa,l):L^2(\Pi_A^-)\to \SS^2(\Pi^-_R)$, $T_{11}(\kappa,l):L^2(\Pi_A)\to L^2(\Pi_A)$, $T_{12}(\kappa,l):
L^2(\Pi_A)\to\SS^1(\Pi_R)$ for any fixed $R$. However, we are not longer allowed to say that these operators are holomorphic in $\kappa$ because of the terms $$\frac{2\Ex^{-\kappa l}}{\sinh \kappa l}\,,\quad
\frac{2\Ex^{-\kappa l}}{\cosh \kappa l}$$ in (\[3.7\]), (\[3.6\]), since these terms have poles at $\kappa=\frac{\pi\mathrm{i}}{l} j$ and $\kappa=\frac{\pi\mathrm{i}}{l} (j+\frac{1}{2})$. Moreover, these terms are also responsible for the fact that the operators have no proper limit as $\kappa\to0$ and $l\to+\infty$. At the same time, restricting the range of $\kappa$ we will be able to show that the operators $T_{10}$, $T_{11}$, and $T_{12}$ are small for small $\kappa$ and large $l$, thus we will be allowed to consider them as small perturbations again. This claim leans on the following lemma.
\[lm3.2\] Let $\k\in(0,\frac{\pi}{2})$ be fixed and $\mathcal{Q}_\k:=
\{\kappa: |\arg{\kappa}\pm\frac{\pi}{2}|\ge\k\}$. Then there is $C>0$ such that for small $\kappa\in \mathcal{Q}_\k$ and large $l$ the following estimate is valid, $$\max\left\{\left|\frac{\kappa\,\Ex^{-\kappa l}}{\sinh{\kappa
l}}\right|,\, \left|\frac{\kappa\,\Ex^{-\kappa l}}{\cosh{\kappa
l}}\right| \right\}\le C\left(|\kappa|+l^{-1}\right)\,.$$
[**Proof.**]{} We will show how to derive the first estimate, the proof of the second one is similar. We start by introducing the function $$P(z):=\frac{z}{\Ex^z-1}\,.$$ Suppose that $z\in\mathcal{Q}_\k$. If we have in addition $|z|\le1$, one can check that $$\label{3.9}
|P(z)|\le C$$ with some $C$ independent on $z$. On the other hand, if $|z|>1$, $z\in\mathcal{Q}_\k$, and $\RE z>0$, then the exponent in the function $P$ increases as $|z|\to\infty$ and we arrive at (\[3.9\]) again (in general with another $C$). Finally, if $|z|>1$, $z\in\mathcal{Q}_\k$, and $\RE z<0$ then the exponent in the function $P$ decreases and we have a uniform estimate, $$|P(z)|\le C|z|.$$ Combining it with (\[3.9\]) we get the inequality $$|P(z)|\le C_1|z|+C_2$$ valid for $z\in\mathcal{Q}$ and suitable $C_1,C_2$. The obvious identity $$\frac{\kappa\,\Ex^{-\kappa l}}{\sinh \kappa l}=\frac{1}{l}\,
P(2\kappa l)$$ then completes the proof of the lemma.
Using this result one can check that the operators $T_{10}$, $T_{11}$, and $T_{12}$ are small for small $\kappa\in
\mathcal{Q}_\k$ and large $l$, holomorphic in $\kappa$, and jointly continuous in $(\kappa,l)$.
Proof of Theorem \[th1\]
========================
In this section we are going to derive the asymptotic expansions for the eigenvalues of $H_l(a)$ separated from the continuum. We will also find the asymptotic behavior of the associated eigenfunctions.
The main idea behind the calculation of the asymptotics is borrowed from [@Ga1; @Ga2; @BEG]. Instead of dealing with eigenvectors of $H_l(a)$ directly we consider here those of the problems (\[1.2\]). In order to find eigenvalues of the latter we should look in accordance with the results of the previous sections for $\l$ such that the operator equation $$\label{4.1}
\Phi+T_4(\l)\Phi+T_{11}(\l,l)\Phi=0$$ has a nontrivial solution. We will deal with eigenvalues which are close to a fixed eigenvalue $\l_j(a)$ of the limiting operator $H(a)$; for simplicity we will denote the latter as $\l_0$ in the following. Also the parameter $\l$ will be assumed to be close to $\l_0$, more specifically, it will be supposed to lie in a neighborhood of $\l_0$ containing neither any other limiting eigenvalue nor the point $\l=1$.
By the definition of $T_{11}$ the term $T_{11}(\l,l)\Phi$ in (\[4.1\]) is supported inside $\Pi_A$. Hence considering it as the right hand side, we arrive at the equation (\[2.6\]) with $f=-T_{11}(\l,l)\Phi$. Choosing $\l\not=\l_0$, we can invert the operator $I+T_4(\l)$ obtaining $$\Phi+(I+T_4(\l))^{-1}T_{11}(\l,l)\Phi=0\,.$$ Using the Lemma \[lm2.2\], we can rewrite the last equation in the form $$\label{4.2}
\Phi-\frac{\phi}{\l-\l_0}(\psi,T_{11}(\l,l)\Phi)_{L^2(\Pi)}
+T_6(\l)T_{11}(\l,l)\Phi=0\,;$$ recall that $\psi\in L^2(\Pi)$ here is the normalized eigenfunction associated with $\l_0$ and $\phi\in L^2(\Pi_A)$ is a function such that $\psi=T_3(\l_0)\phi$.
The operator $T_{11}(\l,l)$ is small in the asymptotic region, $l\to+\infty$, while $T_6(\l)$ is holomorphic in $\l$. Thus we may invert the operator $I+T_6(\l)T_{11}(\l,l)$ and apply the result to the equation (\[4.2\]), which then acquires the form $$\label{4.3}
\Phi-\frac{1}{\l-\l_0}(\psi,T_{11}(\l,l)\Phi)_{L^2(\Pi)}
(I+T_6(\l)T_{11}(\l,l))^{-1}\phi=0\,.$$ The inner product $(\psi,T_{11}(\l,l)\Phi)_{L^2(\Pi)}$ does not vanish. Indeed, otherwise the function $\Phi$ would be zero too, however, we seek a nontrivial solution of the equation (\[4.1\]). With this fact in mind, we express the function $\Phi$ from the equation (\[4.3\]) and then calculate the inner product $(\psi,T_{11}(\l,l)\Phi)_{L^2(\Pi)}$. This procedure leads us to the equation $$1-\frac{1}{\l-\l_0}\left(\psi,T_{11}(\l,l)
(I+T_6(\l)T_{11}(\l,l))^{-1}\phi\right)_{L^2(\Pi)}=0\,,$$ or in a more convenient form $$\label{4.4}
\l-\l_0-\left(\psi,T_{11}(\l,l)
(I+T_6(\l)T_{11}(\l,l))^{-1}\phi\right)_{L^2(\Pi)}=0\,.$$ This is the sought equation determining the perturbed eigenvalues of the problem (\[1.2\]), and, thus, of the operator $H_l(a)$. The associated solution of the equation (\[4.1\]), as it follows from (\[4.3\]), can be written as $$\label{4.5}
\Phi=(I+T_6(\l)T_{11}(\l,l))^{-1}\phi\,;$$ we naturally keep in mind the fact that the eigenfunctions are defined up to a multiplicative constant.
The equation (\[4.4\]) determine all eigenvalues of $H_l(a)$; due to the equivalence between (\[4.1\]) and (\[1.2\]) only the eigenvalues of $H_l(a)$ satisfy this equation. Thus, by Proposition \[lmpr\], for every $T_{11}$ there exists an unique solution of the equation (\[4.4\]) converging to $\l_0$ as $l\to+\infty$.
The desired asymptotic expansions for the perturbed eigenvalues can be calculated directly from the equation (\[4.4\]). First of all we recall the assertion (\[3.10\]) which implies that for $\l$ close to $\l_0$ the norm $T_{11}$ can be estimated by $\mathcal{O}(\Ex^{-(2\sqrt{1-\l_0}-\si)l})$. It allows us first to establish the estimate $$\label{4.14}
\l-\l_0=\mathcal{O}\left(\Ex^{-(2\sqrt{1-\l_0}-\si)l}\right),$$ and secondly to expand the second term in the equation (\[4.4\]) obtaining $$\label{4.6}
\l-\l_0-\left(\psi,T_{11}(\l,l)\phi\right)_{L^2(\Pi)}
+\mathcal{O}(\Ex^{-2(2\sqrt{1-\l_0}-\si)l})=0\,.$$ We can also extract the leading term from the operator $T_{11}(\l,l)$, which obviously comes from the lowest-mode contribution to the sum at the right hand side of (\[3.7\]). We will do that for $hu=u$, in the other case one proceeds analogously.
First we introduce additional notations setting $$\begin{aligned}
\label{4.7}
V(x):=\left\{
\begin{aligned}
{}&-\frac{4\Ex^{-2\kappa_1(\l_0) l}}{\pi \kappa_1(\l_0)}\sinh
\kappa_1(\l_0) x_1\sin x_2\int\limits_{\Pi^-}\sinh \kappa_1(\l_0)
t_1\sin t_2\phi\,d^2t\,,{}& {}&x_1<0\,,
\\
{}&\hphantom{-\frac{4\Ex^{-2\kappa_1(\l) l}}{\pi
\kappa_1(\l)}\sinh \kappa_1(\l_0) x_1}0\,,{}& {}& x_1>0\,.
\end{aligned}\right.\end{aligned}$$ Suppose that a function $W$ solves the problem (\[2.4\]) with $v=V$, then $$T_{11}(\l,l)\phi=-(\D+\l_0)\left(V+\chi(W\!-\!V)\right)
+\mathcal{O}(l\Ex^{-2(2\sqrt{1-\l_0}-\si)l})\quad \text{in
$L^2(\Pi_A)$}\,.$$ Using this identity together with the fact that the function $T_{11}(\l,l)\phi$ is finite, we can calculate the leading term of the second summand in (\[4.6\]), $$\begin{aligned}
{}&\left(\psi,T_{11}(\l,l)\phi\right)_{L^2(\Pi)}=-\int\limits_{\Pi}
(\D+\l_0)\left(V+\chi(W\!-\!V)\right)\,d^2x
+\mathcal{O}(l\Ex^{-2(2\sqrt{1-\l_0}-\si)l})\,,
\\
{}& \int\limits_{\Pi} (\D+\l_0)(V+\chi(W\!-\!V))\,d^2x=
\lim\limits_{R\to+\infty}\!\!
\int\limits_{\{x:\,|x_1|=R,\,0<x_2<\pi\}}\left( \psi\frac{\p V}{\p
\nu}- V\frac{\p \psi}{\p \nu}\right)\,ds
\\
{}&=-\lim\limits_{R\to+\infty}\int\limits_{\{x:\,x_1=-R,\,
0<x_2<\pi\}}\left( \psi\frac{\p V}{\p x_1}- V\frac{\p \psi}{\p
x_1}\right)\,ds\,.
\end{aligned}\label{4.8}$$ In order to calculate the last integral we use the fact that in view of the relation $\psi=T_3(\l_0)\phi$ and the definition of $T_3$ the constant $\a=\a_j$ in (\[1.5\]) is given by $$\begin{aligned}
\a=-\frac{2\rho}{\pi \kappa_1(\l_0)}
\int\limits_{\Pi^-}\sinh(\kappa_1(\l_0) t_1)\sin t_2\,
\phi(t)\,d^2t\,,\nonumber\end{aligned}$$ where $\rho$ is 1 if $\psi$ even and -1 if it is odd. Using this relation together with (\[1.5\]) and (\[4.7\]), we can finish our calculations in (\[4.8\]) arriving at $$\lim\limits_{R\to+\infty}\!\!
\int\limits_{\{x:\,x_1=-R,\,0<x_2<\pi\}}\left( \psi\frac{\p V}{\p
x_1}- V\frac{\p \psi}{\p x_1}\right)\,ds =\pi
\a^2\kappa_1(\l_0)\,\Ex^{-2\kappa_1(\l_0)l}\,.$$ Combining this with (\[4.8\]) and (\[4.6\]) we get the asymptotic (\[asm\]), (\[ltd1\]) for $\l_j^-(a)$. In the case $hu=\frac{\p u}{\p x_1}$ a similar reasoning leads to asymptotics (\[asm\]), (\[ltd1\]) for $\l_j^+(a)$. In order to prove relation (\[ldt\]) it is sufficient to express $\a$ in terms of suitable integrals. Keeping the parity of $\psi$ in mind we compute $$\begin{aligned}
0{}&=\lim\limits_{R\to+\infty}\int\limits_{\Pi_R}\Ex^{\kappa_1(\l_0)
x_1}\sin x_2\,(\D\!+\!\l_0)\psi(x)\,d^2x
\\
{}&=\lim\limits_{R\to+\infty}\int\limits_{\p\Pi_R}\left(\Ex^{
\kappa_1(\l_0) x_1}\sin x_2\frac{\p}{\p\nu}\psi(x)
-\psi(x)\frac{\p}{\p\nu}\,\Ex^{\kappa_1(\l_0) x_1}\sin
x_2\right)\,ds
\\
{}&=\int\limits_{\g(a)}\psi(x)\,
\Ex^{\sqrt{1-\l_0}x_1}\,dx_1-\a\pi\sqrt{1\!-\!\l_0}\,.\end{aligned}$$ This result leads us to formulae (\[ldt\]).
The asymptotics of the eigenfunctions can be derived easily. The definite parity of those associated with $\l_j^\pm(l,a)$ is obvious. The relation (\[4.5\]) tells us that $$\label{4.12}
\Phi^\pm=\phi+\mathcal{O}(\Ex^{-(l\sqrt{1-\l_0}-\si)})\,.$$ The symbol ”$\pm$” indicate here two variants of definition of the operator $T_{11}$. Now in order to prove the expansions for the eigenfunctions one has just to use this expression and to employ the arguments of the previous two section. More precisely, we have $\psi=T_3(\l_0)\phi$ and $\Psi^\pm=(T_3(\l^\pm)+T_{12}(\l^\pm,l))\Phi^\pm$, where $\Psi^\pm$ is the eigenfunction of the problem (\[1.2\]) associated with the chosen eigenvalue and chosen variant of boundary operator $h$. Using (\[4.12\]) and the holomorphy of $T_3$, $T_{12}$, the estimates (\[4.14\]) and (\[3.8\]), we arrive at the asymptotical formula $$\label{4.15}
\Psi^\pm=\phi+\mathcal{O}\left(\Ex^{-l(\sqrt{1-\l_0}-\si)l}\right)$$ in $\SS^1(\Pi^l)$. Recovering now the eigenfunctions of $H_l(a)$ we obtain all their properties stated in Theorem \[th1\].
Let us finally prove that there are no other eigenvalues of $H_l(a)$ in $\mathcal{D}_1$. Consider the equation (\[4.1\]) where $\l$ is close to one and does not lie in real semi-axis $[1,+\infty)$, more specifically, suppose that $\kappa\in
\mathcal{Q}_\k$. Then we can invert the operator $(I+T_4(\kappa))^{-1}$, and arrive at the equation $$\Phi+(I+T_4(\kappa))^{-1}T_{11}(\kappa,l)\Phi=0\,,$$ where the operator $(I+T_4(\kappa))^{-1}$ is uniformly bounded in $\kappa$, because the Neumann segment does not have by assumption a critical size – see Lemma \[lm2.3\] – while $T_{11}(\kappa,l)$ is small for all possible values $\kappa$ and $l$. Hence the operator $(I+T_4(\kappa))^{-1}T_{11}(\kappa,l)$ is also small, and therefore we can invert in turn the operator $(I+(I+T_4(\kappa))^{-1}T_{11}(\kappa,l))$ which immediately leads us to the unique solution $\Phi=0$. Moreover, the operator $H_l(a)$ cannot have eigenvalues corresponding to $\kappa$ satisfying $|\arg \kappa\pm\frac{\pi}{2}|<\k$, $\RE\kappa\not=0$, simply because it is self-adjoint and all its eigenvalues are real, thus there is no other eigenvalues to $H_l(a)$ in $\mathcal{D}_1$. By this the proof of Theorem \[th1\] is complete.
Proof of Theorem \[th3\]
========================
It is sufficient to consider in detail only the eigenvalue $\l_{n+1}$ emerging from the continuum because all the statements related to the other eigenvalues verify in a way completely analogous to the previous section.
We know from Proposition \[lmpr\] that the eigenfunction associated with the indicated eigenvalue is even with respect to $x_1$, thus we have to consider here only the case $hu=\frac{\p
u}{\p x_1}$. Assuming $\kappa\in\mathcal{Q}_\k$, we start with the equation $$\label{6.0}
\Phi+T_4(\kappa)\Phi+T_{13}(\kappa,l)\Phi=0\,,$$ which is how (\[4.1\]) looks like in the present case, with $T_{13}(\kappa,l)$ being the perturbation operator associated with (\[3.6\]). The operator $T_{13}(\kappa,l)$ is small by Lemma \[lm3.2\], and an argument analogous to that which lead us to (\[4.3\]) yields the equation $$\label{6.1}
\Phi+\frac{1}{2\kappa}(\psi,T_{13}(\kappa,l)\Phi)_{L^2(\Pi)}
(I+T_8(\l)T_{13}(\l,l))^{-1}\phi=0\,.$$ Recall that $\psi=\psi^n$ is a solution to the equation $(H(a_n)\!+\!1)\psi^n=0$ which behaves at infinity in accordance with (\[1.1\]) and $\phi\in L^2(\Pi_A)$ such that $\psi=T_3(\kappa=0)\phi$. From this equation one can deduce an analogue of the equation (\[4.4\]), namely $$\label{6.2}
2\kappa+\left(\psi,T_{13}(\kappa,l)(I+T_8(\kappa)
T_{13}(\kappa,l))^{-1}\phi\right)_{L^2(\Pi)}=0\,.$$ The value of $\kappa$ associated with the eigenvalue emerging from the continuum solves this equation and by Proposition \[lmpr\] it tends to zero. Using these two facts we will deduce the asymptotic formula stated in Theorem \[th3\]. First of all, in the following we will consider the equation (\[6.2\]) for real positive $\kappa$ only. This restriction can be justified easily, since for negative $\kappa$ the associated function $u$ given by (\[2.5\]) increases at infinity and thus it does not belong to $L^2(\Pi)$. In order to calculate the asymptotics, we extract the leading part of the second term in the equation (\[6.2\]); for small positive $\kappa$ we have $$\label{6.3}
\left(\psi,T_{13}(\kappa,l)(I\!+\!T_8(\kappa)T_{13}
(\kappa,l))^{-1}\phi\right)_{L^2(\Pi)}=
\left(\psi,T_{13}(\kappa,l)\phi\right)_{L^2(\Pi)}
+T_{14}(\kappa,l)\,,$$ where $T_{13}(\kappa,l):\mathbb{R}^2\to\mathbb{R}$ is a function defined for $(\kappa,l)\in\mathcal{Q}_\k\times[\l_0,+\infty)$ which satisfies the relation $$\label{6.5}
T_{14}(\kappa,l)=\mathcal{O} \left(\frac{\kappa^2\Ex^{-2\kappa
l}}{\cosh^2 \kappa l}+\Ex^{-4\sqrt{3}l}\right)$$ as $(\kappa,l)\to(0,+\infty)$. To get this estimate one has to employ the relation $T_{14}(\kappa,l)=\mathcal{O}(\|T_{13}(\kappa,l)\|^2)$ and the fact that $$\label{6.12}
\|T_{13}(\kappa,l)\|=\mathcal{O}\left(\frac{\kappa\,
\Ex^{-\kappa l}}{\cosh \kappa l}+\Ex^{-2\sqrt{3}l}\right)$$ implied by the definition of $T_{13}$ – see (\[3.6\]). Our next step is to extract the leading term from the first summand at the right hand side of (\[6.3\]). We will do it in the same way as in last section, the only difference is that now we have to take into account also the second transverse-mode contribution to (\[3.6\]).
We introduce the function $V_1$ that is an analogue of (\[4.7\]) by $$\begin{aligned}
\label{6.4}
V_1(x):=\left\{
\begin{aligned}
{}&\frac{2\kappa\,\Ex^{-\kappa l}}{\pi \cosh \kappa l}\, x_1
\sin x_2\int\limits_{\Pi^-} t_1\sin t_2\,\phi(t)\,d^2t\,,{}&
{}&x_1<0\,,
\\
{}&\hphantom{\frac{2m\Ex^{-m l}}{\pi \cosh m l} x_1 \sin
x_2}0\,,{}& {}& x_1>0\,.
\end{aligned}\right.\end{aligned}$$ Let $W_1$ be a solution to the problem (\[2.4\]) with $v=V_1$. We also introduce the function $V_2$ in the following way $$\begin{aligned}
\label{6.11}
V_2(x):=\left\{
\begin{aligned}
{}&\frac{2\Ex^{-\sqrt{3}l}}{\pi\sqrt{3}\cosh \sqrt{3} l}
\sinh\sqrt{3}x_1 \sin 2x_2\int\limits_{\Pi^-}
\sinh\sqrt{3}t_1\sin 2t_2\,\phi(t)\,d^2t\,,{}& {}&x_1<0\,,
\\
{}&\hphantom{\frac{2m\Ex^{-m l}}{\pi \sinh m l} x_1 \sin
x_2}0\,,{}& {}& x_1>0
\end{aligned}\right.\end{aligned}$$ and suppose that $W_2$ is a solution of (\[2.4\]) with $v=V_2$. One can check that $$(\psi,T_{13}(\kappa,l)\phi)_{L^2(\Pi)}=-\left(\psi,
(\D\!+\!1)\left(\t V+\chi(\t W\!-\!\t
V)\right)\right)_{L^2(\Pi)}+T_{15}(\kappa,l)\,,$$ where $\t V=V_1+V_2$, $\t W=W_1+W_2$, and the function $T_{15}(\kappa,l)$ satisfies the estimate $$\label{6.7}
T_{15}(\kappa,l)=\mathcal{O}\left(\frac{\kappa^3\Ex^{-\kappa
l}}{\cosh \kappa
l}+\kappa^2\Ex^{-2\sqrt{3}l}+\Ex^{-2\sqrt{8}l}\right)$$ as $(\kappa,l)\to(0,+\infty)$. Calculating the inner product $\left(\psi,T_{13}(\kappa,l)\phi\right)_{L^2(\Pi)}$ in the same way how we deduced (\[asm\]) and bearing in mind the asymptotics (\[1.1\]) for $\psi$ together with (\[6.3\]), (\[6.5\]), and (\[6.4\])–(\[6.7\]) we obtain the equation $$\label{6.8}
\begin{aligned}
2\kappa+&\rho\frac{\sqrt{2}\kappa\Ex^{-\kappa l}}{\sqrt{\pi}\cosh
\kappa l} \int\limits_{\Pi^-} t_1\sin t_2\,\phi\,d^2t
\\
+&\rho
\frac{\b\,\Ex^{-\sqrt{3}l}}{\cosh\sqrt{3}l}\int\limits_{\Pi^-}
\sinh\sqrt{3}t_1\sin 2t_2\,\phi\,d^2t+T_{16}(\kappa,l)=0\,,
\end{aligned}$$ where $\rho$ is again the parity of $\psi$. The function $T_{16}(\kappa,l)$ satisfies $$\label{6.6}
T_{16}(\kappa,l)=O\left(\frac{\kappa^2\Ex^{-2\kappa l}}{\cosh^2
\kappa l}+\frac{\kappa^3\Ex^{-\kappa l}}{\cosh \kappa
l}+\kappa^2\Ex^{-2\sqrt{3}l}+\Ex^{-2\sqrt{8}l}\right)$$ as $(\kappa,l)\to(0,+\infty)$. Since the function $\phi$ obeys $\psi=T_3(\kappa=0)\phi$, we can take into account the definition of the last operator (see (\[2.13\])) and the asymptotics (\[1.1\]) to conclude that $$\begin{aligned}
{}&\rho\sqrt{\frac{2}{\pi}}=-\frac{2}{\pi}\int\limits_{\Pi^-}
t_1\sin t_2\,\phi(t)\,d^2t\,,
\\
{}&\rho\b=-\frac{2}{\pi\sqrt{3}}\int\limits_{\Pi^-} \sinh
\sqrt{3}t_1\sin 2t_2\,\phi(t)\,d^2t\,,\end{aligned}$$ which together with (\[6.8\]) leads us to ($\b=\b_n$) $$2\kappa-\frac{\kappa\,\Ex^{-\kappa l}}{\cosh \kappa
l}-\frac{\b^2\pi\sqrt{3}}{2}\frac{\Ex^{-\sqrt{3}l}}{\cosh\sqrt{3}l}
+T_{16}(\kappa,l)=0\,,$$ or equivalently, $$\label{6.10}
\frac{\kappa\,\Ex^{\kappa l}}{\cosh \kappa
l}-\frac{\b^2\pi\sqrt{3}}{2}\frac{\Ex^{-\sqrt{3}l}}{\cosh\sqrt{3}l}
+T_{16}(\kappa,l)=0\,.$$ We know that this equation has a positive solution tending to zero as $\kappa\to0$. In view of (\[6.6\]), (\[6.10\]), and the trivial inequality $$1\le\frac{\Ex^{\tau}}{\cosh\tau}\le2$$ we have for this solution the following estimate, $$C_1\Ex^{-2\sqrt{3}l} \le \kappa\le C_2\Ex^{-2\sqrt{3}l}$$ with constants $C_1$, $C_2$ independent on $l$. Using it we can expand the first term in the equation (\[6.10\]) with respect to $\kappa l$, which is small, and to estimate $T_{15}$ – see (\[6.6\]). In this way we arrive at the relation, $$\kappa-\b^2\pi\sqrt{3}\,\Ex^{-2\sqrt{3}l}
+\mathcal{O}(\Ex^{-2\sqrt{8}l})=0\,,$$ which implies the sought asymptotical expansion (\[1.3\]), (\[1.4\]). The second formula for $\mu$ stated in Theorem \[th3\] can be proven completely by analogy with the proof of (\[ldt\]). One just should multiply the equation $(\Delta+1)\psi^n$ by $\Ex^{\sqrt{3}x_1}\sin 2x_2$ and integrate then by parts over $\Pi_R$ passing then to the limit as $R\to+\infty$.
The argument concerning the asymptotics for the associated eigenfunction is completely analogous to that of the previous section. The solution to the equation (\[6.0\]) is given by $$\Phi=(I+T_8(\kappa)T_{13}(\kappa,l))^{-1}\phi\,.$$ Now one has just to perform the expansion using the fact that the operator $T_8(\kappa)T_{13}(\kappa,l)$ is small, then using the obtained asymptotics for $\kappa$, to apply to the remainder the estimate (\[6.12\]), to construct the corresponding eigenfunction of the problem (\[1.2\]) by the scheme described in the Section 5, and finally, to recover the eigenfunctions of $H_l(a_n)$. This completes the proof of the second theorem.
Acknowledgments {#acknowledgments .unnumbered}
---------------
D.B. is grateful for the hospitality in the Department of Theoretical Physics, NPI, Czech Academy of Sciences, where a part of this work was done. The research has been partially supported by GAAS under the contract A1048101, by RFBR under the contracts 02-01-00693, 03-01-06470 and by the program ”Leading scientific schools” (NSh-1446.2003.1).
[99]{}
D. Borisov, P. Exner, and R. Gadyl’shin: Geometric coupling thresholds in a two-dimensional strip, *J. Math. Phys.* **43** (2002), 6265-6278.
Ph. Briet, J.-M. Combes, P. Duclos: Spectral stability under tunneling, *Commun. Math. Phys.* **126** (1989), 133-156.
E.B. Davies: [*Spectral theory and differential operators*]{}, Cambridge University Press 1995.
P. Exner, P. Šeba, M. Tater, D. Vaněk: Bound states and scattering in quantum waveguides coupled laterally through a boundary window, *J. Math. Phys.* **37** (1996), 4867-4887.
Gadyl’shin R.R. Local perturbations of the Schroedinger operator on the axis, *Theor. Math. Phys.* **132** (2002), 976-982
Gadyl’shin R.R. Local perturbations of the Schroedinger operator on the plane, *Theor. Math. Phys.*, to appear; e-print `math-ph/0208025`
T. Kato: [*Perturbation theory for linear operators*]{}, 2nd edition, Springer, Berlin 1976.
O.A.Ladyzhenskaya: *The Boundary Value Problems of Mathematical Physics*, Nauka, Moscow 1973; English translation Springer, New York 1985.
J.T. Londergan, J.P. Carini, D.P. Murdock: *Binding and Scattering in Two-dimensional Systems*, Springer, Berlin 1999.
M. Reed and B. Simon: [*Methods of Modern Mathematical Physics, I. Functional Analysis*]{}, Academic Press, New York 1972.
E. Sanchez-Palencia: *Non-homogeneous Media and Vibration Theory*,\
Springer, New York 1980.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Consistent and reproducible evaluation of Deep Reinforcement Learning (DRL) is not straightforward. In the Arcade Learning Environment (ALE), small changes in environment parameters such as stochasticity or the maximum allowed play time can lead to very different performance. In this work, we discuss the difficulties of comparing different agents trained on ALE. In order to take a step further towards reproducible and comparable DRL, we introduce SABER, a **S**tandardized **A**tari **BE**nchmark for general **R**einforcement learning algorithms. Our methodology extends previous recommendations and contains a complete set of environment parameters as well as train and test procedures. We then use SABER to evaluate the current state of the art, Rainbow. Furthermore, we introduce a *human world records baseline*, and argue that previous claims of *expert or superhuman* performance of DRL might not be accurate. Finally, we propose *Rainbow-IQN* by extending Rainbow with Implicit Quantile Networks (IQN) leading to new state-of-the-art performance. Source code is available for reproducibility.'
author:
- |
Marin Toromanoff\
MINES ParisTech, Valeo DAR, Valeo.ai,\
`[email protected]`\
`[email protected]`\
Emilie Wirbel\
Valeo Driving Assistance Research\
`[email protected]`\
\
Fabien Moutarde\
Center for Robotics, MINES ParisTech, PSL\
`[email protected]`\
bibliography:
- 'rainbowiqn.bib'
title: 'Is Deep Reinforcement Learning Really Superhuman on Atari? Leveling the playing field '
---
Introduction
============
Human intelligence is able to solve many tasks of different natures. In pursuit of generality in artificial intelligence, video games have become an important testing ground: they require a wide set of skills such as perception, exploration and control. Reinforcement Learning (RL) is at the forefront of this development, especially when combined with deep neural networks in DRL. One of the first general approaches reaching reasonable performance on many Atari games while using the exact same hyper-parameters and neural network architecture was Deep Q-Network (DQN) [@mnih2015human], a value based DRL algorithm which directly takes the raw image as input.
This success sparked a lot of research aiming to create better, faster and more stable general algorithms. The ALE [@OriginalALE], featuring more than 60 Atari games (see Figure \[fig:invaders\]), is heavily used in this context. It provides many different tasks ranging from simple paddle control in the ball game Pong to complex labyrinth exploration in Montezuma’s Revenge which remains unsolved by general algorithms up to today.
As the number of contributions is growing fast, it becomes harder and harder to make a proper comparison between different algorithms. In particular, a relevant difference in the training and evaluation procedures exists between available publications. Those issues are exacerbated by the fact that training DRL agents is very time consuming, resulting in a high barrier for reevaluation of previous work. Specifically, even though ALE is fast at runtime, training an agent on one game takes approximately one week on one GPU and thus the equivalent of more than one year to train on all 61 Atari games. A standardization of the evaluation procedure is needed to make DRL *that matters* as pointed out by @Henderson2017 for the Mujoco benchmark [@MuJoCo]: the authors criticize the lack of reproducibility and discuss how to allow for a fair comparison in DRL that is consistent between articles.
[r]{}[0.25]{} {width="25.00000%"}
In this work, we first discuss current issues in the evaluation procedure of different DRL algorithms on ALE and their impact. We then propose an improved evaluation procedure, extending the recommendations of @Machado, named SABER : a Standardized Atari BEnchmark for Reinforcement learning. We suggest benchmarking on the *world records human baseline* and show that RL algorithms are in fact far from solving most of the Atari games. As an illustration of SABER, current state-of-the-art DRL algorithm Rainbow [@Rainbow] is benchmarked. Finally, we introduce and benchmark on SABER a new state-of-the-art agent: a distributable combination of Rainbow and Implicit Quantiles Network (IQN) [@Dabney2018]. The main contributions of this work are :
- The proposal, description and justification of the SABER benchmark.
- Introduction of a *world records human baseline*. We argue it is more representative of the human level than the one used in most of previous works. With this metric, we show that the Atari benchmark is in fact a hard task for current general algorithm.
- A SABER compliant evaluation of current state-of-the art agent Rainbow.
- A new state-of-the-art agent on Atari, Rainbow-IQN, with a comparison on SABER to Rainbow, to give an improvement range for future comparisons.
- For reproducibility sake, an open-source implementation [^1] of Rainbow, Rainbow-IQN, distributed following the idea from @Horgan.
Related Work
------------
#### Reproducibility and comparison in DRL
*Deep Reinforcement Learning that matters* [@Henderson2017] is one of the first works to warn about a reproducibility crisis in the field of DRL. This article relies on the MuJoCo [@MuJoCo] benchmark to illustrate how some common practices can bias reported results. As a continuation to the work of @Henderson2017, J. Pineau introduced a Machine Learning reproducibility checklist [@checklist] to allow for reproducibility and fair comparison. @Machado deal with the Atari benchmark. They describe the divergence in training and evaluation procedures and how this could lead to difficulties to compare different algorithms. A first set of recommendations to standardize them is introduced, constituting the basis of this work and will be summarized in the next section. Finally, the Github Dopamine [@castro18dopamine] provides an open-source implementation of some of the current state-of-the-art algorithms on Atari benchmark, including Rainbow and IQN. An evaluation following almost all guidelines from @Machado are provided in @castro18dopamine. However the implementation of Rainbow is partial, and the recommendation of using the full action set is not applied. This is why our work contains a new evaluation of Rainbow.
#### Value based RL
DQN [@mnih2015human] is the first value based DRL algorithm benchmarked on all Atari games with the exact same set of hyperparameters (although previous work by @hausknecht2014neuroevolution already performed such a benchmark with neural networks). This algorithm relies on the well known Q-Learning algorithm [@QLearning] and incorporates a neural network. Deep Q-learning is quite unstable and the main success of this work is to introduce practical tricks to make it converge. Mainly, transitions are stored in a *replay memory* and sampled to avoid correlation in training batch, and a separate *target network* is used to avoid oscillations. Since then, DQN has been improved and extended to make it more robust, faster and better. Rainbow [@Rainbow] is the combination of 6 of these improvements [@DoubleDQN; @PER; @C51; @Dueling; @NoisyNetwork; @NstepDQNandA3C] implemented in a single algorithm. Some ablations studies showed that the most important components were Prioritized Experience Replay (PER) [@PER] and C51 [@C51]. The idea behind PER is to sample transitions according to their *surprise*, i.e. the worse the network is at predicting the Q-value of a specific transition, the more we sample it. C51 is the first algorithm in *Distributional RL* which predicts the full distribution of the Q-function instead of predicting only the mean of it. Finally, IQN [@Dabney2018] is an improvement over C51. It almost reaches on its own the performance of the full Rainbow with all 6 components. In C51 the distribution of the Q-function is represented as a categorical distribution while in IQN, it is represented by implicit quantiles.
Challenges when Comparing Performance on the Atari Benchmark
============================================================
In this section we discuss several challenges to make a proper comparison between different algorithms trained on the Atari benchmark. First, we briefly summarize the initial problems and their solution as proposed by @Machado. Then we detail a remaining issue not handled by those initial standards, the maximum length time allowed for an episode. Finally, we introduce a readable metric, representative of actual human level and allowing meaningful comparison.
Revisiting ALE: an Initial Step towards Standardization
-------------------------------------------------------
@Machado discuss about divergence of training and evaluation procedures on Atari. They show how those divergences are making comparison extremely difficult. They establish recommendations that should be used in order to standardize the evaluation process.
#### Stochasticity
The ALE environment is fully deterministic, i.e. leading to the exact same state if the exact same actions are taken at each state. This is actually an issue for general algorithm evaluation. For example, an algorithm learning *by heart* good trajectories can actually reach a high score with an open-loop behaviour. To handle this issue, @Machado introduce *sticky actions*: actions coming from the agent are repeated with a given probability $\xi$, leading to a non deterministic behavior. They show that sticky actions are drastically affecting performance of an algorithm exploiting the environment determinism without hurting algorithms learning more robust policies like DQN [@mnih2015human]. We use sticky actions with probability $\xi = 0.25$ [@Machado] in all our experiments.
#### End of an episode: Use actual game over
In most of the Atari games the player has multiple lives and the game is actually over when all lives are lost. But some articles, e.g. DQN, Rainbow, IQN, end a training episode after the loss of the first life but still use the standard game over signal while testing. This can in fact help the agent to learn how to avoid death and is an unfair comparison to agents which are not using this *game-specific* knowledge. @Machado recommend to use only the standard game over signal for all games while training.
#### Action set
Following the recommendation of @Machado we do not use the *minimal useful action set* (the set of actions having an effective impact on the current game) as used by many previous works [@mnih2015human; @Rainbow]. Instead we always use all 18 possible actions on the Atari Console. This removes some specific domain knowledge and reduces the complexity of reproducibility. For some games, the minimal useful action set is different from one version to another of the standard Atari library: an issue to reproduce result on breakout was coming from this [@breakoutBlog].
#### Reporting performance
As proposed by @Machado, we report our score while training by averaging $k$ consecutive episodes (we have set $k = 100$). This gives information about the stability of the training and removes the statistical bias induced when reporting score of the best policy which is today a common practice [@mnih2015human; @Rainbow].
Maximum Episode Length {#sec:max_length}
----------------------
A major parameter is left out of the work of @Machado: the maximum number of frames allowed per episode. This parameter ends the episode after a fixed number of time steps even if the game is not over. In most of recent works [@Rainbow; @Dabney2018], this is set to 30 min of game play and only to 5 min in Revisiting ALE [@Machado]. This means that the reported scores can not be compared fairly. For example, in easy games (e.g. Atlantis, Enduro), the agent never dies and the score is more or less linear to the allowed time: the reported score will be 6 times higher if capped at 30 minutes instead of 5 minutes.
We argue that the time cap can make the performance comparison non significant. On many games (e.g. Atlantis, Video Pinball) the scores reported of Ape-X [@Horgan], Rainbow [@Rainbow] and IQN [@Dabney2018] are almost exactly the same. This is because all agents reach the time limit and get the highest possible score in 30 minutes: the difference in scores is due to minor variations, not algorithmic difference. As a consequence, the more successful agents are, the more games are incomparable because they reach the maximum possible score in the time cap.
This parameter can also be a source of ambiguity and error. The best score on Atlantis (2,311,815) is reported by *Proximal Policy Optimization* by @PPO but this score is almost certainly wrong: it seems impossible to reach it in only 30 minutes! The first distributional paper, C51 [@C51], also did this mistake and reported wrong results before adding an erratum in a later version on ArXiv.
We argue that episodes should not be capped at all. The original ALE article [@OriginalALE pg.3] states that *This functionality is needed for a small number of games to ensure that they always terminate*. On some famously hard games like *Pitfall* and *Tennis*, random exploration leads to much more negative reward than positive and thus the agent effectively learns to do nothing, e.g. not serving in Tennis. We claim that, even with this constraint, agents still end up learning to do nothing, and the drawback of the cap harms the evaluation of all other games. Moreover, the human high scores for Atari games have been achieved in several hours of play, and would have been unreachable if limited to 30 minutes. To summarize, ideally one would not cap at all length of episode while training and testing. However this makes some limitations of the ALE environment appear, as described in the following paragraph.
**Glitch and bug in the ALE environment**
When setting the maximum length of an episode to infinite time, the agent gets stuck on some games, i.e. the episode never ends, because of a bug in the emulator. In this case, even doing random actions for more than 20 hours neither gives any reward nor end the game. This happens consistently on *BattleZone* and less frequently on *Yar’s Revenge*. One unmanaged occurrence of this problem is enough to hamper the whole training of the agent. It is important to note that those bugs were discovered by chance and it is probable that this could happen on some other games.
[r]{}[0.5]{} {width="50.00000%"}
We recommend to set the maximum episode length to infinite (in practice, a limit of 100 hours was used). Additionally we suggest a *maximum stuck time* of 5 minutes. Instead of limiting the allowed time for the agent, we limit the time without receiving any reward. This small trick handles all issues exposed above, and sets all reported scores on the same basis, making comparison to world records possible. Other bugs or particularities harming evaluation were encountered while training on the full Atari benchmark: buffer rollover with sudden negative score, influence of a start key for some games, etc. They are detailed and discussed in the supplementary material and we argue that they can have a drastic impact on performance and explain inconsistencies.
Human World Records Baseline
----------------------------
A common way to evaluate AI for games is to let agents compete against human world champions. Recent examples for DRL include the victory of AlphaGo versus Lee Sedol for Go [@silver2016mastering], OpenAI Five on Dota 2 [@openaifive] or AlphaStar versus Mana for StarCraft 2 [@alphastarblog]. In the same spirit, one of the most used metric for evaluating RL agents on Atari is to compare them to the human baseline introduced by @mnih2015human. Previous works use the normalized human score, i.e. 0% is the score of a random player and 100% is the score of the human baseline, which allows to summarize the performance on the whole Atari set in one number, instead of individually comparing raw scores for each of the 61 games. However we argue that this human baseline is far from being representative of the best human player, which means that using it to claim superhuman performance is misleading. The current world records are available online for 58 of the 61 evaluated Atari game [^2]. Evaluating these world records scores using the usual human normalized score has a median of 4.4k% and a mean of 99.3k% (see Figure \[fig:prov\_vs\_beginner\] for details), to be compared to 200% and 800% of original Rainbow [@Rainbow]. As a consequence, we argue that using a normalized human score with the world records will give a much better indication of the performance of the agents and the margin of improvement. Note that 3 games of the ALE (double dunk, elevator action and tennis) do not have a registered world record, so all following experiments contain 58 games.
SABER : a Standardized Atari BEnchmark for Reinforcement learning
=================================================================
In this section we introduce SABER, a set of training and evaluation procedures on the Atari benchmark allowing for fair comparison and for reproducibility. Moreover, those procedures make it possible to compare with the human world records baseline introduced above and thus to obtain an accurate idea of the gap between general agents and best human players.
Training and Evaluation Procedures
----------------------------------
All recommendations stated in the previous section are summarized in Table \[tab:valid-data\] to constitute the SABER benchmark. It is important to note that those procedures must be used at both training and test time. The recent work Go-Explore [@GoExplore] opened a debate on allowing or not stochasticity at training time. They report state-of-the-art performance on the famously hard game *Montezuma’s Revenge* by removing stochasticity at training time. They conclude that we should have benchmarks with and without it [@GoExploreBlog]. We choose to use same conditions for training and testing general agents: this is more in line with realistic tasks.
Reporting Results {#sec:report_results}
-----------------
[l]{}[0.5]{}
Parameter Value
------------------------------- --------------------------------------
Sticky actions $\xi = 0.25$
Life information Not allowed
Action set 18 actions
Max stuck time 5 min (18000 frames)
Max episode length Infinite (100 hours)
Initial state and random seed Same starting state and varying seed
In accordance with previous guidelines, we advocate to report mean scores of 100 consecutive training episodes at specific time, here 10M, 50M, 100M and 200M frames. This removes the bias of reporting scores of the best agent encountered during training and makes it possible to compare at different data regimes. Due to the complexity of comparing 58 scores in a synthetic manner, we try to provide a single metric to make an effective comparison. Mean and median normalized scores to the records baseline are computed over all games. Note that the median is more relevant: the mean is highly impacted by outliers, in particular by games where the performance is superhuman. For the mean value, games with an infinite game time and score are artificially capped to 200% of the records baseline. We propose to add a histogram of the normalized score, to classify the games according to their performance. We define 5 classes: failing ($<1\%$), poor ($<10\%)$, medium ($<50\%$), fair ($<100\%$) and superhuman ($>100\%$). Medians, means and histograms can be found in Section \[sec:experiments\], and the fully detailed scores are available in the supplementary materials.
Rainbow-IQN
===========
Two different approaches were combined to obtain an improvement over Rainbow [@Rainbow]: Rainbow itself and IQN [@Dabney2018] because of its excellent performance. Implementation details and hyper-parameters are described in the supplementary material. Both our implementations of Rainbow and Rainbow-IQN are distributed [^3], following Ape-X [@Horgan] and based on the implementation of [@castro18dopamine].
IQN is an evolution of the C51 algorithm [@C51] which is one of the 6 components of the full Rainbow, so this is a natural upgrade. After the implementation, preliminary tests highlighted the impact of PER [@PER]: taking the initial hyper-parameters for PER from Rainbow resulted in poor performance. Transitions are sampled from the replay memory proportionally to the training loss to the power of priority exponent $\omega$. Reviewing the distribution of the loss shows that it is significantly more spread for Rainbow-IQN than for Rainbow, thus making the training unstable, because some transitions were over-sampled. To handle this issue, 4 values of $\omega$ were tested on 5 games: 0.1, 0.15, 0.2, 0.25 instead of 0.5 for original Rainbow, with 0.2 giving the best performance. The 5 games were Alien, Battle Zone, Chopper Command, Gopher and Space Invaders. All other parameters were left as is. Rainbow-IQN is evaluated on SABER and compared to Rainbow in the following section.
Experiments {#sec:experiments}
===========
In this section, we describe the experiments performed on SABER. For all parameters not mentioned in SABER (e.g. the action repeat, the network architecture, the image preprocessing, etc) we carefully followed the parameters used in Rainbow [@Rainbow] and IQN [@Dabney2018] papers. Those details and the scores for each agent and individual games can be found in the supplementary materials. Training one agent takes slightly less than a week, which makes a full benchmark use around 1 year-GPU. As a consequence, for each algorithm benchmark, trainings were run with only one seed for the full benchmark, and 5 seeds for 14 of the 61 games. Details on the choice of these games and the associated scores can be found in Section \[sec:stability\]. The combined duration of all experiments conducted for this article is more than 4 years-GPU. Agents are trained using SABER guidelines on the 61 Atari games, and evaluated with the records baseline for 58 games. Scores at both 5 minutes and 30 minutes are kept while training to compare to previous works.
Rainbow Evaluation
------------------
------------- -------- -------- ------------ -------- -------- ------------
Algorithm
Median Mean Superhuman Median Mean Superhuman
Performance 4.20% 24.10% 2 2.61% 17.09% 1
------------- -------- -------- ------------ -------- -------- ------------
: Median and mean human-normalized performance and number of superhuman scores ($>100\%$). Scores are coming from the original Rainbow and from our re-evaluation of Rainbow following recommendations of Machado [et al. ]{}(30 minutes evaluation, at 200M training frames).[]{data-label="tab:rainbow_orig_vs_ale"}
Benchmarking Rainbow makes it possible to measure the impact of the guidelines of Machado et al.: sticky actions, ignore life signal and full action set. Table \[tab:rainbow\_orig\_vs\_ale\] compares the originally reported performance of Rainbow [@Rainbow] to an evaluation following the recommendations of Machado [et al. ]{}The performance is measured with the records baseline, for a 30 minutes evaluation at 200M training frames, to be as close as possible to the conditions of the original Rainbow. The impact of the standardized training procedure is major: as shown in the following paragraph, the difference in median (1.59%) is comparable to the difference between DQN and Rainbow (1.8%, see Figure \[fig:dqn\_vs\_rainbow\_vs\_iqn\]) when both are trained on same training procedures. This demonstrates the importance of explicit and standardized training and evaluation procedures.
Rainbow-IQN: Evaluation and Comparison {#sec:iqn-eval}
--------------------------------------
#### Influence of maximum episode length
[p[2.1cm]{}|\*[3]{}[p[0.8cm]{}]{}|\*[3]{}[p[0.8cm]{}]{}|\*[3]{}[p[0.8cm]{}]{}]{} Time & & &\
& Median & Mean & Super. & Median & Mean & Super. & Median & Mean & Super.\
Rainbow & 2.35% & 14.86% & 0 & 2.61% & 17.09% & 1 & 2.83% & 24.54% & 3\
Rainbow-IQN & 2.61% & 17.62% & 0 & 2.81% & 20.18% & 1 & 3.13% & 30.89% & 4\
Table \[tab:rainbow\_time\_comparison\] studies the influence of the time limit for the evaluation, by reporting performance for Rainbow and Rainbow-IQN depending on the evaluation time. A significant difference can be seen between 5, 30 minutes and without limiting time of evaluation, which confirms the discussion of Section \[sec:max\_length\].
#### Comparison to Rainbow
![Comparison of Rainbow and Rainbow-IQN on SABER: Median normalized scores with regards to training steps.[]{data-label="fig:rainbow_iqn_saber_over_frames"}](images/median_human_normalised_Rainbow_vs_Rainbow_IQN.png){width="\textwidth"}
As introduced in Section \[sec:report\_results\], we compare Rainbow and Rainbow-IQN with median and mean metrics on SABER conditions, and with a classification of the performance of the agents in Figure \[fig:saber\_rainbow\_iqn\_hist\]. Figure \[fig:rainbow\_iqn\_saber\_over\_frames\] shows that Rainbow-IQN performance during training is consistently higher than Rainbow. One can notice on Figure \[fig:saber\_rainbow\_iqn\_hist\] that the majority of agents are in the *poor* and *failing* categories, showing the gap that must be crossed to achieve superhuman performance on the ALE.
![Comparison of Rainbow and Rainbow-IQN on SABER: classifying performance of agents relatively to the records baseline (at 200M training frames).[]{data-label="fig:saber_rainbow_iqn_hist"}](images/rainbow_iqn_saber_comparison_hist.png){width="80.00000%"}
#### Comparison to DQN
Figure \[fig:dqn\_vs\_rainbow\_vs\_iqn\] provides a comparison between DQN, Rainbow and Rainbow-IQN. The evaluation time is set at 5 minutes to be consistent with the reported score of DQN by @Machado. As expected, DQN is outperformed for all training steps. As aforementioned, the difference between DQN and Rainbow is in the same range as the difference coming from divergent training procedures, showing again the necessity for standardization.
![Median performance comparison for DQN, Rainbow and Rainbow-IQN with regards to training frames. Evaluation time is set at 5 minutes to allow a comparison to DQN.[]{data-label="fig:dqn_vs_rainbow_vs_iqn"}](images/5min_dqn_rainbow_iqn_mean.png){width="\textwidth"}
Stability of both Rainbow and Rainbow-IQN {#sec:stability}
-----------------------------------------
[@Machado] use 5 different seeds for training to check that the results are reproducible and stable. For this article, these 5 runs are conducted on both Rainbow and Rainbow-IQN for 14 games (around 25% of the whole benchmark). It would be best to have the whole benchmark on 5 seeds but this was way above our computational resources. Still, these 14 games allow us to make a first step of stability studies. They are chosen according to the results of the first seed, with the idea of prioritizing games on which scores were most notably different between Rainbow and Rainbow-IQN. We also try to choose diverse games from different categories (from failing to superhuman) and we removed the 5 games used for the hyperparameter tuning. Games that were either too hard (such as *Montezuma’s Revenge* or *Pitfall*) or too simple (such as *Pong* or *Atlantis*) are intentionally excluded to make the additional tests as significant as possible. For each game with 5 seeds conducted, we computed the median and mean human-normalized performance averaged over the 5 trials. This way, we can both have a reasonable estimation of the stability of the trainings, and a comparison as fair as possible between Rainbow and Rainbow-IQN.
![Median normalized scores with regards to training steps averaged over 5 seeds for both Rainbow and Rainbow-IQN. Only the 14 games on which 5 seeds have been conducted were used for this figure.[]{data-label="fig:5_seeds_rainbow_riqn"}](images/median_human_normalised_5_seeds_both_rainbow_riqn.png){width="\textwidth"}
Figure \[fig:5\_seeds\_rainbow\_riqn\] shows the median averaged over 5 trials for both Rainbow and Rainbow-IQN. We also plot each seed separately and the standard deviation over the 5 seeds. This has been computed only on the 14 games on which we succeeded to conduct 5 runs. This shows that standard deviations are roughly similar for Rainbow and Rainbow-IQN, around 0.2 % on the world record baseline. As these standard deviations are rather small for 25% of the Atari games, we can assume they would be still small on the whole benchmark. We think that this reveals that both Rainbow and Rainbow-IQN are quite stable on Atari and strengthens our confidence on Rainbow-IQN being the new state-of-the-art on the Atari benchmark. In particular, Rainbow-IQN reaches *infinite game time* on *Asteroids* on all 5 trials whereas Rainbow fails for each seed.
Conclusion: why is RL that Bad at Atari Games?
==============================================
In the current work, we confirm the impact of standardized guidelines for DRL evaluation, and build a consolidated benchmark, SABER. The importance of the play time is highlighted: agents should be trained and evaluated with no time limitation. To provide a more significant comparison, a new baseline is built, based on human world records. Following these recommendations, we show that the state-of-the-art Rainbow agent is in fact far from human world records performance. As a further illustration, we provide an improvement, Rainbow-IQN, and use it to measure the impact of the evaluation time over performance.
The striking information from these results is that general DRL algorithms are far from best human performance. The median of world records human normalized score for Rainbow-IQN is 3,1%, meaning that for half of the games, the agent is only 3% of the way from random play to the actual best human play. There are many possible reasons for this failure, which we will briefly discuss here to give an intuition of the current limitations of general DRL algorithm.
#### Reward clipping
In some games the optimal play for the RL algorithm is not the same as for the human player. Indeed, all rewards are clipped between -1 and 1 so RL agents will prefer to obtain many small rewards over a single large one. This problem is well represented in the game *Bowling*: the agent learns to avoid striking or sparing. Indeed the actual optimal play is to perform 10 strikes in a row leading to one big reward of 300 (clipped to 1 for the RL agent) but the optimal play for the RL agent is to knock off bowling pins one by one. This shows the need of a better way to handle reward of different magnitude, by using an invertible value function as suggested by @OberveAndLook or using Pop-Art normalization [@PopArt].
#### Exploration
Another common reason for failure is a lack of exploration, resulting in the agent getting stuck in a local minimum. Random exploration or Noisy Networks [@NoisyNetwork] are far from being enough to solve most of Atari games. In *Kangaroo* for example, the agent learns to obtain rewards easily on the first level but never tries to go to the next level. This problem might be exacerbated by the reward clipping: changing level may yield a higher reward, but for the RL algorithm all rewards are the same. Exploration is one of the most studied field in Reinforcement Learning, so possible solutions could rely on curiosity [@Curiosity] or count-based exploration [@Ostrovski].
#### Human basic knowledge
Atari games are designed for human players, so they rely on implicit prior knowledge. This will give a human player information on actions that are probably positive, but with no immediate score reward (climbing a ladder, avoiding a skull etc). The most representative example can be seen in *Riverraid*: shooting a fuel container gives an immediate score reward, but taking it makes it possible to play longer. Current general RL agents do not identify it as a potential bonus, and so die quickly. Even with smart exploration, this remains an open challenge for any general agent.
#### Loop on a sub-optimal policy
Finally, we discovered that on some games the agent finds quickly a loop continuously giving a small amount of reward and spends the whole training on this loop. In *Bank Heist* for example, the agent understood that bonus were respawning when changing level. Therefore the agent learned to just take over and over the same bonus until game timeout, failing to reach a good score. A very similar behaviour was discovered on *Elevator Action, Kangaroo, Krull* and *Tutankham*.
Acknowledgements {#acknowledgements .unnumbered}
================
The authors would like to thank Maximilian Jaritz for his thorough review and help on writing the final version of this paper.
We also thank Gabriel de Marmiesse for his assistance to make the open-source implementation on ValeoAI Github [^4].
Finally, most of computing resources were kindly provided by Centre de Calcul Recherche et Technologie (CCRT) of CEA.
Supplementary materials: Implementation details
===============================================
Rainbow Ape-X
-------------
Practically, we started with the PyTorch [@PyTorch] open source implementation of Rainbow coming from Kaikhin [@RainbowKaixhin]. We tested this initial implementation on some games with the exact same training conditions as in the original Rainbow to ensure our results were consistent. After this sanity check, we implemented a distributed version of Rainbow following the paper Distributed Prioritized Experience Replay (Ape-X) [@Horgan]. Ape-X [@Horgan] is a distributed version of Prioritized Experience Replay (PER) but which can be adapted on any value-based RL algorithm including PER, e.g. Rainbow. There is no study of this in the main article because we lacked time and computing resources to run experiments on whole Atari set with distributed actors. However, some experiments were conducted to ensure our distributed implementation was working as expected. These experiments are detailed in the next section. We claim that our Ape-X implementation is an important practical improvement compared to the single agent implementation of both Dopamine [@castro18dopamine] and Kaikhin [@RainbowKaixhin]. It is important to note that all the experiments detailed in the main paper have been made with a single actor and thus do not really show the interest of distributed Rainbow Ape-X. A lock was added to synchronize all single-agent experiments to ensure that one step of learner is done every 4 steps of actor as in the original Rainbow [@Rainbow]. All our hyperparameter values match closely those reported in Rainbow [@Rainbow]. There is still one difference coming from our Ape-X implementation (even using a single actor). Indeed, we compute priorities before putting transitions in memory instead of putting new transitions with the maximum priorities seen as in the original Rainbow [@Rainbow]. We argue that this should not have much impact on single-actor setting and that it is straightforward to implement for each algorithm using Prioritized Experience Replay [@PER].
For the distributed memory implementation, we use a key-memory database with REDIS [@redis]. The database is kept in RAM, which makes access faster and is possible for the ALE considering the size of the images and the replay memory size.
Rainbow-IQN Ape-X
-----------------
We combined our Rainbow Ape-X implementation with IQN [@Dabney2018] coming from the TensorFlow [@Tensorflow] open source implementation of Dopamine [@castro18dopamine] to obtain a PyTorch [@PyTorch] implementation of *Rainbow-IQN Ape-X* [^5]. All our hyperparameter values match closely those reported in IQN. As indicated in the main paper, we had to tune the *priority exponent* coming from Prioritized Experience Replay [@PER] in order to make the training stable. We tested both value of learning rate and epsilon of the adam optimizer from Rainbow and from IQN. A minor improvement in performance was found with the learning rate of IQN [@Dabney2018] (tested only on 3 games for computational reasons), which was then used for all our experiments.
Experiments {#experiments}
===========
Image preprocessing and architecture
------------------------------------
We used the same preprocessing procedure used in Rainbow and IQN, i.e an action repeat of 4, frames are converted to grayscale, resized to 84\*84 with a bilinear interpolation [^6] and max-pooled over 2 adjacent frames. The actual input to our network consists in 4 stacked frames.
Our architecture followed carefully the one from the original DQN for the main branch which was also used in Rainbow and IQN. The branch responsible of implicit quantiles is made exactly as the one from the original implementation section of IQN [@Dabney2018 p.g. 5]
Training infrastructure
-----------------------
The training of the agents was split over several computers and GPUs, containing in total:
- 3 Nvidia Titan X and 1 Nvidia Titan V (training computer)
- 1 Nvidia 1080 Ti (local workstation)
- 2 Nvidia 1080 (local workstations)
- 3 Nvidia 2080 (training computer)
- 4 Nvidia P100 (in a remote supercomputer)
- 2 Nvidia V100 (in a remote supercomputer)
- 4 Nvidia Tesla V-100 (DGX station)
- 4 Nvidia Quadro M2000 (local workstations)
Rainbow-IQN Ape-X
-----------------
To ascertain our distributed implementation of Rainbow-IQN was functional, 3 experiments were conducted with multiple actors (10 actors instead of one). All locks and synchronization processes are removed to let actors fill the replay memory as fast as possible. The experiments are stopped when the learner reaches the same number of steps as in our single-agent experiments.
Table \[tab:expe-apex\] reports the raw scores obtained by the agents on the selected games. Although the same number of batches is used in the training, there is a huge improvement in performances for the 3 games tested over the single agent version. This confirms the results coming from the Ape-X [@Horgan] paper. Even at same learner step, the agent can benefit greatly from more experiences coming from multiple actors. Thanks to PER, the learner focuses on the most important transitions in the replay memory. Moreover this could avoid being stuck in a local minimum as assumed in Ape-X [@Horgan]. For the 3 experiments done, all actors together played around 6 times more than in our single-agent setup, leading to 1,2B frames instead of 200M.
---------------- ------------- --------------
Raw score Multi-agent Single agent
Game
Asterix 274,491 28,015
Ms Pacman 9,901 6,090.74
Space Invaders 24,183 7,385.4
---------------- ------------- --------------
: Raw agents scores after training Rainbow-IQN Ape-X with 10 actors or a single synchronized actor[]{data-label="tab:expe-apex"}
Glitch and bug in the ALE
=========================
Inconsistent game behaviors and bugs were encountered while benchmarking Rainbow and Rainbow-IQN on all Atari games. The most damageable is the one described in the main article: games getting stuck forever even doing random actions. This is one of the main reasons why the *maximum stuck length* parameter is introduced.
Another issue is the *buffer rollover*: the emulator sends a reward of -1M when reaching 1M, effectively making the agent goes to 0 score over and over. For example, for our first implementation of Rainbow on Asterix, the scores were going up to 1M, then suddenly collapsing to random values between 0 and 1M. However, the trained agent was in fact playing almost perfectly and was indeed resolving the game many times before dying. This can also be observed in the reported score of Asterix by both Ape-X [@Horgan] and Rainbow [@Rainbow]: the score goes up to 1M and then varies randomly. This is an issue to compare agent, because a weaker agent could actually be reported with a higher score. We found this kind of *buffer rollover* bug in 2 others games: Video Pinball and Defender. To detect this in potential other games, we advocate to keep track of really high negative rewards. Indeed on the 61 games evaluated, there are no game on which there is reward inferior to -1000. And if it happens, most probably this is a buffer rollover and this reward should be ignored.
Additionally, on many games (such as Breakout for example), a specific key must be pressed to start the game (most of the time the Fire button). This means that agent can easily get stuck for long time because it does not press the key. This impacts the stability of the training because the replay memory is filled with useless transitions. We argue that this problem is exacerbated by not finishing episode as loss of life. Indeed there are many games where a specific key must be pressed, but only after losing a life to continue the game. Moreover this is probably harder to learn with the whole action set available, because the number of actions to iterate on is higher than with the minimal useful action set. This is definitely not a bug, and a general agent should learn to press fire to restart or start game.
Detailed experimental figures
=============================
In this section, we provide more detailed versions of the figures in the main article. The structure of this section follows the one of Section 5 of the main article.
As a reminder, all *normalized world record baseline* scores $s$ are reported according to the following equation, where we note $r$ the score of a random agent, $w$ the score of the world record, and $a$ the score of the agent to be evaluated:
$$s = \frac{a - r}{|w - r|}$$
Rainbow evaluation
------------------
Figure \[fig:rainbow\_vs\_ale\_hist\] illustrates in more details the difference between the reported original performance of [@Rainbow] (reported in the world record baseline), and the one obtained when applying the recommendations of [@Machado]. In particular, the number of failing games is much lower for the original implementation. Figure \[fig:rainbow\_vs\_ale\_games\] gives the breakdown for each game of the ALE.
![Agents performance comparison for the original Rainbow [@Rainbow] versus Rainbow trained with [@Machado] guidelines (30 minutes evaluation time to align with original conditions)[]{data-label="fig:rainbow_vs_ale_hist"}](images/rainbow_orig_vs_ale_hist.png){width="90.00000%"}
![Performance comparison per game between the original Rainbow [@Rainbow] versus Rainbow trained with [@Machado] guidelines (30 minutes evaluation time to align with original conditions)[]{data-label="fig:rainbow_vs_ale_games"}](images/rainbow_orig_vs_ale_per_game.png){width="90.00000%"}
Rainbow-IQN: evaluation and comparison {#rainbow-iqn-evaluation-and-comparison}
--------------------------------------
#### Influence of maximum episode length
Figure \[fig:rainbow\_with\_time\] details the influence of evaluation time over the performance range of the agents. As expected and discussed in the main article, evaluation time has a strong impact on the normalized performance of the agents. In particular, no agent reaches superhuman performance before 30 minutes evaluation. More agents reach superhuman performance when the evaluation time is not capped (in particular the ones that never stop playing, see next paragraph).
![Evolution of agents performance classification with evaluation time: Rainbow-IQN, 200M training frames, evaluation time ranging from 5min to SABER conditions[]{data-label="fig:rainbow_with_time"}](images/rainbow_with_time.png){width="\textwidth"}
#### Comparison of Rainbow and Rainbow-IQN
Figure \[fig:saber\_rainbow\_iqn\_games\] details the difference in performance between Rainbow and Rainbow-IQN on SABER conditions, at 200M training frames. Note that superhuman, never ending scores are artificially capped at 200% of the baseline. The most drastic difference is found on the game *asteroids*, which goes from failing to superhuman performance.
![Performance comparison per game between Rainbow and Rainbow-IQN on SABER conditions (200M training frames)[]{data-label="fig:saber_rainbow_iqn_games"}](images/rainbow_iqn_saber_comparison_full.png){width="90.00000%"}
Some failing games are still significantly improved: for example, *space invaders* is increased of roughly a factor of 3. To highlight these improvements, we compare Rainbow-IQN to Rainbow by using a normalized baseline similar to the world record baseline, but using Rainbow scores as a reference. So if we note $r$ the score of a random agent, $R$ the score of a Rainbow agent and $I$ the score of a Rainbow-IQN agent, then the normalized score $s$ is:
$$s = \frac{I - r}{|R - r|}$$
Note that we use the absolute value because in the game Skiing, the Rainbow agent is worse than the random agent. The details per game can be found in Figure \[fig:rainbow\_baseline\]. Note that games that are already superhuman in Rainbow are skipped, and that the Asteroids games, which is failing in Rainbow, becomes superhuman and is skipped in the figure for visualization purposes.
![Rainbow-IQN normalized with regards to a Rainbow baseline for each game[]{data-label="fig:rainbow_baseline"}](images/rainbow_iqn_saber_comparison_full_increase.png){width="\textwidth"}
Stability of both Rainbow and Rainbow-IQN {#stability-of-both-rainbow-and-rainbow-iqn}
-----------------------------------------
The 14 games on which we ran 5 trials for both Rainbow and Rainbow-IQN are: Asteroids, Centipede, Demon Attack, Frostbite, Gravitar, Jamesbond, Krull, Kung Fu Master, Ms Pacman, Private Eye, Seaquest, Up N Down, Yars Revenge and Zaxxon.
Raw scores
==========
For verification purposes, we provide tables containing all relevant agent scores used to build the figures from the principal article.
#### Baseline scores
Table \[tab:raw\_baseline\] contains all raw game scores for ALE games, both for the previous human baseline [@mnih2015human] and the new proposed world record baseline from TwinGalaxies. Note that some of the scores are missing for some games (marked as NA). For the world record baseline scores, some of them were extrapolated from the reported world record and are marked with a $\ast$. Indeed, some world records report the play time or other metrics (e.g. the distance travelled for *Enduro*) instead of the raw score of the game. Note that all agents are trained and reported on all games of the ALE, even if the world record baseline is computed for 58 games.
------------------- ----------- ------------------ --------------
Game Name Random [@mnih2015human] World Record
air raid 579.25 NA 23050.0
alien 211.9 7127.7 251916.0
amidar 2.34 1719.5 104159.0
assault 283.5 742.0 8647.0
asterix 268.5 8503.3 1000000.0
asteroids 1008.6 47388.7 10506650.0
atlantis 22188.0 29028.1 10604840.0
bank heist 14.0 753.1 82058.0
battle zone 3000.0 37187.5 801000.0
beam rider 414.32 16926.5 999999.0
berzerk 165.6 2630.4 1057940.0
bowling 23.48 160.7 300.0
boxing -0.69 12.1 100.0\*
breakout 1.5 30.5 864.0
carnival 700.8 NA 2541440.0
centipede 2064.77 12017.0 1301709.0
chopper command 794.0 7387.8 999999.0
crazy climber 8043.0 35829.4 219900.0
defender 4142.0 18688.9 6010500.0
demon attack 162.25 1971.0 1556345.0
double dunk -18.14 -16.4 NA
elevator action 4387.0 NA NA
enduro 0.01 860.5 9500.0\*
fishing derby -93.06 -38.7 71.0
freeway 0.01 29.6 38.0
frostbite 73.2 4334.7 454830.0
gopher 364.0 2412.5 355040.0
gravitar 226.5 3351.4 162850.0
hero 551.0 30826.4 1000000.0
ice hockey -10.03 0.9 36.0
jamesbond 27.0 302.8 45550.0
journey escape -19977.0 NA 4317804.0
kangaroo 54.0 3035.0 1424600.0
krull 1566.59 2665.5 104100.0
kung fu master 451.0 22736.3 1000000.0
montezuma revenge 0.0 4753.3 1219200.0
ms pacman 242.6 6951.6 290090.0
name this game 2404.9 8049.0 25220.0
phoenix 757.2 7242.6 4014440.0
pitfall -265.0 6463.7 114000.0
pong -20.34 14.6 21.0\*
pooyan 371.2 NA 13025.0
private eye 34.49 69571.3 101800.0
qbert 188.75 13455.0 2400000.0
riverraid 1575.4 17118.0 1000000.0
road runner 7.0 7845.0 2038100.0
robotank 2.24 11.9 76.0
seaquest 88.2 42054.7 999999.0
skiing -16267.91 -4336.9 -3272.0\*
solaris 2346.6 12326.7 111420.0
space invaders 136.15 1668.7 621535.0
star gunner 631.0 10250.0 77400.0
tennis -23.92 -8.3 NA
time pilot 3682.0 5229.2 65300.0
tutankham 15.56 167.6 5384.0
up n down 604.7 11693.2 82840.0
venture 0.0 1187.5 38900.0
video pinball 15720.98 17667.9 89218328.0
wizard of wor 534.0 4756.5 395300.0
yars revenge 3271.42 54576.9 15000105.0
zaxxon 8.0 9173.3 83700.0
------------------- ----------- ------------------ --------------
: Raw scores for ALE games, for a random agent, the beginner baseline and the world records. $\ast$ indicates games on which score has been extrapolated from the reported world record (the time of visit was 25 July 2019).[]{data-label="tab:raw_baseline"}
#### SABER raw scores for Rainbow-IQN
Table \[tab:raw\_saber\] contains all raw agents scores for ALE games for Rainbow-IQN. A few of these games (Atlantis and Defender and Asteroids for Rainbow-IQN) successfully keep playing with a positive score increase after 100 hours, so their raw scores are infinite. They are marked as *infinite gameplay* in the table, and capped at 200% of the world record baseline for the mean computation.
------------------- -------------------- --------------------- -------------------- ----------------------
Game name 10M 50M 100M 200M
air raid 7765.25 11690.0 13434.25 12289.75
alien 2740.6 1878.1 5223.0 7046.4
amidar 347.13 1554.84 2129.27 3092.05
assault 966.87 2783.49 4443.03 6372.7
asterix 3467.0 9280.0 16344.5 28015.0
asteroids 1194.16 (98.25) 3261.88 (2602.48) Infinite gameplay Infinite gameplay
atlantis Infinite gameplay Infinite gameplay Infinite gameplay Infinite gameplay
bank heist 756.4 1325.3 1402.2 1412.4
battle zone 33000.0 36730.0 33480.0 44410.0
beam rider 11510.78 11900.7 10042.74 9826.62
berzerk 546.7 697.0 640.2 892.9
bowling 29.64 30.0 29.86 29.92
boxing 92.71 98.62 98.92 98.7
breakout 53.77 121.83 132.56 175.47
carnival 5148.7 4824.1 4851.3 4566.3
centipede 2241.70 (251.56) 4099.89 (405.19) 4720.54 (626.31) 5260.96 (920.10)
chopper command 3018.0 6523.0 9053.0 11405.0
crazy climber 86310.0 118038.0 133114.0 144437.0
defender Infinite gameplay Infinite gameplay Infinite gameplay Infinite gameplay
demon attack 3433.13 (656.97) 6616.96 (2949.90) 8267.82 (3065.27) 24599.31 (17441.86)
double dunk -5.54 0.3 1.52 1.3
elevator action 2.0 0.0 43490.0 77010.0
enduro 1380.05 3867.42 5014.49 5146.73
fishing derby 22.11 34.82 48.11 49.08
freeway 32.65 33.9 33.95 33.96
frostbite 4351.54 (1456.01) 9135.10 (1611.78) 9768.28 (1742.88) 10002.78 (1752.75)
gopher 4798.4 15629.8 14136.0 15797.6
gravitar 283.70 (56.37) 1258.90 (228.31) 1725.90 (471.00) 1973.60 (614.80)
hero 13728.55 27450.65 28759.85 28957.4
ice hockey -2.43 1.8 -0.72 -0.07
jamesbond 445.70 (33.50) 609.70 (46.97) 605.00 (37.52) 870.80 (171.30)
journey escape -2096.0 -1116.0 -780.0 -736.0
kangaroo 1740.0 4416.0 7088.0 9567.0
krull 6780.10 (467.12) 8804.04 (97.77) 9132.15 (207.84) 9409.73 (98.14)
kung fu master 24102.80 (6513.61) 27867.00 (5783.19) 28905.80 (6570.13) 33312.00 (4119.74)
montezuma revenge 0.0 0.0 0.0 0.0
ms pacman 2276.30 (144.50) 5058.96 (602.27) 5871.52 (454.28) 6755.47 (555.18)
name this game 10702.2 9702.9 10094.5 9946.4
phoenix 4586.7 5145.4 5370.6 5505.8
pitfall 0.0 -3.95 -2.74 -21.34
pong 6.76 19.77 19.86 20.35
pooyan 4989.7 6334.05 6339.2 6776.7
private eye 99.40 (1.20) 144.64 (46.57) 173.02 (39.13) 164.31 (42.75)
qbert 4343.75 14809.5 16812.5 18736.25
riverraid 3955.9 15068.6 15891.3 15655.7
road runner 32737.0 51383.0 54599.0 67962.0
robotank 30.66 53.55 57.18 62.68
seaquest 3077.86 (131.08) 21853.50 (4243.86) 29694.50 (6157.97) 46735.26 (10631.30)
skiing -27031.73 -20930.88 -21053.79 -12295.78
solaris 2027.2 2770.2 2205.2 1495.4
space invaders 695.15 1748.45 3365.2 10110.4
star gunner 13345.0 52961.0 59574.0 72441.0
tennis -3.19 -0.02 -0.07 -0.03
time pilot 6501.0 11598.0 13550.0 19050.0
tutankham 128.7 177.96 284.42 288.41
up n down 18544.78 (3272.37) 44569.10 (12243.70) 56722.56 (9966.49) 110907.70 (10256.62)
venture 0.0 1046.0 1486.0 1679.0
video pinball 40107.82 1784770.52 3008620.51 1254569.69
wizard of wor 4133.0 7441.0 7466.0 9369.0
yars revenge 11077.61 (1366.42) 72860.33 (7560.21) 84238.64 (7721.16) 93144.71 (5251.19)
zaxxon 8319.00 (557.20) 12494.80 (282.63) 14077.60 (917.33) 13913.40 (585.68)
------------------- -------------------- --------------------- -------------------- ----------------------
: Raw scores for ALE game agents trained with Rainbow-IQN on SABER at 10M, 50M, 100M and 200M training frames. For the 14 games ran on 5 seeds, we also show the standard deviation. []{data-label="tab:raw_saber"}
#### Evolution of scores with time
Table \[tab:raw\_with\_time\] compares agents scores with increasing evaluation times for Rainbow and Rainbow-IQN, at 200M training frames.
------------------- ----------------- ------------------ ------------------- ----------- ------------ -------------------
Game name 5 minutes 30 minutes SABER 5 minutes 30 minutes SABER
air raid 10549 12308.25 12308.25 11107.25 12289.75 12289.75
alien 3458.5 3458.5 3458.5 7046.4 7046.4 7046.4
amidar 2835.53 2952.43 2952.43 2601.82 3092.05 3092.05
assault 3779.98 3986.1 3986.1 5178.41 6372.7 6372.7
asterix 29269 29269 29269 28015.0 28015.0 28015
asteroids 1716.90 (238) 1716.90 (238) 1716.90 (238) 30838.86 159426.4 Infinite gameplay
atlantis 129392 858765 Infinite gameplay 130475.0 839433.0 Infinite gameplay
bank heist 1563.2 1563.2 1563.2 1412.4 1412.4 1412.4
battle zone 45610 45610 45610 44410.0 44410.0 44410
beam rider 5437.14 5542.22 5542.22 8165.14 9826.62 9826.62
berzerk 1049.3 1049.3 1049.3 888.0 892.9 892.9
bowling 29.92 29.92 29.92 29.92 29.92 29.92
boxing 98.7 98.7 98.7 98.7 98.7 98.7
breakout 173.01 173.01 173.01 175.39 175.47 175.47
carnival 4163.5 4163.5 4163.5 4566.3 4566.3 4566.3
centipede 7267.82 (265) 7267.82 (265) 7267.82 (265) 5260.96 5260.96 5260.96
chopper command 7973 7973 7973 11405.0 11405.0 11405
crazy climber 133756 144373 144373 137299.0 144437.0 144437
defender 18524.71 30976.24 Infinite gameplay 19004.03 24926.15 Infinite gameplay
demon attack 10234.20 (415) 14617.11 (2215) 14617.11 (2215) 10294.51 24596.37 24599.31
double dunk 0 0 0 1.1 1.3 1.3
elevator action 13421 85499 85499 12455.0 77010.0 77010
enduro 369.87 2332.63 6044.36 373.3 2316.67 5146.73
fishing derby 43.57 43.57 43.57 49.08 49.08 49.08
freeway 33.96 33.96 33.96 33.96 33.96 33.96
frostbite 7075.14 (656) 7075.14 (656) 7075.14 (656) 10002.78 10002.78 10002.78
gopher 12405 16736.4 16736.4 11724.8 15797.6 15797.6
gravitar 2647.50 (398) 2647.50 (398) 2647.50 (398) 1973.6 1973.6 1973.6
hero 28911.15 28911.15 28911.15 28957.4 28957.4 28957.4
ice hockey -0.69 -0.69 -0.69 -0.07 -0.07 -0.07
jamesbond 1421.00 (502) 1434.00 (509) 1434.00 (509) 870.8 870.8 870.8
journey escape -645 -645 -645 -736.0 -736.0 -736
kangaroo 13242 13242 13242 9567.0 9567.0 9567
krull 4697.19 (273) 4697.19 (273) 4697.19 (273) 9409.73 9409.73 9409.73
kung fu master 32265.20 (6476) 32692.80 (6790) 32692.80 (6790) 32934.8 33312.0 33312
montezuma revenge 0 0 0 0.0 0.0 0
ms pacman 4738.30 (265) 4738.30 (265) 4738.30 (265) 6755.47 6755.47 6755.47
name this game 8187.4 11787.7 11787.7 7579.8 9946.4 9946.4
phoenix 5943.9 5943.9 5943.9 5505.8 5505.8 5505.8
pitfall 0 0 0 -11.11 -21.34 -21.34
pong 20.35 20.35 20.35 20.35 20.35 20.35
pooyan 4766.3 4788.5 4788.5 6466.6 6776.7 6776.7
private eye 100.00 (0) 100.00 (0) 100.00 (0) 164.31 164.31 164.31
qbert 26116 26171.75 26171.75 18736.25 18736.25 18736.25
riverraid 18456 18456 18456 15655.7 15655.7 15655.7
road runner 66593 66593 66593 67962.0 67962.0 67962
robotank 52.34 62.99 62.99 51.35 62.68 62.68
seaquest 12281.82 (7018) 20670.40 (17377) 20670.40 (17377) 28554.0 46735.26 46735.26
skiing -28105.83 -28134.23 -28134.23 -12294.58 -12295.78 -12295.78
solaris 2299.4 2779.4 2779.4 819.0 1495.4 1495.4
space invaders 2764.55 2764.55 2764.55 4718.2 10110.4 10110.4
star gunner 72944 73331 73331 71705.0 72441.0 72441
tennis 0 0 0 -0.03 -0.03 -0.03
time pilot 20198 20198 20198 19050.0 19050.0 19050
tutankham 177.17 177.42 177.42 288.41 288.41 288.41
up n down 52599 (4454) 105213 (23843) 105213 (23843) 56646.0 110655.76 110907.7
venture 1781 1781 1781 1679.0 1679.0 1679
video pinball 96345.36 656571.52 2197677.95 76587.14 465419.66 1254569.69
wizard of wor 9913 9943 9943 9369.0 9369.0 9369
yars revenge 60913 (2342) 60913 (2342) 60913 (2342) 93144.71 93144.71 93144.71
zaxxon 19017 (1228) 19060 (1238) 19060 (1238) 13913.4 13913.4 13913.4
------------------- ----------------- ------------------ ------------------- ----------- ------------ -------------------
: Agent scores for Rainbow and Rainbow-IQN at 200M training frames, reported for 5min, 30min and SABER (no limit) evaluation time. Standard deviation are showed for Rainbow (for Rainbow-IQN it can be found on next tables).[]{data-label="tab:raw_with_time"}
#### Evolution of scores with training frames
Table \[tab:raw\_with\_train\_frames\_5min\] (resp. Table \[tab:raw\_with\_train\_frames\_30min\]) contains all raw agents scores for ALE games for Rainbow-IQN, with an evaluation time of 5 minutes (resp. 30 minutes), after 10M, 50M, 100M and finally 200M training frames.
------------------- -------------------- -------------------- --------------------- --------------------
Game name 10M 50M 100M 200M
air raid 7549.0 9168.75 10272.75 11107.25
alien 2740.6 1878.1 5223.0 7046.4
amidar 347.13 1554.84 2129.27 2601.82
assault 966.87 2783.49 4103.89 5178.41
asterix 3467.0 9280.0 16344.5 28015.0
asteroids 1194.16 (98.25) 3251.64 (2582.07) 12261.36 (12251.18) 30838.86 (5427.63)
atlantis 101945.0 118844.0 125696.0 130475.0
bank heist 756.4 1325.3 1402.2 1412.4
battle zone 33000.0 36730.0 33480.0 44410.0
beam rider 6764.82 8554.82 7818.72 8165.14
berzerk 546.7 697.0 640.2 888.0
bowling 29.64 30.0 29.86 29.92
boxing 92.71 98.62 98.92 98.7
breakout 53.77 121.83 132.56 175.39
carnival 5148.7 4824.1 4851.3 4566.3
centipede 2241.70 (251.56) 4099.89 (405.19) 4720.54 (626.31) 5260.96 (920.10)
chopper command 3018.0 6523.0 9053.0 11405.0
crazy climber 86085.0 117582.0 130559.0 137299.0
defender 36353.98 19608.36 18915.17 19004.03
demon attack 3383.66 (648.57) 5833.77 (1542.65) 7161.13 (1364.32) 10294.51 (1868.54)
double dunk -5.24 0.3 1.52 1.1
elevator action 2.0 0.0 7360.0 12455.0
enduro 340.68 379.34 382.91 373.3
fishing derby 22.11 34.82 48.11 49.08
freeway 32.65 33.9 33.95 33.96
frostbite 4351.54 (1456.01) 9135.10 (1611.78) 9768.28 (1742.88) 10002.78 (1752.75)
gopher 4798.4 11561.0 10944.4 11724.8
gravitar 283.70 (56.37) 1258.90 (228.31) 1725.90 (471.00) 1973.60 (614.80)
hero 13728.55 27450.65 28759.85 28957.4
ice hockey -2.43 1.8 -0.72 -0.07
jamesbond 445.70 (33.50) 609.70 (46.97) 605.00 (37.52) 870.80 (171.30)
journey escape -2096.0 -1116.0 -780.0 -736.0
kangaroo 1740.0 4416.0 7088.0 9567.0
krull 6780.10 (467.12) 8804.04 (97.77) 9132.15 (207.84) 9409.73 (98.14)
kung fu master 23970.80 (6513.13) 27701.20 (5814.09) 28708.80 (6580.12) 32934.80 (4170.04)
montezuma revenge 0.0 0.0 0.0 0.0
ms pacman 2276.30 (144.50) 5058.96 (602.27) 5871.52 (454.28) 6755.47 (555.18)
name this game 8212.4 7790.3 7754.6 7579.8
phoenix 4586.7 5145.4 5370.6 5505.8
pitfall 0.0 -3.95 -2.58 -11.11
pong 6.29 19.77 19.86 20.35
pooyan 4956.6 6233.55 6183.95 6466.6
private eye 99.40 (1.20) 144.64 (46.57) 173.02 (39.13) 164.31 (42.75)
qbert 4343.75 14809.5 16812.5 18736.25
riverraid 3955.9 15068.6 15891.3 15655.7
road runner 32737.0 51383.0 54426.0 67962.0
robotank 25.0 42.14 45.56 51.35
seaquest 3077.86 (131.08) 18200.66 (2114.98) 21750.36 (1891.98) 28554.00 (3617.42)
skiing -27012.53 -20923.28 -21046.99 -12294.58
solaris 1210.6 1552.4 1338.0 819.0
space invaders 695.15 1748.45 3347.25 4718.2
star gunner 13345.0 52961.0 59572.0 71705.0
tennis -3.19 -0.02 -0.04 -0.03
time pilot 6501.0 11598.0 13550.0 19050.0
tutankham 128.7 177.71 284.42 288.41
up n down 14722.22 (2551.46) 35663.92 (7724.72) 42380.48 (4978.69) 56646.00 (2541.90)
venture 0.0 1046.0 1486.0 1679.0
video pinball 29524.06 122029.58 79508.52 76587.14
wizard of wor 4133.0 7441.0 7466.0 9369.0
yars revenge 11077.61 (1366.42) 72860.33 (7560.21) 84238.64 (7721.16) 93144.71 (5251.19)
zaxxon 8319.00 (557.20) 12494.80 (282.63) 14073.20 (910.64) 13913.40 (585.68)
------------------- -------------------- -------------------- --------------------- --------------------
: Raw scores for ALE game agents trained for Rainbow-IQN at 10M, 50M, 100M and 200M training frames for 5 minutes evaluation. For the 14 games ran on 5 seeds, we also show the standard deviation.[]{data-label="tab:raw_with_train_frames_5min"}
------------------- -------------------- --------------------- --------------------- ----------------------
Game name 10M 50M 100M 200M
air raid 7765.25 11690.0 13434.25 12289.75
alien 2740.6 1878.1 5223.0 7046.4
amidar 347.13 1554.84 2129.27 3092.05
assault 966.87 2783.49 4443.03 6372.7
asterix 3467.0 9280.0 16344.5 28015.0
asteroids 1194.16 (98.25) 3261.88 (2602.48) 48027.06 (56599.58) 159426.40 (56987.42)
atlantis 261697.0 788006.0 817118.0 839433.0
bank heist 756.4 1325.3 1402.2 1412.4
battle zone 33000.0 36730.0 33480.0 44410.0
beam rider 11510.78 11900.7 10042.74 9826.62
berzerk 546.7 697.0 640.2 892.9
bowling 29.64 30.0 29.86 29.92
boxing 92.71 98.62 98.92 98.7
breakout 53.77 121.83 132.56 175.47
carnival 5148.7 4824.1 4851.3 4566.3
centipede 2241.70 (251.56) 4099.89 (405.19) 4720.54 (626.31) 5260.96 (920.10)
chopper command 3018.0 6523.0 9053.0 11405.0
crazy climber 86310.0 118038.0 133114.0 144437.0
defender 49409.81 35899.7 24663.27 24926.15
demon attack 3433.13 (656.97) 6616.96 (2949.90) 8267.82 (3065.27) 24596.37 (17442.46)
double dunk -5.54 0.3 1.52 1.3
elevator action 2.0 0.0 43490.0 77010.0
enduro 1378.3 2242.11 2307.42 2316.67
fishing derby 22.11 34.82 48.11 49.08
freeway 32.65 33.9 33.95 33.96
frostbite 4351.54 (1456.01) 9135.10 (1611.78) 9768.28 (1742.88) 10002.78 (1752.75)
gopher 4798.4 15629.8 14136.0 15797.6
gravitar 283.70 (56.37) 1258.90 (228.31) 1725.90 (471.00) 1973.60 (614.80)
hero 13728.55 27450.65 28759.85 28957.4
ice hockey -2.43 1.8 -0.72 -0.07
jamesbond 445.70 (33.50) 609.70 (46.97) 605.00 (37.52) 870.80 (171.30)
journey escape -2096.0 -1116.0 -780.0 -736.0
kangaroo 1740.0 4416.0 7088.0 9567.0
krull 6780.10 (467.12) 8804.04 (97.77) 9132.15 (207.84) 9409.73 (98.14)
kung fu master 24102.80 (6513.61) 27867.00 (5783.19) 28905.80 (6570.13) 33312.00 (4119.74)
montezuma revenge 0.0 0.0 0.0 0.0
ms pacman 2276.30 (144.50) 5058.96 (602.27) 5871.52 (454.28) 6755.47 (555.18)
name this game 10702.2 9702.9 10094.5 9946.4
phoenix 4586.7 5145.4 5370.6 5505.8
pitfall 0.0 -3.95 -2.74 -21.34
pong 6.76 19.77 19.86 20.35
pooyan 4989.7 6334.05 6339.2 6776.7
private eye 99.40 (1.20) 144.64 (46.57) 173.02 (39.13) 164.31 (42.75)
qbert 4343.75 14809.5 16812.5 18736.25
riverraid 3955.9 15068.6 15891.3 15655.7
road runner 32737.0 51383.0 54599.0 67962.0
robotank 30.66 53.55 57.18 62.68
seaquest 3077.86 (131.08) 21853.50 (4243.86) 29694.50 (6157.97) 46735.26 (10631.30)
skiing -27031.73 -20930.88 -21053.79 -12295.78
solaris 2027.2 2770.2 2205.2 1495.4
space invaders 695.15 1748.45 3365.2 10110.4
star gunner 13345.0 52961.0 59574.0 72441.0
tennis -3.19 -0.02 -0.07 -0.03
time pilot 6501.0 11598.0 13550.0 19050.0
tutankham 128.7 177.96 284.42 288.41
up n down 18516.40 (3286.13) 44569.10 (12243.70) 56722.56 (9966.49) 110655.76 (10325.07)
venture 0.0 1046.0 1486.0 1679.0
video pinball 40107.82 798642.24 565903.18 465419.66
wizard of wor 4133.0 7441.0 7466.0 9369.0
yars revenge 11077.61 (1366.42) 72860.33 (7560.21) 84238.64 (7721.16) 93144.71 (5251.19)
zaxxon 8319.00 (557.20) 12494.80 (282.63) 14077.60 (917.33) 13913.40 (585.68)
------------------- -------------------- --------------------- --------------------- ----------------------
: Raw scores for ALE game agents trained for Rainbow-IQN at 10M, 50M, 100M and 200M training frames for 30 minutes evaluation. For the 14 games ran on 5 seeds, we also show the standard deviation.[]{data-label="tab:raw_with_train_frames_30min"}
[^1]: Code available at https://github.com/valeoai/rainbow-iqn-apex
[^2]: on the TwinGalaxies website <https://www.twingalaxies.com/games.php?platformid=5>
[^3]: See supplementary materials for details
[^4]: https://github.com/valeoai/rainbow-iqn-apex
[^5]: Code available at https://github.com/valeoai/rainbow-iqn-apex
[^6]: for some experiments we made this interpolation using the Python image library PIL instead of OpenCV because OpenCV was not available on the remote supercomputer. This was leading to small differences in the final resized image.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Dai [Aoki]{}$^{1,2}$[^1], Georg [Knebel]{}$^2$, and Jacques [Flouquet]{}$^2$'
title: Fermi Surface Instabilities in Ferromagnetic Superconductor URhGe
---
Introduction
============
The coexistence of ferromagnetism (FM) and superconductivity (SC) attracts much interest because unconventional superconductivity is expected. [@Aok12_JPSJ_review] In the conventional view, SC competes against FM, since the strong internal field due to the FM order easily destroys Cooper pairs.
The first case of the microscopic coexistence of FM and SC was found in UGe$_2$, [@Sax00] where SC appears in the FM phase just below the FM critical pressure $P_{\rm c}\sim 1.5\,{\rm GPa}$, where the FM state changes into the paramagnetic (PM) state. After the discovery of UGe$_2$, two other uranium ferromagnets, URhGe and UCoGe, were found to be superconductors even at ambient pressure. [@Aok01; @Huy07] It is considered that the triplet state of Cooper pairs is responsible for SC, because it can survive even in the strong internal field due to FM. One of the most spectacular characteristics is the field-reinforced (field-reentrant) superconductivity (RSC). [@Lev05; @Aok09_UCoGe] In URhGe and UCoGe, when the field is applied along the hard-magnetization axis ($b$-axis), the FM Curie temperature $T_{\rm Curie}$ is suppressed to $0\,{\rm K}$. The simple image is that the effective mass of conduction electrons increases in the region of $T_{\rm Curie}\to 0$ and then SC is reinforced under a magnetic field. Our previous results of resistivity, magnetization, and specific heat measurements clearly indicate that the resistivity $A$ coefficient and the specific heat $\gamma$-value increase at high fields for $H\parallel b$-axis, whereas they decrease for $H\parallel c$-axis (easy-magnetization axis). [@Miy08; @Har11; @Aok11_ICHE; @Aok09_UCoGe] Correspondingly, the suppression of the FM longitudinal fluctuation is observed in UCoGe by NMR experiments when the field is applied along the easy-magnetization axis ($c$-axis). [@Hat12]
Up to now, the Fermi surface has been assumed to be unchanged under a magnetic field. In reality, the Fermi surface can be affected by the magnetic field owing to the polarization of the moment between the majority- and minority-spin bands, or by the change of the ground state itself. In the triplet equal-spin pairing, $H_{\rm c2}$, which is governed by the orbital limit, is linked to the Fermi velocity $v_{\rm F}$ by the relation $H_{\rm c2}\propto (T_{\rm sc}/v_{\rm F})^2$. Thus, the enhancement of $H_{\rm c2}$ can be induced by ether the collapse of the Fermi wave vector $k_{\rm F}$ or the enhancement of the effective mass $m^\ast$. Therefore, it is important to clarify the interplay between the Fermi surface instability and the superconductivity.
A clear example is UGe$_2$, in which the Fermi surfaces markedly change among FM1 (weakly polarized phase), FM2 (strongly polarized phase), and PM, as detected by de Haas-van Alphen (dHvA) experiments. [@Ter01; @Set02] Corresponding to the change of Fermi surfaces, the $H_{\rm c2}$ curve shows an S-shape at a pressure of $P_{\rm x} < P <P_{\rm c}$, where $P_{\rm x}$ is the critical pressure between FM2 and FM1, and $P_{\rm c}$ is that between FM1 and PM. [@She01] In UCoGe, the S-shaped $H_{\rm c2}$ curve for $H\parallel b$-axis was qualitatively explained by the results of thermopower measurement, which is a sensitive probe for the Fermi surface change [@Mal12]. Furthermore, the Shubnikov-de Haas (SdH) experiments show the modification of the Fermi surface at high fields above $20\,{\rm T}$, indicating that UCoGe is a low carrier system associated with a large effective mass, [@Aok11_UCoGe] which is favorable for the field-induced Fermi surface change. In URhGe, the SdH experiments reveal the collapse of a small pocket Fermi surface around the spin reorientation field $H_{\rm R}$. [@Yel11] However, in both UCoGe and URhGe, the Fermi surface is not fully determined experimentally because of the insufficient sample quality and heavy effective mass.
Thus, in order to study the RSC and Fermi surface instabilities, we have chosen URhGe and measured the Hall effect at low temperatures at high fields with the precise tuning of field directions, using high-quality single crystals.
URhGe crystallizes in the TiNiSi-type orthorhombic crystal structure. The FM order occurs at $T_{\rm Curie}=9.5\,{\rm K}$ with the ordered moment of $0.42\,\mu_{\rm B}/{\rm U}$, as shown in Fig. \[fig:Mag\_phase\](a). The moment is directed along the $c$-axis with a collinear structure. The SC appears below $T_{\rm sc}=0.25\,{\rm K}$ at zero field. The electronic specific heat coefficient is $160\,{\rm mJ\, K^{-2} mol^{-1}}$, indicating that URhGe is a moderately enhanced heavy fermion system. When the field is applied along the $b$-axis (hard-magnetization axis), the moment starts to tilt from the $c$-axis to the $b$-axis with the field, and finally the moment is completely directed along the $b$-axis above the spin reorientation field $H_{\rm R}\sim 12\,{\rm T}$, which is connected to the decrease in $T_{\rm Curie}$ under a magnetic field. Interestingly, the RSC appears around the field window approximately from $9$ to $13\,{\rm T}$. The temperature-field phase diagram [@Aok12_JPSJ_review] and magnetization curves [@Har11] are shown in Fig. \[fig:Mag\_phase\](b) and the bottom-right inset of Fig. \[fig:Mag\_phase\](a), respectively.
In this paper, we present the results of Hall effect measurements in URhGe for $H \parallel a$-, $b$-, and $c$-axes. From the low-field measurements for $H\parallel c$-axis, it is found that the Hall coefficient in the FM state changes from that in the PM state, indicating the reconstruction of the Fermi surface below $T_{\rm Curie}$. A large jump of Hall resistivity at $H_{\rm R}$ for $H\parallel b$-axis implies that the Fermi surfaces markedly change also through $H_{\rm R}$. The first-order transition at $H_{\rm R}$ for $H\parallel b$-axis was clearly detected in Hall resistivity with a hysteresis, which is immediately suppressed by tilting the field direction slightly to the $c$-axis. In addition, an anisotropic response of magnetoresistance for $H\parallel b$-axis between $J \parallel a$-, $b$- $c$-axes is found. These results suggest that two effects are favorable for RSC in URhGe. One is the ferromagnetic fluctuation, which was already observed as the enhancement of the effective mass $m^{\ast}$. The other is the Fermi surface instability, which is detected by the present Hall effect measurements and previous SdH experiments [@Yel11].
Experimental
============
High-quality single crystals of URhGe were grown by the Czochralski method in a tetra-arc furnace. The grown single crystals were annealed under ultra high vacuum at high temperatures. The single-crystal ingot was then oriented by taking X-ray Laue photographs and cut using a spark cutter. The quality of the single crystals was checked by resistivity measurements at low temperatures down to $0.1\,{\rm K}$ using a homemade adiabatic demagnetization refrigerator (ADR) combined with a commercial PPMS. All samples in the present studies show superconductivity at $\sim 0.25\,{\rm K}$, and RSC was confirmed for $H\parallel b$-axis. The residual resistivity ratio (RRR) is 20–40. The thin samples for Hall effect measurements with a thickness of $0.15$–$0.05\,{\rm mm}$ were prepared for $H\parallel a$-, $b$- and, $c$-axes. The Hall effect was measured by the four-probe AC method ($f\sim 17\,{\rm Hz}$) at high fields up to $16\,{\rm T}$ and at low temperatures down to $0.1\,{\rm K}$. The field was applied for both positive and negative directions to eliminate the effect of magnetoresistance. In addition, the magetoresistance was measured by the four-probe AC method under the same experimental conditions using the same samples. The magnetization and susceptibility, which were used for the analysis of Hall effect measurements, were measured by a commercial SQUID magnetometer at temperature down to $2\,{\rm K}$ and at high fields up to $5.5\,{\rm T}$. For the analysis of high-field Hall effect data for $H\parallel b$-axis at low temperatures, the magnetization data in Ref. were used.
Results and Discussion
======================
Figure \[fig:Hall\_Tdep\] shows the temperature dependences of the Hall resistivity for $H\parallel a$- and $b$-, and $c$-axes in URhGe. As shown in Fig. \[fig:Hall\_Tdep\](a), when a small field ($0.1\,{\rm T}$) is applied along the easy-magnetization axis ($H\parallel c$-axis), the Hall resistivity increases on cooling from room temperature and shows a peak just below $T_{\rm Curie}$ ($=9.5\,{\rm K}$). By applying a higher field ($1\,{\rm T}$), a broad and larger maximum is observed at approximately $12\,{\rm K}$. These results display typical behaviors of the Hall effect in ferromagnets. The Hall resistivity $\rho_{xy}$ can be described by $$\rho_{xy} = R_{\rm 0} H + R_{\rm s} M,
\label{eq1}$$ where $R_0$ is the normal Hall coefficient and the second term is attributed to the anomalous Hall effect with the magnetization $M$. The anomalous Hall effect originates from skew scattering, side jump scattering, and the Berry phase. In general, the anomalous Hall effect is very large in ferromagnets. The decrease in $\rho_{xy}$ with decreasing temperature below $9\,{\rm K}$ is mainly due to the strong decrease in resistivity in the FM state, which plays an important role in the anomalous Hall effect.
To extract the Hall coefficient $R_0$, the Hall resistivity data are plotted in the form of $\rho_{xy}/H$ vs $\rho M/H$, assuming that the anomalous Hall effect mainly originates from skew scattering, namely, $R_{\rm s}\propto \rho$, as shown in Fig. \[fig:Hall\_analysis\](a). A good linear relation is found in the wide temperature range from $300$ to $12\,{\rm K}$. The intercept for $\rho M/H \to 0$ gives $R_0=-5.6\times 10^{-9}\,{\rm m^3/C}$ in the paramagnetic state. Here, we assume that the normal Hall coefficient is constant above $T_{\rm Curie}$. Assuming the single-band model with $|R_0|=1/(ne)$, we obtain the carrier number $n=1.1\times 10^{27}\,{\rm /m^3}$, which is equal to $0.25$ electrons/unit cell. The negative sign of $R_0$ indicates that the carrier is dominated by electrons.
It should be noted that $\rho_{xy}$ in Fig. \[fig:Hall\_Tdep\](a) is always positive because of the large positive contribution of the anomalous Hall effect, plus the small negative contribution of the normal Hall effect. For example, at $100\,{\rm K}$ at $0.1\,{\rm T}$, the contribution of the normal Hall effect is $R_0 H = -0.056 \,\mu\Omega\!\cdot\!{\rm cm}$, while the anomalous Hall effect will give $R_{\rm s}M=0.104\,\mu\Omega\!\cdot \!{\rm cm}$. Thus, the Hall resistivity in total has a positive sign with $\rho_{xy}=0.10\,\mu\Omega\!\cdot\!{\rm cm}$.
In the FM state well below $T_{\rm Curie}$, a good linear relation was also found between $2$ and $3\,{\rm K}$, as shown in Fig. \[fig:Hall\_analysis\](b), which gives $R_0=-2.2\times 10^{-8}\,{\rm m^3/C}$. Since the linear fit as a function of $\rho M/H$ is only limited, the field dependence of $\rho_{xy}$ was also measured at $2\,{\rm K}$, as shown in Fig. \[fig:Hall\_analysis\](c). Following the same method mentioned above, a good linear relation is again obtained, as shown in Fig. \[fig:Hall\_analysis\](d). The obtained $R_0$ is $-5.4\times 10^{-8}\,{\rm m^3/C}$ at $2\,{\rm K}$ in the ferromagnetic state. This value is not very far from that obtained from the temperature scan in Fig. \[fig:Hall\_analysis\](b), supporting the validity of the fitting. The large change in $R_0$ between the PM and FM states with a one order magnitude difference implies that the Fermi surface is reconstructed at the FM transition.
For $H\parallel b$-axis, which corresponds to the hard-magnetization axis, a sharp kink is observed at $T_{\rm Curie}$, as shown in Fig. \[fig:Hall\_Tdep\](b). At room temperature, $\rho_{xy}$ is positive and smoothly increases with decreasing temperature. A sharp minimum with a negative sign appears at $T_{\rm Curie}$ and then $\rho_{xy}$ becomes closer to zero at lower temperatures.
On the other hand, $\rho_{xy}$ for $H\parallel a$-axis decreases below $T_{\rm Curie}$ with a tiny but sharp minimum at $T_{\rm Curie}$, retaining the positive sign. The sharp kinks at $T_{\rm Curie}$ at $1\,{\rm T}$ in both cases may correspond to the sharp anomaly of susceptibility for $H\parallel b$- and $a$-axes at $T_{\rm Curie}$, which can be defined even at high fields when the field is parallel to the hard axes.
Figure \[fig:Hall\_Hdep\_b\](a) shows the field dependence of the Hall resistivity for $H\parallel b$-axis at low temperatures below $3\,{\rm K}$. At $0.21\,{\rm K}$, $\rho_{xy}$ is almost constant, crossing zero up to $13\,{\rm T}$. RSC is observed in the field range from $11$ to $12.5\,{\rm T}$, as indicated by small downward arrows in Fig. \[fig:Hall\_Hdep\_b\](a). The sharp positive jump at $13\,{\rm T}$ corresponds to the recovery of the normal state after the spin reorientation at $H_{\rm R}\sim 12.5\,{\rm T}$. With further increasing field, $\rho_{xy}$ rapidly decreases with a sign change from positive to negative.
Fine structures are also found at $0.21\,{\rm K}$, as indicated by small upward arrows in Fig. \[fig:Hall\_Hdep\_b\](a). These anomalies are immediately smeared out by increasing the temperature. This behavior seems to be similar to the results obtained by the thermopower measurements in UCoGe, URu$_2$Si$_2$, and YbRh$_2$Si$_2$, where many anomalies are detected as a function of field only at low temperatures. [@Mal12; @Pou13; @Pou13_YbRh2Si2; @Pfa13]
No RSC is observed at higher temperatures ($T \geq 0.4\,{\rm K}$). Note that the field dependence of $\rho_{xy}$ with a poor-quality sample (RRR $\sim 5$, not shown here) is highly different from that with a high-quality sample (RRR $\sim 40$) shown in Fig. \[fig:Hall\_Hdep\_b\](a), although the anomaly due to the spin reorientation is clearly observed in both cases. This is most likely due to the large contribution of the anomalous Hall effect, which includes the magnetoresistance.
The first-order transition at $H_{\rm R}$ was clearly detected in Hall resistivity. Figure \[fig:Hall\_Hdep\_b\](b) shows the field dependence of $\rho_{xy}$ near $H_{\rm R}$ for $H\parallel b$-axis, using a different sample with the fine tuning of the field direction by rotating the sample. The temperature was maintained at $0.8\,{\rm K}$ to avoid any trace of RSC. A clear hysteresis between up- and down-sweep fields is observed at $0.8\,{\rm K}$, indicating the first-order transition. When the field direction is slightly tilted by 3 deg from the $b$ to $c$-axes, no hysteresis is found within the experimental precision, as shown in the inset of Fig. \[fig:Hall\_Hdep\_b\](b). The broad jump of $\rho_{xy}$ is related to the jump of magnetization, as shown in the inset of Fig. \[fig:Mag\_phase\](a).
Figures \[fig:MR\](a)-\[fig:MR\](c) show the magnetoresistance for $H\parallel b$-axis with different current directions $J\parallel a$-, $b$-, and $c$-axes for different temperatures. All data show the SC and RSC, although the magnetoresistance for $J\parallel a$-axis shows only a drop of resistivity instead of zero resistivity because of the sample quality. The critical fields of SC and RSC slightly differ among $J\parallel a$-, $b$-, and $c$-axes, because of the sample quality and small misorientation against the field direction within 1 deg.
Above $0.6\,{\rm K}$, all magnetoresistances with different current directions show a peak at $H_{\rm R}\sim 12\,{\rm T}$, which corresponds to the spin reorientation. For $J\parallel a$-axis with transverse configuration, the positive magnetoresistance is observed, and magnetoresistance shows a slightly higher value above $H_{\rm R}$ than below $H_{\rm R}$, as we previously reported in Ref. . For $J\parallel b$-axis corresponding to the longitudinal configuration, the magnetoresistance is positive and almost constant above $5\,{\rm T}$. At high fields above $H_{\rm R}$, the magnetoresistance is slightly smaller than that below $H_{\rm R}$. For $J\parallel c$-axis, the initial positive magnetoresistance changes into the negative magnetoresistance above $H_{\rm R}$ with a large decrease in $\Delta \rho/\rho_0$, as shown in Fig. \[fig:MR\](d)
The response of the magnetoresistance in heavy fermion compounds can have different contributions such as the enhancement of the elastic and inelastic resistivity terms on crossing magnetic instability, valence instability or Fermi surface instability with feedbacks on $k_{\rm F}$ and $m^\ast$. Qualitatively, the response of the magnetoresistance is a mark of electronic instability with emerging maxima regardless of the current direction for the three configurations. It is clearly related to the extrapolated enhancement of the $\gamma$-value observed at $H_{\rm R}$.
In $3d$-electron systems, it is known that the anisotropic magnetoresistance with different current directions is mainly due to the anisotropic spin-orbit coupling in the ferromagnets, where the different densities of states between up and down spins contribute to the different magnetoresistances. [@Cam82] When the current direction is perpendicular to the direction of the moment, $J \perp M$, the magnetoresistance decreases, while the magnetoresistance can increase for $J\parallel M$. This behavior is also observed in the $5f$-electron system, such as UCoAl [@Mat00].
In URhGe, the magnetoresistance for $J\parallel c$-axis can be explained by the spin reorientation, where $J \perp M$ is realized above $H_{\rm R}$, showing the decrease in magnetoresistance, while at low fields below $H_{\rm R}$, the moment starts to tilt gradually from the $c$ to $b$-axes; thus, the behavior of the magnetoresistance is not simple.
To analyze the field dependence of $\rho_{xy}$ for $H\parallel b$-axis, the data of magnetoresistance in Fig. \[fig:MR\](c) and magnetization data in the inset of Fig. \[fig:Mag\_phase\](a) were used, following Eq. (\[eq1\]) with the skew scattering $R_{\rm s}M \propto \rho M$. The inset of Fig. \[fig:Hall\_Hdep\_b\](a) shows the plot in the form of $\rho_{xy}/H$ vs $\rho M/H$ in a low-field range from $2.5$ to $9\,{\rm T}$ and in a high-field range from $13.8$ to $16\,{\rm T}$. The extrapolation of $\rho_{xy}/H$ for $\rho M/H \to 0$ corresponds to the Hall coefficient. As a rough estimate, $R_0 = 2.2 \times 10^{-11}\,{\rm m^3/C}$ at low fields below $H_{\rm R}$ and $R_0 = -1.6\times 10^{-9}\,{\rm m^3/C}$ at high fields above $H_{\rm R}$, which was extracted from the quadratic extrapolation for $\rho M/H \to 0$. Using these values, we obtain $R_0 H = 0.0055\,{\mu\Omega\!\cdot\!{\rm cm}}$ and $R_{\rm s}M =-0.026 \,{\mu\Omega\!\cdot\!{\rm cm}}$ for the low-field FM phase at $2.5\,{\rm T}$. For high fields in the polarized PM phase, $R_0 H = -2.56\,{\mu\Omega\!\cdot\!{\rm cm}}$ and $R_{\rm s}M = 2.94\,{\mu\Omega\!\cdot\!{\rm cm}}$ at $16\,{\rm T}$. Note that the sign of the anomalous Hall effect also changes below/above $H_{\rm R}$.
It is difficult to estimate the precise carrier number; nonetheless the large difference in $R_0$ between $H < H_{\rm R}$ and $H > H_{\rm R}$ is indicative of Fermi surface reconstruction at $H_{\rm R}$. Assuming a simple one-band model, the Fermi surface is smaller above $H_{\rm R}$ than below $H_{\rm R}$.
However, URhGe is a multiband system; thus, the interpretation of the Hall effect is not simple. The large mobility, that is, light carrier and long scattering lifetime, mainly contributes to the normal Hall coefficient. Thus, complementary experiments, such as thermoelectric power measurements, which are dominantly sensitive for the heavy band, or quantum oscillation measurements as a microscopic probe, are required to determine the Fermi surface change more precisely. However, the present experimental results indicate that at least part of the Fermi surface is strongly modified at $H_{\rm R}$.
URhGe is a compensated metal with equal carrier numbers of electrons and holes in both the FM and PM states. The Fermi surfaces in the FM state consist of four different bands, according to the band structure calculation based on the spin-polarized LAPW method with the 5$f$-itinerant model [@Yamagami]. The calculated Fermi surfaces in the FM state are highly different from those in the PM state. The Fermi surfaces in the PM state also consist of four different bands, but the shape of the Fermi surfaces differs from that in the FM state. Furthermore the total volume of the calculated Fermi surface corresponding to the carrier number is larger in the FM state than in the PM state.
In recent ARPES experiments [@Fuj14], the 5$f$ electron is found to be itinerant. The change of the electronic structure in the FM state is also found, although the observed bands are not fully in agreement with those obtained from the calculations.
For $H\parallel b$-axis, $T_{\rm Curie}$ decreases with the field, as shown in Fig. \[fig:Mag\_phase\](b). It is connected with the first order transition at $H_{\rm R}$ with a large change of the sublattice magnetization, as observed in UGe$_2$ from the FM2 phase to the FM1 phase or from the FM1 phase to the PM phase. When a first-order transition occurs with large change of the FM sublattice magnetization, a marked change of the Fermi surface is generally expected, in agreement with the present results showing the large change in the Hall coefficient. In URhGe, the change of the Fermi surface occurs between the low-field PM and FM phases at $T_{\rm Curie}$ on cooling. By entering into the new phase above $H_{\rm R}$ for $H \parallel b$-axis, the extrapolation of the magnetization suggests a zero-field FM component $M_0^b\sim 0.1\,\mu_{\rm B}$, which is far lower than the zero-field FM moment $M_0^c\sim 0.4\,\mu_{\rm B}$ directly obtained for $H\parallel c$-axis, as shown in the inset of Fig. \[fig:Mag\_phase\](a). Thus, the change of the Fermi surface topology at $H_{\rm R}$ can be regarded as an analogous case to UGe$_2$ in which the ground state changes either from FM2 ($M_0\sim 1.5\,\mu_{\rm B}$) to FM1 ($M_0\sim 1\,\mu_{\rm B}$), or from FM1 to PM if the component of the magnetic moment for $H\parallel b$-axis is zero ($M_0^b \sim 0$).
In URhGe, the FM structure is collinear. The Brillouin zone in the FM state is not modified from that in the PM state. Thus, the change of the Fermi surface at $H_{\rm R}$ is not due to a change of Brillouin zone. One can consider that the Lifshitz-type transition occurs at $H_{\rm R}$ and modifies the Fermi surfaces. In SdH experiments, one small pocket Fermi surface ($F=5\times 10^6\,{\rm Oe}$) with heavy cyclotron mass ($m^\ast \sim 20\,m_0$) is detected below $H_{\rm R}$, [@Yel11] which seems to collapse at $H_{\rm R}$. This restricted observation suggests the fact that the Lifshitz transition has the driving mechanism for RSC. However, the detected Fermi surface carries only $1.5\,{\%}$ of the total $\gamma$-value. It is difficult to explain the large change in $\gamma$-value from $160$ to $220\,{\rm mJ\,K^{-2} mol^{-1}}$ at $H_{\rm R}$, as experimentally observed in Ref. .
Our results, which show a large change in the Hall coefficient at $H_{\rm R}$, is consistent with the Fermi surface reconstruction at $H_{\rm R}$ observed by SdH experiments. It should be noted that the Lifshitz-type transition is not necessarily associated with the first-order transition in general, because it is basically a “continuous” evolution of the Fermi surface due to the Zeeman effect. In thermopower measurements at low temperatures, many anomalies are often observed as a function of field, for example in UCoGe, URu$_2$Si$_2$, and YbRh$_2$Si$_2$ [@Mal12; @Pou13; @Pou13_YbRh2Si2; @Pfa13]. The other sign changes in $\rho_{xy}$ at low temperatures below $H_{\rm R}$ in Fig. \[fig:Hall\_Hdep\_b\](a) may indicate the precursor effect of the Fermi surface evolution with the field.
Our experiments confirm that the Fermi surface is strongly modified when the ordered moment is changed. For comparison, the Hall resistivities of three ferromagnetic superconductors are shown in Fig. \[fig:Hall\_compare\]. As shown in Fig. \[fig:Hall\_compare\](a), the Hall resistivity in the ferromagnetic superconductor UGe$_2$ changes markedly at $H_{\rm c}$ which separates the PM and the FM1 states through a first-order transition, indicating a marked change of the Fermi surface.
In UCoGe, surprisingly, no clear anomaly is observed in the Hall resistivity for $H \parallel b$-axis [@Aok14_SCES], however, the thermopower measurements detect the anomaly at approximately $12\,{\rm T}$, implying the Fermi surface change. [@Mal12] In the magnetization measurements of UCoGe, no clear anomaly was found so far for $H\parallel b$-axis, because the ordered moment ($m_0\simeq 0.05\,\mu_{\rm B}$) is one order of magnitude smaller than that in URhGe, and the ferromagnetic transition is not very clearly detected compared with that in UGe$_2$ or URhGe. Furthermore, the initial slope of magnetization for $H \parallel b$-axis, namely, $dM/dH |_{H\to 0}$ is not very large compared with that for $H\parallel c$-axis [@Huy08; @Kna12], which is not favorable for spin reorientation.
Summary
=======
We measured the Hall resistivity of URhGe using high-quality single crystals. The Fermi surface change between PM and FM at low fields was observed from the change in the Hall coefficient. The Fermi surface further changes when a field is applied along the $b$-axis across the spin-reorientation field $H_{\rm R}$. This change is most likely explained by the Lifshitz-type transition associated with magnetic instabilities related to the marked change of the FM sublattice magnetization.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank S. Araki, S. Fujimori, H. Harima, L. Malone, K. Miyake, A. Pourret, and H. Yamagami for useful discussions. This work was supported by ERC starting grant (NewHeavyFermion), French ANR project (CORMAT, SINUS, DELICE), KAKENHI, ICC-IMR, and REIMEI.
[10]{}
D. Aoki and J. Flouquet, J. Phys. Soc. Jpn. [**81**]{}, 011003 (2012).
S. S. Saxena, P. Agarwal, K. Ahilan, F. M. Grosche, R. K. W. Haselwimmer, M. J. Steiner, E. Pugh, I. R. Walker, S. R. Julian, P. Monthoux, G. G. Lonzarich, A. Huxley, I. Sheikin, D. Braithwaite, and J. Flouquet, Nature [**406**]{}, 587 (2000).
D. Aoki, A. Huxley, E. Ressouche, D. Braithwaite, J. Flouquet, J.-P. Brison, E. Lhotel, and C. Paulsen, Nature [**413**]{}, 613 (2001).
N. T. Huy, A. Gasparini, D. E. [de Nijs]{}, Y. Huang, J. C. P. Klaasse, T. Gortenmulder, A. [de Visser]{}, A. Hamann, T. [Görlach]{}, and H. v. [Löhneysen]{}, Phys. Rev. Lett. [**99**]{}, 067006 (2007).
F. Lévy, I. Sheikin, B. Grenier, and A. D. Huxley, Science [**309**]{}, 1343 (2005).
D. Aoki, T. D. Matsuda, V. Taufour, E. Hassinger, G. Knebel, and J. Flouquet, J. Phys. Soc. Jpn. [**78**]{}, 113709 (2009).
A. Miyake, D. Aoki, and J. Flouquet, J. Phys. Soc. Jpn. [**77**]{}, 094709 (2008).
F. Hardy, D. Aoki, C. Meingast, P. Schweiss, P. Burger, H. v. Loehneysen, and J. Flouquet, Phys. Rev. B [**83**]{}, 195107 (2011).
D. Aoki, T. D. Matsuda, F. Hardy, C. Meingast, V. Taufour, E. Hassinger, I. Sheikin, C. Paulsen, G. Knebel, H. Kotegawa, and J. Flouquet, J. Phys. Soc. Jpn. [**80**]{}, SA008 (2011).
T. Hattori, Y. Ihara, Y. Nakai, K. Ishida, Y. Tada, S. Fujimoto, N. Kawakami, E. Osaki, K. Deguchi, N. K. Sato, and I. Satoh, Phys. Rev. Lett. [**108**]{}, 066403 (2012).
T. Terashima, T. Matsumoto, C. Terakura, S. Uji, N. Kimura, M. Endo, T. Komatsubara, and H. Aoki, Phys. Rev. Lett. [**87**]{}, 166401 (2001).
R. Settai, M. Nakashima, S. Araki, Y. Haga, T. C. Kobayashi, N. Tateiwa, H. Yamagami, and Y. Ōnuki, J. Phys., Condens. Matter [**14**]{}, L29 (2002).
I. Sheikin, A. Huxley, D. Braithwaite, J. P. Brison, S. Watanabe, K. Miyake, and J. Flouquet, Phys. Rev. B [**64**]{}, 220503 (2001).
L. Malone, L. Howald, A. Pourret, D. Aoki, V. Taufour, G. Knebel, and J. Flouquet, Phys. Rev. B [**85**]{}, 024526 (2012).
D. Aoki, I. Sheikin, T. D. Matsuda, V. Taufour, G. Knebel, and J. Flouquet, J. Phys. Soc. Jpn. [**80**]{}, 013705 (2011).
E. A. Yelland, J. M. Barraclough, W. Wang, K. V. Kamenev, and A. D. Huxley, Nature Phys. [**7**]{}, 890 (2011).
A. Pourret, A. Palacio-Morales, S. [Krämer]{}, L. Malone, M. Nardone, D. Aoki, G. Knebel, and J. Flouquet, J. Phys. Soc. Jpn. [**82**]{}, 034706 (2013).
A. Pourret, G. Knebel, T. D. Matsuda, G. Lapertot, and J. Flouquet, J. Phys. Soc. Jpn. [**82**]{}, 053704 (2013).
H. Pfau, R. Daou, S. Lausberg, H. R. Naren, M. Brando, S. Friedemann, S. Wirth, T. Westerkamp, U. Stockert, P. Gegenwart, C. Krellner, C. Geibel, G. Zwicknagl, and F. Steglich, Phys. Rev. Lett. [**110**]{}, 256403 (2013).
I. A. Campbell and A. Fert, [*Ferromagnetic Materials*]{} (North-Holland, Amsterdam, 1982) Vol. 3, Chap. 9, p. 747.
T. D. Matsuda, H. Sugawara, Y. Aoki, H. Sato, A. V. Andreev, Y. Shiokawa, V. Sechovsky, and L. Havela, Phys. Rev. B [**62**]{}, 13852 (2000).
H. Yamagami, to be published.
S. Fujimori, I. Kawasaki, A. Yasui, Y. Takeda, T. Okane, Y. Saitoh, A. Fujimori, H. Yamagami, Y. Haga, E. Yamamoto, and Y. [Ō]{}nuki, Phys. Rev. B [**89**]{}, 104518 (2014).
D. Aoki and J. Flouquet, J. Phys. Soc. Jpn. [**83**]{}, 061011 (2014).
N. T. Huy, D. E. [de Nijs]{}, Y. K. Huang, and A. [de Visser]{}, Phys. Rev. Lett. [**100**]{}, 077002 (2008).
W. Knafo, T. D. Matsuda, D. Aoki, F. Hardy, G. W. Scheerer, G. Ballon, M. Nardone, A. Zitouni, C. Meingast, and J. Flouquet, Phys. Rev. B [**86**]{}, 184416 (2012).
H. Kotegawa, V. Taufour, D. Aoki, G. Knebel, and J. Flouquet, J. Phys. Soc. Jpn. [**80**]{}, 083703 (2011).
[^1]: E-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Observational tests of stellar and Galactic chemical evolution call for the joint knowledge of a star’s physical parameters, detailed element abundances, and precise age. For cool main-sequence (MS) stars the abundances of many elements can be measured from spectroscopy, but ages are very hard to determine. The situation is different if the MS star has a white dwarf (WD) companion and a known distance, as the age of such a binary system can then be determined precisely from the photometric properties of the cooling WD. As a pilot study for obtaining precise age determinations of field MS stars, we identify nearly one hundred candidate for such wide binary systems: a faint WD whose GPS1 proper motion matches that of a brighter MS star in Gaia/TGAS with a good parallax ($\sigma_\varpi/\varpi\le 0.05$). We model the WD’s multi-band photometry with the BASE-9 code using this precise distance (assumed to be common for the pair) and infer ages for each binary system. The resulting age estimates are precise to $\le 10\%$ ($\le 20\%$) for $42$ ($67$) MS-WD systems. Our analysis more than doubles the number of MS-WD systems with precise distances known to date, and it boosts the number of such systems with precise age determination by an order of magnitude. With the advent of the Gaia DR2 data, this approach will be applicable to a far larger sample, providing ages for many MS stars (that can yield detailed abundances for over 20 elements), especially in the age range 2 to 8[$\rm Gyr$]{}, where there are only few known star clusters.'
author:
- 'M. Fouesneau'
- 'H-W. Rix'
- 'T. von Hippel'
- 'D. W. Hogg'
- H Tian
bibliography:
- 'tc.bib'
title: Precise Ages of Field Stars from White Dwarf Companions
---
Introduction {#sec:introduction}
============
The two members of a binary star systems are stars born at nearly the same time from the material of the same element composition, but usually with different masses. Binary stars are not only interesting in themselves but offer a wide range of avenues to measure stellar properties and learn about stellar physics. These opportunities include the dynamical and geometrical calibration of their masses and radii [@Torres2010], or the cross-check of age or abundance estimates.
Binaries are also systems where some physical characteristics (e.g. age) are far more easily or precisely estimated from one component, while other characteristics (e.g. element composition) are far more easily estimated from the other one; yet they should be near-identical among them: this is in particular the case for wide well-resolved binary systems that consist of a main-sequence (MS) stars and a white dwarf (WD). If we have the distance, the magnitude, the color, and the atmospheric type information for a WD, we can precisely and accurately age-date that object [@Bergeron2001], yielding $\tau_{age}$. This age-dating draws on well-understood WD cooling curves and initial-final mass relations (IFMR), which have been calibrated using star clusters [e.g. @Salaris2009]. We can then safely assume that the MS primary component must be co-eval, which provides us $\tau_{age}$ of this MS field star, a quantity that would be difficult or impossible to determine (unless the star were near the MS turn-off). For MS stars, their (photospheric) element abundances $[\vec{X}/H]$ can be estimated straightforwardly from spectra, at least if they are FGK stars. The binary system as a whole then provides us with a joint estimate of temperature ${\ensuremath{{T}_{\rm eff}}}$, luminosity $L$, abundances $[\vec{X}/H]$, and a precise age $\tau_{age}$, which is fundamental input for Galactic chemical evolution studies and tests of stellar evolution. At the moment, we have excellent parallaxes for many MS stars from Gaia DR1 TGAS [@Gaia2016], but we have good direct parallax distances for only a few WDs.
In this work, we set out to identify previously unknown wide binaries consisting of MS primaries with good TGAS parallaxes, and common proper motion WD secondaries; those secondaries are equidistant, which gives us their luminosity, thereby enabling the age determination for the whole binary system. This is the same approach that @Tremblay2017 pursued, who focused on the masses and radii of their WD sample and did not determine ages.
Exploiting WD-MS binaries is by no means the only approach to determining the ages of MS field stars [e.g. @soderblom2010]. For example, for stars near the MS turn-off the precise determination of ${\ensuremath{\log{g}}}$, ${\ensuremath{{T}_{\rm eff}}}$, and ${\ensuremath{\rm [Fe/H]}}$ constrains the age well. Further, asteroseismology [@Chaplin2014] and gyrochronology [@Angus2015] have been recently proven powerful tools in practice. But those approaches are largely restricted to stars of $\gtrsim 1\,{\ensuremath{{\rm M}_\odot}}$ and yield typical age uncertainties of 30% [@Chaplin2014]. For Galactic (chemical) evolution, however, consistent tracers that exist across all relevant ages (1-13 [$\rm Gyr$]{}) are crucial: on the MS that applies to stars with $\lesssim 0.8\,{\ensuremath{{\rm M}_\odot}}$, where asteroseismic and gyrochronological approaches are difficult and far less tested. In this regime, WD-MS wide binaries may be the best way forward to reach $\sim 10\% $ age precision.
This paper is organized as follows: in Section \[sec:CandidateSelection\] we describe the identification of likely WD-MS binary systems that have TGAS information on the MS component; in Section \[sec:Ages\] we then exploit the resulting precise luminosity information of the WD to derive its cooling and overall age. In Section \[sec:Outlook\] we then discuss follow-up of our analysis and the prospects of this approach with Gaia DR2 data.
Identification of Candidate WD - MS Wide Binaries {#sec:CandidateSelection}
=================================================
We aim to identify WD-MS wide binary candidates without using the actual luminosity (or apparent magnitude) or detailed color of the possible WD component, as these quantities should subsequently serve as constraints on the WD’s age. We cannot also rely on only spectroscopically confirmed WDs, as this would severely limit the sample in sky-coverage and apparent (WD) magnitude. Requiring a precise parallax-based distance for at least one of the components (almost inevitably the MS star) limits us to MS stars with “good” parallaxes from TGAS (we adopt relative precisions $\le 5\%$). Possible WD companions to these stars have to be nearby on the sky ($\le 50$ arcsec), and we arbitrarily restrict these further to angular separations that correspond to $\le 10,000\,{\ensuremath{\rm AU}}$ at the distance of the MS primary, $D_{MS}$. Any wide but gravitationally bound WD companions will be co-moving (typically within $\le 1$ km/s) in their proper motions, $\vec{\mu}$ (at separations $\Delta\theta \ll 1$ radian). This means that as a first step we need to identify the binary components as co-moving pairs of stars (one of them in TGAS) that are projected to within $\le 10,000\,{\ensuremath{\rm AU}}$ on the sky (at $D_{MS}$).
![Color-(absolute) magnitude diagram of candidate binary companions to TGAS primary stars (blue). The WDs analyzed in this study are highlighted in red. These candidates were selected to be within $200\,{\ensuremath{\rm pc}}$, to have separations $<10,000\,{\ensuremath{\rm AU}}$, to have identical joint proper motions within 5$\sigma$. We also eliminated TGAS primaries with very small proper motions to reduce background contamination (see Appendix \[appendix:gps1query\]). Assuming the candidate secondaries to be equidistant to the TGAS primaries, we can place them on a color-magnitude diagram. The comparison with MESA isochrones (gray dots, @Dotter2016) shows a clear main sequence, and a very clear WD sequence, with some remaining contaminants (that are far from any isochrone or cooling curve). For the present paper, we only consider the candidate WD companions, identified from this diagram (red). (See Appendix \[appendix:gps1query\] for database query.) []{data-label="fig:fig_selection"}](figure1.pdf){width="\columnwidth"}
The WD secondaries will generally be much fainter than the MS primaries from TGAS. Therefore, we cannot draw on TGAS for their proper motions. Combining extensive sky coverage ($3\pi$) with proper motion precision and accuracy, the GPS1 catalog [@Tian2017] may be the best current source of such proper motions. Specifically, we queried (see Appendix \[appendix:gps1query\]) the GPS1 catalog to return the possible companions to all $\sim 100,000$ TGAS stars that had parallax measurements better than $5\%$ and parallax estimates greater than $5\,{\ensuremath{\rm mas}}$ (i.e. $<200\,{\ensuremath{\rm pc}}$ in the limit of exact parallaxes); we also required that the projected separation corresponded to less than $\le 10,000\,{\ensuremath{\rm AU}}$ and that the proper motions among the potential pair were consistent at the $5\sigma$ level. We further required that the PS1 photometry for the companion was $\sigma < 0.05$ mag in $girz$, that the sources had colors consistent with the $(g-r)~vs.~(r-i)$ color-color locus of WDs. Finally, we eliminated candidates that had very wide separation, yet low proper motions, as they are particularly susceptible to (background) contamination. The specifics are detailed in Appendix \[appendix:gps1query\].
This above selection left us with a wide binary sample of about 150 objects, where we expect the companions to the TGAS MS stars to be either fainter MS stars or WDs. Adopting the parallax-distance to the primary MS, we can construct a color – absolute magnitude diagram for the candidate companions, which is shown in Fig.\[fig:fig\_selection\]. It shows both a clear MS and a WD sequence, attesting to the fact that for the most part, we have selected equidistant (and presumably bound) companions; there are few interlopers, apparent in Fig. \[fig:fig\_selection\] as objects whose color-magnitude position is inconsistent with stellar isochrones of WD cooling curves. Some of these objects are MS-MS binaries, others may just be background contaminants. For the present paper, we are not interested in the MS secondary components and the obvious interlopers, so we eliminate them from further consideration.
Age Constraints on the Wide Binary Systems {#sec:Ages}
==========================================
We are now left with a set of 91 candidate WDs (cWD), whose distances are precisely constrained by the parallaxes to their companions. Of those, 15 are brighter (Figure \[fig:fig\_data\_cooling\_curve\], red circles) than the predictions from the $0.5\,{\ensuremath{{\rm M}_\odot}}$ cooling curve of @Bergeron1995, which implies they have masses that are too low to be consistent with single-star evolution during the age of the Universe. Thus, these objects are either the result of common envelope evolution, or are themselves unresolved binary WDs, or the photometry is contaminated, e.g. by a background source. We conservatively eliminate these objects from further consideration in this preliminary work.
To now infer precisely the ages of these WDs, we need to know and compare their trigonometric parallaxes, their spectral energy distributions (SEDs), and their atmospheric types (DA, DB, etc.) to models. Such modeling requires an understanding of WD cooling processes, of the initial-final mass relation (IFMR) of WDs, and an understanding of the precursor stars’ lifetimes as a function of mass and metallicity. In practice, this inference can be accomplished via the software suite BASE-9 [@vonHippel2006; @DeGennaro2008; @vanDyk2009; @Stein2013; @Stenning2016], which fits the SED of each cWD, using the Gaia trigonometric parallax for the MS star as prior information.
For the present context, BASE-9 serves as a flexible software package that combines stellar evolution models [e.g. @dotter2008], an IFMR [e.g. @Salaris2009; @williams2009], WD interior cooling models (e.g., @Althaus1998; @Montgomery1999, updated and expanded for our use in 2011; @Renedo2010), and WD atmosphere models [e.g. @Bergeron1995 updated regularly on-line], with photometric constraints in a wide range of possible passbands. BASE-9 accounts for the individual uncertainties for all data; the ancillary information (e.g. parallax) and astrophysical knowledge are incorporated through the prior distributions. @OMalley2013 demonstrated BASE-9 derives reliable posterior age distributions for individual field WDs and [von Hippel et al (2018, in prep)]{} show how the derived WD age precision depends on WD masses, number and quality of photometric bands, and parallax precision.
![Comparison of the color-(absolute) magnitude distribution of our candidate WDs to a set of cooling curves for DA (dashed; and DB, dotted) WDs with masses between $0.5$ and $0.9\,{\ensuremath{{\rm M}_\odot}}$. The plot also indicates (red circles) the candidates for which BASE-9 modeling could not find acceptable solutions, presumably because they are not single WDs at the primary star’s distance.[]{data-label="fig:fig_data_cooling_curve"}](fig_HRD.pdf){width="\columnwidth"}
The WD ages we derive below will indicate that these systems are most likely to be disk or thick disk stars. Because we do not yet have spectroscopic abundances (of the MS primary), we set the prior distribution on metallicity to be a broad Gaussian with a mean $\langle{\ensuremath{\rm [Fe/H]}}\rangle = -0.5\,{\ensuremath{\rm dex}}$ and a dispersion $\sigma({\ensuremath{\rm [Fe/H]}}) = 1.0\,{\ensuremath{\rm dex}}$. While we also do not have the line-of-sight absorption for these stars, they are all closer than $200\,{\ensuremath{\rm pc}}$, with most being nearer than $100\,{\ensuremath{\rm pc}}$, so we set a strong prior on the absorption of $A_{\rm 0} \approx 0$mag.
Using these input data and constraints, we ran BASE-9 on each cWD individually, without further knowledge of the properties of its MS companion, employing @dotter2008 precursor models, the @williams2009 IFMR, @Montgomery1999 WD interiors, and @Bergeron1995 WD atmospheres. Without spectroscopy, we do not know which objects are H-atmosphere (DA) WDs and which are DBs. Fortunately for our analysis, nature makes predominantly DA WDs ($\sim$75%; @Tremblay2008), and it is therefore a good initial assumption that those cWDs that have posterior distance probabilities consistent with their candidate MS companion Gaia parallaxes, are indeed DAs.
Figure \[fig:fig\_example\] presents the joint posterior distributions ([PDF]{}) for eight example WDs. Panels show the zero-age main sequence (ZAMS) mass vs.age plane, with each dot presenting a [PDF]{} sample. The panels are sorted in order of increasing mass. The first panel, for `WD 1`, shows an example where the parallax prior mean is [*inconsistent*]{} with the posterior distance distribution: models would like to predict a star older than the age of the Universe. This star is one of the 15 candidate WDs whose luminosities are above the $0.5\,{\ensuremath{{\rm M}_\odot}}$ model in Figure \[fig:fig\_data\_cooling\_curve\]. For the other seven WDs presented Fig.\[fig:fig\_example\] and for all but the 15 problematical objects identified in Figure \[fig:fig\_data\_cooling\_curve\] (red circles), their posterior distance distributions are consistent with their companion parallax prior, indicating that the model star could readily fit the data at the appropriate luminosity. The age precisions among the eight cases in Figure \[fig:fig\_data\_cooling\_curve\] range from $90\,{\ensuremath{\rm Myr}}$ to $1.46\,{\ensuremath{\rm Gyr}}$. Four of these eight WDs have fractional age errors of only 3%, and the WD with the poorest age constraint (`WD 42`, with a ZAMS mass of $1.75 \pm 0.15\,{\ensuremath{{\rm M}_\odot}}$ and age = $2.1 \pm 0.5$ [$\rm Gyr$]{}) still provides meaningful age information. This figure also indicates that a more constraining parallax prior, which would in turn further constrain the WD mass and thereby its ZAMS mass, would additionally improve the age precision for these WDs.
![The joint mass vs. age posterior distribution derived from BASE-9 modeling for eight example WD candidates. The panels list the specific WD and are ordered by increase ZAMS mass of the WD. The panels show that there are precise (though covariant) constraints on both the ages and the precursor mass. Note that the first panel presents the case of an overly low inferred WD mass that in practice could not be fit by BASE-9 in a manner consistent with its parallax. The only hint of that issue in this particular diagram is that age is running up against the age of the Universe.[]{data-label="fig:fig_example"}](figs_mass_ag.pdf){width="\columnwidth"}
The formal uncertainties in the fitted WD ages are dominated by the parallax precision. While WD models are mature and have benefited from substantial tests in star clusters, nearby binaries, and asteroseismology, the [*accuracy*]{} of the ages may still be poorer than the precision in certain regions of parameter space. Particularly WDs with ZAMS masses $\la 2\,{\ensuremath{{\rm M}_\odot}}$ or WDs with surface effective temperatures lower than about 5000K are challenging. Gaia parallaxes tightly constrain the present mass of cool WDs. But when that mass is mapped back onto the ZAMS, small uncertainties in mass transform to large uncertainties in the time a WD spent evolving as a MS star. Additionally, the IFMR is not known perfectly, and small adjustments in the IFMR may change the precursor mass values and thus the pre-WD ages, especially for low-mass precursors. Thus, for those objects, we can derive a precise cooling age, but not a precise total age. For WDs with ${\ensuremath{{T}_{\rm eff}}}\leq 5000$K, issues arise both in our present understanding of their atmospheres and possibly with additional sources of energy release during crystallization [e.g. @Horowitz2010]. We can avoid most of these problems by focusing on the WDs in a suitable mass and temperature range. Nevertheless, formal tests on WD ages have not yet been performed at the level of the best of these WD age precisions; we will have to await tests that can be performed in open clusters and WD-WD pairs with Gaia DR2. At this point, we would like to emphasize that the WD ages we derive should be highly precise and deliver excellent [*relative*]{} ages. These ages are likely to be accurate at the 5-10% level, subject to further testing.
![Age in Gyrs vs. ZAMS mass of the precursor in solar units for the 91 WD candidates, all of which are companions to TGAS main sequence stars. Error bars indicate the marginalized $\pm 1\sigma$ age and mass uncertainties; note that in many cases the age uncertainties are smaller than the symbols. Age uncertainties for WDs with precursor ZAMS masses $\leq 1.7\,{\ensuremath{{\rm M}_\odot}}$ are large, because of the increasing fraction of the system lifetime spent on the MS (rather than as a cooling WD). The color indicates the mass of the MS primary, showing that most of them are low-mass stars ($M\lesssim 1\,{\ensuremath{{\rm M}_\odot}}$), whose ages could not be inferred well from isochrones or asteroseismology. []{data-label="fig:fig_mass_age"}](age_mass_mprim.pdf){width="\columnwidth"}
The 91 WDs that BASE-9 fit consistently with the parallaxes are plotted in Figure \[fig:fig\_mass\_age\]. The error bars represent $\pm$1 standard deviations in age and ZAMS mass, respectively. Their colors indicate the approximative initial mass of their MS companion using their 2MASS photometry (and the strong prior that they live on the main-sequence). Age uncertainties drop rapidly with ZAMS masses $\ge 1.7\,{\ensuremath{{\rm M}_\odot}}$. The relative age uncertainties, in the sense $\sigma(\tau_{age})/\langle \tau_{age} \rangle$, range from the highly precise value of 1.9% to as poor as 54.5% at the low ZAMS mass end. Of these 91 WDs, 42 have relative age precisions better than 10% and 67 have relative age precisions of better than 20%. The objects plotted in Figure \[fig:fig\_mass\_age\] are both the largest sample of field WDs and the largest sample of WD - MS pairs with precise ages.
Discussion and Outlook {#sec:Outlook}
======================
In this paper we carried out a pilot study for one of the many applications of using Gaia data to constrain stellar properties. We identified systems where Gaia parallaxes gave us distances to nearby ($<200\,{\ensuremath{\rm pc}}$) main sequence stars, and where common proper motion information from the GPS1 catalog provided strong evidence for a wide (and equidistant) WD companion. Our analysis nearly doubles the number of such known wide binaries with parallax distances.
We applied the BASE-9 modeling to infer ages for the white dwarfs, which must be the same as those of the MS stars. Achieving better than 10% age precision for 42 systems, and better than 20% for another 25 systems (67 in total) constitutes an order of magnitude increase in the number of low-mass ($\sim
1~N_\odot$) MS field stars for which ages are known with that precision. This approach seems particularly suited to obtain precise ages for low-mass ($<1\,{\ensuremath{{\rm M}_\odot}}$) MS stars, where most other methods fail for field stars. The majority of our systems have ages of 1-8 [$\rm Gyr$]{}, an age range that is poorly sampled by known clusters.
To realize the scientific potential of the sample at hand, spectroscopic follow-up is necessary in two respects. First, simple low-resolution spectroscopy needs to verify which of these WDs are actually DA WD’s, as assumed in the modeling. Second, higher-resolution spectroscopy of the bright ($m<11$mag) MS stars should be used to determine their detailed abundance pattern, to increase well-calibrated constraints of $[\vec{X}/Fe]$, i.e., $\tau_{age}$ for chemical evolution modeling. We are currently pursuing this follow-up.
While this particular sample will of course be superseded by the data from Gaia’s DR2 (in April 2018), this overall approach will be particularly powerful in light of the full Gaia data. For studying the WD’s themselves, precise parallaxes will be paramount, especially for the oldest and faintest WDs. In these case, the boost in parallax precision transferred from the MS star, will aid the analysis. In turn, identifying WD companions to MS stars mostly by their common proper motion, will greatly enlarge the volumes over which this analysis can be done (compared to insisting on precise parallaxes for both the MS and the WD).
This project was developed in part at the 2017 Heidelberg Gaia Sprint, hosted by the Max-Planck-Institut für Astronomie, Heidelberg.
This work has made use of data from the European Space Agency (ESA) mission Gaia (http://www.cosmos.esa.int/gaia), processed by the Gaia Data Processing and Analysis Consortium (DPAC, http://www.cosmos.esa.int/web/gaia/dpac/consortium). Funding for the DPAC has been provided by national institutions, in particular the institutions participating in the Gaia Multilateral Agreement.
Parts of the work has used the German Astrophysical Virtual Observatory (GAVO) team at the Zentrum für Astronomie Heidelberg. GAVO is funded through the Verbundforschung of the German Ministry for Research (BMBF).
H.-W.R.’s research contribution is supported by the European Research Council under the European Union’s Seventh Framework Programme (FP 7) ERC Grant Agreement n.$\,$\[321035\]. T.v.H.’s research contribution is supported by the National Science Foundation Award AST-1715718.
H.-W.R. gratefully acknowledges early discussion with Dan Maoz that proved seminal for this paper.
GPS1$\times$TGAS Query {#appendix:gps1query}
======================
In this section, we detail the selection query we performed on TGAS and GPS1 catalogs.
Matching GPS1 against TGAS will report all the stars from GPS1 within some radius that could potentially be associated with a TGAS bright star. If we also filter on parallax and motion similarity this will only give co-moving pairs. We consider nearby objects according to TGAS parallaxes as $$\label{eq:distance}
\begin{split}
\textrm{distance}((\alpha, \delta)_{GPS1}, &(\alpha, \delta)_{TGAS})\,[\rm deg] \\
& < 10.3 \times \frac{\varpi [{\ensuremath{\rm mas}}]}{3600}.
\end{split}$$ Further tuning can be done by adding a contamination model, though this is out of the proof-of-concept scope of this paper. In addition, we need to only conserve good parallaxes within a $200\,{\ensuremath{\rm pc}}$ ($5\,{\ensuremath{\rm mas}}$) volume around the Sun as $$\label{eq:good_parallax}
\begin{split}
& \varpi \geq 5 \, {\ensuremath{\rm mas}}\\
& \frac{\varpi}{\sigma_\varpi} > 20,
\end{split}$$ and relatively good motion precision in GPS1 $$\label{eq:good_motion}
\sqrt{\sigma^2_{\mu,\alpha} + \sigma^2_{\mu,\delta}} < 6\,{\ensuremath{\rm mas.yr}^{-1}}$$
Additionally, we want pairs of objects that are co-moving according to both surveys (within their uncertainties). Therefore we select pairs that appear co-moving within $5-\sigma$ uncertainties: $$\begin{split}
& \frac{\left((\mu_\alpha^\star)_{GPS1} - (\mu_\alpha^\star)_{TGAS}\right)^2}{\left((\sigma_{\mu,\alpha})_{GPS1}^2 + (\sigma_{\mu,\alpha})_{TGAS}^2 \right)}\\
\quad & +
\frac{\left((\mu_\delta)_{GPS1} - (\mu_\delta)_{TGAS}\right)^2}{\left((\sigma_{\mu,\delta})_{GPS1}^2 + (\sigma_{\mu,\delta})_{TGAS}^2 \right)}
\leq (5 \,{\ensuremath{\rm mas.yr}^{-1}}) ^2.
\end{split}$$
However, many objects with small motion where actually contaminant or main-sequence objects. Therefore we also include a revised cut that rejects objects with small motions (despite leading to incompleteness): $$\begin{split}
& \sqrt{(\mu_\alpha^\star)_{TGAS}^2 + (\mu_\delta)_{TGAS}^2}~[{\ensuremath{\rm mas.yr}^{-1}}] \\
\quad & > {25}\left(\frac{1000}{0.3\,\varpi [{\ensuremath{\rm mas}}]} \times \textrm{distance}((\alpha, \delta)_{GPS1}, (\alpha, \delta)_{TGAS})\right)^{0.7}.
\end{split}$$ Note that the constant and power of the above equation are results of an empirical inspection. Finally, we also added color terms that avoid having main-sequence objects and we also select good photometry for their SED analysis. Based on empirical definitions we added the following selections:
$$\begin{split}
& \left|(g - i) -1.6 \times (g - r) + 0.1 \right| < 0.15\, \mathrm{mag},\\
& \left(\sigma_g,\, \sigma_r ,\, \sigma_i,\, \sigma_z\right) < 0.05\, \mathrm{mag},
\end{split}$$
This selection translates into the following ADQL query. As GAVO is currently the only service providing the GPS1 catalog, the field names correspond to their definition, and may vary when using other sources (e.g., VizieR, Gaia Archive).
Note that on Fig.\[fig:fig\_selection\], the red selection corresponds to this query, while the blue selection results from the same query were we only apply the `JOIN` and the two first `WHERE` conditions.
Catalogs
========
In this section we describe the content of the catalog generated during this study.
The catalog contains the photometric and astrometric data for all of the WD candidates of this study. For each star, we also provide the mean, median and standard deviation of the posterior PDF of the WD properties, esp. age and ZAMS mass. In addition, the catalog contains the matched MS component 2MASS (J, H, K), and WISE (W1, W2, W3, W4) photometry as well as our mass estimates and uncertainties.
Column Units Description Column Units Description
------------------ -------------------------- ---------------------------------------------------------- --------------------- ---------------------------------- -------------------------------------
`source_id` Gaia DR1 identifier `AllWISE` AllWise identifier
`magg` mag Gaia DR1 $G$ magnitude (of the WD) `gps1_ra` deg right ascension from GPS1
`e_magg` mag Gaia $G$ magnitude uncertainty `gps1_dec` deg declination from GPS1
`magr` mag GPS1 $r$ magnitude `gps1_e_ra` deg GPS1 RA uncertainty
`e_magr` mag GPS1 $r$ uncertainty `gps1_e_dec` deg GPS1 DEC uncertainty
`magi` mag GPS1 $i$ magnitude `gps1_pmra` deg/yr$^{-1}$ GPS1 $\mu_\alpha^\star$
`e_magi` mag GPS1 $i$ uncertainty `gps1_pmde` deg/yr$^{-1}$ GPS1 $\mu_\delta$
`magz` mag GPS1 $z$ magnitude `gps1_e_pmde` deg/yr$^{-1}$ GPS1 $\mu_\alpha^\star$ uncertainty
`e_magz` mag GPS1 $z$ uncertainty `gps1_e_pmra` deg/yr$^{-1}$ GPS1 $\mu_\delta$ uncertainty
`magy` mag GPS1 $y$ magnitude `primary_Hmag` mag primary $H$ photometry
`e_magy` mag GPS1 $y$ uncertainty `primary_Jmag` mag primary $J$ photometry
`magj` mag GPS1 $J$ magnitude `primary_Kmag` mag primary $K$ photometry
`e_magj` mag GPS1 $J$ uncertainty `primary_W1mag` mag primary $W1$ photometry
`magh` mag GPS1 $H$ magnitude `primary_W2mag` mag primary $W2$ photometry
`e_magh` mag GPS1 $H$ uncertainty `primary_W3mag` mag primary $W3$ photometry
`magk` mag GPS1 $K$ magnitude `primary_W4mag` mag primary $W4$ photometry
`e_magk` mag GPS1 $K$ uncertainty `primary_e_Hmag` mag primary $H$ uncertainty
`maggaia` mag GPS1 Gaia $G$ magnitude `primary_e_Jmag` mag primary $J$ uncertainty
`e_maggaia` mag GPS1 converted Gaia $G$ uncertainty `primary_e_Kmag` mag primary $K$ uncertainty
`parallax` [$\rm mas$]{} Gaia DR1 Parallax (Primary) `primary_e_W1mag` mag primary $W1$ uncertainty
`parallax_error` [$\rm mas$]{} Gaia DR1 parallax uncertainty `primary_e_W2mag` mag primary $W2$ uncertainty
`mn_Age` ${\ensuremath{\rm Gyr}}$ posterior mean WD age `primary_e_W3mag` mag primary $W3$ uncertainty
`mn_fe` dex posterior mean ${\ensuremath{\rm [Fe/H]}}$ `primary_e_W4mag` mag primary $W4$ uncertainty
`mn_mod` mag posterior mean distance modulus `primary_mass_p16` [${\rm M}_\odot$]{} 16th mass percentile
`mn_mass` [${\rm M}_\odot$]{} posterior mean WD mass `primary_mass_p50` [${\rm M}_\odot$]{} 50th mass percentile
`mn_cAge` [$\rm Gyr$]{} posterior mean WD cooling age `primary_mass_p84` [${\rm M}_\odot$]{} 84th mass percentile
`mn_pAge` [$\rm Gyr$]{} posterior mean WD precusor’s age `tgas_ra` deg right ascension from TGAS
`md_Age` ${\ensuremath{\rm Gyr}}$ posterior median WD age `tgas_ra_error` [$\rm mas$]{} TGAS right ascension uncertainty
`md_fe` dex posterior median ${\ensuremath{\rm [Fe/H]}}$ `tgas_dec` deg declination from TGAS
`md_mod` mag posterior median distance modulus `tgas_dec_error` [$\rm mas$]{} TGAS declination uncertainty
`md_mass` [${\rm M}_\odot$]{} posterior median WD mass `tgas_b` deg Galactic latitude from TGAS
`md_cAge` [$\rm Gyr$]{} posterior median WD cooling age `tgas_l` deg Galactic longitude from TGAS
`md_pAge` [$\rm Gyr$]{} posterior median WD precusor’s age `tgas_Gmag` mag primary TGAS G magnitude
`st_Age` ${\ensuremath{\rm Gyr}}$ posterior standard deviation WD age `tgas_pmdec` ${\ensuremath{\rm mas.yr}^{-1}}$ TGAS $\mu_\alpha^\star$
`st_fe` dex posterior standard deviation ${\ensuremath{\rm [Fe/H]}}$ `tgas_pmdec_error` ${\ensuremath{\rm mas.yr}^{-1}}$ TGAS $\mu_\alpha^\star$ uncertainty
`st_mod` mag posterior standard deviation distance modulus `tgas_pmra` ${\ensuremath{\rm mas.yr}^{-1}}$ TGAS $\mu_\delta$
`st_mass` [${\rm M}_\odot$]{} posterior standard deviation WD mass `tgas_pmra_error` ${\ensuremath{\rm mas.yr}^{-1}}$ TGAS $\mu_\delta$ uncertainty
`st_cAge` [$\rm Gyr$]{} posterior standard deviation WD cooling age `primary_source_id` primary TGAS DR1 identifier
`st_pAge` [$\rm Gyr$]{} posterior standard deviation WD precusor’s age
| {
"pile_set_name": "ArXiv"
} |
[**Investment and Consumption without Commitment**]{}\
Ivar Ekeland [^1]\
Department of Mathematics\
The University of British Columbia\
Vancouver, BC, V6T1Z2\
[email protected]\
Traian A. Pirvu\
Department of Mathematics\
The University of British Columbia\
Vancouver, BC, V6T1Z2\
[email protected]\
\
[**Abstract.**]{} In this paper, we investigate the Merton portfolio management problem in the context of non-exponential discounting. This gives rise to time-inconsistency of the decision-maker. If the decision-maker at time $t=0$ can commit his/her successors, he/she can choose the policy that is optimal from his/her point of view, and constrain the others to abide by it, although they do not see it as optimal for them. If there is no commitment mechanism, one must seek a subgame-perfect equilibrium strategy between the successive decision-makers. In the line of the earlier work by Ekeland and Lazrak [@EkeLaz] we give a precise definition of equilibrium strategies in the context of the portfolio management problem, with finite horizon, we characterize it by a system of partial differential equations, and we show existence in the case when the utility is CRRA and the terminal time T is small. We also investigate the infinite-horizon case and we give two different explicit solutions in the case when the utility is CRRA (in contrast with the case of exponential discount, where there is only one). Some of our results are proved under the assumption that the discount function $h(t)$ is a linear combination of two exponentials, or is the product of an exponential by a linear function.
[**Key words:**]{} Portfolio optimization, Merton problem, Equilibrium policies
[**JEL classification:** ]{}[G11]{}\
[**Mathematics Subject Classification (2000):** ]{} [91B30, 60H30, 60G44]{}
Introduction
============
The discounted utility model (DU) has been in use since the beginning of economic theory. Landmark papers ar the ones by Ramsey in $1928$ and Samuelson in $1937$. There is by now a very rich literature, the common assumption being that the discount rate is constant over time so the discount function is exponential. The model makes it possible to compare outcomes occurring at different times by discounting future utility by some constant factor. A decision maker with high discount rates exhibit more impatience (care more about the future) than one with low discount rates. Most of financial-economics works have considered that the rate of time preference is constant (exponential discounting). However there is growing evidence to suggest that this may not be the case. Ainslie [@Ain] performed empirical studies on human and animal behavior and found that discount functions are almost hyperbolic. Loewenstein and Prelec [@LoPre] show four drawbacks of exponential discounting and propose a model which accounts for them. They discuss implications for savings behavior and estimation of discount rates.
As soon as discounting is non-exponential, the decision-maker becomes time-inconsistent: a policy, to be implemented after time $t>0$, which is optimal when discounted at time $0$, no longer is optimal if it is discounted at a later time, for instance $t$. If the decision-maker at time $0$ can commit the later ones, that is, constrain them to follow the policy he/she has decided upon, then the policy which is optimal from his/her perspective can be implemented. But, apart for the dubious validity of doing that (why should his/her perspective be better than the one of decision-maker at time $t$, who, after all, will have to carry out a policy decided long before, and will be the one to bear its consequences ?) it is not often the case that management decisions are irreversible; there will usually be many opportunities to reverse a decision which, as times goes by, seems ill-advised.
[**[Existing Research]{}**]{}
Dynamic inconsistent behavior was first formalized analytically by Strotz [@STRO]. Further work by Pollak [@Pol], Peleg and Yaari [@PeYa], Goldmann [@Gol] on this issue advocates that the policies to be followed should be the output of an intra-personal game among different temporal selves (today’s self is a different player from tomorrow’s self). Laibson [@Lai] considers a discrete time consumption-investment economy without uncertainty. An agent (self) observes past consumption and financial wealth levels and chooses the consumption level for period $t.$ They establish existence of a unique subgame perfect equilibrium. It is characterized by time-dependent consumption rules which are linear in wealth and should satisfy an Euler type equation. The work of Barro [@Bar] is in a non-stochastic Ramsey model with logarithmic utility, and the discount function as a special form, whereby the rate of time preference is high in the near future but almost constant in the distant future. Krusell and Smith [@KruSm] also consider a discrete non-stochastic Ramsey paradigm with quasi-geometric discounting. They seek the equilibrium policy as the solution of a subgame-perfect equilibrium where the players are the agent and her future selves, and they show that there are several solutions to this problem. All this work is in the deterministic case, usually with discrete time.
Ekeland and Lazrak [@EkeLaz] consider a deterministic problem with continuous time, namely the Ramsey problem of economic growth with non-exponential discounting. They define an subgame perfect equilibrium strategies by letting the decision maker at time $t$ build a coalition with his/her immediate successors $s,$ with $s\in[t,t+\epsilon]$ and by letting $\epsilon\rightarrow0$. They show that these strategies are characterized by two equivalent equations: a partial differential equation with a non-local term and an integral equation. The PDE coincide with the Hamilton-Jacobi-Bellmann equation of optimal control if the discounting is exponential.
[**[Our Contributions]{}**]{}
Dynamic asset allocation has been an area in finance which received a lot of attention lately. The ground-breaking paper in this literature is Merton [@Mer71]. He considered a model consisting of a risk-free asset with constant rate of return and one or more stocks, each with constant mean rate of return and volatility. An agent invests in this market and consume to maximize her expected utility of intertemporal consumption and final wealth. Merton was able to derive closed form solution for Constant Relative Risk Aversion (CRRA) and Constant Absolute Risk Aversion (CARA) preferences. It turns out for CRRA preferences, the optimal consumption and investment in the risky asset are a constant proportion of wealth. This is not the case for CARA although they are still linear in wealth. Karatzas et al. [@KarLehShr87], Cox and Huang [@CoxHua89] also solved the optimal investment and consumption problem by the static martingale method. In another paper Merton [@Mer69] also investigates the infinite horizon case. Although it yields the same quantitative results it is easier to handle - provided the constant discount rate satisfies a condition known in the literature as the transversality condition. It does not seem to have much empirical support, and what happens if it is not satisfied is a matter of debate.
All the papers mentioned above are in the exponential discounting paradigm which as we shown earlier was challenged by some economic literature.
The goal of this paper is to understand how non-exponential discounting affects an agent’s investment-consumption policies in a Merton model. First we show that doing naive optimization does not work in the absence of commitment technologies. Then we follow the approach of [@EkeLaz] and introduce the concept of equilibrium policies in a stochastic context. In the finite horizon case we give a first description of the equilibrium policies for a general discount function through the solutions of a flow of BSDEs. If the discount function is exponential, these conditions reduce to the classical HJB equation, so the equilibrium policy coincide with the optimal one given by dynamic programming. We then introduce special classes of discount functions, type I which is a linear combination of two exponentials, and type II which is the product of an exponential by a polynomial of degree one, and we show that for discount functions in that class, equilibrium solutions can be characterized by a system of HJB equations. We go one step further and characterize them through a parabolic PDE system. This system does not seem to have been considered before. Existence of a solution can be established in the case of CRRA utility.
It is perhaps surprising that the same integral equation (IE) from [@EkeLaz] holds in this model for the special discounts. This suggests that (IE) holds for any discounts. We can see this for other types of discount but we are not concern with a general result in this direction. In an infinite horizon model the novel feature is stationarity and the key observation is, to borrow a sentence from [@EkeLaz], the decision maker at time $t$ resets her watch so that time $s$ becomes $s-t,$ so she faces the same problem as the decision-maker at time $0.$ Keeping this in mind we define the equilibrium policies in this context. We follow the same approach as for finite horizon: first we describe them through the solutions of a infinite horizon BSDE and then for the special discounts through an ODE system and an integral equation. We end this case with the study of CRRA preferences were we find interesting results. As one might probably expected the equilibrium policies and optimal ones coincide if the discounting is exponential given the transversality condition of [@Mer69]. However we find a weaker condition to grant existence of an equilibrium policy so there may be circumstances when the equilibrium policy exists but the optimal one may not (in the sense that one can not perform the verification argument). Even more interesting for type I discounts we find instances when there are two equilibrium policies. Moreover transversality-like conditions were worked out. The equilibrium policies for CRRA preferences resemble the optimal ones. More precisely the agent invests the same constant proportion of her wealth in the risky asset as in the Merton problem and consumes also a constant proportion of her wealth which is different than the Merton one for non-exponential discounting.
[**[Organization of the paper]{}**]{}
The reminder of this paper is organized as follow. In section $2$ we describe the model and formulate the objective. Section $3$ and $4$ treat the finite horizon and infinite horizon problem. The conclusions are summarized in section $5.$ The paper ends with Appendix containing the proofs.
The model and problem formulation
=================================
Financial Market
----------------
We adopt a model for the financial market consisting of a saving account and one stock (risky asset). The inclusion of more risky assets can be achieved by notational changes. The saving account accrues interest at the riskless rate $r>0.$ The stock price per share follows an exponential Brownian motion $$dS(t)=S(t)\left[\alpha\,dt
+\sigma\,dW(t)\right],\quad0\leq
t\leq\infty,$$ where $\{W(t)\}_{t\in[0,\infty)}$ is a $1-$dimensional Brownian motion on a filtered probability space $(\Omega,\{\mathcal{F}_t\}_{0\leq t\leq T},\mathcal{F},P).$ The filtration $\{\mathcal{F}_t\}_{0\leq t\leq T}$ is the completed filtration generated by $\{W(t)\}_{t\in[0,\infty)}.$ As usual $\alpha$ is *the mean rate of return* and $\sigma>0$ is [*the volatility*]{}. Let us denote $\mu\triangleq\alpha-r>0$ [*the excess return*]{}.
Investment-consumption policies and wealth processes
----------------------------------------------------
A decision-maker in this market is continuously investing her wealth in the stock/bond and is consuming. An investment-consumption policy is determined by the proportion of current wealth her/she invests in the bond/stock and consume (consumption rate). Formally we have:
\[def:portfolio-proportions\] An ${{\mathbb R}}^{2}$-valued stochastic process $\{\zeta(t), c(t)\}_{t\in[0,\infty}$ is called an [ *admissible policy process*]{} if it is progressively measurable, $c(t)\geq 0,\,\,\mbox{for all}\,\,t\in[0,\infty)$ and it satisfies $$\label{kj}
\sup_{0\leq t\leq T }\mathbb{E}[|\zeta(t)|^{m}+c^{m}(t)]<\infty\quad \forall m=1,2,\cdots,$$ This condition originates from [@Pen] and is needed since we are using their result.
Given a policy process $\{\zeta(t), c(t)\}_{t\in[0,\infty)}$, the proportion of the wealth $X^{\zeta,c}(t),$ at time $t$ invested in the stock is $\zeta(t)$ and the consumption rate is $c(t).$ The equation describing the dynamics of wealth ${X^{\zeta,c}(t)}$ is given by $$\begin{aligned}
\notag
dX^{\zeta,c}(t)&=&X^{\zeta,c}(t)\left((\alpha\zeta(t)-c(t))\,dt+\sigma\zeta(t)\,dW(t)\right)
+(1-\zeta(t))X^{\zeta,c}(t)r\,dt\\\label{equ:wealth-one}&=&X^{\zeta,c}(t)(\left(r+\mu\zeta(t)-c(t))\,dt
+\sigma\zeta(t)\,dW(t)\right).\end{aligned}$$ It simply says that the changes in wealth over time are due solely to gains/loses from investing in stock, from consumption and there is no cashflow coming in or out. This is usually referred to as the self-financing condition.
Under the regularity condition imposed on $\zeta(t)$ and $c(t)$ above, , admits a unique strong solution given by the explicit expression $$\notag \begin{split}
X^{\zeta,c}(t)=X(0)\exp\left( \int_0^t
\Big(r+\mu\zeta(u)-c(u)-|\sigma\zeta(u)|^2\Big)\, du+
\int_0^t \sigma\zeta(u)\, dW(u) \right),
\end{split}$$ The initial wealth $X^{\zeta,c}(0)=X(0)\in (0,\infty)$, is exogenously specified.
Utility Function
----------------
All decision-makers have the same von Neumann-Morgenstern utility. This is crucial for understanding the model: time-inconsistency arise, not from a change in preferences, but from the way the future is discounted. All decision-makers try to maximize the discounted expectation of a function $U:(0,\infty)\rightarrow \mathbb{R}$ strictly increasing and strictly concave, which is their (common) utility. We restrict ourselves to utility functions which are continuous differentiable and satisfy the Inada conditions $$\label{In}
U'(0+)\triangleq\lim_{x\downarrow 0}U'(x)=\infty,\quad U'(\infty)\triangleq\lim_{x\uparrow{\infty}}U'(x)=0.$$ We shall denote by $I(\cdot)$ the (continuous, strictly decreasing) inverse of the marginal utility function $U'(\cdot),$ and by $$\label{In1}
I(0+)\triangleq\lim_{x\downarrow 0}I(x)=\infty,\quad I(\infty)\triangleq\lim_{x\uparrow{\infty}}I(x)=0.$$ The agent is deriving utility from intertemporal consumption and final wealth. Let $U$ be the utility of intertemporal consumption and $\hat{U}$ the utility of the terminal wealth at some non-random horizon $T$ (which is a primitive of the model and we may alow it to be infinite).
Discount Function
-----------------
Unlike other works in this area we do not restrict ourselves to the framework of exponential discounting. Following [@EkeLaz], a discount function $h:[0,\infty]\rightarrow\mathbb{R}$ is assumed to be continuously differentiable, with: $$h(0)=1,\,\,h(s)\geq 0 ,$$ and $$\int_{0}^{\infty}h(t)\,dt<\infty.$$ The definition and characterization of equilibrium strategies (Theorems \[Hamil1\] and \[Hamil\]) hold for general discount functions. We will then particularize them to two special cases, which we will call [*[pseudo-exponential]{}*]{} discount functions. They are of two types, type I: $$h_{1}(t)=\lambda\exp(-\rho_{1}t)+(1-\lambda)\exp(-\rho_{2}t)$$ and type II: $$h_{2}(t)=(1+\lambda t)\exp(-\rho t)$$
Pseudo-exponential discount rates were first considered in the context of time-inconsistency in [@EkeLaz]. Note that, in contrast to the more studied case of hyperbolic discount, decision-makers discount the distant future less heavily than the immediate future. It will be clear from our approach how to handle other cases.
Objective
---------
We conclude this section by formulating our problem. The objective is to find time consistent policies and the optimal ones may fail to have this feature. Indeed, if the agent starts with a given positive wealth $x,$ at some instant $t,$ her optimal policy process $\{\tilde{\zeta}_{t}(s),\tilde{c}_{t}(s)\}_{u\in[t,T]}$ is chosen such that $$\sup_{\zeta,c}\mathbb{E}\left[\int_{t}^{T}h(u-t)U(c(u)X^{\zeta,c}(u))\,du+h(T-t)\hat{U}(X^{\zeta,c}(T))\right]=$$ $$=\mathbb{E}\left[\int_{t}^{T}h(u-t)U(\tilde{c}(u)X^{\tilde{\zeta},\tilde{c}}(u))\,du+h(T-t)\hat{U}(X^{\tilde{\zeta},\tilde{c}}(T))\right].$$
The value function associated with this stochastic control problem is
$$\notag
V(t,s,x)\triangleq\sup_{\zeta,c}\mathbb{E}\left[\int_{s}^{T}h(u-t)U(c(u)X^{\zeta,c}(u))\,du
+h(T-s)\hat{U}(X^{\zeta,c}(T))\bigg|X(s)=x\right],$$
$t\leq s\leq T,$ and it solves the following Hamilton-Jacobi-Bellman equation $$\frac{\partial V}{\partial s}(t,s,x)+\sup_{\zeta,c}\left[(r+\mu\zeta-c)x\frac{\partial V}{\partial x}(t,s,x)+\frac{1}{2}\sigma^{2}\zeta^{2}x^{2}\frac{\partial^{2} V}{\partial x^{2}}(t,s,x)\right]$$$$+
\frac{h'(s-t)}{h(s-t)}V(t,s,x)+U(xc)=0,$$ with the boundary condition $$\label{boundarycondition}
V(t,T,x)=\hat{U}(x).$$
The first order necessary conditions yield the $t-$optimal policy $\{\tilde{\zeta}_{t}(s),\tilde{c}_{t}(s)\}_{s\in[t,T]}$ $$\label{11}
\tilde{\zeta}_{t}(s,x)=-\frac{ \mu\frac{\partial V}{\partial x}(t,s,x) }{\sigma^{2}x\frac{\partial^{2} V}{\partial x^{2}}(t,s,x)},\quad t\leq s\leq T,$$ $$\label{c11}
\tilde{c}_{t}(s,x)=\frac{I(\frac{\partial V}{\partial x}(t,s,x))}{x},\quad t\leq s\leq T.$$
Therefore, unless the discounting is exponential (in which case $\frac{h'}{h}=\mbox{constant},$ so there is no $t$ dependence in the HJB), the $t-$optimal policy may not be optimal after $t.$ That is $$\{\tilde{\zeta}_{t}(s),\tilde{c}_{t}(s)\}_{s\in[t',T]}\neq{\operatorname{argmax}}_{\zeta, c}\mathbb{E}\left[\int_{t'}^{T}h(u-t')U(c(u)X^{\zeta,c}(u))\,du+h(T-t')\hat{U}(X^{\zeta,c}(T))\right],$$ for some subsequent instant $t',$ so the decision-maker would implement the $t-$optimal policy at later times only if she is constrained to do so. This failure to remain optimal across times can be regarded as time inconsistency.
Because of time-inconsistency, optimal solutions are irrelevant in practice (although they do exist mathematically) and one must look for an alternative way to solve the problem. This will be done by considering equilibrium policies, that is, policies such that, given that they will be implemented in the future, it is individually optimal to implement them right now. Following [@EkeLaz] consider:
$$\label{eq}
\bar{\zeta}(s,x)=\frac{F_{1}(s,x)}{x},\quad\bar{c}(s,x)=\frac{F_{2}(s,x)}{x},$$
for some functions $F_{1},$ $F_{2}$ and the equilibrium wealth process $\{\bar{X}(s)\}_{s\in[0,T]}$ evolves according to $$\label{dyn}
d\bar{X}(s)=[r\bar{X}(s)+\mu F_{1}(s,\bar{X}(s))-F_{2}(s,\bar{X}(s))]ds+\sigma F_{1}(s,\bar{X}(s))dW(s).$$ The functions $F_{1}, F_{2}$ are chosen such that on $[t,t+\epsilon]$ it is optimal (this is made precise in our formal definition of equilibrium policies) to pick $\bar{\zeta}(t,x)=\frac{F_{1}(t,x)}{x},\quad\bar{c}(t,x)=\frac{F_{2}(t,x)}{x},$ given the agent’s wealth at time $t$ is $x,$ and for every subsequent instance $s\geq t+\epsilon$ she follows .
Finite Horizon
===============
General Discount Function
-------------------------
Let $T$ be a finite time horizon exogenously specified. In general, for a policy process $\{{\zeta}(s),{c}(s)\}_{s\in[0,T]}$ satisfying and its corresponding wealth process $\{X(s)\}_{s\in[0,T]}$ (see ) we denote the expected utility functional $$\label{01FUNCT}
J(t,x,\zeta,c)\triangleq\mathbb{E}\left[\int_{t}^{T}h(s-t)U(c(s)X^{t,x}(s))\,ds+h(T-t)\hat{U}(X^{t,x}(T))\right].$$ We shall give a rigorous mathematical formulation of the equilibrium policies in the formal definition below.
\[finiteh\] A map $F=(F_{1},F_{2}):(0,\infty)\times[0,T]\rightarrow\mathbb{R}\times[0,\infty)$ is an equilibrium policy for the finite horizon investment-consumption problem, if for any $t,x>0$ $$\label{opt}
{\lim_{\epsilon\downarrow 0}}\frac{J(t,x,F_{1},F_{2})-J(t,x,\zeta_{\epsilon},c_{\epsilon})}{\epsilon}\geq 0,$$ where $$J(t,x,F_{1},F_{2})\triangleq J(t,x,\bar{\zeta},\bar{c}),$$ $$\label{0eq}
\bar{\zeta}(s)=\frac{F_{1}(s,\bar{X}(s))}{\bar{X}(s)},\quad\bar{c}(s)=\frac{F_{2}(s,\bar{X}(s))}{\bar{X}(s)},$$ and $\{\bar{\zeta}(s),\bar{c}(s)\}_{s\in[t,T]}$ should satisfy . The equilibrium wealth process $\{\bar{X}(s)\}_{s\in[t,T]}$ is a solution of the stochastic differential equation (SDE) $$\label{0dyn}
d{X}(s)=[r{X}(s)+\mu F_{1}(s,{X}(s))-F_{2}(s,{X}(s))]ds+\sigma F_{1}(s,{X}(s))dW(s).$$
The process $\{{\zeta}_{\epsilon}(s),{c}_{\epsilon}(s)\}_{s\in[t,T]}$ is another investment-consumption policy defined by $$\label{1e}
\zeta_{\epsilon}(s)=\begin{cases} \bar{\zeta}(s),\quad s\in[t,T]\backslash E_{\epsilon,t}\\
\zeta(s), \quad s\in E_{\epsilon,t}, \end{cases}$$
$$\label{2e}
c_{\epsilon}(s)=\begin{cases} \bar{c}(s),\quad s\in[t,T]\backslash E_{\epsilon,t}\\
c(s), \quad s\in E_{\epsilon,t}, \end{cases}$$
with $E_{\epsilon,t}=[t,t+\epsilon],$ and $\{{\zeta}(s),{c}(s)\}_{s\in[t,T]}$ is any policy for which $\{{\zeta}_{\epsilon}(s),{c}_{\epsilon}(s)\}_{s\in[t,T]}$ satisfy .
The basic idea is the following: the agent (self) has a different rate of impatience $\frac{h'(t)}{h(t)}$ as times goes by (unless the discounting is exponential) and can be regarded as a continuum of agents (selves); at every instant $t$ she is building a coalition with her immediate selves $s,$ with $s\in[t,t+\epsilon]$ and try to maximize expected utility of intertemporal consumption and terminal wealth given that the selves on $[t+\epsilon,T]$ agreed upon an equilibrium strategy.
Our next item in the agenda is to characterize the equilibrium policies by a means of some adjoint processes defined by a flow (one for every instant $t$) of backward stochastic differential equations (BSDE). More precisely for every $0\leq t\leq T$ the processes $\{M(t,s), N(t,s)\}_{s\in[t,T]}$ are a solution of the BSDE
$$\label{BSDE1}
\begin{cases}
dM(t,s)=-\bigg(M(t,s)\left(r+\mu\frac{\partial{F_{1}}}{\partial{x}}(s,\bar{X}(s))-\frac{\partial{F_{2}}}{\partial{x}}(s,\bar{X}(s))\right)+\sigma \frac{\partial{F_{1}}}{\partial{x}}(s,\bar{X}(s))N(t,s)+\\+h(s-t)\frac{\partial{F_{2}}}{\partial{x}}(s,\bar{X}(s))U'(F_{2}(s,\bar{X}(s)))\bigg)ds+N(t,s)dW(s)\\
M(t,T)=h(T-t)\hat{U}'(\bar{X}(T)),
\end{cases}$$
where the equilibrium wealth process $\{\bar{X}(s)\}_{s\in[0,T]}$ follows .
The next central result does not depend on the choice of the discount function $h.$
\[Hamil1\] Assume there exists a map $F=(F_{1},F_{2}):(0,\infty)\times[0,T]\rightarrow\mathbb{R}\times(0,\infty),$ continuously differentiable with respect to $x$ such that for every $t\in[0,T]$ there exists a solution $\{M(t,s), N(t,s)\}_{s\in[t,T]}$ of which satisfies $$\label{0au1}
\mu M(t,t)+\sigma N(t,t)=0,$$ and $$\label{09p}
F_{2}(t,x)=I(M(t,t)|X(t)=x).$$
Then $F$ is an equilibrium strategy.
Appendix A proves this Theorem.
Special Discount Functions
--------------------------
Following [@EkeLaz] we would like to give a characterization of the equilibrium policies in terms of both a partial differential equation and an integral equation. We restrict ourselves to three types of discounting although our method goes far beyond. We consider exponential, and pseudo-exponential discounting. Let us introduce the Legendre transform of $-U(-x)$ $$\label{Leg}
\tilde{U}(y)\triangleq\sup_{x>0}[U(x)-xy]=U(I(y))-yI(y),\quad 0<y<\infty.$$ The function $\tilde{U}(\cdot)$ is strictly decreasing, strictly convex and satisfies the dual relationships $$\label{In12}
\tilde{U}'(y)=-x\quad\mbox{iff}\quad U'(x)=-y.$$
The next Theorem is our main result. It describes the equilibrium policies through a coupled system of parabolic equations.
\[PDE\] Assume there exist two functions $v(t,x)$ and $w(t,x)$ three times continuously differentiable which satisfy $$\notag
\frac{\partial v}{\partial t}(t,x)+rx\frac{\partial v}{\partial x}(t,x)-
\frac{\mu^{2}}{2\sigma^{2}}\frac{{[\frac{\partial v}{\partial x}}(t,x)]^{2}}{\frac{\partial^{2} v}{\partial x^{2}}(t,x)}+\tilde{U}\left(\frac{\partial v}{\partial x}(t,x)\right) =\alpha_{1j}v(t,x)+\beta_{1j}w(t,x),$$ $$\frac{\partial w}{\partial t}(t,x)+\left(rx-I\left(\frac{\partial v}{\partial x}(t,x)\right)\right)\frac{\partial w}{\partial x}(t,x)
-\frac{\mu^{2}}{\sigma^{2}}\frac{{\frac{\partial v}{\partial x}}(t,x)
{\frac{\partial w}{\partial x}}(t,x)}{\frac{\partial^{2} v}{\partial x^{2}}(t,x)}$$$$+
\frac{\mu^{2}}{2\sigma^{2}}\frac{{[\frac{\partial v}{\partial x}}(t,x)]^{2}\frac{\partial^{2}w}{\partial x^{2}}(t,x)}{[\frac{\partial^{2} v}{\partial x^{2}}(t,x)]^{2}}
=\alpha_{2j}v(t,x)+\beta_{2j}w(t,x),$$ for all $(t,x)\in[0,T]\times(0,\infty),$ with boundary condition $$v(T,x)=\hat{U}(x),\qquad w(T,x)=0.$$ Then $F=(F_{1},F_{2})$ given by $$\label{09con}
F_{1}(t,x)=-\frac{\mu\frac{\partial v}{\partial x}(t,x)}{\sigma^{2}\frac{\partial^{2} v}{\partial x^{2}}(t,x)},\,\,
F_{2}(t,x)=I\left(\frac{\partial v}{\partial x}(t,x)\right),\,\,\,t\in[0,T],$$ is an equilibrium policy. The coefficients $\alpha_{ij},\, \beta_{ij}$ corresponds to different choices of discount functions. Thus for exponential discounting $$\alpha_{10}=\delta,\quad \alpha_{20}=0,\quad \beta_{10}=0,\quad \beta_{20}=0,$$ for type I $$\alpha_{11}=\lambda\rho_{1}+(1-\lambda)\rho_{2},\quad \alpha_{21}=\rho_{1}-\rho_{2},\quad \beta_{11}=
\lambda(1-\lambda)(\rho_{1}-\rho_{2}),\quad \beta_{21}=\lambda\rho_{2}+(1-\lambda)\rho_{1},$$ and for type II $$\alpha_{12}=\rho-\lambda,\quad \alpha_{22}=-\lambda,\quad \beta_{12}=\lambda,\quad \beta_{22}=\rho+\lambda.$$
Appendix B proves this Theorem.
Let us point out that for the case of exponential discounting the equilibrium policy coincide with the optimal one given by dynamic programming. Note that the value function $v$ $$\label{tr}
v(t,x)=\sup_{\zeta,c}\mathbb{E}\left[\int_{t}^{T}e^{-\delta(s-t)}U(c(s)X^{\zeta,c}(s))\,ds+
e^{-\delta(T-t)}\hat{U}(X^{\zeta,c}(T))\bigg|X(t)=x\right],$$ and $w=0$ satisfies the PDE system of Theorem \[PDE\].
The next Proposition is in the spirit of Theorem $2$ in [@EkeLaz] and gives a description of the equilibrium policy through an integral equation. It is stated only for the special discounts.
\[IE\] Assume there exist two functions $v(t,x)$ and $w(t,x)$ three times continuously differentiable which solve the PDE system of Theorem \[PDE\]. Then $v(t,x)$ satisfy the integral equation $$\label{0ie1}
v(t,x)=\mathbb{E}\left[\int_{t}^{T}h(s-t)U(F_{2}(s,\bar{X}^{t,x}(s)))\,ds+h(T-t)\hat{U}(\bar{X}^{t,x}(T))\right],$$ Recall that $\{\bar{X}(s)\}_{s\in[0,T]}$ is the equilibrium wealth process and it satisfies $$\label{10dyn}
d\bar{X}(s)=[r\bar{X}(s)+\mu F_{1}(s,\bar{X}(s))-F_{2}(s,\bar{X}(s))]ds+\sigma F_{1}(s,\bar{X}(s))dW(s).$$
Appendix C proves this Proposition.
This Proposition suggests that for a general discount function $h(t)$ the equilibrium policies are of the form for a function $v$ as in . It is not hard to see it holds for $h(t)=(1+\lambda_{11} t+\lambda_{12}t^{2}+\cdots)\exp(-\rho_{1} t)+(1+\lambda_{21} t+\lambda_{22}t^{2}+\cdots)\exp(-\rho_{2} t).$
In the next Proposition stochastic representations for $v$ and $w$ are given.
Assume there exist two functions $v(t,x)$ and $w(t,x)$ three times continuously differentiable which solve the PDE system of Theorem \[PDE\]. Then $$\notag
w(t,x)=\mathbb{E}\left[\alpha_{2j}\int_{t}^{T}{\exp(-\beta_{2j}(s-t))}v(s,\bar{X}^{t,x}(s))\,ds\right],$$
$$\!\!\!\!\!\!\!\!\!\!{v}(t,x)=\mathbb{E}\bigg[\int_{t}^{T}{\exp(-\alpha_{1j}(s-t))}\bigg[U\bigg(I\bigg(\frac{\partial v}{\partial x}(s,\bar{X}^{t,x}(s)
\bigg)
\bigg)-$$$$\beta_{1j}w(s,\bar{X}^{t,x}(s))\bigg]\,ds+h(T-t)\hat{U}(\bar{X}^{t,x}(T))\bigg].$$ Therefore $${v}(t,x)=\mathbb{E}\bigg[\int_{t}^{T}{\exp(-\alpha_{1j}(s-t))}\bigg[U\bigg(I\bigg(\frac{\partial v}{\partial x}(s,\bar{X}^{t,x}(s)\bigg)\bigg) -$$ $$-\alpha_{2j}\beta_{1j}\int_{s}^{T}
\exp(-\alpha_{2j}(z-s))v(z,\bar{X}^{t,x}(z))\,dz\bigg]\,ds+h(T-t)\hat{U}({X}^{t,x}(T))\bigg].$$
Proof: It is a direct consequence of Feynman-Kac’s formula.
$\square$
The main question (since all the results of this section are based on it) is: when does the system of Theorem \[PDE\] have solutions? Of course the answer is known for exponential discounting. Although there is work in progress, we only have a partial answer at the moment and that is for CRRA preferences, $U(x)=\hat{U}(x)=\frac{x^{p}}{p}.$ In this case one can disentangle time and wealth and look for $v(t,x)=f(t)\frac{x^{p}}{p}$ and $w(t,x)=g(t)\frac{x^{p}}{p},$ where $f(t),$ $g(t)$ solve the ODE system $$\label{on3}
f'(s)+Kf(s)+(1-p)[f(s)]^{\frac{p}{p-1}}= \alpha_{1j}f(s)+ \beta_{1j} g(s),$$
$$\label{on4}
g'(s)+Kg(s)-pg(s)[f(s)]^{\frac{1}{p-1}}= \beta_{2j}f(s)+ \beta_{2j}g(s),$$
for all $t\in[0,T]$ with boundary condition $$f(T)=1,\qquad g(T)=0,$$
where $K=rp+\frac{\mu^{2}p}{2\sigma^{2}(p-1)}.$ Existence of this is obvious on small intervals (up to explosions) $[T-\epsilon, T]$ for some $\epsilon>0.$ One can get a global result for small $\beta_{ij}$ (small $\lambda$) by the Implicit Function Theorem.
Infinite Horizon
================
We next investigate stationary equilibrium policies and this is done in the infinite horizon framework. Before engaging into the formal definition let us point the following key fact. For a time homogenous policy process $\{{\zeta}(t),{c}(t)\}_{t\in[0,\infty)}$ satisfying and its corresponding wealth process $\{X(t)\}_{t\in[0,\infty)}$ (see ) the expected utility functional $J(x,\zeta,c)$ satisfies $$\begin{aligned}
\label{FUNCT}
J(x,\zeta,c)&\triangleq&\mathbb{E}\left[\int_{t}^{\infty}h(s-t)U(c(s)X^{t,x}(s))\,ds\right]
\\\notag
&=&\mathbb{E}\left[\int_{0}^{\infty}h(s)U(c(s)X^{t,x}(t+s))\,ds\right].\\\notag
&=&\mathbb{E}\left[\int_{0}^{\infty}h(s)U(c(s)X^{0,x}(s))\,ds\right].\end{aligned}$$ This is due to the fact that the processes $\{{X}^{t,x}(t+s)\}_{s\in[0,\infty)}$ and $\{{X}^{0,x}(s)\}_{s\in[0,\infty)}$ have the same $\mathbb{P}$ distribution. Following the case of finite horizon we have:
\[infiniteh\] A map $F=(F_{1},F_{2}):(0,\infty)\rightarrow\mathbb{R}\times[0,\infty)$ is an equilibrium policy for the infinite horizon investment-consumption problem, if for any $x>0$ $$\label{1opt}
{\lim_{\epsilon\downarrow 0}}\frac{J(x,F_{1},F_{2})-J(x,\zeta_{\epsilon},c_{\epsilon})}{\epsilon}\geq 0,$$ where $$J(x,F_{1},F_{2})\triangleq J(x,\bar{\zeta},\bar{c}),$$ $$\label{0eq}
\bar{\zeta}(t)=\frac{F_{1}(\bar{X}(t))}{\bar{X}(t)},\quad\bar{c}(t)=\frac{F_{2}(\bar{X}(t))}{\bar{X}(t)},$$ and $\{\bar{\zeta}(t),\bar{c}(t)\}_{t\in[0,\infty)}$ should satisfy . The equilibrium wealth process $\{\bar{X}(t)\}_{t\in[0,\infty)}$ satisfies $$\label{00dyn}
d\bar{X}(t)=[r\bar{X}(t)+\mu F_{1}(\bar{X}(t))-F_{2}(\bar{X}(t))]dt+\sigma F_{1}(\bar{X}(t))dW(t).$$
The process $\{{\zeta}_{\epsilon}(t),{c}_{\epsilon}(t)\}_{t\in[0,\infty)}$ is another time homogenous investment-consumption policy defined by $$\label{71e}
\zeta_{\epsilon}(t)=\begin{cases} \bar{\zeta}(t),\quad t\in[0,\infty)\backslash E_{\epsilon}\\
\zeta(t), \quad t\in E_{\epsilon}, \end{cases}$$
$$\label{72e}
c_{\epsilon}(t)=\begin{cases} \bar{c}(t),\quad t\in[0,\infty)\backslash E_{\epsilon}\\
c(t), \quad t\in E_{\epsilon}. \end{cases}$$
Here $E_{\epsilon}\subset[0,\infty)$ is a measurable set with Lebesque measure $|E_{\epsilon}|=\epsilon,$ and $\{{\zeta}(t),{c}(t)\}_{t\in[0,\infty)}$ is any time homogenous policy for which $\{{\zeta}_{\epsilon}(t),{c}_{\epsilon}(t)\}_{t\in[0,\infty)}$ satisfies .
In a first step we would like to describe the equilibrium policies in terms of the process $\{M(t), N(t)\}_{t\in[0,\infty)},$ the solution of the infinite horizon BSDE $$\label{BSDE}
\begin{cases}
dM(t)=-(M(t)(\mu F'_{1}(\bar{X}(t))-F'_{2}(\bar{X}(t))+\sigma F'_{1}(\bar{X}(t))N(t)\\
+h(t)F'_{2}(\bar{X}(t))U'(F_{2}(\bar{X}(t)))dt+N(t)dW(t)\\
M(\infty)=0,\quad t\in[0,\infty),
\end{cases}$$ and the equilibrium wealth process $\{\bar{X}(t)\}_{t\in[0,\infty)}$ it is given by . For more about infinite horizon BSDE see [@Chen]. The next result is the infinite horizon counterpart of Theorem \[Hamil1\].
\[Hamil\] Assume there exists a map $F=(F_{1},F_{2}):(0,\infty)\rightarrow\mathbb{R}\times(0,\infty),$ continuously differentiable with respect to $x$ such that there exists a solution $\{M(t), N(t)\}_{t\in[0,\infty)}$ of which satisfy $$\label{00au1}
\mu M(0)+\sigma N(0)=0,$$ and $$\label{009p}
F_{2}(x)=I(M(0)|X(t)=x).$$
Then $F$ is an equilibrium strategy.
The proof is similar to Theorem \[Hamil\], so it is skipped.
$\square$
As in the preceding Section we find the equilibrium policies for exponential, type I and type II discounting. With the coefficients $\alpha_{ij}$ and $\beta_{ij}$ as in Theorem \[PDE\], we have the following formal result.
\[ODE\] Assume there exist two functions $v(x)$ and $w(x)$ three times continuously differentiable which solve the following ODE system $$\notag
rxv'(x)-\frac{\mu^{2}}{2\sigma^{2}}\frac{v'^{2}(x)}{v''(x)}+\tilde{U}\left(v'(x)\right) =\alpha_{1j}v(x)+\beta_{1j}w(x),$$ $$(rx-I(v'(x)))w'(x)
-\frac{\mu^{2}}{\sigma^{2}}\frac{v'(x)w'(x)}{w''(x)}+
\frac{\mu^{2}}{2\sigma^{2}}\frac{[v'(x)]^{2}w''(x)}{[v''(x)]^{2}}
=\alpha_{2j}v(x)+\beta_{2j}w(x),$$ for all $x\in(0,\infty).$ If in addition the following transversality condition $$\label{transv}
M(\infty)=0,$$ is satisfied, then $F=(F_{1},F_{2})$ given by $$\label{009con}
F_{1}(x)=-\frac{\mu v'(x)}{\sigma^{2}v''(x)},\,\,\,\,
F_{2}(x)=I(v'(x)),$$ is an equilibrium policy.
The proof follows as in Theorem \[PDE\].
$\square$
The next Proposition gives the description of equilibrium policies through an integral equation (IE) as in the Proposition \[IE\].
\[1IE\] Assume there exist two functions $v(x)$ and $w(x)$ three times continuously differentiable which solve the ODE system of Theorem \[ODE\]. Then $v(x)$ satisfies the integral equation $$\label{9000ie1}
v(x)=\mathbb{E}\left[\int_{0}^{\infty}h(t)U(F_{2}(\bar{X}^{0,x}(t)))\,dt\right].$$
The proof follows as in Proposition \[ODE\].
$\square$
CRRA Preferences
----------------
In this subsection we are still in the paradigm of exponential, type I, type II discounts and further investigate the case of $U(x)=\frac{x^{p}}{p}.$ Lets look for the function $v$ of the form $ v(x)=\frac{k{x^{p}}}{p},$ for a constant $k$ which is to be found. The equilibrium policies are linear in wealth $$\label{lk}
F_{1}(x)=\frac{\mu x}{(1-p)\sigma^{2}},\,\,\,F_{2}(x)=k^{\frac{1}{p-1}}x,$$ and the corresponding wealth process is $$\label{as}
\bar{X}(t)=X(0)\exp\left(\left(r+\frac{(1-2p)\mu^{2}}{2(1-p)^{2}\sigma^{2}}-k^{\frac{1}{p-1}}\right)t+ \frac{\mu}{(1-p)\sigma} W(t)\right).$$ According to Proposition \[1IE\] the value function $v$ should satisfy the integral equation (and this is also sufficient to grant that $F_{1}, F_{2}$ of is an equilibrium policy), which in this context becomes $$\label{26}
k^{\frac{1}{1-p}}=\int_{0}^{\infty}h(u)e^{\tilde{k}u}\,du,$$ where $$\label{k52}
\tilde{k}=p\left(r+\frac{\mu^{2}}{2(1-p)\sigma^{2}}-k^{\frac{1}{p-1}}\right).$$
Let us treat the three cases independently.
### Exponential Discounting
With $h(t)=\exp(-\delta t),$ reads
$$\label{36}
k^{\frac{1}{1-p}}(\delta-\tilde{k})=1,$$
given that $$\label{io1}
\delta>\tilde{k}.$$
Thus $$\label{0008}
k=\left[\frac{1}{1-p}\left(\delta-rp-\frac{p\mu^{2}}{2(1-p)\sigma^{2}}\right)\right]^{p-1}.$$
The condition
$$\label{trans}
\delta>(p\vee0)\left[\frac{\mu^{2}}{2(1-p)\sigma^{2}}+r\right],$$
should hold true in view of positivity of $k.$
The adjoint process $\{M(t)\}_{t\in[0,\infty)}$ is given by $$M(t)=\exp(-\delta t)v'(\bar{X}(t))$$ $$=kX(0)\left[\exp\left(\left(r(p-1)+\frac{(2p-1)\mu^{2}}{2(1-p)\sigma^{2}}-(p-1)k^{\frac{1}{p-1}}-\delta\right)t+ \frac{\mu}{\sigma} W(t)\right)\right]=$$ $$kX(0)\left[\exp\left(-\left(r+\frac{\mu^{2}}{2\sigma^{2}}\right)t+ \frac{\mu}{\sigma} W(t)\right)\right],$$ whence the condition $M(\infty)=0$ is automatically satisfied.
Condition is weaker than the transversality condition $$\label{trans0}
\delta>(p\vee0)\left[\frac{(2-p)\mu^{2}}{2(1-p)\sigma^{2}}+r\right],$$ of [@Mer69]. If is met the equilibrium policy coincide with the optimal policy given by dynamic programming. However if $$(p\vee0)\left[\frac{\mu^{2}}{2(1-p)\sigma^{2}}+r\right]<\delta\leq (p\vee0)\left[\frac{(2-p)\mu^{2}}{2(1-p)\sigma^{2}}+r\right],$$ equilibrium policy still exists but one cannot prove verification for the optimal policy.
### Type I Discounting
If $h(t)=\lambda\exp(-\rho_{1}t)+(1-\lambda)\exp(-\rho_{2}t),$ equation leads to
$$\label{86}
k^{\frac{1}{1-p}}=\frac{\lambda}{\rho_{1}-\tilde{k}}+\frac{1-\lambda}{\rho_{2}-\tilde{k}},$$
given that $$\label{086}
\rho_{i}-\tilde{k}>0,\,\,\, i=1,2.$$
This can be expressed as the positive root of the quadratic equation for $z=k^{\frac{1}{p-1}}$ $$\label{7rt}
Q(z)=Az^{2}+Bz+C=0,$$ where
$$\label{34rt}
A\triangleq 1-p,$$
$$\label{pk0}
B\triangleq(2p-1)\left(\frac{\mu^{2}}{2(1-p)\sigma^{2}}+r\right)+\frac{\lambda\rho_{2}+(1-\lambda)\rho_{1}}{p}-(\rho_{1}+\rho_{2}),$$
$$\label{pk1}
C\triangleq-\frac{1}{p}\left(\rho_{1}-rp-\frac{p\mu^{2}}{2(1-p)\sigma^{2}}\right)\left(\rho_{2}-rp-\frac{p\mu^{2}}{2(1-p)\sigma^{2}}\right).$$
When $\rho_{1}=\rho_{2}=\delta,$ the equation becomes $$\label{rt0}
\left[(1-p)z-\left(\delta-pr-\frac{p\mu^{2}}{
2(1-p)\sigma^{2}}\right)\right]\left[z+\frac{1}{p}\left(\delta-pr-\frac{p\mu^{2}}{
2(1-p)\sigma^{2}}\right)\right]=0,$$ which compered to brings in a new solution $$\label{artif}
z_{\delta}=\left[-\frac{1}{p}\left(\delta-rp-\frac{p\mu^{2}}{2(1-p)\sigma^{2}}\right)\right].$$ Thus when $\rho_{1},\,\rho_{2}$ are close to $\delta$ the equation has at least a positive solution provided that $\delta$ satisfies . It has two positive solutions if $p<0.$
Let us search for the transversality condition sufficient to grant ($ M(\infty)=0$). The process $\{M(t)\}_{t\in[0,\infty)}$ is given by $$\label{1q1}
M(t)\triangleq \lambda\exp(-\rho_{1}t)v'_{1}(\bar{X}(t))+(1-\lambda)\exp(-\rho_{2}t)v'_{2}(\bar{X}(t)),$$ where $$\label{00ie1}
v_{i}(x)=\mathbb{E}\left[\int_{t}^{\infty}e^{-\rho_{i}(s-t)}U(F_{2}(\bar{X}^{t,x}_{s}))\,ds\right],\,\,\,i=1,2.$$
Consequently $$\label{08}
v_{i}(x)=\frac{k^{\frac{p}{1-p}}x^{p}}{p(\rho_{i}-\tilde{k})},\,\,\,i=1,2.$$
Furthermore, for $i=1,2$
$$\exp(-\rho_{i}s)v'_{i}(\bar{X}(t))$$$$=\frac{X(0)k^{\frac{p}{1-p}}}{(\rho_{i}-\tilde{k})}
\left[\exp\left(\left(r(p-1)+\frac{(2p-1)\mu^{2}}{2(1-p)\sigma^{2}}-(p-1)k^{\frac{1}{p-1}}-\rho_{i}\right)t+ \frac{\mu}{\sigma} W(t)\right)\right].$$
Therefore the transversality condition reads
$$\label{o8}
k^{\frac{1}{p-1}}< r+ \frac{1}{1-p}\left[\rho_{i}-\frac{(2p-1)\mu^{2}}{2(1-p)\sigma^{2}}\right],\,\,\, i=1,2.$$
In general it is not obvious if the equation has a positive solutions for which the transversality condition and are met. However this is the case when the $\rho_{i}$ are close to $\delta$ and $\delta$ satisfies .
It is interesting to note that in some cases there may be two equilibrium policies. Let us illustrate this in an example: we take $p>\frac{1}{2},\,\, p\simeq\frac{1}{2} $ and denote $y\triangleq r+\frac{\mu^{2}}{
2(1-p)\sigma^{2}}.$ Moreover, for some small $\epsilon>0$ let $\rho_{1}=p{y}+\epsilon,$ $\rho_{2}=p{y}-\epsilon,$ so that $B=\frac{\epsilon(1-2\lambda)}{p},$ and $C=\frac{\epsilon^{2}}{p}.$ If $\lambda>\frac{1+\sqrt{4p(1-p)}}{2}, \,\,\lambda\simeq\frac{1+\sqrt{4p(1-p)}}{2}$ the roots of , $z_{1,2}\simeq\frac{\epsilon}{\sqrt{p(1-p)}}>\frac{\epsilon}{p}$ so that holds true. The transversality condition is satisfied if $\epsilon$ is chosen small enough.
### Type II Discounting
If $h(t)=(1+\lambda t)\exp(-\rho t),$ equation leads to
$$\label{186}
k^{\frac{1}{1-p}}=\frac{1}{\rho-\tilde{k}}+\frac{\lambda}{(\rho-\tilde{k})^{2}},$$
given that $$\label{1086}
\rho-\tilde{k}>0.$$
This can be expressed as the positive root of the quadratic equation for $z=k^{\frac{1}{p-1}}$ $$\label{1rt}
Q(z)=Az^{2}+Bz+C=0,$$ with
$$\label{1pk0}
A\triangleq 1-p,$$
$$\label{91pk0}
B\triangleq(2p-1)\left(\frac{\mu^{2}}{2(1-p)\sigma^{2}}+r\right)+\frac{\rho(1-2p)}{p}+\frac{\lambda}{p},$$
$$\label{1pk1}
C=-\frac{1}{p}\left(\rho-rp-\frac{p\mu^{2}}{2(1-p)\sigma^{2}}\right)^{2}.$$
The transversality condition in this case is
$$\label{1o8}
k^{\frac{1}{p-1}}< r+\frac{1}{1-p}\left[\rho-\frac{(2p-1)\mu^{2}}{2(1-p)\sigma^{2}}\right].$$
Similar to the type I discounting it is possible to establish a positive solution of which satisfy and when $\lambda$ is close to zero.
Concluding Remarks
==================
This paper introduced a novel concept in stochastic optimization namely the notion of equilibrium policies. We analyze the Merton portfolio management problem in the context of exponential and non-exponential discounting. Although type I, II and exponential discounting are considered only, our methodology can be extended for more general discount functions. The optimal policies are characterized by both an integral equation and a system of partial differential equations. The infinite horizon case is covered as well and it is shown that in some situations there are more equilibrium policies. Moreover for CRRA preferences the equilibrium policies are to consume and invest in the risky asset a constant proportion of the corresponding wealth, which is similar to the optimal policy for exponential discounting, but the constants are different.
Appendix
========
[**A Proof of Theorem \[Hamil1\]**]{}
We assume $\hat{U}=r=0$ to ease the presentation. Let $\{X^{\epsilon}(s)\}_{s\in[0,T]}$ be the wealth corresponding to $\{{\zeta}_{\epsilon}(s),{c}_{\epsilon}(s)\}_{s\in[0,T]}$ i.e., $$\label{eq2}
dX^{\epsilon}(s)= X^{\epsilon}(s)((r+\mu\zeta_{\epsilon}(s)-c_{\epsilon}(s))\, ds+\sigma\zeta_{\epsilon}(s)\,dW(s)).$$ The processes $\{Y^{\epsilon}(s)\}_{s\in[0,T]}$ and $\{Z^{\epsilon}(s)\}_{s\in[0,T]}$ defined by the SDE $$\label{09}
\begin{cases}
dY^{\epsilon}(s)=Y^{\epsilon}(s)(\mu \frac{\partial F_{1}}{\partial x}(s,\bar{X}(s))-
\frac{\partial F_{2}}{\partial x}(s,\bar{X}(s)))\,ds+
\sigma[Y^{\epsilon}(t)\frac{\partial F_{1}}{\partial x}(s,\bar{X}(s)+\\+(\bar{X}(s)\zeta(s)
-F_{1}(s,\bar{X}(s)){\chi_{E_{\epsilon}}(s)}]\,dW(s)\\
Y^{\epsilon}(0)=0
\end{cases}$$ and
$$\begin{cases}$$dZ\^(s)=\[Z\^(s)((s,|[X]{}(s))-(s,|[X]{}(s)))+\
+(|[X]{}(s)((s) -(s,|[X]{}(s)) )-(c(s)|[X]{}(s)-F\_[2]{}(s,|[X]{}(s))[\_[E\_]{}(s)]{}\]ds+$$\\$$+ \[Z\^(s)(s,|[X]{}(s)) +(Y\^(s)((s) -(s,|[X]{}(s))[\_[E\_]{}(s)]{}\]dW(s) $$\\$$Z\^(0)=0$$\end{cases}$$
can be regarded as first order and second order variation of $\{X(s)\}_{s\in[0,T]}$
At this point we need the following Lemma from [@Pen].
\[Peng\] For any $k\geq 1$
$$\label{1q}
\sup_{s\in[0,\infty]}\mathbb{E}|X^{\epsilon}(s)-\bar{X}(s)|^{2k}=O(\epsilon^{k}),$$
$$\label{2q}
\sup_{s\in[0,\infty]}\mathbb{E}|Y^{\epsilon}(s)|^{2k}=O(\epsilon^{k}),$$
$$\label{3q}
\sup_{s\in[0,\infty]}\mathbb{E}|Z^{\epsilon}(s)|^{2k}=O(\epsilon^{2k}),$$
$$\label{4q}
\sup_{s\in[0,\infty]}\mathbb{E}|X^{\epsilon}(s)-\bar{X}(s)-Y^{\epsilon}(s)|^{2k}=O(\epsilon^{2k})$$
$$\label{5q}
\sup_{s\in[0,\infty]}\mathbb{E}|X^{\epsilon}(s)-\bar{X}(s)-Y^{\epsilon}(s)-Z^{\epsilon}(s)|^{2k}=o(\epsilon^{2k})$$
In the light of this Lemma the following expansion holds:
$$J(t,x,\zeta_{\epsilon},c_{\epsilon})= J(t,x,F_{1},F_{2})$$$$+\mathbb{E}\int_{t}^{T}\bigg(h(s-t)(U(c(s)\bar{X}^{t,x}(s))-
U(F_{2}(\bar{X}^{t,x}(s))){\chi_{E_{\epsilon}}(s)}$$$$+h(s-t)(Y^{\epsilon}(s)+
Z^{\epsilon}(s))\frac{\partial F_{2}}{\partial x}(s,\bar{X}^{t,x}(s))U'\left(\frac{\partial F_{2}}{\partial x}(s,\bar{X}^{t,x}(s))\right)\bigg)\,ds+o(\epsilon)$$
In the above equation we would like to get rid of $\{Y^{\epsilon}(s)\}_{s\in[t,T]},$ $\{Z^{\epsilon}(s)\}_{s\in[t,T]}$ and the way to accomplish this is by using the adjoint processes $\{M(t,s), N(t,s)\}_{s\in[t,T]}$ and integration by parts. It turns out $$\!\!\!\!\!\!\!\!\!\!Y^{\epsilon}(s)\frac{\partial F_{2}}{\partial x}(s,\bar{X}(s))U'\left(\frac{\partial F_{2}}{\partial x}(s,\bar{X}(s))\right)\,ds=-Y^{\epsilon}(s)dM(t,s)$$$$-\bigg[Y^{\epsilon}(s)M(t,s)\bigg(\mu\frac{\partial F_{1}}{\partial x}(s,\bar{X}(s))-
\frac{\partial F_{2}}{\partial x}(s,\bar{X}(s)\bigg)
+\sigma Y^{\epsilon}(s) N(t,s)\frac{\partial F_{2}}{\partial x}(s,\bar{X}(s))\bigg]\,ds$$$$+Y^{\epsilon}(s)N(t,s)\,dW(s),$$
and $$\!\!\!\!\!\!\!\!\!\!Z^{\epsilon}(s)\frac{\partial F_{2}}{\partial x}(s,\bar{X}(s))U'\left(\frac{\partial F_{2}}{\partial x}(s,\bar{X}(s))\right)\,ds=-Z^{\epsilon}(s)dM(t,s)$$$$-\bigg[Z^{\epsilon}(s)M(t,s)\bigg(\mu\frac{\partial F_{1}}{\partial x}(s,\bar{X}(s))
-\frac{\partial F_{2}}{\partial x}(s,\bar{X}(s)\bigg)
+\sigma Z^{\epsilon}(s) N(t,s)\frac{\partial F_{2}}{\partial x}(s,\bar{X}(s))\bigg]\,ds$$$$+Z^{\epsilon}(s)N(t,s)\,dW(s),$$
Consequently
$$\mathbb{E}\int_{t}^{T}h(s-t)(Y^{\epsilon}(s)+Z^{\epsilon}(s))\frac{\partial F_{2}}{\partial x}(s,\bar{X}(s))U'\left(\frac{\partial F_{2}}{\partial x}(s,\bar{X}(s))\right)\,ds=$$$$-\mathbb{E}\int_{t}^{T}(Y^{\epsilon}(s)+Z^{\epsilon}(s))dM(t,s)$$$$\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!-
\mathbb{E}\int_{t}^{T}\bigg[(Y^{\epsilon}(s)+Z^{\epsilon}(s))M(t,s)\bigg(\mu\frac{\partial F_{1}}{\partial x}(s,\bar{X}(s))-\frac{\partial F_{2}}{\partial x}(s,\bar{X}(s)\bigg) +$$$$\sigma (Y^{\epsilon}(s)+Z^{\epsilon}(s))\frac{\partial F_{2}}{\partial x}(s,\bar{X}(s))N(s)\bigg]\,ds.$$
Moreover by and the equation following it
$$\mathbb{E}\int_{t}^{T}(Y^{\epsilon}(s)+Z^{\epsilon}(s))M(t,s)\bigg(\mu\frac{\partial F_{1}}{\partial x}(s,\bar{X}(s))-\frac{\partial F_{2}}{\partial x}(s,\bar{X}(s)\bigg)\,ds=$$$$\mathbb{E}\int_{t}^{T}
M(t,s)d(Y^{\epsilon}(s)+Z^{\epsilon}(s))$$$$-\mathbb{E}\int_{t}^{T}[M(t,s)(\mu\bar{X}(s)(\zeta(s) -F_{1}(s,\bar{X}(s)))-(\bar{X}(s)c(s)-F_{2}(s,\bar{X}(s)))]{\chi_{E_{\epsilon}}(s)})\,ds$$
Thanks to Itô’s formula $$\begin{aligned}
\notag
d(M(t,s)(Y^{\epsilon}(s)+Z^{\epsilon}(s))&=&M(t,s)d(Y^{\epsilon}(s)+Z^{\epsilon}(s))+(Y^{\epsilon}(s)+Z^{\epsilon}(s))dM(t,s)
\\\notag&+&d(Y^{\epsilon}(s)+Z^{\epsilon}(s))dM(t,s),\end{aligned}$$
which in conjunction with the previous equations yield
$$\!\!\!\!\!\!\!\!\!\!\mathbb{E}\int_{t}^{T}h(s-t)\frac{\partial F_{2}}{\partial x}(s,\bar{X}(s))U'(F_{2}(s,\bar{X}(s))) (Y^{\epsilon}(s)+Z^{\epsilon}(s))\,ds$$$$=\mathbb{E}\int_{t}^{T} [M(t,s)\mu(\bar{X}(s)\zeta(s)-F_{1}(\bar{X}(s))]\,ds$$$$+\mathbb{E}\int_{t}^{T}[(\bar{X}(s)c(s)-F_{2}(\bar{X}(s)))-(
\sigma N(t,s)(\bar{X}(s)\zeta(s)
+F_{1}(\bar{X}(s)))]{\chi_{E_{\epsilon}}(s)}\,ds$$$$-\mathbb{E}\int_{0}^{\infty}\sigma M(t,s)Y^{\epsilon}(s)((\zeta(s)
+F_{1}(\bar{X}(s))){\chi_{E_{\epsilon}}(s)}\,ds.$$
Next we introduce the Hamiltonian function $H$ by $$\label{Hamiltonian}
H(t,s,x,u,m,n)=mx(\zeta\mu-c)+nx\zeta\sigma+h(s-t)U(cx),\quad u=(\zeta,c).$$
The asymptotical expansion leads to
$$J(t,x,F_{1},F_{2})-J(t,x,\zeta_{\epsilon},c_{\epsilon})=\mathbb{E}\int_{t}^{T}\sigma M(t,s)Y^{\epsilon}(s)\bigg(\frac{\partial F_{2}}{\partial x}(s,\bar{X}^{t,x}(s))
-{\zeta}(s)\bigg){\chi_{E_{\epsilon}}(s)}\,ds$$$$\!\!\!\!\!\!\!\!\!\!+
\mathbb{E}\int_{t}^{T}[H(t,s,\bar{X}^{t,x}(s),\bar{\zeta}(s),\bar{c}(s),M(t,s),N(t,s))$$$$-H(t,s,\bar{X}^{t,x}(s),\zeta(s),c(s),M(t,s),N(t,s))]{\chi_{E_{\epsilon}}(s)}\,ds+o(\epsilon).$$
Additionally, with sufficient integrability assumptions and $$\label{2w}
\lim_{\epsilon\downarrow0}\frac{\mathbb{E}\int_{t}^{T}\sigma M(t,s)Y^{\epsilon}(s)\bigg(\frac{\partial F_{2}}{\partial x}(s,\bar{X}^{t,x}(s))
-{\zeta}(s)\bigg){\chi_{E_{\epsilon}}(s)}\,ds}{\epsilon}=0,$$
whence
$$\lim_{\epsilon\downarrow0}\frac{J(t,x,F_{1},F_{2})-J(t,x,\zeta_{\epsilon},c_{\epsilon})}{\epsilon}$$ $$=\mathbb{E}\bigg[H\bigg(t,t,x,\frac{F_{1}(t,x)}{x},\frac{F_{2}(t,x)}{x},M(t,t),N(t,t)\bigg)-H(t,t,x,\zeta(t),c(t),M(t,t),N(t,t))\bigg].$$
Therefore a sufficient condition for $F=(F_{1},F_{2})$ to be a equilibrium policy is
$$\label{op2}
\left(\frac{F_{1}(t,x)}{x},\frac{F_{2}(t,x)}{x}\right)={{\operatorname{argmax}}_{\zeta,c}}H(t,t,x,\zeta,c,M(t,t),N(t,t)),$$
Finally, the linearity of $H$ in $\zeta$ implies
$$\label{10au1}
\mu M(t,t)+\sigma N(t,t)=0,$$
and the first order conditions for $c$ (which are also sufficient due to concavity of $U$)
$$\label{109p}
F_{2}(t,x)=I(M(t,t)|X(t)=x).$$
$\square$
[**B Proof of Theorem \[PDE\]**]{}
The result is established for all three cases (exponential, type I, type II) separately.
[**Exponential discounting: $h(t)=e^{-\delta t}$**]{}
Let $v(t,x)$ be the solution of the classical HJB $$\label{0HJB3}
\frac{\partial v}{\partial t}(t,x)+rx\frac{\partial v}{\partial x}(t,x)-\frac{\mu^{2}}{2\sigma^{2}}\frac{{\frac{\partial^{2} v}{\partial x}(t,x)}}{\frac{\partial^{2} v}{\partial x^{2}}(t,x)}+\tilde{U}\left(\frac{\partial v}{\partial x}(t,x)\right)=\delta v(t,x),$$ for all $(t,x)\in[0,T]\times(0,\infty),$ with boundary condition $$v(T,x)=\hat{U}(x).$$
Then $v(t,x)$ and $w(t,x)=0$ solve the parabolic system.
We show the equilibrium strategies are
$$\label{0con}
F_{1}(t,x)=-\frac{\mu\frac{\partial v}{\partial x}(t,x)}{\sigma^{2}\frac{\partial^{2} v}{\partial x^{2}}(t,x)},\,\,
F_{2}(t,x)=I\left(\frac{\partial v}{\partial x}(t,x)\right),\,\,\,t\in[0,T].$$
Indeed lets consider the processes
$$\label{9op0}
M(t,s)=e^{-\delta(s-t)}\frac{\partial v}{\partial x}(s,\bar{X}(s)),\,\, N(t,s)=\sigma e^{-\delta(s-t)} F_{1}(s,\bar{X}(s))\frac{\partial^{2} v}{\partial x^{2}}(s,\bar{X}(s)),$$
$s\in[t,T].$ Recall that the equilibrium wealth process $\{\bar{X}(s)\}_{s\in[0,T]},$ is defined by $$\label{10dyn}
d\bar{X}(s)=[r\bar{X}(s)+\mu F_{1}(s,\bar{X}(s))-F_{2}(s,\bar{X}(s))]ds+\sigma F_{1}(s,\bar{X}(s))dW(s).$$ It is a matter of direct calculations to prove that $\{M(t,s), N(t,s)\}_{s\in[t,T]}$ solves BSDE . Next we observe $$\label{10au1}
\mu M(t,t)+\sigma N(t,t)=0,$$ and $$\label{109p}
F_{2}(t,x)=I(M(t,t)|X(t)=x),$$ so by Theorem \[Hamil1\], $F=(F_{1},F_{2})$ is an equilibrium strategy.
$\square$
[**[ Type I discounting: **[$h(t)=\lambda\exp(-\rho_{1}t)+(1-\lambda)\exp(-\rho_{2}t) $]{}**]{}**]{}
Let $v$ and $w$ be a solution of the PDE system. We show the equilibrium strategies are
$$\label{10con}
F_{1}(t,x)=-\frac{\mu\frac{\partial v}{\partial x}(t,x)}{\sigma^{2}\frac{\partial^{2} v}{\partial x^{2}}(t,x)},\,\,
F_{2}(t,x)=I\left(\frac{\partial v}{\partial x}(t,x)\right),\,\,\,t\in[0,T].$$
Let us define the process $\{M(t,s), N(t,s)\}_{s\in[t,T]}$ by
$$\notag
M(t,s)\triangleq \lambda\exp(-\rho_{1}(s-t))\frac{\partial v_{1}}{\partial x}(s,\bar{X}(s)) +(1-\lambda)\exp(-\rho_{2}(s-t))\frac{\partial v_{2}}{\partial x}(s,\bar{X}(s)),$$
and $$\!\!\!\!\!\!\!\!\!\!\!N(t,s)\triangleq \lambda\sigma\exp(-\rho_{1}(s-t))F_{1}(s,\bar{X}(s))\frac{\partial^{2} v_{1}}{\partial x^{2}}(s,\bar{X}(s))$$$$+(1-\lambda)\sigma\exp(-\rho_{2}(s-t))F_{1}(s,\bar{X}(s))\frac{\partial^{2} v_{2}}{\partial x^{2}}(s,\bar{X}(s)),$$
for some functions $v_{1}$ and $v_{2}$ which will be specified later on and $\{\bar{X}(s)\}_{s\in[0,T]}$ is the equilibrium wealth process of . By requesting $\{M(t,s), N(t,s)\}_{s\in[t,T]}$ to solve BSDE , one finds a PDE system for $(v_{1},v_{2}).$
Indeed, on one hand by Itô’s formula
$$dM(t,s)=\bigg(\lambda\exp(-\rho_{1}(s-t))\bigg[-\rho_{1}\frac{\partial v_{1}}{\partial x}(s,\bar{X}(s))
+\frac{\partial^{2} v_{1}}{\partial x\partial s}(s,\bar{X}(s))$$$$+(\mu F_{1}(s,\bar{X}(s))-F_{2}(s,\bar{X}(s)))\frac{\partial^{2} v_{1}}{\partial x^{2}}(s,\bar{X}(s))+
\frac{\sigma^{2}}{2}F^{2}_{1}(s,\bar{X}(s))\frac{\partial^{3} v_{1}}{\partial x^{3}}(s,\bar{X}(s))\bigg]$$$$+(1-\lambda)\exp(-\rho_{2}(s-t))\bigg[-\rho_{2}\frac{\partial v_{2}}{\partial x}(s,\bar{X}(s))+\frac{\partial^{2} v_{2}}{\partial x\partial s}(s,\bar{X}(s))
+$$$$+(\mu F_{1}(s,\bar{X}(s))-F_{2}(s,\bar{X}(s)))\frac{\partial^{2} v_{2}}{\partial x^{2}}(s,\bar{X}(s))
+\frac{\sigma^{2}}{2}F^{2}_{1}(s,\bar{X}(s))\frac{\partial^{3} v_{2}}{\partial x^{3}}(s,\bar{X}(s)) \bigg]\bigg)ds$$$$+N(t,s)dW(s).$$
On the other hand from
$$dM(t,s)=\bigg(\lambda\exp(-\rho_{1}(s-t))\bigg[\frac{\partial v_{1}}{\partial x}(s,\bar{X}(s))\left(\mu\frac{\partial{F_{1}}}{\partial{x}}(s,\bar{X}(s))-\frac{\partial{F_{2}}}{\partial{x}}(s,\bar{X}(s)\right)+$$$$+\sigma^{2}F_{1}(s,\bar{X}(s))\frac{\partial{F_{1}}}{\partial{x}}(s,\bar{X}(s))\frac{\partial^{2} v_{1}}{\partial x^{2}}(s,\bar{X}(s))+\frac{\partial{F_{2}}}{\partial{x}}(s,\bar{X}(s))U'(F_{2}(s,\bar{X}(s))) \bigg]+$$ $$+(1-\lambda)\exp(-\rho_{2}(s-t))\bigg[\frac{\partial v_{2}}{\partial x}(s,\bar{X}(s))\left(\mu\frac{\partial{F_{1}}}{\partial{x}}(s,\bar{X}(s))-\frac{\partial{F_{2}}}{\partial{x}}(s,\bar{X}(s)\right)+$$$$+\sigma^{2}F_{1}(s,\bar{X}(s))\frac{\partial{F_{1}}}{\partial{x}}(s,\bar{X}(s))\frac{\partial^{2} v_{2}}{\partial x^{2}}(s,\bar{X}(s))+\frac{\partial{F_{2}}}{\partial{x}}(s,\bar{X}(s))U'(F_{2}(s,\bar{X}(s)))\bigg]\bigg)ds+$$ $$+N(t,s)dW(s).$$
The two representations of $dM(t,s)$ yield
$$-\rho_{1}\frac{\partial v_{1}}{\partial x}(s,x)+\frac{\partial^{2} v_{1}}{\partial x\partial s}(s,x) +(\mu F_{1}(s,x)-F_{2}(s,x))\frac{\partial^{2} v_{1}}{\partial x^{2}}(s,x)$$$$+\frac{\sigma^{2}}{2}
F^{2}_{1}(s,x)\frac{\partial^{3} v_{1}}{\partial x^{3}}(s,x) = -\left(\mu \frac{\partial F_{1}}{\partial x}(s,x)-\frac{\partial F_{2}}{\partial x}(s,x)\right)\frac{\partial v}{\partial x}(s,x)$$$$-\sigma^{2}F_{1}(s,x)\frac{\partial F_{1}}{\partial x}(s,x)\frac{\partial^{2} v_{1}}{\partial x^{2}}(s,x) - \frac{\partial F_{2}}{\partial x}(s,x)\frac{\partial v}{\partial x}(s,x),$$
and
$$-\rho_{2}\frac{\partial v_{2}}{\partial x}(s,x)+\frac{\partial^{2} v_{2}}{\partial x\partial s}(s,x) +(\mu F_{1}(s,x)-F_{2}(s,x))\frac{\partial^{2} v_{2}}{\partial x^{2}}(s,x)$$$$+\frac{\sigma^{2}}{2}
F^{2}_{1}(s,x)\frac{\partial^{3} v_{2}}{\partial x^{3}}(s,x)= -\left(\mu \frac{\partial F_{1}}{\partial x}(s,x)-\frac{\partial F_{2}}{\partial x}(s,x)\right)\frac{\partial v_{2}}{\partial x}(s,x)$$$$-\sigma^{2}F_{1}(s,x)\frac{\partial F_{1}}{\partial x}(s,x)\frac{\partial^{2} v_{2}}{\partial x^{2}}(s,x) - \frac{\partial F_{2}}{\partial x}(s,x)\frac{\partial v}{\partial x}(s,x).$$
This can be rewritten as $$\frac{\partial}{\partial x}\bigg[\frac{\partial v_{1}}{\partial s}(s,x)-\rho_{1}v_{1}(s,x)
+(\mu F_{1}(s,x)-F_{2}(s,x))\frac{\partial v_{1}}{\partial x}(s,x)$$$$+
\frac{\sigma^{2}}{2}
F^{2}_{1}(s,x)\frac{\partial^{2} v_{1}}{\partial x^{2}}(s,x)+U(F_{2}(s,x)) \bigg]=0,$$ and $$\frac{\partial}{\partial x}\bigg[\frac{\partial v_{2}}{\partial s}(s,x)-\rho_{2}v_{2}(s,x)
+(\mu F_{1}(s,x)-F_{2}(s,x))\frac{\partial v_{2}}{\partial x}(s,x)$$$$+
\frac{\sigma^{2}}{2}
F^{2}_{1}(s,x)\frac{\partial^{2} v_{2}}{\partial x^{2}}(s,x)+U(F_{2}(s,x)) \bigg]=0.$$
Recall that $v$ and $w$ is a solution of the PDE system. Thus $$\label{9ei0}
v_{1}\triangleq v+(1-\lambda)w,\qquad v_{2}\triangleq v-\lambda w$$ satisfy the above PDEs. Moreover $$\label{110au1}
\mu M(t,t)+\sigma N(t,t)=0,$$ and $$\label{1109p}
F_{2}(t,x)=I(M(t,t)|X(t)=x),$$ so by Theorem \[Hamil1\], $F=(F_{1},F_{2})$ is an equilibrium strategy.
$\square$
[**[ Type II discounting: **[$h(t)=(1+\lambda t)\exp(-\rho t) $]{}**]{}**]{}
Let $v$ and $w$ be a solution of the PDE system. We show the equilibrium strategies are
$$\label{10con}
F_{1}(t,x)=-\frac{\mu\frac{\partial v}{\partial x}(t,x)}{\sigma^{2}\frac{\partial^{2} v}{\partial x^{2}}(t,x)},\,\,
F_{2}(t,x)=I\left(\frac{\partial v}{\partial x}(t,x)\right),\,\,\,t\in[0,T].$$
Let us define the process $\{M(t,s), N(t,s)\}_{s\in[t,T]}$ by
$$\notag
M(t,s)\triangleq \exp(-\rho(s-t))\frac{\partial v_{1}}{\partial x}(s,\bar{X}(s)) +\lambda t\exp(-\rho(s-t))\frac{\partial v_{2}}{\partial x}(s,\bar{X}(s)),$$
and $$\!\!\!\!\!\!\!\!\!\!\!N(t,s)\triangleq \sigma\exp(-\rho_{1}(s-t))F_{1}(s,\bar{X}(s))\frac{\partial^{2} v_{1}}{\partial x^{2}}(s,\bar{X}(s))$$$$+\sigma\lambda t \exp(-\rho_{2}(s-t))F_{1}(s,\bar{X}(s))\frac{\partial^{2} v_{2}}{\partial x^{2}}(s,\bar{X}(s)),$$
for the functions $$\label{71ei0}
v_{1}\triangleq v,\qquad v_{2}\triangleq v- w.$$
As for the case of type I one can check that $\{M(t,s), N(t,s)\}_{s\in[t,T]}$ solves BSDE . Moreover $$\label{1110au1}
\mu M(t,t)+\sigma N(t,t)=0,$$ and $$\label{11109p}
F_{2}(t,x)=I(M(t,t)|X(t)=x),$$ so by Theorem \[Hamil1\], $F=(F_{1},F_{2})$ is an equilibrium strategy.
$\square$
[**C Proof of Proposition \[IE\]**]{}
If the discounting is exponential the result is a direct consequence of Feynman-Kac’s formula. For type I, the functions $v_{1}$ and $v_{2}$ of admit the following stochastic representations
$$\label{000ie1}
v_{i}(t,x)=\mathbb{E}\left[\int_{t}^{T}e^{-\rho_{i}(s-t)}U(F_{2}(\bar{X}^{t,x}_{s}))\,ds\right],\,\,\,i=1,2,$$
by Feynman-Kac’s formula. Consequently $$\label{0000ie1}
v(t,x)=\lambda v_{1}(t,x)+(1-\lambda )v_{2}(t,x)
=\mathbb{E}\left[\int_{t}^{T}h(s-t)U(F_{2}(\bar{X}^{t,x}_{s}))\,ds\right].$$ Similarly for type II, holds true.
$\square$
[**[Acknowledgements]{}**]{}
The authors would like to thank Ali Lazrak for helpful discussions and comments.
[99]{} (1992) [*Picoeconomics*]{}, [*Cambridge University Press*]{}.
\(1999) [*Ramsey Meets Laibson in the Neoclassical Growth Model*]{}, [*The Quarterly Journal of Economics*]{}, [**114**]{}, 1125-1152.
\(1998) Existence and uniqueness of BSDE with stopping time, [*Chinese Science Bulletin*]{} [**43**]{}, 69-99.
\(1989) Optimal consumption and portfolio policies when asset prices follow a diffusion process, [*J. Economic Theory*]{} [**49**]{}, 33-83.
\(2006) [*Being serious about non-commitment: subgame perfect equilibrium in continuous time*]{}, [*Preprint*]{}.
\(1980) [*Consistent Plans*]{}, [*Rev. Financial Stud.*]{}, [**47**]{}, 533-537.
\(1987) Optimal portfolio and consumption decisions for a “small investor" on a finite horizon. , 1557–1586.
\(2003) [*Consumption and Savings Decisions with Quasi-geometric Discounting*]{}, [*Econometrica*]{} [**71**]{}, 365-375.
\(1992) [*Anomalies in Intertemporal Choice; Evidence and an Inter pretation*]{}, [*The Quarterly Journal of Economics*]{}, [**107**]{}, 573-597.
\(1997) [*Golden Eggs and Hyperbolic Discounting*]{}, [*The Quarterly Journal of Economics*]{}, [**112**]{}, 443-477.
\(1969) [*Lifetime portfolio selection under uncertainty: the continuous-time case*]{}, [*Rev. Econom. Statist.*]{} [**51**]{}, 247-257.
( 1971) Optimum consumption and portfolio rules in a conitinuous-time model. , [**3**]{}, 373-413.
\(1973) [*On the existence of a Consistent Course of Action when Tastes are Changing*]{}, [*Rev. Financial Stud.*]{}, [**40**]{}, 391-401.
\(1990) [*A general stochastic maximum principle for optimal control problems*]{}, [*SIAM Journal of Control and Optimisation*]{} [**28**]{}, 966-979.
\(1968) [*Consistent Planning*]{}, [*Rev. Financial Stud.*]{}, [**35**]{}, 185-199.
\(1955) [*Myopia and inconsistency in Dynamic Utility Maximization*]{}, [*Rev. Financial Stud.*]{}, [**23**]{}, 165-180.
[^1]: Work supported by PIMS under NSERC grant 298427-04.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In recent Letters Marzlin and Sanders [@marzlin] and Tong [*et al.*]{} [@tong] study an adiabatically varying Hamiltonian $h(t)$ that generates the time evolution $U(t)$ and its dual $H(t)$ that generates the evolution $U^{\dagger}(t)$. Marzlin and Sanders show that inconsistent results are obtained if an adiabatic approximation is used to calculate $H(t)$; Tong [*et al.*]{} show that the adiabatic approximation can be very inaccurate when applied to the exact dual Hamiltonian $H(t)$ even if it is an excellent approximation for $h(t)$. We show that these two observations are equivalent and are not inconsistent with the adiabatic theorem because in general, even if $h(t)$ satisfies the conditions of the adiabatic theorem, $H(t)$ will likely violate those conditions.'
author:
- Solomon Duki$^1$
- 'H. Mathur$^1$'
- Onuttom Narayan$^2$
title: 'Is the Adiabatic Approximation Inconsistent?'
---
The adiabatic theorem is the basis of an approximation scheme that was discovered at the dawn of quantum mechanics [@born] and that has been in widespread and continuous use ever since. Applications range from two-level systems (such as nuclei undergoing magnetic resonance or atoms interacting resonantly with a laser field) to quantum field theory (where a low-energy effective theory is derived by integrating out fast, high-energy degrees of freedom). Two decades ago, Berry uncovered the beautiful geometric structure underlying the adiabatic approximation [@berry], leading to a resurgence of interest in the subject and to new applications [@shapere; @anandan]. More recently, it has been proposed that Berry phase effects lead to quantum phase transitions that lie outside the usual Landau-Ginzburg-Wilson paradigm [@senthil]. The adiabatic theorem is also the basis of a newly proposed quantum computing scheme [@farhi]. Considering the significance of the adiabatic approximation to quantum physics, the discovery of an inconsistency would be most disturbing. In a recent Letter Marzlin and Sanders [@marzlin] ask whether such an inconsistency might exist, at least for a class of Hamiltonians. That question has been further studied by Tong [*et al.*]{} [@tong] and subsequently commented on in Refs [@lidar; @pati]. The purpose of this Letter is to show that there is no inconsistency.
Refs [@marzlin] and [@tong] start with a time-dependent Hamiltonian $h(t)$ for which it is presumed that the adiabatic approximation is accurate. The evolution operator for $h(t)$ is denoted by $U(t),$ the solution to $$i \frac{\partial}{\partial t} U(t) = h(t) U(t),
\label{eq:uschrodinger}$$ with $U(0) = {\cal I},$ where ${\cal I}$ is the identity operator. Next, they consider the dual Hamiltonian $H(t)$ for which the evolution operator is $U^{\dagger}(t)$. Evidently $$H(t) = - U^{\dagger}(t) i \frac{\partial}{\partial t} U(t) =
- U^{\dagger}(t) h(t) U(t).
\label{eq:dual}$$ Tong [*et al.*]{} now formulate the putative inconsistency as follows: First they argue that $H(t)$ satisfies the conditions for the adiabatic theorem as well as does $h(t)$. Then they demonstrate that the adiabatic approximation can be very inaccurate for the dual Hamiltonian. These conflicting observations constitute the “inconsistency”. Marzlin and Sanders [@marzlin] originally formulated the inconsistency in a different form; below we will show the equivalence of the two formulations.
In this Letter we examine the dual Hamiltonian $H(t)$ and find it violates well known adiabaticity conditions; hence there is no inconsistency. The chief difficulty in determining whether $H(t)$ satisfies adiabatic conditions is that we do not have an explicit expression for $H(t)$ except in the special cases where the dynamics of $h(t)$ are simple enough to allow evaluation of $U(t)$. Nonetheless, we are able to give a general argument that $H(t)$ violates the conditions of the adiabatic theorem. As an illustration, we apply our general arguments to a solvable two-level model also studied by ref [@tong]. In this case it is possible to obtain the explicit form of $H(t)$; inspection of this form is sufficient to show immediately that $H(t)$ violates the conditions of the adiabatic theorem, consistent with our result.
It is helpful to first review the conditions under which the adiabatic approximation is accurate. Consider a two-level system with the Schrödinger equation $$i \frac{\partial}{\partial t}
\left(
\begin{array}{c}
c_1 \\
c_2 \\
\end{array}
\right)
= \frac{1}{2}
\left(
\begin{array}{cc}
\omega_0 & \Omega e^{-i \omega t} \\
\Omega e^{i \omega t} & - \omega_0 \\
\end{array}
\right)
\left(
\begin{array}{c}
c_1 \\
c_2 \\
\end{array}
\right).
\label{eq:twolevel}$$ We regard the off-diagonal terms as a perturbation and ask when the perturbation may be neglected. Evidently, the perturbation must be small in magnitude, but even if it is, it can have a big effect on resonance when $\omega \approx \omega_0$. This is more transparent if we go over to the interaction picture by writing $c_1 = a_1 \exp( - i \omega_0 t/2 )$, $c_2 = a_2
\exp( i \omega_0 t/2 )$. In the interaction picture $$i \frac{\partial}{\partial t}
\left(
\begin{array}{c}
a_1 \\
a_2 \\
\end{array}
\right)
=
\frac{1}{2}
\left(
\begin{array}{cc}
0 & \Omega e^{-i(\omega - \omega_0)t} \\
\Omega e^{i (\omega - \omega_0) t} & 0 \\
\end{array}
\right)
\left(
\begin{array}{c}
a_1 \\
a_2 \\
\end{array}
\right).
\label{eq:2levelinteraction}$$ Away from resonance the off-diagonal terms oscillate rapidly and average to zero. Near resonance, the perturbation varies slowly in the interaction picture and can have a big effect, producing Rabi oscillations. The precise condition to be off-resonance is $|\omega - \omega_0| \gg \Omega$. For this problem, the condition can be derived by transforming to another rotating frame via $a_1 = b_1 \exp[ - i (\omega - \omega_0)t/2 ]$, $a_2 = b_2 \exp[i (\omega - \omega_0)t/2 ]$. In this frame, the Hamiltonian is time independent and equal to $(1/2)[ (\omega - \omega_0) \sigma_z + \Omega \sigma_x ]$. In summary, we need the perturbation to be off-resonance ($\Omega \ll | \omega -
\omega_0 |$) for it to be truly negligible.
Now let us consider the problem of a general time dependent hamiltonian $h(t).$ It is helpful to use a slightly different approach from the one above, which is useful for adiabatic perturbations. To this end, we introduce the instantaneous eigenstates $|n(t) \rangle$ that satisfy $$h(t) |n(t) \rangle = \epsilon_n(t) |n(t) \rangle.
\label{eq:instestates}$$ We choose the phases of $|n(t) \rangle$ to satisfy $ \langle n(t) | \partial/\partial t | n(t) \rangle = 0 $, a convention called the parallel transport gauge [@anholonomy]. We expand the state of the system $ | \psi \rangle $ in this time-dependent basis. Thus $$| \psi (t) \rangle =
\sum_{n} \phi_n (t) \exp[ - i \int_0^t d t' \varepsilon_n(t') ]
| n(t) \rangle.
\label{eq:psin}$$ In this moving frame the time-dependent Schrödinger equation has the form $$i \frac{ \partial }{ \partial t } \phi_n (t) =
\sum_{n \neq m} A_{nm} (t)
\exp( i \int_{0}^{t} d t' [ \varepsilon_n(t') -
\varepsilon_m(t') ] ) \phi_n (t).
\label{eq:adiabaticinteraction}$$ Here $$A_{nm} (t) = -i \langle n | \frac{ \partial}{ \partial t} | m \rangle.
\label{eq:anm}$$ The adiabatic approximation amounts to neglect of the terms on the right hand side of eq (\[eq:adiabaticinteraction\]). For this to be justified, by analogy to eq (\[eq:2levelinteraction\]), we see that the neglected terms must be off-resonance. Roughly the condition to be off-resonance is that the neglected terms should [*not*]{} vary slowly; i.e. the terms should be small in magnitude compared to their predominant frequency. Note that this is somewhat different from the adiabatic condition $ |A_{nm}| \ll
| \varepsilon_n - \varepsilon_m | $ used by refs [@marzlin; @tong]. More precisely, if the Hamiltonian varies predominantly at a frequency $\omega,$ and the typical spacing between instantaneous eigenvalues is $\Delta,$ our adiabatic condition states $| A_{nm} | \ll (\Delta - \omega)$. Alternatively, if the Hamiltonian varies on a time scale $T$ for which $1/\Delta T\ll 1,$ and $A_{nm}\sim 1/T,$ the correction to the adiabatic approximation is $\sim 1/\Delta T$ which is very small. This condition is more restrictive than neccessary, but it agrees with the intuitive expectation that for the adiabatic approximation to apply, the Hamiltonian must “ vary slowly”.
Sampling the literature we find that the graduate text by Schiff [@schiff] gives a rather complete discussion of the adiabatic approximation, emphasizing that the neglected terms must be non-resonant. On the other hand, Landau and Lifshitz [@landau] give the more restrictive adiabatic condition $T \rightarrow \infty$. Moody, Shapere and Wilczek [@moody] compute non-perturbative corrections to adiabatic evolution using $1/\Delta T$ as the small parameter. Berry [@iteration] also regards $1/\Delta T$ as the adiabatic parameter and suggests that the corrections vanish as $\exp( - \Delta T )$, also proposed by Hwang and Pechukas [@hwang]. Thus it appears to be generally accepted that it is sufficient for the Hamiltonian to be slowly varying, but less restrictive conditions are also discussed.
For the dual Hamiltonian $H(t),$ denote the instantaneous eigenstates $ |n(t); H \rangle$ and the eigenvalues $\varepsilon_n^H (t)$. With the parallel transport gauge, in the adiabatic frame $$| \psi (t) \rangle = \sum_{n} \phi_n^H (t)
\exp[ - i \int_0^t d t' \varepsilon_n^H (t') ] | n(t); H \rangle.
\label{eq:Hansatz}$$ By analogy to eq (\[eq:adiabaticinteraction\]) the Schrödinger equation obeyed by the amplitudes $\phi_n^H(t)$ is $$i \frac{ \partial }{\partial t} \phi_n^H (t) =
\sum_{m \neq n} A_{nm}^H (t)
\exp( i \int_0^t d t' [\varepsilon_n^H (t') - \varepsilon_m^H (t')] )
\phi_m^H(t).
\label{eq:Hinteraction}$$
Although, as noted above, we do not have an explicit expression for $H(t)$, it is easy to relate the eigenstates of $H$ to those of $h$. Evidently $U^{\dagger}(t) | n (t) \rangle$ is an eigenstate of $H(t)$ with eigenvalue $ - \varepsilon_n (t)$. Hence we write $$\begin{aligned}
| n(t); H \rangle & = & U^{\dagger} (t) | n(t) \rangle
\exp[ - i \int_0^t d t' \varepsilon_n(t')] ;
\nonumber \\
\varepsilon_n^H (t) & = & - \varepsilon_n (t).
\label{eq:relate}\end{aligned}$$ The phase of $|n(t); H \rangle$ has been chosen to ensure parallel transport, $ \langle n; H | \partial_t | n; H \rangle
= 0$. To verify this, it is helpful to recall that $i U \partial_t U^{\dagger} = - h,$ which follows from the Schrödinger equation (\[eq:uschrodinger\]). Using eq (\[eq:relate\]) and this result, one can show that $$A_{nm}^H (t) = A_{nm}(t) \exp( i \int_0^t d t'
[ \varepsilon_m (t') - \varepsilon_n (t') ]).
\label{eq:arelation}$$ Substituting eq (\[eq:arelation\]) into eq (\[eq:Hinteraction\]) we obtain the final form of the Schrödinger equation in the adiabatic frame for the dual Hamiltonian $H(t)$, $$i \frac{\partial}{\partial t} \phi_n^H (t) =
\sum_{m \neq n} A_{nm} (t) \phi_m^H (t).
\label{eq:clearview}$$ Eq (\[eq:clearview\]) is the main result of our general analysis.
Eq (\[eq:clearview\]) shows that the terms that would be neglected in the adiabatic approximation $A_{nm}(t)$ vary slowly. The typical frequency of these terms is $1/T$, the same as their magnitude. Thus $H(t)$ does not fulfil the adiabatic condition, and using the adiabatic approximation for $H(t)$ leads to inaccurate results.
This concludes our general analysis of $H(t)$. We turn to a solvable example. Take the two-level Hamiltonian $$h(t) = - \frac{1}{2} \omega_0
\left(
\begin{array}{cc}
\cos \theta & \sin \theta e^{-i \omega t} \\
\sin \theta e^{i \omega t} & - \cos \theta \\
\end{array}
\right)
\label{eq:solvableh}$$ Physically we can picture this as a spin $\frac{1}{2}$ particle in a magnetic field tilted at an angle $\theta$ to the $z$-axis and rotating at a frequency $\omega$. Essentially this model was also studied in ref [@tong]. The instantaneous eigenvalues of this Hamiltonian are $ \varepsilon_{\pm} (t) =
\pm \omega_0/2$; the corresponding eigenspinors in the parallel transport gauge are [@notetwo] $$\begin{aligned}
|+(t) \rangle & = &
\left(
\begin{array}{c}
\cos \frac{ \theta }{2} \\
\sin \frac{\theta}{2} e^{i \omega t} \\
\end{array}
\right)
\exp[ - i \frac{ \omega t}{2} (1 - \cos \theta)],
\nonumber \\
|-(t) \rangle & = &
\left(
\begin{array}{c}
- \sin \frac{\theta}{2} e^{- i \omega t} \\
\cos \frac{\theta}{2} \\
\end{array}
\right)
\exp[ i \frac{\omega t}{2} (1 - \cos \theta)].
\label{eq:hspinors}\end{aligned}$$ A straightforward computation reveals that $$A_{+ -} (t) =
\langle + | \partial_t | - \rangle =
\frac{ \omega }{2} \sin \theta
e^{- i \omega t \cos \theta }.
\label{eq:aplusminus}$$
It follows from eq (\[eq:adiabaticinteraction\]) that the adiabatic frame Schrödinger equation for $h(t)$ is $$i \frac{ \partial }{ \partial t }
\left(
\begin{array}{c}
\phi_+ \\
\phi_- \\
\end{array}
\right) = \left(
\begin{array}{cc}
0 & A_{+-} e^{ i \omega_0 t }\\
A_{+-}^* e^{- i \omega_0 t} & 0 \\
\end{array}
\right)
\left(
\begin{array}{c}
\phi_+ \\
\phi_- \\
\end{array}
\right).
\label{eq:hinteraction}$$ Clearly for the off-diagonal term to be off-resonance, we need $| \omega \cos \theta - \omega_0 | \gg \omega \sin \theta$. Evidently, this condition is satisfied if $ \omega \ll \omega_0 $, [ *i.e.*]{}, $h(t)$ varies slowly.
It follows from eq (\[eq:clearview\]) that the adiabatic frame Schrödinger equation for $H(t)$ is $$i \frac{\partial}{\partial t}
\left(
\begin{array}{c}
\phi_+ \\
\phi_- \\
\end{array}
\right) =
\left(
\begin{array}{cc}
0 & A_{+-}(t) \\
A_{+-}^*(t) & 0 \\
\end{array}
\right)
\left(
\begin{array}{c}
\phi_+ \\
\phi_- \\
\end{array}
\right).
\label{eq:Hinteractiontwo}$$ Clearly, for the off-diagonal term to be off-resonance we need $ | \omega \sin \theta | \ll | \omega \cos \theta|$, a condition met only when $\theta \ll 1$ (or $ \pi - \theta \ll 1$).
In summary we find that for $h(t)$ to be adiabatic it is sufficient that $\omega \ll \omega_0$; but for $H(t)$ to be adiabatic we also need $ \theta \ll 1 $. In this limit, $\theta \rightarrow 0$, the fidelity of the adiabatic solution to $H(t)$ (the overlap of the adiabatic and exact solutions) computed by [@marzlin; @tong] approaches unity, consistent with our finding.
We note that in this solvable problem it is possible to compute $U$ and explicitly obtain $H(t)$. This column is far too narrow to write the entire expression, but it includes terms that oscillate at a frequency $$\nu = \sqrt{\omega_0^2 + \omega^2 + 2 \omega_0 \omega \cos \theta}.
\label{eq:nu}$$ In the limit $\omega \ll \omega_0$, needed for the Hamiltonian $h(t)$ to be adiabatically varying, $ \nu \rightarrow \omega_0$. Thus, even without going to to the adiabatic frame, a cursory inspection of $H(t)$ is sufficient to show it is not slowly varying and is unlikely to satisfy the adiabatic condition.
We briefly comment on cases where the adiabatic approximation applies to $h(t)$ and $H(t).$ This happens when there is a parameter other than $T$ which can be tuned to make the magnitude of the off-diagonal term in eq (\[eq:clearview\]) small compared to the predominant frequency $1/T$. However, for $h(t)$ the approximation becomes more accurate as $T \rightarrow
\infty$ (with exponentially small corrections according to refs [@moody; @iteration; @hwang]), whereas for for $H(t)$ the distance from resonance is essentially independent of $T$ and is controlled by the additional parameter. This can be seen in the example above, where $T \rightarrow 2 \pi/\omega$ and $\theta$ is the additional parameter.
We make a few observations in passing here about the discussion of the adiabatic conditions after Eq.(\[eq:anm\]). First, the effects of $A_{nm}$ are implicitly integrated over a finite time window. This is appropriate, since in experiments $h(t)$ is typically varied only within a finite time window. However, it is the discontinuity in $dh/dt$ that causes the $\sim 1/(\Delta T)$ correction to the adiabatic approximation for large $T;$ if all derivatives of $h(t)$ are continuous for all $t,$ it is known [@moody; @iteration; @hwang] that the correction will be exponentially small in $T\Delta.$ Second, $\Delta$ was effectively taken to be time independent. When both $A_{nm}$ and $\Delta$ vary sinusoidally at frequency $\omega,$ higher order resonances result when $\omega = \Delta_0/k$ for integer $k.$ This is essentially the same mechanism that causes resonances going beyond first order perturbation theory. Finally we discuss the equivalence between the formulations of refs [@marzlin] and [@tong]. They assume that the adiabatic approximation is accurate for $h(t):$ $U(t) \approx U_{{\rm adia}}(t)$ where $$U_{{\rm adia}} (t) = \sum_{n} |n(t) \rangle \langle n(0)|
\exp \left[ - i \int_{0}^{t} d t' \varepsilon_n(t') \right]
\label{eq:uadia}$$is the adiabatic approximation to the exact evolution $U(t)$. Marzlin [*et al.*]{} [@marzlin] then develop an approximation to $U^{\dagger}(t)$ that we denote $V^{\dagger}(t)$. They approximate the dual Hamiltonian \[defined by eq (\[eq:dual\])\] as $H^{(1)}_{{\rm adia}} (t) = - U^{\dagger}_{{\rm adia}} (t) h(t)
U_{{\rm adia}} (t).$ They compute the evolution $V^{\dagger}(t)$ corresponding to $H^{(1)}_{{\rm adia}}(t),$ obtaining $$V^{\dagger}(t) = \sum_{n} | n(0) \rangle \langle n(0) |
\exp \left[ i \int_{0}^{t} d t' \varepsilon_n(t') \right].
\label{eq:vdagger}$$ In contrast, Tong [*et al.*]{} compute a different approximation to $U^{\dagger} (t)$ that we denote $W^{\dagger}(t)$. They work with the exact dual Hamiltonian $H(t)$, but work out its evolution using the adiabatic approximation. Analogy to eq (\[eq:uadia\]) and use of eq (\[eq:relate\]) leads to the result $$W^{\dagger}(t) = \sum_{n} U^{\dagger} (t) | n(t) \rangle \langle n(0) |.
\label{eq:wdagger}$$ The operators $V^{\dagger}$ and $W^{\dagger}$ are different in appearance and in the approximations that lead to them, but we will show they are equivalent to the extent that the adiabatic approximation $U \approx U_{{\rm adia}}$ is valid.
With this notation established we now turn to the inconsistencies. Marzlin and Sanders [@marzlin] consider the identity $U U^{\dagger}
= {\cal I}$ and replace $U \rightarrow U_{{\rm adia}}$ and $U^{\dagger} \rightarrow V^{\dagger}$ with the disastrous result that $U U^{\dagger} \rightarrow \sum_{n} | n (t) \rangle \langle n(0) |
\neq {\cal I}.$ Tong [*et al.*]{} derive the same inconsistency by replacing $U^{\dagger}$ with $W^{\dagger}:$ $U U^{\dagger} \rightarrow U W^\dagger$ and using Eq.(\[eq:wdagger\]). The equivalence of the two approaches can be seen by starting with the trivial identity $U^{\dagger} = U^{\dagger} U U^{\dagger}$. If on the right hand side we replace $U^{\dagger} U U^{\dagger} \rightarrow
U^{\dagger} U_{{\rm adia}} V^{\dagger}$ we obtain $W^{\dagger}$ by use of eqs (\[eq:uadia\]), (\[eq:vdagger\]) and (\[eq:wdagger\]). In other words, the approximation of Marzlin and Sanders with the adiabatic approximation is equivalent to the approximation of Tong [*et al.*]{}
The resolution of the inconsistencies of Refs.[@marzlin; @tong] is that $V^{\dagger} \approx U^{\dagger}$ and $W^{\dagger} \approx U^{\dagger}$ are not good approximations; the adiabatic approximation $U_{{\rm adia}} \approx U$ is not at fault. In this Letter we have explained the failure of the approximation $W^{\dagger} \approx U^{\dagger}$. This may also be considered a resolution of the Marzlin and Sanders form of the inconsistency, due to the equivalence shown above. An alternative resolution was provided in ref [@marzlin] who used the adiabatic approximation to calculate a second approximation to the dual Hamiltonian $H^{(2)}_{{\rm adia}} (t) = - i U^{\dagger}_{{\rm adia}}
\partial_t U_{{\rm adia}}.$ By construction, the evolution operator corresponding to $H^{(2)}_{{\rm adia}}(t)$ is $U^{\dagger}_{{\rm adia}}(t)$ which is indeed a good approximation to $U^{\dagger}$. Since the trustworthy adiabatic approximation has been used to compute both $H^{(1)}_{{\rm adia}}(t)$ and $H^{(2)}_{{\rm adia}}(t)$ we may presume that $H^{(1)}_{{\rm adia}}(t) \approx
H^{(2)}_{{\rm adia}} (t) \approx H(t)$. This does not mean that the three Hamiltonians will generate essentially the same evolution since small errors can grow upon time exponentiation. Indeed ref [@marzlin] finds that $H^{(1)}_{{\rm adia}}$ and $H^{(2)}_{{\rm adia}}$ differ by a small resonant perturbation leading to the inference that $V^{\dagger}$ may be very different from $U^{\dagger}_{{\rm adia}}$. Since $U^{\dagger}_{{\rm adia}}
\approx U^{\dagger}$ this may be considered to explain the failure of the approximation $V^{\dagger} \approx U^{\dagger}.$
To summarise, we have studied the dual pair of Hamiltonians $h(t)$ and $H(t)$ that generate time evolution $U(t)$ and $U^{\dagger} (t)$ respectively. Marzlin and Sanders [@marzlin] showed that if an adiabatic approximation is used to compute $H(t)$ an inconsistency results. Tong [*et al.*]{} [@tong] showed that the same inconsistency results if the adiabatic approximation is applied to the exact dual Hamiltonian $H(t)$. We show that these observations are essentially equivalent. Our main finding is that even if $h(t)$ satisfies the conditions of the adiabatic theorem, $H(t)$ will not (except in the trivial case when the total change in $h(t)$ is small) because the terms neglected in the adiabatic approximation are resonant for $H(t).$ Thus the breakdown of the adiabatic approximation for $H$ is not inconsistent with the adiabatic theorem.
It is a pleasure to acknowledge helpful discussions with Francesc Ferrer.
[99]{}
M. Born and V. Fock, Z Physik [**51**]{}, 165 (1928).
M.V. Berry, Proc. Roy. Soc. London [**A392**]{}, 45 (1984).
For a collection of early reprints with valuable commentary, see A. Shapere and F. Wilczek, [*Geometric Phases in Physics*]{} (World Scientific, Singapore, 1989).
For a bibliography see [*e.g.*]{} J. Anandan, J. Christian and K. Wanelik, [*Resource Letter GPP-1: Geometric Phases in Physics*]{}, Am J Phys [**65**]{}, 180 (1997).
For an introduction, see S. Sachdev, [*Quantum Phase Transitions*]{} (Cambridge University Press, 1999). For a recent review, see T. Senthil, cond-mat/0411275.
E. Farhi [*et al.*]{}, Science [**292**]{}, 472 (2001).
K.-P. Marzlin and B.C. Sanders, Phys. Rev. Lett. [**93**]{}, 160408 (2004).
D.M. Tong, K. Singh, L.C. Kweh and C. H. Oh, Phys. Rev. Lett. [**95**]{}, 110407 (2005).
M.S. Sarandy, L.-A. Wu and D.A. Lidar, Quant. Info. Proc. [**3**]{}, 331 (2004); quant-ph/040509v3.
A.K. Pati and A.K. Rajagopal, quant-ph/0405129v1.
Suppose $h(t_2) = h(t_1)$. With our phase convention $ |n(t_2) \rangle = |n(t_1) \rangle \exp( - i \gamma )$ where $\gamma $ is Berry’s phase for adiabatic evolution between $t_1$ and $t_2$. An explicit example will be discussed below.
L.I. Schiff, [*Quantum Mechanics*]{} (McGraw-Hill, New York, 1955).
L.D. Landau and E.M. Lifshitz, [*Course of Theoretical Physics, vol 3, Non-Relativistic Quantum Mechanics*]{} (Pergamon Press, Oxford, 1977).
See the article by J. Moody, A. Shapere and F. Wilczek in ref [@shapere], p 160-183.
M.V. Berry, Proc. Roy. Soc. London, [**A414**]{}, 31 (1987), reprinted in [@shapere].
J.-T. Hwang and P. Pechukas, J. Chem. Phys. [**67**]{}, 4640 (1977).
For the benefit of readers of endnote [@anholonomy] we observe that the Hamiltonian in eq (\[eq:solvableh\]) is periodic under $ t \rightarrow t + 2 \pi/\omega$. Under this time translation, the instantaneous eigenstates, eq (\[eq:hspinors\]), change according to $| \pm \rangle \rightarrow
| \pm \rangle \exp( \mp i \gamma )$. Here $\gamma = \pi (1 - \cos \theta)$ is one half the solid angle traced by the magnetic field, the well-known Berry phase for this problem [@berry].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Let $H_0$ and $H$ be a pair of self-adjoint operators satisfying some standard assumptions of scattering theory. It is known from previous work that if $\lambda$ belongs to the absolutely continuous spectrum of $H_0$ and $H$, then the difference of spectral projections $$D(\lambda)={\mathbbm{1}}_{(-\infty,0)}(H-\lambda)-{\mathbbm{1}}_{(-\infty,0)}(H_0-\lambda)$$ in general is not compact and has non-trivial absolutely continuous spectrum. In this paper we consider the compact approximations $D_\varepsilon(\lambda)$ of $D(\lambda)$, given by $$D_\varepsilon(\lambda)
=
\psi_\varepsilon(H-\lambda)-\psi_\varepsilon(H_0-\lambda),$$ where $\psi_\varepsilon(x)=\psi(x/\varepsilon)$ and $\psi(x)$ is a smooth real-valued function which tends to $\mp1/2$ as $x\to\pm\infty$. We prove that the eigenvalues of $D_\varepsilon(\lambda)$ concentrate to the absolutely continuous spectrum of $D(\lambda)$ as ${\varepsilon}\to+0$. We show that the rate of concentration is proportional to $|\log\varepsilon|$ and give an explicit formula for the asymptotic density of these eigenvalues. It turns out that this density is independent of $\psi$. The proof relies on the analysis of Hankel operators.
address: 'Department of Mathematics, King’s College London, Strand, London, WC2R 2LS, U.K.'
author:
- Alexander Pushnitski
title: The spectral density of a difference of spectral projections
---
Introduction {#sec.a}
============
Background
----------
Let $H_0$ and $H$ be self-adjoint operators in a Hilbert space such that the difference $H-H_0$ is compact. Then it is not difficult to show that for any continuous function ${\varphi}:{{\mathbb R}}\to{{\mathbb R}}$ which tends to zero at infinity, the difference $${\varphi}(H)-{\varphi}(H_0)
\label{a0}$$ is also compact. However, if ${\varphi}$ has discontinuities on the essential spectrum of $H_0$ and $H$, then the difference may acquire non-trivial absolutely continuous (a.c.) spectrum. The first example of this kind was constructed by M. G. Krein in [@Krein]. He was interested in the difference of spectral projections $$D(\lambda)
=
{\mathbbm{1}}_{(-\infty,0)}(H-\lambda)-{\mathbbm{1}}_{(-\infty,0)}(H_0-\lambda);
\label{a3}$$ here ${\mathbbm{1}}_{(-\infty,0)}$ is the characteristic function of the interval $(-\infty,0)$. Krein exhibited an explicit pair of bounded operators $H_0$, $H$ with $\operatorname{rank}(H-H_0)=1$; for $\lambda$ in the a.c. spectrum of $H_0$, he computed the difference and showed that it is not in the Hilbert-Schmidt class, which sufficed for his purposes. Later in [@KM], using methods of the theory of Hankel operators, it was shown that the operator $D(\lambda)$ in Krein’s example is not even compact and has non-trivial a.c. spectrum.
In [@Push1; @PY1; @PY2] this phenomenon was studied in a general setting. Let us briefly recall the results of this work. Suppose that the pair $H_0$, $H$ satisfies some standard assumptions of smooth scattering theory. It was shown that for $\lambda$ in the a.c. spectrum of $H_0$ and $H$, the spectral structure of $D(\lambda)$ can be described in terms of the scattering matrix $S(\lambda)$ for the pair $(H_0,H)$. The scattering matrix is a unitary operator in an auxiliary Hilbert space ${\mathcal{N}}$, which is a fiber space in the spectral representation of $H_0$. The space ${\mathcal{N}}$ may be finite or infinite dimensional; we denote $N=\dim{\mathcal{N}}\leq\infty$. Let $\{e^{i\theta_n(\lambda)}\}_{n=1}^N$ be the eigenvalues of $S(\lambda)$ and let us denote $$a_n(\lambda)
=
\frac12{\lverte^{i\theta_n(\lambda)}-1\rvert}
=
{\lvert\sin(\theta_n(\lambda)/2)\rvert}$$ for $\lambda$ in the a.c. spectrum of $H_0$. In [@Push1; @PY1] it was proven that the a.c. spectrum of $D(\lambda)$ can be characterised as the union of the intervals $$\sigma_{\mathrm{ac}}(D(\lambda))
=
\bigcup_{n=1}^N [-a_n(\lambda),a_n(\lambda)],
\label{a5}$$ where each interval contributes multiplicity one to the a.c. spectrum. It was also proven that the singular continuous spectrum of $D(\lambda)$ is absent and some partial information about the eigenvalues of $D(\lambda)$ was obtained. Finally, in [@PY2], the spectrum of ${\varphi}(H)-{\varphi}(H_0)$ was studied for functions ${\varphi}$ with several jump discontinuities.
Informal description of the main result
---------------------------------------
In this paper, we study the regularizations of the difference $D(\lambda)$ obtained by replacing the characteristic function ${\mathbbm{1}}_{(-\infty,0)}$ in the definition by a smooth function $\psi_{\varepsilon}$ which approaches ${\mathbbm{1}}_{(-\infty,0)}$ as ${\varepsilon}\to+0$. More precisely, let $\psi\in C^\infty({{\mathbb R}})$ be a real valued function such that $$\psi(x)
\to
\begin{cases}
1/2 & \text{ as } x\to-\infty,
\\
-1/2 & \text{ as } x\to\infty.
\end{cases}
\label{a6}$$ For ${\varepsilon}>0$, we denote $\psi_{\varepsilon}(x)=\psi(x/{\varepsilon})$ and consider the difference $$D_{\varepsilon}(\lambda)
=
\psi_{\varepsilon}(H-\lambda)
-
\psi_{\varepsilon}(H_0-\lambda).
\label{a7}$$ Clearly, we have $$\psi_{\varepsilon}(x)\to{\mathbbm{1}}_{(-\infty,0)}(x)-\tfrac12,
\quad {\varepsilon}\to+0, \quad x\not=0$$ and therefore, if $\lambda$ is not an eigenvalue of $H_0$ or $H$, then $$D_{\varepsilon}(\lambda)\to D(\lambda)
\quad\text{ strongly as ${\varepsilon}\to+0$.}$$ Fix $\lambda$ in the a.c. spectrum of $H_0$. By the results of [@Push1; @PY1], the a.c. spectrum of $D(\lambda)$ is described by the union of the bands . On the other hand, under our assumptions, the operator $D_{\varepsilon}(\lambda)$ is compact (see Lemma \[lma.b0\]) and so it has pure point spectrum. One expects that the eigenvalues of $D_{\varepsilon}(\lambda)$ concentrate to the spectral bands as ${\varepsilon}\to+0$. We show that this is indeed the case and give a quantitative description of this concentration.
Let us briefly and somewhat informally describe our assumptions on the self-adjoint operators $H_0$ and $H$; precise statements will be given in Section \[sec.b\]. We assume that $H_0$ is lower semi-bounded and $H=H_0+V$, where $V$ is $H_0$-form compact. (Lower semi-boundedness is not essential for our construction but it allows us to avoid some unimportant technical issues.) We assume that for some $b\in{{\mathbb R}}$ and for some $k\in{{\mathbb N}}$ $$(H_0+bI)^{-k}-(H+bI)^{-k}\in{\mathbf{S}}_p,$$ where ${\mathbf{S}}_p$ is a Schatten class with an exponent $p<\infty$. Finally, and most importantly, we make (in the terminology of [@Yafaev1]) a strong smoothness assumption. Let us fix an open interval $\delta\subset {{\mathbb R}}$; in what follows, the parameter $\lambda$ will be taken inside this interval. Roughly speaking, the strong smoothness assumption means that
(i) $H_0$ has a purely a.c. spectrum of a constant multiplicity $N\leq\infty$ on $\delta$;
(ii) the operator ${\mathbbm{1}}_\delta(H_0)V{\mathbbm{1}}_\delta(H_0)$ can be represented as an integral operator with a sufficiently regular kernel in the spectral representation of $H_0$.
*Our main result is as follows:* Let $g\in C({{\mathbb R}})$ be a function that vanishes identically in a neighbourhood of zero. Then for every ${\varepsilon}>0$, the operator $g(D_{\varepsilon}(\lambda))$ has a finite rank and so its trace is well defined; we prove that for any $\lambda\in\delta$ one has the asymptotic relation $$\operatorname{Tr}g(D_{\varepsilon}(\lambda))
=
{\lvert\log{\varepsilon}\rvert}
\int_{-1}^1 g(y)\mu_\lambda(y)dy
+
o({\lvert\log{\varepsilon}\rvert}),
\quad
{\varepsilon}\to+0,
\label{a8}$$ where the density function $\mu_\lambda$ is given by $$\mu_\lambda(y)
=
\frac1{\pi^2}\sum_{n=1}^N
\frac{{\mathbbm{1}}_{(-a_n(\lambda),a_n(\lambda))}(y)}{{\lverty\rvert}\sqrt{1-y^2/a_n^2(\lambda)}},
\quad
y\in(-1,1).
\label{a9}$$
Discussion
----------
*Universality:* Observe that $\mu_\lambda$ is *independent of the choice of $\psi$*, as long as $\psi$ satisfies . Further, the density $\mu_\lambda$ is the sum of the functions each of which is supported on a single band $[-a_n(\lambda), a_n(\lambda)]$. Each of these functions is a scaled version of the explicit function $$\frac1{\pi^2}
\frac{{\mathbbm{1}}_{(-1,1)}(y)}{{\lverty\rvert}\sqrt{1-y^2}},
\quad
y\in(-1,1).$$ This can be interpreted as a certain universality phenomenon in this spectral problem.
We also note that by shifts and scaling, it is easy to obtain analogous results in the case when the function $\psi\in C^\infty({{\mathbb R}})$ satisfies $$\psi(x)\to A_\pm\quad \text{ as $x\to\pm\infty$,}
\label{a9d}$$ for any given values $A_+\not=A_-$.
*Symmetry of $\mu_\lambda$:* Observe that $\mu_\lambda$ is even, $\mu_\lambda(-y)=\mu_\lambda(y)$. In particular, yields $$\operatorname{Tr}g(D_{\varepsilon}(\lambda))
=
o({\lvert\log{\varepsilon}\rvert}),
\quad
{\varepsilon}\to+0,
\quad
\text{ $g$ odd.}
\label{a10}$$ We shall give some explanation of this in Section \[sec.c2\].
*Logarithmic rate:* Let us present a heuristic argument that provides some intuition into the appearance of the logarithmic term ${\lvert\log{\varepsilon}\rvert}$ in . We use the formalism of the double operator integrals, see e.g. [@BSdoi] and references therein. Fix $\lambda\in\delta$; we have $$\psi_{\varepsilon}(H-\lambda)-\psi_{\varepsilon}(H_0-\lambda)
=
\int_{{\mathbb R}}\int_{{\mathbb R}}\frac{\psi_{\varepsilon}(x-\lambda)-\psi_{\varepsilon}(y-\lambda)}{x-y}d{\mathcal{E}}(x,y),
\label{a9a}$$ where ${\mathcal{E}}(x,y)$ is the operator valued measure on ${{\mathbb R}}\times{{\mathbb R}}$ given by $${\mathcal{E}}(\Delta,\Delta_0)
=
{\mathbbm{1}}_\Delta(H)V{\mathbbm{1}}_{\Delta_0}(H_0), \quad \Delta,\Delta_0\subset{{\mathbb R}}.$$ Roughly speaking, our strong smoothness assumption on $V$ ensures that the measure ${\mathcal{E}}$ is sufficiently regular on $\delta_0\times\delta_0$, where $\delta_0\subset \delta$ is an open set which contains $\lambda$. “Sufficiently regular” in this context means that $d{\mathcal{E}}(x,y)={\mathcal{E}}'(x,y)dxdy$, where the norm ${\lVert{\mathcal{E}}'(x,y)\rVert}_p$ in an appropriate Schatten class ${\mathbf{S}}_p$ is bounded uniformly in $(x,y)\in\delta_0\times\delta_0$.
From the regularity of ${\mathcal{E}}$ it follows that the singular behaviour of the operator as ${\varepsilon}\to+0$ is determined entirely by the behaviour of the function $\frac{\psi_{\varepsilon}(x-\lambda)-\psi_{\varepsilon}(y-\lambda)}{x-y}$ near $x=y=\lambda$ (and not by the measure ${\mathcal{E}}$). To see why ${\lvert\log {\varepsilon}\rvert}$ appears in the asymptotics , let us compute the Hilbert-Schmidt norm of the operator in : $$\begin{gathered}
\operatorname{Tr}(D_{\varepsilon}(0))^2
=
{\lVert\psi_{\varepsilon}(H-\lambda)-\psi_{\varepsilon}(H_0-\lambda)\rVert}_2^2
\\
=
\int_{\delta_0}\int_{\delta_0}
{\left\lvert\frac{\psi_{\varepsilon}(x-\lambda)-\psi_{\varepsilon}(y-\lambda)}{x-y}\right\rvert}^2{\lVert{\mathcal{E}}'(x,y)\rVert}_2^2 \, dx dy+O(1)\end{gathered}$$ as ${\varepsilon}\to+0$. Assuming that ${\lVert{\mathcal{E}}'(x,y)\rVert}_2$ is bounded uniformly in $(x,y)\in\delta_0\times\delta_0$, we end up with estimating the integral $$\int_{\delta_0}\int_{\delta_0}
{\left\lvert\frac{\psi_{\varepsilon}(x-\lambda)-\psi_{\varepsilon}(y-\lambda)}{x-y}\right\rvert}^2 dx dy.$$ An elementary calculation (using our assumption on the asymptotic behaviour of $\psi$) shows that this integral has the asymptotics $2{\lvert\log {\varepsilon}\rvert}+O(1)$ as ${\varepsilon}\to+0$.
*Comparison with other estimates:* Under somewhat more restrictive assumptions on $\psi$ (see ), the asymptotics is valid for $g(t)=t^m$, where $m$ is a sufficiently large integer (this follows from the first step of the proof in Section \[sec.j\]). Taking $m$ even, directly from we obtain an estimate in the Schatten class ${\mathbf{S}}_m$: $${\lVert\psi_{\varepsilon}(H-\lambda)-\psi_{\varepsilon}(H_0-\lambda)\rVert}_m^m
=
\operatorname{Tr}(D_{\varepsilon}(\lambda))^m
=
O({\lvert\log{\varepsilon}\rvert}),
\quad {\varepsilon}\to+0.
\label{a9b}$$ On the other hand, one can find estimates of the type $${\lVert{\varphi}(H)-{\varphi}(H_0)\rVert}_m
\leq C{\lVert{\varphi}\rVert}_* {\lVertH-H_0\rVert}_m
\label{a9c}$$ in the literature; here ${\lVert{\varphi}\rVert}_*$ is the norm of ${\varphi}$ in an appropriate function class. There are several variations of : with a different Schatten norm on the right-hand side (r.h.s.), with a different power of ${\lVertH-H_0\rVert}$ in the r.h.s., etc; see e.g. [@AP; @PS] and references therein. Substituting ${\varphi}(x)=\psi_{\varepsilon}(x-\lambda)$ into any of the estimates of the type , one *does not* recover the logarithmic behaviour . In fact, the best one can get in this way is $O({\varepsilon}^{-\alpha})$ with $\alpha>0$. This is not surprising, because estimates of the type are valid for *all* pairs of operators $H_0$, $H$, whereas we use the crucial strong smoothness assumption. This aspect is further illustrated in the following example.
*Example:* To show what can happen with $\operatorname{Tr}g(D_{\varepsilon}(\lambda))$ without any structural assumptions on $H_0$ and $H$, let us consider the following example. Let $H_0=0$ and let $H\in{\mathbf{S}}_p$ be a compact self-adjoint operator with the eigenvalues $\{\lambda_n\}_{n=1}^\infty$. Assume $\psi(0)=0$; then $$\operatorname{Tr}g(D_{\varepsilon}(0))
=
\operatorname{Tr}g(\psi(H/{\varepsilon}))
=
\sum_{n=1}^\infty g(\psi(\lambda_n/{\varepsilon})).$$ Suppose that $\psi(t)=-1/2$ for $t\geq 1$ and $g\geq0$, $g(-1/2)=1$. Then the r.h.s. can be estimated below as $$\sum_{n=1}^\infty g(\psi(\lambda_n/{\varepsilon}))
\geq
\#\{n: \lambda_n>{\varepsilon}\}.$$ Thus, by choosing the sequence $\{\lambda_n\}_{n=1}^\infty$ appropriately, we can make the r.h.s. behave as ${\varepsilon}^{-\alpha}$ with any $\alpha<p$.
*Applications:* In Section \[sec.x\] we give some examples of applications of the main result to the Schrödinger operator.
*Connection with the trace formula:* Suppose that $H-H_0$ is a trace class operator. In [@Krein], Krein proved that there exists a real-valued function $\xi\in L^1({{\mathbb R}})$ (called the spectral shift function) such that the following trace formula holds true: $$\operatorname{Tr}({\varphi}(H)-{\varphi}(H_0))
=
\int_{-\infty}^\infty \xi(x){\varphi}'(x)dx,$$ for all smooth functions ${\varphi}$ of a certain class. Taking *formally* $\varphi(x)={\mathbbm{1}}_{(\infty,0)}(x-\lambda)$ and observing that in this case $\varphi'(x)=-\delta(x-\lambda)$, we obtain the “naive trace formula” $$\operatorname{Tr}D(\lambda)=-\xi(\lambda).
\label{a1b}$$ Since $D(\lambda)$ in general fails to belong to trace class, the naive formula does not make sense as it is. However, it remains a source of inspiration in this area and several regularisations have been considered in the literature (see e.g. [@Nakamura; @FLLS]). One regularisation is to take $\varphi(x)=\psi_{\varepsilon}(x-\lambda)$; since $-\psi_{\varepsilon}'(x-\lambda)$ converges to the delta function $\delta(x-\lambda)$, we obtain $$\operatorname{Tr}D_{\varepsilon}(\lambda)
=
\int_{-\infty}^\infty \xi(x)\psi_{\varepsilon}'(x-\lambda)dx
\to
-\xi(\lambda),
\label{a1a}$$ if $\lambda$ is a Lebesgue point of $\xi$. The main result of this paper is a step towards a better understanding of .
*A conjecture:* Taking formally $g(t)=t$ in (or in ) and comparing with suggests that in the trace class case, one can hope to replace the error term $o({\lvert\log{\varepsilon}\rvert})$ in (or in ) by $\operatorname{const}+o(1)$, where the constant is related to the spectral shift function.
*Related work:* Much of our construction uses the ideas of [@PY2]. However, the nature of the results is quite different: [@PY2] describes the a.c. spectrum of $D(\lambda)$, whereas here we deal exclusively with the point spectrum.
In [@Push2], the spectrum of $D_{\varepsilon}(\lambda)$ was considered for functions $\psi(x)$ that tend to zero as ${\lvertx\rvert}\to\infty$. In this case, no spectral concentration for $D_{\varepsilon}(\lambda)$ occurs. Instead, the eigenvalues of $D_{\varepsilon}(\lambda)$ converge to some “limiting spectrum”, which is described as the spectrum of a certain *compact* model operator. This model operator depends on the scattering matrix $S(\lambda)$ and also *depends on the choice of $\psi$.* Thus, in this case the universality phenomenon discussed above does not hold. To comment on this, we note that the case $\psi(x)\to0$ as ${\lvertx\rvert}\to\infty$ can be considered as a “degenerate case” $A_+=A_-=0$ of the function of the type . Thus, roughly speaking, the problem discussed in [@Push2] corresponds to the (conjectural) next term in the asymptotics .
In [@FP] the products of spectral projections $$\Pi_{\varepsilon}(\lambda)
=
{\mathbbm{1}}_{(-\infty,-{\varepsilon})}(H_0-\lambda)
{\mathbbm{1}}_{({\varepsilon},\infty)}(H-\lambda)
{\mathbbm{1}}_{(-\infty,-{\varepsilon})}(H_0-\lambda)$$ are considered under some assumptions similar to the ones of this paper. These products are compact, while the limiting product $$\Pi(\lambda)
=
{\mathbbm{1}}_{(-\infty,0)}(H_0-\lambda)
{\mathbbm{1}}_{(0,\infty)}(H-\lambda)
{\mathbbm{1}}_{(-\infty,0)}(H_0-\lambda)$$ in general has a non-trivial a.c. spectrum. Similarly to , this spectrum can be described as the union of bands $$\sigma_{\mathrm{ac}}(\Pi(\lambda))
=
\bigcup_{n=1}^N [0,a_n(\lambda)^2];
\label{a12c}$$ this fact was established in [@Push1; @PY1]. In [@FP] it is proved that the eigenvalues of $\Pi_{\varepsilon}(\lambda)$ accumulate to the spectral bands in a manner similar to . The technique of [@FP] is quite different from the one of this paper, although it also relies on the analysis of Hankel operators.
Key ideas of the proof
----------------------
We study the operator $D_{\varepsilon}(\lambda)$ in the spectral representation of $H_0$. For the simplicity of notation, let us take $\lambda=0$ in (the general case reduces to this one by a shift). Let $\chi_0\in C_0^\infty({{\mathbb R}})$ be a real valued function equal to $1$ in a neighbourhood of zero. Roughly speaking, after a number of reductions we show that the spectrum of $D_{\varepsilon}(0)$ is accurately approximated by the spectrum of the compact self-adjoint operator $$P_-{{\pmb{\psi}}}_{\varepsilon}{{\pmb{\chi}}}_0 P_+\otimes (S(0)-I)
+
P_+{{\pmb{\psi}}}_{\varepsilon}{{\pmb{\chi}}}_0 P_-\otimes (S(0)^*-I)
\quad \text{ in } L^2({{\mathbb R}})\otimes {\mathcal{N}}.
\label{a13}$$ Here $P_\pm$ are the orthogonal projections onto the Hardy classes $H^2_\pm({{\mathbb R}})\subset L^2({{\mathbb R}})$ (see Section \[sec.a5\]), ${\mathcal{N}}$ is the Hilbert space where the scattering matrix $S(0)$ acts, and ${{\pmb{\psi}}}_{\varepsilon}$ (resp. ${{\pmb{\chi}}}_0$) is the operator of multiplication by the function $\psi_{\varepsilon}(x)$ (resp. $\chi_0(x)$) in $L^2({{\mathbb R}},dx)$. The operators with the structure are called *symmetrised Hankel operators* (SHO) in [@PY2]; they were introduced in connection with the study of the spectrum of ${\varphi}(H)-{\varphi}(H_0)$ with piecewise continuous ${\varphi}$.
Formula already depends on $S(0)$ in an explicit way. In order to deal with the part of the operator acting on $L^2({{\mathbb R}})$, we show, roughly speaking, that the relevant spectral asymptotics is independent of the choice of the function $\psi$, as long as $\psi(x)$ approaches $\mp 1/2$ as $x\to\pm\infty$. This allows us to replace $\psi$ by the explicit function $$-\tfrac1\pi\tan^{-1}(x).$$ In this case, we are able to determine the spectral asymptotics by reducing the problem to the analysis of a simple explicit Hankel operator. See Sections \[sec.c7\], \[sec.c8\] for the details of the last step.
Notation {#sec.a5}
--------
We denote by ${\mathbf{S}}_p$, $p\geq1$, the standard Schatten class and by ${\lVert\cdot\rVert}_p$ the norm in this class. We will frequently use the Hölder inequality for Schatten classes: $${\lVertXY\rVert}_r\leq {\lVertX\rVert}_p{\lVertY\rVert}_q,
\quad
\tfrac1r=\tfrac1p+\tfrac1q.
\label{a14}$$ $\mathbf B$ denotes the class of all bounded operators, ${\mathbf{S}}_\infty$ is the class of all compact operators and ${\lVert\cdot\rVert}$ is the operator norm. For a set $\delta\subset {{\mathbb R}}$, ${\mathbbm{1}}_\delta$ denotes the characteristic function of $\delta$. If $X$ is a Banach space, we denote by $L^p({{\mathbb R}},X)$, $C({{\mathbb R}},X)$ etc. the classes of $X$-valued functions on ${{\mathbb R}}$.
Let ${{\mathfrak{h}}}_1$, ${{\mathfrak{h}}}_2$ be Hilbert spaces, and let $\Omega$ be a function on ${{\mathbb R}}$ with values in the set of bounded operators acting from ${{\mathfrak{h}}}_1$ to ${{\mathfrak{h}}}_2$. Assume that $\Omega\in L^\infty({{\mathbb R}},\mathbf B)$. We associate with $\Omega$ the bounded operator (which will be denoted by the corresponding boldface symbol) $${{\pmb{\Omega}}}: L^2({{\mathbb R}},{{\mathfrak{h}}}_1)\to L^2({{\mathbb R}},{{\mathfrak{h}}}_2)$$ acting as “the multiplication by $\Omega$": $$({{\pmb{\Omega}}}f)(x)=\Omega(x)f(x),
\quad
x\in{{\mathbb R}},
\quad
f\in L^2({{\mathbb R}},{{\mathfrak{h}}}_1).
\label{a12}$$
We will work with the standard Hardy classes $H^2_\pm({{\mathbb R}})\subset L^2({{\mathbb R}})$ defined as the classes of functions $f\in L^2({{\mathbb R}})$ that admit an analytic continuation into the half-plane ${{\mathbb C}}_\pm=\{z\in{{\mathbb C}}: \pm{\hbox{{\rm Im}}\,}z>0\}$ and satisfy the estimate $$\sup_{y>0}\int_{-\infty}^\infty{\lvertf(x\pm iy)\rvert}^2 dx<\infty.$$ We denote by $P_\pm: L^2({{\mathbb R}})\to H^2_\pm({{\mathbb R}})$ the Hardy projections. Recall that the explicit formula for $P_\pm$ is $$(P_\pm f)(x)
=
\mp \frac1{2\pi i} \lim_{\epsilon\to+0}
\int_{-\infty}^\infty
\frac{f(x')}{x-x'\pm i\epsilon}dx',
\quad
\text{ a.e. $x\in{{\mathbb R}}$.}
\label{a11}$$
The structure of the paper
--------------------------
In Section \[sec.b\] we state our assumptions and present the main result of the paper (Theorem \[thm.b1\]). In Section \[sec.c\] we describe the plan of the proof, define all the main objects appearing in our construction and state the main steps of the proof as lemmas. These lemmas are proven in Sections \[sec.d\]–\[sec.i\]. The proof is concluded in Section \[sec.j\]. Applications are discussed in Section \[sec.x\].
Main result {#sec.b}
===========
Assumptions {#sec.b1}
-----------
Let $H_0$ be a self-adjoint lower semi-bounded operator in a Hilbert space ${{\mathcal H}}$. Let the perturbation $V$ be of the form $$V=G^*V_0G \quad \text{ in ${{\mathcal H}}$.}$$ Here $G$ is an operator from ${{\mathcal H}}$ to an auxiliary Hilbert space ${{\mathcal K}}$ and $V_0=V_0^*$ is a bounded operator in ${{\mathcal K}}$ (of course, the simplest case is ${{\mathcal K}}={{\mathcal H}}$, $G={\lvertV\rvert}^{1/2}$ and $V_0=\operatorname{sign}(V)$). We assume that the operator $G$ satisfies $$G(H_0+bI)^{-1/2}\in {\mathbf{S}}_{\infty}, \quad
b>-\inf \sigma(H_0).
\label{b2}$$ Condition ensures that $V$ is $H_0$-form compact, and so the perturbed operator $$H=H_0+V\quad \text{ in ${{\mathcal H}}$.}$$ may be defined as a form sum. From it is easy to derive
\[lma.b0\] Let hold true and let $\psi\in C^\infty({{\mathbb R}})$ satisfy . Then for any $\lambda\in{{\mathbb R}}$ and for any ${\varepsilon}>0$, the operator $D_{\varepsilon}(\lambda)$ is compact.
The proof is given in Section \[sec.d\].
Next, we assume that for some $p<\infty$, some $b>-\min\{\inf\sigma(H_0),\inf\sigma(H)\}$ and some $k\in{{\mathbb N}}$, one has $$(H+bI)^{-k}-(H_0+bI)^{-k}\in {\mathbf{S}}_{p}.
\label{b2a}$$ Further, we describe the strong smoothness assumption. Let $\delta\subset {{\mathbb R}}$ be an open interval and let $\overline{\delta}$ be the closure of $\delta$. Assume that the spectrum of $H_0$ on $\delta$ is purely a.c. with a constant multiplicity $N\leq\infty$. More explicitly, we assume that for some auxiliary Hilbert space ${\mathcal{N}}$ with $\dim{\mathcal{N}}=N$, there exists a unitary operator $${\mathcal{F}}:\operatorname{Ran}{\mathbbm{1}}_{\delta}(H_0)\to L^2(\delta,{\mathcal{N}})$$ such that ${\mathcal{F}}$ diagonalises $H_0$: $$({\mathcal{F}}H_0 f)(\lambda)=\lambda {\widetilde}f(\lambda), \quad
\lambda\in \delta,
\quad
{\widetilde}f={\mathcal{F}}f,$$ for any $f\in\operatorname{Ran}{\mathbbm{1}}_{\delta}(H_0)$. Further, for $\lambda\in \delta$, let $Z(\lambda): {{\mathcal K}}\to{\mathcal{N}}$ be the operator formally defined by the relation $$Z(\lambda)u=({\mathcal{F}}G^* u)(\lambda), \quad \lambda\in\delta, \quad u\in{{\mathcal K}}.
\label{b3a}$$ We assume that $Z(\lambda)\in{\mathbf{S}}_{2p}$ and that the estimates $${\lVertZ(\lambda)\rVert}_{2p}\leq C,
\quad
{\lVertZ(\lambda)-Z(\lambda')\rVert}_{2p}\leq C{\lvert\lambda-\lambda'\rvert}^\gamma,
\quad
\lambda, \lambda'\in \overline{\delta}
\label{b5}$$ are satisfied with some exponent $\gamma\in(0,1)$. In other words, $Z\in C^\gamma(\overline{\delta},{\mathbf{S}}_{2p})$. Formula implies that $G$ acts upon any $f\in\operatorname{Ran}{\mathbbm{1}}_{\delta}(H_0)$ according to the formula $$Gf = \int_{\delta} Z(\lambda)^*{\widetilde}f(\lambda)d\lambda,
\quad
{\widetilde}f={\mathcal{F}}f\in L^2(\delta,{\mathcal{N}}).
\label{b4}$$ This is slightly stronger than what is called strong smoothness in [@Yafaev1 Section 4.4]; strong smoothness refers to the case when the norm in is the operator norm. It is natural to call the above assumption the ${\mathbf{S}}_{2p}$-valued strong smoothness. In applications, this assumption is easy to verify, see Section \[sec.x\].
Let us summarize our assumptions:
\[ass1\]
1. $H_0$ is lower semi-bounded and $H=H_0+V$, where $V=G^*V_0G$ satisfies .
2. For some $b>-\min\{\inf\sigma(H_0),\inf\sigma(H)\}$, $p<\infty$ and $k\in{{\mathbb N}}$, we have the inclusion .
3. $H_0$ has a purely a.c. spectrum with multiplicity $N$ on the interval $\delta$.
4. $G$ satisfies the ${\mathbf{S}}_{2p}$-valued strong smoothness assumption , on $\delta$.
The assumption that $H_0$ and $H$ are lower semi-bounded is not essential for our construction. We choose to impose this assumption here simply because it allows us to avoid several irrelevant technical issues and to make the exposition more transparent. In [@PY2], related analysis of $D(\lambda)$ is carried out without the lower semi-boundedness assumption.
Scattering theory {#sec.b2}
-----------------
For ${\hbox{{\rm Im}}\,}z\not=0$, we set $$R(z)=(H-z)^{-1},
\quad
R_0(z)=(H_0-z)^{-1},
\quad
T(z)=GR_0(z)G^*.
\label{b6}$$ Since $G$ is not assumed bounded, the precise definition of $T(z)$ is $$T(z)=G(H_0+bI)^{-1/2}\frac{H_0+bI}{H_0-zI}(G(H_0+bI)^{-1/2})^*.$$ By the assumption , it follows that $T(z)$ is compact. We need the following well-known results, see e.g. [@Yafaev1 Section 4].
\[pr.b1\] Let Assumption \[ass1\] hold. Then the operator-valued function $T(z)$ defined by is Hölder continuous (with the exponent $\gamma$) in the operator norm for ${\hbox{{\rm Re}}\,}z\in \delta$, ${\hbox{{\rm Im}}\,}z\geq 0$; in particular, the limits $T(\lambda+i0)$ exist in ${\mathbf{S}}_\infty$ and are Hölder continuous in $\lambda\in \delta$. Let $\delta_*\subset \delta$ be the set of $\lambda$ such that the equation $$f+T(\lambda+ i0)V_0f=0
\label{b10}$$ has a non-trivial solution $f\in{{\mathcal K}}$, and let $\delta_0=\delta\setminus\delta_*$. Then $\delta_*$ is closed in $\delta$ and has the Lebesgue measure zero. The inverse operator $(I+T(\lambda+i0)V_0)^{-1}$, $\lambda\in\delta_0$, exists, is bounded and is a Hölder continuous function of $\lambda\in\delta_0$. Finally, the local wave operators $$W_\pm=
W_\pm(H,H_0;\delta_0)
=
\operatorname*{s-lim}_{t\to \pm \infty}e^{iHt}e^{-iH_{0}t}{\mathbbm{1}}_{\delta_0}(H_0)$$ exist and are complete, i.e. $\operatorname{Ran}W_\pm=\operatorname{Ran}{\mathbbm{1}}_{\delta_0}(H)$.
The local scattering operator $$\mathbf S=W_+^*W_-$$ is unitary on $\operatorname{Ran}{\mathbbm{1}}_{\delta_0}(H_0)$ and commutes with $H_0$. Thus, we have a representation $$({\mathcal{F}}\mathbf S {\mathcal{F}}^* u)(\lambda)
=
S(\lambda)u(\lambda),
\quad \text{ a.e. $\lambda\in \delta_0$, \ $u\in L^2(\delta_0,{\mathcal{N}})$}$$ where the operator $S(\lambda)$ in ${\mathcal{N}}$ is called the scattering matrix for the pair $H_0$, $H$. The scattering matrix is unitary in ${\mathcal{N}}$. The difference $S(\lambda)-I$ is a compact operator (in fact, it belongs to ${\mathbf{S}}_p$, see ). Thus, the spectrum of $S(\lambda)$ consists of eigenvalues on the unit circle accumulating possibly only to the point $1$. As in Section \[sec.a\], we denote the eigenvalues of $S(\lambda)$ by $\{e^{i\theta_n(\lambda)}\}_{n=1}^N$, $N\leq\infty$, and use the notation $a_n(\lambda)=\tfrac12{\lverte^{i\theta_n(\lambda)}-1\rvert}$.
Main result {#sec.b3}
-----------
\[thm.b1\] Let Assumption \[ass1\] hold true, and let $\lambda\in\delta_0$ (the set $\delta_0$ is defined in Proposition \[pr.b1\]). Let $\psi\in C^\infty({{\mathbb R}})$ satisfy and let $D_{\varepsilon}(\lambda)$ be defined by . Let $g={\mathbbm{1}}_\omega$, where $\omega\subset{{\mathbb R}}$ is an open interval such that $0\notin\overline{\omega}$. Then the asymptotic relations $$\begin{aligned}
\lim_{{\varepsilon}\to+0}\,
{\lvert\log{\varepsilon}\rvert}^{-1}\operatorname{Tr}g(D_{\varepsilon}(\lambda))
&=
\frac1{2\pi^2}\sum_{n=1}^N
\int_{-\infty}^\infty
\biggl\{
g\biggl(\frac{a_n(\lambda)}{\cosh x}\biggr)
+
g\biggl(\frac{-a_n(\lambda)}{\cosh x}\biggr)
\biggr\}
dx
\label{b7}
\\
&=
\int_{-1}^1 g(y)\mu_\lambda(y)dy,
\label{b8}\end{aligned}$$ hold true, where $\mu_\lambda$ is given by . The relations , also hold true for any $g\in C({{\mathbb R}})$ that vanishes in a neighbourhood of the origin.
Of course, follows from by a change of variable $y=a_n/\cosh x$.
The outline of the proof {#sec.c}
========================
Here we present the key steps of the proof of Theorem \[thm.b1\] and introduce all relevant objects. Details are filled in in Sections \[sec.d\]–\[sec.j\].
The strategy {#sec.c1}
------------
In what follows for the simplicity of notation we set the parameter $\lambda\in\delta_0$ in Theorem \[thm.b1\] to be equal to zero and denote $D_{\varepsilon}:=D_{\varepsilon}(0)$. We will initially assume that $\psi\in C^\infty({{\mathbb R}})$ satisfies a stronger assumption than , viz. $$\psi(x)
=
\begin{cases}
1/2 & \text{ for } x<-R
\\
-1/2 & \text{ for } x>R
\end{cases}
\label{a6aa}$$ with some $R>0$. Reduction of the general case to this one is done at the very end of the proof in Section \[sec.j\] by using a simple variational argument.
Our strategy is to transform $D_{\varepsilon}$ through the sequence of steps $$D_{\varepsilon}=:D_{\varepsilon}^{(0)}
\to D_{\varepsilon}^{(1)}
\to D_{\varepsilon}^{(2)}
\to D_{\varepsilon}^{(3)}
\to D_{\varepsilon}^{(4)}
\to D_{\varepsilon}^{(5)}
\label{c1}$$ until we arrive at a “sufficiently simple” operator $D_{\varepsilon}^{(5)}$. At each step (apart from $D_{\varepsilon}^{(1)}\to D_{\varepsilon}^{(2)}$, which is a unitary equivalence) we are able to control the error term as follows: $${\lVertD_{\varepsilon}^{(i)}-D_{\varepsilon}^{(i-1)}\rVert}_q=O(1), \quad {\varepsilon}\to+0,
\label{c2}$$ where $q$ is a sufficiently large exponent. The precise restrictions on $q$ will vary with $i$, but the choice $q\geq2p$, $q>1/\gamma$ will work for all $i$; here $\gamma$ is the Hölder exponent in the ${\mathbf{S}}_{2p}$ valued strong smoothness assumption .
After the estimates have been established, the proof proceeds as follows. The main task is to prove the asymptotics for $g(t)=t^m$ for all integers $m\geq q$. (The general case is easily obtained from here by a standard application of the Weierstrass approximation theorem.) Denote by $\Delta_m$ be the r.h.s. of for $g(t)=t^m$: $$\Delta_m
=
\frac{1+(-1)^m}{2\pi^2}
\sum_{n=1}^N a_n(0)^m
\int_{-\infty}^\infty (\cosh x)^{-m} dx.$$ Let $P^{(i)}$ be the statement $$P^{(i)}:
\qquad
{\lvert\log{\varepsilon}\rvert}^{-1}\operatorname{Tr}(D_{\varepsilon}^{(i)})^m\to\Delta_m,
\quad
{\varepsilon}\to+0,
\quad
\forall m\geq q.
\label{c4}$$ Our aim is to prove $P^{(0)}$. But we start from the other end of the chain : the operator $D^{(5)}_{\varepsilon}$ is sufficiently simple so we are able to establish $P^{(5)}$, see Section \[sec.c8\] below. After that, using an operator theoretic argument (see Section \[sec.j\]), we prove that $P^{(i)}$ together with the estimate gives $P^{(i-1)}$. Thus, moving backwards along the chain , we arrive at the required statement $P^{(0)}$.
One exception is the step $D_{\varepsilon}^{(1)}\to D_{\varepsilon}^{(2)}$; here the operators $D_{\varepsilon}^{(1)}$ and $D_{\varepsilon}^{(2)}$ are unitarily equivalent. Thus, we have $\operatorname{Tr}(D_{\varepsilon}^{(1)})^m=\operatorname{Tr}(D_{\varepsilon}^{(2)})^m$ and so the statements $P^{(1)}$ and $P^{(2)}$ are equivalent.
The operators $D_{\varepsilon}^{(i)}$ with $i=3,4,5$ are constructed as operator valued *symmetrised Hankel operators*. In the next subsection we introduce and briefly discuss this concept. In the rest of this section, we describe each of the operators $D_{\varepsilon}^{(i)}$ in the chain .
Symmetrised Hankel operators {#sec.c2}
----------------------------
Let ${{\mathfrak{h}}}$ be a Hilbert space (the case $\dim{{\mathfrak{h}}}<\infty$ is not excluded). We will denote by $P_\pm$ the orthogonal projection in $L^2({{\mathbb R}},{{\mathfrak{h}}})$ onto the vector-valued Hardy class $H^2_\pm({{\mathbb R}},{{\mathfrak{h}}})$. Of course, this projection is given by the same formula as in the scalar-valued case. For $\Omega\in L^\infty({{\mathbb R}},{{\mathfrak{h}}})$ we call $$P_-{{\pmb{\Omega}}}P_+: H^2_+({{\mathbb R}},{{\mathfrak{h}}})\to H^2_-({{\mathbb R}},{{\mathfrak{h}}})$$ the *Hankel operator* (HO) with the symbol $\Omega$, and we call $$\operatorname{SHO}(\Omega)
=
P_-{{\pmb{\Omega}}}P_+
+
P_+{{\pmb{\Omega}}}^* P_-
: L^2({{\mathbb R}},{{\mathfrak{h}}})\to L^2({{\mathbb R}},{{\mathfrak{h}}})
\label{c6}$$ the *symmetrized Hankel operator* (SHO) with the symbol $\Omega$. Of course, the notion of HO is standard, while the notion of SHO is not; to the best of the author’s knowledge, SHOs were introduced in [@PY1] as models for the operators of the type $D(\lambda)$.
By definition, SHOs are self-adjoint. It is well known (see e.g. [@Peller Section 2.4]) that if the symbol $\Omega\in C({{\mathbb R}},{\mathbf{S}}_\infty)$ and ${\lVert\Omega(\lambda)\rVert}\to0$ as ${\lvert\lambda\rvert}\to\infty$, then the corresponding HO (and therefore also the SHO) is compact. In this paper, we will only deal with symbols of this class.
In order to comment on the nature of the spectrum of SHOs, we recall without proof a simple operator theoretic statement.
\[lma.c1\] Let ${{\mathcal H}}_1$, ${{\mathcal H}}_2$ be Hilbert spaces and let $X:{{\mathcal H}}_1\to{{\mathcal H}}_2$ be a compact operator. Then the non-zero eigenvalues of the “block-matrix” $$\begin{pmatrix}
0 & X^*
\\
X & 0
\end{pmatrix}
\quad
\text{ in }
{{\mathcal H}}_1\oplus {{\mathcal H}}_2$$ are given by $\{\pm s_n(X)\}$, where $\{s_n(X)\}$ are the non-zero singular values of $X$.
The operator $\operatorname{SHO}(\Omega)$ can be written as $$\operatorname{SHO}(\Omega)
=
\begin{pmatrix}
0 & (P_-{{\pmb{\Omega}}}P_+)^*
\\
(P_ - {{\pmb{\Omega}}}P_+) & 0
\end{pmatrix}
\quad
\text{ in }
L^2({{\mathbb R}},{{\mathfrak{h}}})=H^2_+({{\mathbb R}},{{\mathfrak{h}}})\oplus H^2_-({{\mathbb R}},{{\mathfrak{h}}}).$$ Thus, Lemma \[lma.c1\] reduces the analysis of the spectrum of $\operatorname{SHO}(\Omega)$ to computing the singular values of the HO $P_- {{\pmb{\Omega}}}P_+$. We will use this idea below.
Lemma \[lma.c1\] also shows that the spectrum of $\operatorname{SHO}(\Omega)$ is symmetric with respect to the reflection around zero. In particular, if $\operatorname{Tr}g(\operatorname{SHO}(\Omega))$ exists for some odd function $g$, then it equals zero. This gives some insight into the symmetry of the density function $\mu_\lambda$.
Spectral localization: $D_{\varepsilon}^{(0)}\to D_{\varepsilon}^{(1)}$ {#sec.c3}
-----------------------------------------------------------------------
Let $\chi_0\in C_0^\infty({{\mathbb R}})$ be a real valued function such that $\operatorname{supp}\chi_0\subset \delta_0$ and such that $\chi_0(\lambda)=1$ for $\lambda$ in a neighbourhood of $\lambda=0$. We set $$D_{\varepsilon}^{(1)}
=
\chi_0(H_0)
(\chi_0(H)^2\psi_{\varepsilon}(H)-\chi_0(H_0)^2\psi_{\varepsilon}(H_0))\chi_0(H_0).$$ Thus, we regularize $D_{\varepsilon}^{(0)}$ in two ways: (i) we introduce the “spectral cutoff” by replacing $D_{\varepsilon}^{(0)}$ with $\chi_0(H_0)D_{\varepsilon}^{(0)}\chi_0(H_0)$ and (ii) we replace $\psi_{\varepsilon}$ in the definition of $D_{\varepsilon}^{(0)}$ by $\psi_{\varepsilon}\chi_0^2$. (We write $\chi_0^2$ rather than $\chi_0$ here for a trivial reason: it will be convenient later to split this term into a product of two cutoff functions.) Item (i) above is purely technical; item (ii) highlights the fact that the only important aspect of $\psi_{\varepsilon}$ is that this function “approaches a jump” at zero, and the behaviour of $\psi_{\varepsilon}$ outside a neighbourhood of zero is irrelevant. We will prove
\[lma.c2\] Let Assumption \[ass1\](1), (2) hold true, and assume . Then $${\lVertD_{\varepsilon}^{(0)}-D_{\varepsilon}^{(1)}\rVert}_p=O(1), \quad {\varepsilon}\to+0.$$
Note that we do not need the strong smoothness assumption here. The proof of the Lemma will be given in Section \[sec.d\]; it involves only some simple $C^\infty$ functional calculus for $H_0$ and $H$.
Application of the resolvent identity: $D_{\varepsilon}^{(1)}\to D_{\varepsilon}^{(2)}$ {#sec.c4}
---------------------------------------------------------------------------------------
First we need some notation. Let $$Y(z)=V_0(I+T(z)V_0)^{-1}, \quad {\hbox{{\rm Im}}\,}z>0
\label{c10}$$ (recall that $T(z)$ is defined in ). Let $\chi_0$ be as in the previous subsection. Using the notation $Z(\lambda)$ (see ), we set $$Z_0(\lambda)
=
\begin{cases}
Z(\lambda)\chi_0(\lambda), & \lambda\in \delta_0,
\\
0, & \lambda\in {{\mathbb R}}\setminus\delta_0,
\end{cases}
\quad
Y_0(\lambda)
=
\begin{cases}
Y(\lambda+i0)\chi_0(\lambda), & \lambda\in \delta_0,
\\
0, & \lambda\in {{\mathbb R}}\setminus\delta_0.
\end{cases}$$ Thus, $Z_0$ and $Y_0$ are operator valued functions on ${{\mathbb R}}$ and by Assumption \[ass1\](4) and by Proposition \[pr.b1\] they are Hölder continuous: $$Z_0\in C^\gamma({{\mathbb R}},{\mathbf{S}}_{2p}),
\quad
Y_0\in C^\gamma({{\mathbb R}},\mathbf B).
\label{c12}$$ We will use the corresponding “multiplication operators” $${{\mathbf{Z}}}_0: L^2({{\mathbb R}},{{\mathcal K}})\to L^2({{\mathbb R}},{\mathcal{N}})
\quad \text{ and }\quad
{{\mathbf{Y}}}_0: L^2({{\mathbb R}},{{\mathcal K}})\to L^2({{\mathbb R}},{{\mathcal K}}),$$ defined as in . Consider the operator $$D_{\varepsilon}^{(2)}:={\mathcal{F}}D_{\varepsilon}^{(1)}{\mathcal{F}}^*
\quad\text{ in }\quad
L^2(\delta_0,{\mathcal{N}})\subset L^2({{\mathbb R}},{\mathcal{N}}).$$ It will be convenient to consider this operator as an operator acting on $L^2({{\mathbb R}},{\mathcal{N}})$, extending it by zero to $L^2({{\mathbb R}}\setminus\delta_0,{\mathcal{N}})$. Using the resolvent identity for $H_0$ and $H$, we will prove
\[lma.c3\] Let Assumption \[ass1\] hold true. Then $$D_{\varepsilon}^{(2)}
=
4\pi{\hbox{{\rm Im}}\,}({{\mathbf{Z}}}_0 P_- {{\mathbf{Y}}}_0 {{\pmb{\psi}}}_{\varepsilon}{{\pmb{\chi}}}_0 P_+ {{\mathbf{Z}}}_0^*)
\quad
\text{ in } L^2({{\mathbb R}},{\mathcal{N}}).
\label{c14}$$
In , the projections $P_+$ and $P_-$ and the multiplication operators ${{\pmb{\psi}}}_{\varepsilon}$ and ${{\pmb{\chi}}}_0$ act in $L^2({{\mathbb R}},{{\mathcal K}})$.
Swapping ${{\mathbf{Z}}}_0$ and $P_\pm$: $D_{\varepsilon}^{(2)}\to D_{\varepsilon}^{(3)}$ {#sec.c5}
-----------------------------------------------------------------------------------------
Our next step is to swap ${{\mathbf{Z}}}_0$ with $P_-$ and $P_+$ with ${{\mathbf{Z}}}_0^*$ in the representation . Set $$D_{\varepsilon}^{(3)}
=
4\pi{\hbox{{\rm Im}}\,}(P_- {{\pmb{\psi}}}_{\varepsilon}{{\pmb{\chi}}}_0 {{\mathbf{Z}}}_0 {{\mathbf{Y}}}_0 {{\mathbf{Z}}}_0^* P_+ )
\quad
\text{ in } L^2({{\mathbb R}},{\mathcal{N}}).
\label{c15}$$ In , $P_\pm$ act in $L^2({{\mathbb R}},{\mathcal{N}})$. We will prove
\[lma.c4\] Let Assumption \[ass1\] hold true and let the exponent $q$ satisfy $q\geq 2p$, $q>1/\gamma$. Then $${\lVertD_{\varepsilon}^{(2)}-D_{\varepsilon}^{(3)}\rVert}_q=O(1), \quad {\varepsilon}\to+0.$$
The proof will be achieved by a straightforward application of the results of [@Peller Section 6.9], where Schatten norm estimates for commutators of $P_\pm$ with operator-valued functions are given.
Comparing with the definition of SHO, we find that the operator $D_{\varepsilon}^{(3)}$ is in fact a SHO in $L^2({{\mathbb R}},{\mathcal{N}})$: $$D_{\varepsilon}^{(3)}=\operatorname{SHO}(\Omega_{\varepsilon}^{(3)}), \quad
\Omega_{\varepsilon}^{(3)}=-2\pi i \psi_{\varepsilon}\chi_0 Z_0 Y_0 Z_0^*.$$ This already shows (see the discussion after Lemma \[lma.c1\]) that $\operatorname{Tr}(D_{\varepsilon}^{(3)})^m=0$ for all odd $m$, whenever the trace exists.
It is important that we can rewrite the symbol $\Omega_{\varepsilon}^{(3)}$ in terms of the scattering matrix $S(\lambda)$ for the pair $H_0$, $H$. Recall the stationary representation for $S(\lambda)$ (see e.g. [@Yafaev1 Section 5.5]): $$S(\lambda)=I-2\pi i Z(\lambda)Y(\lambda+i0)Z(\lambda)^*, \quad \lambda\in \delta_0.
\label{c18}$$ We denote $$S_0(\lambda)=I-2\pi i Z_0(\lambda)Y_0(\lambda)Z_0(\lambda)^*, \quad \lambda\in {{\mathbb R}}.$$ Thus, $S_0(\lambda)=S(\lambda)$ in a neighbourhood of $\lambda=0$ and by , $$S_0-I\in C^\gamma({{\mathbb R}},{\mathbf{S}}_p).$$ With this notation, we can rewrite the symbol $\Omega_{\varepsilon}^{(3)}$ as $$\Omega_{\varepsilon}^{(3)}(\lambda)
=
\psi_{\varepsilon}(\lambda)\chi_0(\lambda)(S_0(\lambda)-I), \quad \lambda\in{{\mathbb R}}.$$
Replacing $S_0(\lambda)$ by $S(0)$: $D_{\varepsilon}^{(3)}\to D_{\varepsilon}^{(4)}$ {#sec.c6}
------------------------------------------------------------------------------------
At this step, we replace the symbol $\Omega^{(3)}_{\varepsilon}$ with $$\Omega_{\varepsilon}^{(4)}(\lambda)
=
\psi_{\varepsilon}(\lambda)\chi_0(\lambda)(S(0)-I), \quad \lambda\in{{\mathbb R}}.$$ Set $D_{\varepsilon}^{(4)}=\operatorname{SHO}(\Omega^{(4)}_{\varepsilon})$; we prove
\[lma.c5\] Let Assumption \[ass1\] hold true, and assume . Let the exponent $q$ satisfy $q\geq 2p$, $q>1/\gamma$. Then $${\lVertD_{\varepsilon}^{(3)}-D_{\varepsilon}^{(4)}\rVert}_q=O(1), \quad {\varepsilon}\to+0.$$
The proof is based on the Hölder continuity of $S_0(\lambda)$ and again uses the estimates from [@Peller Section 6.9].
It is important that $\Omega_{\varepsilon}^{(4)}(\lambda)$ is a scalar multiple of a single operator $S(0)-I$ in ${\mathcal{N}}$. Identifying $L^2({{\mathbb R}},{\mathcal{N}})$ with $L^2({{\mathbb R}})\otimes {\mathcal{N}}$, we can write the SHO with the symbol $\Omega_{\varepsilon}^{(4)}$ as $$\operatorname{SHO}(\Omega_{\varepsilon}^{(4)})
=
(P_- 2{{\pmb{\psi}}}_{\varepsilon}{{\pmb{\chi}}}_0 P_+)\otimes \tfrac12 (S(0)-I)
+
(P_- 2{{\pmb{\psi}}}_{\varepsilon}{{\pmb{\chi}}}_0 P_+)^*\otimes \tfrac12 (S(0)^*-I).$$ Using Lemma \[lma.c1\], we see that for even $m$ $$\operatorname{Tr}(\operatorname{SHO}(\Omega_{\varepsilon}^{(4)}))^m
=
2\operatorname{Tr}{\lvertP_- {{\pmb{\Omega}}}_{\varepsilon}^{(4)} P_+\rvert}^m
=
2\operatorname{Tr}{\lvertP_- 2 {{\pmb{\psi}}}_{\varepsilon}{{\pmb{\chi}}}_0 P_+\rvert}^m
\sum_{n=1}^N a_n(0)^m
\label{c25}$$ (recall that $a_n(0)$ are the eigenvalues of $\tfrac12{\lvertS(0)-I\rvert}$). Thus, the problem reduces to analysing the *scalar* SHO with the symbol $2\psi_{\varepsilon}\chi_0$.
Replacing $2\psi_{\varepsilon}\chi_0$ by a model symbol : $D_{\varepsilon}^{(4)}\to D_{\varepsilon}^{(5)}$ {#sec.c7}
----------------------------------------------------------------------------------------------------------
It turns out that the leading term of the asymptotics of the trace in the r.h.s. of is independent of the details of the behaviour of the function $\psi(\lambda)$, as long as it converges sufficiently fast to the limits $\mp1/2$ as $\lambda\to\pm\infty$. Thus, we are going to replace the symbol $2\psi_{\varepsilon}\chi_0$ by an explicit model symbol. Let $$\zeta(\lambda)=-\frac2\pi \tan^{-1}(\lambda),
\quad
\zeta_{\varepsilon}(\lambda)=\zeta(\lambda/{\varepsilon}),
\quad
\lambda\in {{\mathbb R}}.
\label{c26}$$ We set $$D_{\varepsilon}^{(5)}=\operatorname{SHO}(\Omega_{\varepsilon}^{(5)}),
\quad
\Omega_{\varepsilon}^{(5)}(\lambda)
=
(\zeta_{\varepsilon}(\lambda)-\zeta(\lambda))\tfrac12(S(0)-I),
\quad
\lambda\in{{\mathbb R}}$$ and prove
\[cr.c7\] Let Assumption \[ass1\] hold true, and assume . Let $\zeta$ be given by . Then for $q=\max\{2,p\}$ one has $${\lVertD_{\varepsilon}^{(4)}-D_{\varepsilon}^{(5)}\rVert}_q=O(1), \quad {\varepsilon}\to+0.$$
The proof is based on an elementary Hibert-Schmidt estimate for scalar HOs.
Computing $\operatorname{Tr}(D_{\varepsilon}^{(5)})^m$ {#sec.c8}
------------------------------------------------------
Just as in , for even $m$ we have $$\operatorname{Tr}(D_{\varepsilon}^{(5)})^m
=
2\operatorname{Tr}{\lvertP_- ({{\pmb{\zeta}}}_{\varepsilon}-{{\pmb{\zeta}}}) P_+\rvert}^m
\sum_{n=1}^N a_n (0)^m.
\label{c30}$$ It turns out that the operator ${\lvertP_- ({{\pmb{\zeta}}}_{\varepsilon}-{{\pmb{\zeta}}}) P_+\rvert}$ can be explicitly identified and the asymptotics of its traces can be computed. Let $J$ be the involution in $L^2({{\mathbb R}})$ given by $(Jf)(x)=f(-x)$. Evidently, $J$ maps $H^2_+({{\mathbb R}})$ into $H^2_-({{\mathbb R}})$ and vice versa. Consider the operator $$K_{\varepsilon}=-iP_+({{\pmb{\zeta}}}_{\varepsilon}-{{\pmb{\zeta}}})JP_+
\quad
\text{ in $H^2_+({{\mathbb R}})$}$$ for $0<{\varepsilon}<1$. Since $\zeta$ is odd, we have $({{\pmb{\zeta}}}_{\varepsilon}-{{\pmb{\zeta}}})J=-J({{\pmb{\zeta}}}_{\varepsilon}-{{\pmb{\zeta}}})$ and so $K_{\varepsilon}$ is self-adjoint. We prove
\[lma.c8\] For ${\varepsilon}\in(0,1)$, we have $K_{\varepsilon}={\lvertP_- ({{\pmb{\zeta}}}_{\varepsilon}-{{\pmb{\zeta}}}) P_+\rvert}$. This operator belongs to the trace class and for all integers $m\geq1$ $$\operatorname{Tr}K_{\varepsilon}^m
=
{\lvert\log {\varepsilon}\rvert}\frac1{2\pi^2}
\int_{-\infty}^\infty (\cosh x)^{-m} dx+O(1)
\quad
{\varepsilon}\to+0.
\label{c31}$$
In the proof of this lemma, we use a calculation from [@FP]. Of course, and give the statement $P^{(5)}$, see .
Spectral localization {#sec.d}
=====================
Here we prove Lemmas \[lma.c2\] and \[lma.b0\]. We start with a proposition which is essentially well known:
\[lma.d1\]
1. Let Assumption \[ass1\](1) hold true and let ${\varphi}\in C({{\mathbb R}})$ be such that ${\varphi}(x)\to\operatorname{const}$ as $x\to+\infty$. Then $${\varphi}(H)-{\varphi}(H_0)\in{\mathbf{S}}_\infty.$$ In particular, under the assumption on $\psi$, the operator $D_{\varepsilon}(\lambda)$ is compact for any $\lambda\in{{\mathbb R}}$ and ${\varepsilon}>0$.
2. Let Assumption \[ass1\](1), (2) hold true and let ${\varphi}\in C^\infty({{\mathbb R}})$ be such that ${\varphi}(x)=\operatorname{const}$ for all sufficiently large $x>0$. Then $${\varphi}(H)-{\varphi}(H_0)\in {\mathbf{S}}_p.$$ In particular, under the assumption on $\psi$, the operator $D_{\varepsilon}(\lambda)$ belongs to ${\mathbf{S}}_p$ for any $\lambda\in{{\mathbb R}}$ and ${\varepsilon}>0$.
(Item (i) proves Lemma \[lma.b0\].)
\(i) We first note that the values ${\varphi}(x)$ for $x<\min\{\inf\sigma(H),\inf\sigma(H_0)\}$ do not matter, and we may modify the definition of ${\varphi}$ for such $x$ as we wish. Thus, we can represent ${\varphi}$ as ${\varphi}(x)=\operatorname{const}+{\varphi}_0(x)$, where ${\varphi}_0(x)\to0$ as ${\lvertx\rvert}\to\infty$. So it suffices to prove the statement of the Lemma for the case ${\varphi}\in C({{\mathbb R}})$ with ${\varphi}(x)\to0$ as ${\lvertx\rvert}\to\infty$. Next, let $b>-\min\{\inf\sigma(H_0),\inf\sigma(H)\}$. By the resolvent identity (see ), we have $$R(-b)-R_0(-b)\in{\mathbf{S}}_\infty.$$ It follows that $${\widetilde}{\varphi}(R(-b))
-
{\widetilde}{\varphi}(R_0(-b))
\in{\mathbf{S}}_\infty$$ for all polynomials ${\widetilde}{\varphi}$. By the Weierstrass approximation theorem, the same is true for all continuous functions ${\widetilde}{\varphi}$. Now choosing ${\widetilde}{\varphi}$ such that ${\widetilde}{\varphi}(1/(x+b))={\varphi}(x)$, we obtain $${\varphi}(H)-{\varphi}(H_0)
=
{\widetilde}{\varphi}(R(-b))
-
{\widetilde}{\varphi}(R_0(-b))
\in{\mathbf{S}}_\infty,$$ as required.
\(ii) Similarly to part (i), by subtracting a constant from ${\varphi}$ we reduce the situation to the case ${\varphi}\in C_0^\infty({{\mathbb R}})$. From the inclusion it is not difficult to deduce that $${\widetilde}{\varphi}((H+bI)^{-k})
-
{\widetilde}{\varphi}((H_0+bI)^{-k})
\in{\mathbf{S}}_p,
\quad
\forall {\widetilde}{\varphi}\in C_0^\infty({{\mathbb R}}).
\label{d3}$$ The implication $\Rightarrow$ can be obtained by any of a number of standard methods. For example, it follows from [@BS Theorem 10]; alternatively, one can use the functional calculus based on the almost analytic continuation of ${\widetilde}{\varphi}$, see e.g. [@DSj Section 8]. Now choosing ${\widetilde}{\varphi}$ such that ${\widetilde}{\varphi}(1/(x+b)^k)={\varphi}(x)$, we obtain $${\varphi}(H)-{\varphi}(H_0)
=
{\widetilde}{\varphi}((H+bI)^{-k})
-
{\widetilde}{\varphi}((H_0+bI)^{-k})
\in{\mathbf{S}}_p,$$ as required.
Let $\chi_\infty=1-\chi_0$. First consider the product $$\begin{gathered}
\chi_\infty(H_0)D_{\varepsilon}=
\chi_\infty(H_0)(\psi_{\varepsilon}(H)-\psi_{\varepsilon}(H_0))
\\
=
(\chi_\infty(H_0)-\chi_\infty(H))\psi_{\varepsilon}(H)
+
(\chi_\infty(H)\psi_{\varepsilon}(H)-\chi_\infty(H_0)\psi_{\varepsilon}(H_0)).
\label{d4}\end{gathered}$$ For the first term in the r.h.s. here we get $${\lVert(\chi_\infty(H_0)-\chi_\infty(H))\psi_{\varepsilon}(H)\rVert}_p
\leq
{\lVert\chi_\infty(H_0)-\chi_\infty(H)\rVert}_p
{\lVert\psi\rVert}_{L^\infty}
<\infty
\label{d5}$$ by Lemma \[lma.d1\](ii). Consider the second term in the r.h.s. of . Denote $\psi_0(x)=-\tfrac12\operatorname{sign}(x)$. By our assumption on $\psi$, we have $$\chi_\infty\psi_{\varepsilon}=\chi_\infty\psi_0
\quad
\text{ for all sufficiently small ${\varepsilon}>0$,}
\label{d6}$$ and therefore for such ${\varepsilon}$ the second term in the r.h.s. of becomes $$\chi_\infty(H)\psi_0(H)-\chi_\infty(H_0)\psi_0(H_0),$$ which is in ${\mathbf{S}}_p$ by Lemma \[lma.d1\](ii). Together with , this yields $${\lVertD_{\varepsilon}-\chi_0(H_0)D_{\varepsilon}\rVert}_p
=
{\lVert\chi_\infty(H_0)D_{\varepsilon}\rVert}_p
=
O(1)
\label{d7}$$ as ${\varepsilon}\to+0$. Similarly, $$\begin{gathered}
{\lVert\chi_0(H_0)D_{\varepsilon}-\chi_0(H_0)D_{\varepsilon}\chi_0(H_0)\rVert}_p
=
{\lVert\chi_0(H_0)D_{\varepsilon}\chi_\infty(H_0)\rVert}_p
\\
\leq
{\lVert\chi_0\rVert}_{L^\infty}
{\lVertD_{\varepsilon}\chi_\infty(H_0)\rVert}_p
=
O(1)
\label{d8}\end{gathered}$$ as ${\varepsilon}\to+0$. Finally, denote ${\widetilde}\chi_\infty=1-\chi_0^2$; we have, using , $$\begin{gathered}
\chi_0(H_0)D_{\varepsilon}\chi_0(H_0)-D_{\varepsilon}^{(1)}
=
\chi_0(H_0)({\widetilde}\chi_\infty(H)\psi_{\varepsilon}(H)-{\widetilde}\chi_\infty(H_0)\psi_{\varepsilon}(H_0))\chi_0(H_0)
\\
=
\chi_0(H_0)({\widetilde}\chi_\infty(H)\psi_0(H)-{\widetilde}\chi_\infty(H_0)\psi_0(H_0))\chi_0(H_0)
\label{d9}\end{gathered}$$ for all sufficiently small ${\varepsilon}$. By Lemma \[lma.d1\](ii), the expression in brackets in the r.h.s. of is in ${\mathbf{S}}_p$, and so we obtain $${\lVert\chi_0(H_0)D_{\varepsilon}\chi_0(H_0)-D_{\varepsilon}^{(1)}\rVert}_p=O(1)$$ as ${\varepsilon}\to+0$. Combining this with and , we obtain the claim of the lemma.
Application of the resolvent identity {#sec.e}
=====================================
In fact, this calculation has appeared before in [@Push2]. It is based on the iterated resolvent identity written in the form $$R(z)-R_0(z)
=
-(GR_0(\overline{z}))^*Y(z)GR_0(z),
\quad
{\hbox{{\rm Im}}\,}z>0.
\label{e1}$$ Let us recall the derivation of (see e.g. [@Yafaev1 Section 1.9]). In order to avoid inessential technical explanations related to the operator $G^*$, let us assume here that $G$ is bounded. Iterating the usual resolvent identity, we get $$\begin{aligned}
R(z)-R_0(z)
=
-R(z)VR_0(z)
&=
-R_0(z)VR_0(z)
+R_0(z)VR(z)VR_0(z)
\label{e1a}
\\
&=-R_0(z)G^*V_0(I-GR(z)G^*V_0)GR_0(z).
\label{e1b}\end{aligned}$$ We also have the identity $$(I-GR(z)G^*V_0)(I+GR_0(z)G^*V_0)=I,$$ which can be verified by expanding and using . Writing $(I+GR_0(z)G^*V_0)^{-1}$ instead of $(I-GR(z)G^*V_0)$ in and recalling the definition of $Y(z)$, we obtain .
Let ${\varphi}\in C_0^\infty (\delta_0)$ (we will eventually take ${\varphi}=\psi_{\varepsilon}\chi_0^2$). By a version of Stone’s formula, for any $u\in{{\mathcal H}}$ we have $$({\varphi}(H)u,u)
=
\frac1\pi
\lim_{\epsilon\to+0}
{\hbox{{\rm Im}}\,}\int_{-\infty}^\infty(R(x+i\epsilon)u,u){\varphi}(x)dx.
\label{e2}$$ Subtracting the analogous formula for $({\varphi}(H_0)u,u)$ from and using , we obtain $$\begin{gathered}
(({\varphi}(H)-{\varphi}(H_0))u,u)
\\
=
-\frac1\pi \lim_{\epsilon\to+0}
{\hbox{{\rm Im}}\,}\int_{-\infty}^\infty
(Y(x+i\epsilon)GR_0(x+i\epsilon)u,GR_0(x-i\epsilon)u){\varphi}(x)dx.\end{gathered}$$ Now let us apply this to $u=\chi_0(H_0)f$, where $f\in\operatorname{Ran}{\mathbbm{1}}_{\delta_0}(H_0)$. By the strong smoothness assumption , we have $$GR_0(z)\chi_0(H_0)f
=
\int_{\delta_0}\frac{Z(t)^*{\widetilde}f(t)}{t-z}\chi_0(t)dt
=
\int_{{{\mathbb R}}}\frac{Z_0(t)^*{\widetilde}f(t)}{t-z}dt,
\quad
{\widetilde}f={\mathcal{F}}f.$$ Combining these formulas, we obtain $$\begin{gathered}
(({\varphi}(H)-{\varphi}(H_0))\chi_0(H_0)f,\chi_0(H_0)f)
\\
=
4\pi \lim_{\epsilon\to+0} {\hbox{{\rm Im}}\,}\int_{{\mathbb R}}dx \int_{{\mathbb R}}dt \int_{{\mathbb R}}ds
(M_\epsilon(x,t,s){\widetilde}f(t),{\widetilde}f(s)),\end{gathered}$$ where $$M_\epsilon(x,t,s)
=
Z_0(s)\frac{1}{2\pi i}\frac{1}{s-x-i\epsilon} Y(x+i\epsilon){\varphi}(x)\biggl(-\frac1{2\pi i}\biggr)
\frac1{x-t+i\epsilon}Z_0(t)^*.$$ Now let us use this identity with ${\varphi}\chi_0$ instead of ${\varphi}$. Recalling the formulas for $P_\pm$, we obtain $$\chi_0(H_0)
({\varphi}(H)\chi_0(H)-{\varphi}(H_0)\chi_0(H_0))
\chi_0(H_0)
=
4\pi {\hbox{{\rm Im}}\,}({{\mathbf{Z}}}_0 P_- {{\mathbf{Y}}}_0 \pmb{\varphi}P_+ {{\mathbf{Z}}}_0^*).$$ Finally, we substitute ${\varphi}=\psi_{\varepsilon}\chi_0$ to obtain the required identity.
Swapping $Z_0$ and $P_\pm$ {#sec.f}
===========================
We start by giving the following corollary of the general results of [@Peller Section 6.9]:
\[lma.f1\] Let ${{\mathfrak{h}}}_1$, ${{\mathfrak{h}}}_2$ be Hilbert spaces and let $\Omega$ be a function on ${{\mathbb R}}$ with values in the set of compact operators acting from ${{\mathfrak{h}}}_1$ to ${{\mathfrak{h}}}_2$. Let $q>1$ and $\gamma>1/q$; assume that $\Omega\in C^\gamma({{\mathbb R}},{\mathbf{S}}_q)$ and $\operatorname{supp}\Omega\subset(-r,r)$. Then the operators $$P_\mp{{\pmb{\Omega}}}P_\pm: L^2({{\mathbb R}},{{\mathfrak{h}}}_1)\to L^2({{\mathbb R}},{{\mathfrak{h}}}_2)$$ belong to ${\mathbf{S}}_q$ with the norm bound $${\lVertP_\mp {{\pmb{\Omega}}}P_\pm\rVert}_q
\leq
C(q,\gamma,r) {\lVert\Omega\rVert}_{C^\gamma({{\mathbb R}},{\mathbf{S}}_q)}.
\label{f1}$$
In the product $P_\mp{{\pmb{\Omega}}}P_\pm$, the projection $P_\pm$ on the right acts in $L^2({{\mathbb R}},{{\mathfrak{h}}}_1)$ and the projection $P_\mp$ on the left acts in $L^2({{\mathbb R}},{{\mathfrak{h}}}_2)$. Notation ${\mathbf{S}}_q$ in the statement of the lemma is used in two different senses: in $C^\gamma({{\mathbb R}},{\mathbf{S}}_q)$ this is the Schatten class of operators acting from ${{\mathfrak{h}}}_1$ to ${{\mathfrak{h}}}_2$ and in the line above it is the Schatten class of operators acting from $L^2({{\mathbb R}},{{\mathfrak{h}}}_1)$ to $L^2({{\mathbb R}},{{\mathfrak{h}}}_2)$.
The relevant results in [@Peller] are stated for functions on the unit circle ${{\mathbb T}}$ rather than on the real line. In order to make the connection, consider the unitary operator $U_j:L^2({{\mathbb T}},{{\mathfrak{h}}}_j)\to L^2({{\mathbb R}},{{\mathfrak{h}}}_j)$, $j=1,2$, corresponding to the standard conformal map from the unit circle to the real line: $$(U_jf)(x)=\frac1{\sqrt{\pi}}\frac1{x+i}f\left(\frac{x-i}{x+i}\right),
\quad x\in{{\mathbb R}},
\quad f\in L^2({{\mathbb T}},{{\mathfrak{h}}}_j).$$ Then $$U_j^* P_\pm U_j=p_\pm,$$ where $p_\pm$ are the orthogonal projections in $L^2({{\mathbb T}},{{\mathfrak{h}}}_j)$ onto the Hardy classes $H^2_\pm({{\mathbb T}},{{\mathfrak{h}}}_j)$, and $$U_2^* {{\pmb{\Omega}}}U_1={{\pmb{\omega}}},$$ where ${{\pmb{\omega}}}$ is the operator of multiplication by the function $\omega$ on ${{\mathbb T}}$ obtained from $\Omega$ by the change of variable: $$\omega(e^{i\theta})=\Omega\biggl(i\frac{1+e^{i\theta}}{1-e^{i\theta}}\biggr),
\quad
e^{i\theta}\in{{\mathbb T}}.
\label{f2}$$ Corollary 9.4 from [@Peller] states that $p_\mp{{\pmb{\omega}}}p_\pm\in{\mathbf{S}}_q$ if and only if ${{\pmb{\omega}}}$ belongs to the operator-valued Besov class $B^{1/q}_{q,q}({{\mathbb T}},{\mathbf{S}}_q)$. This class for $q>1$ is defined as the class of all ${\mathbf{S}}_q$-valued functions $\omega$ on ${{\mathbb T}}$ such that the norm given by $${\lVert\omega\rVert}_{B^{1/q}_{q,q}}^q
=
\int_{-\pi}^\pi
\int_{-\pi}^\pi
\frac{{\lVert\omega(e^{i(\theta+\tau)})-\omega(e^{i\theta})\rVert}_q^q}{{\lverte^{i\tau}-1\rvert}^2}
d\theta d\tau
\label{f3}$$ is finite. (In fact, the norm vanishes on constant functions, so the precise definition of this Besov class involves taking a quotient over constants.) One also has the corresponding norm bound $${\lVertp_\mp {{\pmb{\omega}}}p_\pm\rVert}_q
\leq
C_q {\lVert\omega\rVert}_{B^{1/q}_{q,q}}.$$ By our assumption and by , we have $\omega\in C^\gamma({{\mathbb T}},{\mathbf{S}}_q)$: $$\sup_\theta{\lVert\omega(e^{i(\theta+\tau)})-\omega(e^{i\theta})\rVert}_q
\leq
C{\lverte^{i\tau}-1\rvert}^\gamma.$$ Thus, the Besov norm in is finite if $q\gamma>1$, and we obtain .
We have $$\begin{gathered}
D_{\varepsilon}^{(2)}-D_{\varepsilon}^{(3)}
=
4\pi{\hbox{{\rm Im}}\,}(({{\mathbf{Z}}}_0 P_--P_-{{\mathbf{Z}}}_0){{\mathbf{Y}}}_0{{\pmb{\psi}}}_{\varepsilon}{{\pmb{\chi}}}_0 P_+{{\mathbf{Z}}}_0^*)
\\
+
4\pi{\hbox{{\rm Im}}\,}(P_- {{\mathbf{Z}}}_0{{\mathbf{Y}}}_0{{\pmb{\psi}}}_{\varepsilon}{{\pmb{\chi}}}_0(P_+{{\mathbf{Z}}}_0^*-{{\mathbf{Z}}}_0^*P_+)).
\label{f4}\end{gathered}$$ Consider the first term in the r.h.s.; we have $${{\mathbf{Z}}}_0 P_--P_-{{\mathbf{Z}}}_0
=
P_+{{\mathbf{Z}}}_0 P_- - P_-{{\mathbf{Z}}}_0 P_+.
\label{f5}$$ Since $Z_0\in C^\gamma({{\mathbb R}},{\mathbf{S}}_{2p})\subset C^\gamma({{\mathbb R}},{\mathbf{S}}_q)$ and $Z_0$ has a compact support, we can apply Lemma \[lma.f1\] to conclude that the operator belongs to ${\mathbf{S}}_q$. Thus, we have $$\sup_{{\varepsilon}>0}
{\lVert({{\mathbf{Z}}}_0 P_--P_-{{\mathbf{Z}}}_0){{\mathbf{Y}}}_0{{\pmb{\psi}}}_{\varepsilon}{{\pmb{\chi}}}_0 P_+{{\mathbf{Z}}}_0^*\rVert}_q
\leq
{\lVert{{\mathbf{Z}}}_0 P_--P_-{{\mathbf{Z}}}_0\rVert}_q
\sup_{{\varepsilon}>0}
{\lVert{{\mathbf{Y}}}_0{{\pmb{\psi}}}_{\varepsilon}{{\pmb{\chi}}}_0 P_+{{\mathbf{Z}}}_0^*\rVert}<\infty.$$ Similar reasoning applies to the second term in the r.h.s. of .
Replacing $S_0(\lambda)$ by $S(0)$ {#sec.g}
===================================
We need to prove the estimate $${\lVertP_-(\mathbf S_0-S(0)){{\pmb{\psi}}}_{\varepsilon}{{\pmb{\chi}}}_0 P_+\rVert}_q=O(1), \quad {\varepsilon}\to+0.$$ By Lemma \[lma.f1\], it suffices to prove that $${\lVert(S_0-S(0))\psi_{\varepsilon}\chi_0\rVert}_{C^\gamma({{\mathbb R}},{\mathbf{S}}_q)}=O(1), \quad {\varepsilon}\to+0.$$ In fact, we will prove a uniform estimate in $C^\gamma({{\mathbb R}},{\mathbf{S}}_p)\subset C^\gamma({{\mathbb R}},{\mathbf{S}}_q)$. The proof of this estimate is an elementary argument which involves only the Hölder continuity of $S_0$ and the definition of $\psi$. Denote ${\widetilde}S(\lambda)=(S_0(\lambda)-S(0))\chi_0(\lambda)$. Let $\lambda_1, \lambda_2\in{{\mathbb R}}$ with ${\lvert\lambda_1\rvert}\leq{\lvert\lambda_2\rvert}$. We need to prove the estimate $${\lVert
{\widetilde}S(\lambda_2)\psi(\lambda_2/{\varepsilon})
-
{\widetilde}S(\lambda_1)\psi(\lambda_1/{\varepsilon})\rVert}_p
\leq
C{\lvert\lambda_1-\lambda_2\rvert}^\gamma
\label{g3}$$ with $C$ independent of ${\varepsilon}$. We have $$\begin{gathered}
{\widetilde}S(\lambda_2)\psi(\lambda_2/{\varepsilon})
-
{\widetilde}S(\lambda_1)\psi(\lambda_1/{\varepsilon})
\\
=
({\widetilde}S(\lambda_2)-{\widetilde}S(\lambda_1))\psi(\lambda_2/{\varepsilon})
+
{\widetilde}S(\lambda_1)(\psi(\lambda_2/{\varepsilon})-\psi(\lambda_1/{\varepsilon})).
\label{g2}\end{gathered}$$ For the first term in the r.h.s., by the Hölder continuity of ${\widetilde}S$ we immediately obtain the required uniform bound: $${\lVert({\widetilde}S(\lambda_2)-{\widetilde}S(\lambda_1))\psi(\lambda_2/{\varepsilon})\rVert}_p
\leq
{\lVert{\widetilde}S(\lambda_2)-{\widetilde}S(\lambda_1)\rVert}_p
{\lVert\psi\rVert}_{L^\infty}
\leq
C{\lvert\lambda_2-\lambda_1\rvert}^\gamma.$$ For the second term in the r.h.s. of , we have $$\begin{gathered}
{\lVert{\widetilde}S(\lambda_1)(\psi(\lambda_2/{\varepsilon})-\psi(\lambda_1/{\varepsilon}))\rVert}_p
\leq
{\lVert{\widetilde}S(\lambda_1)\rVert}_p
{\lvert\psi(\lambda_2/{\varepsilon})-\psi(\lambda_1/{\varepsilon})\rvert}
\\
\leq
C{\lvert\lambda_1\rvert}^\gamma
{\lvert\psi(\lambda_2/{\varepsilon})-\psi(\lambda_1/{\varepsilon})\rvert}.\end{gathered}$$ In order to estimate the expression in the r.h.s., let us consider three cases:
*Case 1: $\lambda_1\geq0$ and $\lambda_2\geq0$.* Recall that we have also assumed ${\lvert\lambda_1\rvert}\leq{\lvert\lambda_2\rvert}$, so in this case we have $0\leq \lambda_1\leq \lambda_2$. Since $\psi(\lambda)=-1/2$ for $\lambda\geq R$, the case $\lambda_1\geq {\varepsilon}R$ is trivial (the difference $\psi(\lambda_2/{\varepsilon})-\psi(\lambda_1/{\varepsilon})$ vanishes). So let us assume that $\lambda_1< {\varepsilon}R$. Then, by the Hölder continuity of $\psi$ (our function $\psi$ is $C^\infty$ smooth), $${\lvert\lambda_1\rvert}^\gamma {\lvert\psi(\lambda_2/{\varepsilon})-\psi(\lambda_1/{\varepsilon})\rvert}
\leq
C{\varepsilon}^\gamma R^\gamma{\lvert\lambda_2/{\varepsilon}-\lambda_1/{\varepsilon}\rvert}^\gamma
\leq
CR^\gamma{\lvert\lambda_2-\lambda_1\rvert}^\gamma.$$
*Case 2: $\lambda_1\leq0$ and $\lambda_2\leq0$.* This case can be treated exactly as Case 1.
*Case 3: $\lambda_1\lambda_2<0$.* We have $${\lvert\lambda_1\rvert}^\gamma {\lvert\psi(\lambda_2/{\varepsilon})-\psi(\lambda_1/{\varepsilon})\rvert}
\leq
2{\lVert\psi\rVert}_{L^\infty}{\lvert\lambda_1\rvert}^\gamma
\leq
2{\lVert\psi\rVert}_{L^\infty}{\lvert\lambda_2-\lambda_1\rvert}^\gamma.$$ Thus, we have proven the uniform bound .
Replacing $2\psi_{\varepsilon}\chi_0$ by a model symbol {#sec.h}
=======================================================
First let us prove an elementary estimate for scalar-valued symbols:
\[lma.c6\] Let $\psi$ satisfy and let $\zeta$ be given by . Then $${\lVertP_- (2{{\pmb{\psi}}}_{\varepsilon}{{\pmb{\chi}}}_0-({{\pmb{\zeta}}}_{\varepsilon}-{{\pmb{\zeta}}}))P_+\rVert}_2=O(1), \quad {\varepsilon}\to+0.$$
Denote, as in Section \[sec.d\], $\chi_\infty=1-\chi_0$ and $\psi_0(\lambda)=-\tfrac12\operatorname{sign}(\lambda)$. We have (using ) $$2\psi_{\varepsilon}\chi_0-(\zeta_{\varepsilon}-\zeta)
=
(2\psi_{\varepsilon}-\zeta_{\varepsilon})+(\zeta-2\psi_{\varepsilon}\chi_\infty)
=
(2\psi_{\varepsilon}-\zeta_{\varepsilon})+(\zeta-2\psi_0\chi_\infty)
\label{h1}$$ for all sufficiently small ${\varepsilon}$. First consider the second term in the r.h.s. of . Denote ${\varphi}=\zeta-2\psi_0\chi_0$; we have $$P_-\pmb{\varphi}P_+=P_-(\pmb{\varphi}P_+ - P_+\pmb{\varphi}),
\label{h1a}$$ and therefore it suffices to prove the inclusion $$\pmb{\varphi}P_+ - P_+\pmb{\varphi}\in{\mathbf{S}}_2.
\label{h2}$$ We have $${\varphi}\in C^\infty({{\mathbb R}}), \quad
{\varphi}(x)=O(1/x), \quad {\varphi}'(x)=O(1/x^2), \quad {\lvertx\rvert}\to\infty.
\label{h3}$$ The integral kernel of $\pmb{\varphi}P_+ - P_+\pmb{\varphi}$ is $$-\frac1{2\pi i}\frac{{\varphi}(x)-{\varphi}(y)}{x-y}, \quad x,y\in {{\mathbb R}}.
\label{h4}$$ It is an elementary calculation to check that conditions imply that the kernel belongs to $L^2({{\mathbb R}}^2,dx dy)$. Thus, we obtain the inclusion .
Next, consider the first term in the r.h.s. of . Denote ${\varphi}=2\psi-\zeta$; by it suffices to prove that the norm ${\lVert\pmb {\varphi}_{\varepsilon}P_+-P_+\pmb {\varphi}_{\varepsilon}\rVert}_2$ is uniformly bounded. Again, ${\varphi}$ satisfies and we have $$\int_{{\mathbb R}}\int_{{\mathbb R}}{\left\lvert\frac{{\varphi}(x/{\varepsilon})-{\varphi}(y/{\varepsilon})}{x-y}\right\rvert}^2 dxdy
=
\int_{{\mathbb R}}\int_{{\mathbb R}}{\left\lvert\frac{{\varphi}(x)-{\varphi}(y)}{x-y}\right\rvert}^2 dxdy,$$ which is finite and independent of ${\varepsilon}$.
Since $$D_{\varepsilon}^{(4)}-D_{\varepsilon}^{(5)}=\operatorname{SHO}(\Omega_{\varepsilon}^{(4)}-\Omega_{\varepsilon}^{(5)}),$$ and $$\Omega_{\varepsilon}^{(4)}-\Omega_{\varepsilon}^{(5)}
=
\bigl(2\psi_{\varepsilon}\chi_0-
(\zeta_{\varepsilon}-\zeta )\bigr)
\tfrac12(S(0)-I),$$ we obtain (for $q=\max\{p,2\}$) $$\begin{gathered}
{\lVertD_{\varepsilon}^{(4)}-D_{\varepsilon}^{(5)}\rVert}_q
\leq
2{\lVertP_-({{\pmb{\Omega}}}_{\varepsilon}^{(4)}-{{\pmb{\Omega}}}_{\varepsilon}^{(5)})P_+\rVert}_q
\\
\leq
2{\lVertP_-(2{{\pmb{\psi}}}_{\varepsilon}{{\pmb{\chi}}}_0-({{\pmb{\zeta}}}_{\varepsilon}-{{\pmb{\zeta}}}))P_+\rVert}_q
{\lVert\tfrac12(S(0)-I)\rVert}_q
\\
\leq
2{\lVertP_-(2{{\pmb{\psi}}}_{\varepsilon}{{\pmb{\chi}}}_0-({{\pmb{\zeta}}}_{\varepsilon}-{{\pmb{\zeta}}}))P_+\rVert}_2
{\lVert\tfrac12(S(0)-I)\rVert}_p,\end{gathered}$$ which is uniformly bounded by Lemma \[lma.c6\].
Computing $\operatorname{Tr}(D_{\varepsilon}^{(5)})^m$ {#sec.i}
======================================================
*Step 1:* let us prove that ${\lvertP_-({{\pmb{\zeta}}}_{\varepsilon}-{{\pmb{\zeta}}})P_+\rvert}=K_{\varepsilon}$ for all ${\varepsilon}\in(0,1)$ and that $K_{\varepsilon}$ is trace class. We denote by $\Phi$ the standard unitary Fourier transform in $L^2({{\mathbb R}})$, $$(\Phi f)(t)=\frac1{\sqrt{2\pi}}\int_{-\infty}^\infty f(x)e^{-itx}dx.$$ We have $\Phi (H^2_+({{\mathbb R}}))=L^2({{\mathbb R}}_+)$. Denote $${\widehat}K_{\varepsilon}=\Phi K_{\varepsilon}\Phi^*
\quad\text{ in $L^2({{\mathbb R}}_+)$.}$$ The operator ${\widehat}K_{\varepsilon}$ is an integral operator with the kernel $k_{\varepsilon}(t+s)$, $t,s\in{{\mathbb R}}_+$, where $$k_{\varepsilon}(t)=-\frac{i}{2\pi}\int_{-\infty}^\infty(\zeta_{\varepsilon}(x)-\zeta(x))e^{-ixt}dx,
\quad
t>0.$$ Recalling the explicit formula for $\zeta$ and integrating by parts, we obtain $$\begin{gathered}
k_{\varepsilon}(t)
=
\frac{i}{\pi^2}\int_{-\infty}^\infty
(\tan^{-1}(x/{\varepsilon})-\tan^{-1}(x))
e^{-ixt}dx
\\
=
-\frac{1}{\pi^2 t}\int_{-\infty}^\infty
(\tan^{-1}(x/{\varepsilon})-\tan^{-1}(x))
(\tfrac{d}{dx}e^{-ixt})dx
\\
=
\frac{1}{\pi^2 t}\int_{-\infty}^\infty
\left(\frac{1/{\varepsilon}}{1+x^2/{\varepsilon}^2}-\frac{1}{1+x^2}\right)e^{-ixt}dx
=
\frac{e^{-{\varepsilon}t}-e^{-t}}{\pi t}.\end{gathered}$$ Observe that the operator with this kernel may be represented as $${\widehat}K_{\varepsilon}=
\tfrac1\pi {\mathcal{L}}{\mathbbm{1}}_{({\varepsilon},1)}{\mathcal{L}}=
\tfrac1\pi
({\mathbbm{1}}_{({\varepsilon},1)}{\mathcal{L}})^*({\mathbbm{1}}_{({\varepsilon},1)}{\mathcal{L}}),
\label{i4}$$ where ${\mathcal{L}}$ is the Laplace transform in $L^2({{\mathbb R}}_+)$, $$({\mathcal{L}}f)(x)=\int_0^\infty e^{-tx}f(t)dt,$$ and ${\mathbbm{1}}_{({\varepsilon},1)}$ is the operator of multiplication by the characteristic function of $({\varepsilon},1)$ in $L^2({{\mathbb R}}_+)$. From it follows that ${\widehat}K_{\varepsilon}\geq0$ and therefore $K_{\varepsilon}\geq0$. In particular, this means that ${\lvertK_{\varepsilon}\rvert}=\sqrt{K_{\varepsilon}^*K_{\varepsilon}}=K_{\varepsilon}$. Since $$P_-({{\pmb{\zeta}}}_{\varepsilon}-{{\pmb{\zeta}}})P_+
=
JP_+J({{\pmb{\zeta}}}_{\varepsilon}-{{\pmb{\zeta}}})P_+
=
-JP_+({{\pmb{\zeta}}}_{\varepsilon}-{{\pmb{\zeta}}})JP_+
=-iJK_{\varepsilon},$$ it follows that $${\lvertP_-({{\pmb{\zeta}}}_{\varepsilon}-{{\pmb{\zeta}}})P_+\rvert}={\lvert-iJK_{\varepsilon}\rvert}
=
\sqrt{(-iJK_{\varepsilon})^*(-iJK_{\varepsilon})}={\lvertK_{\varepsilon}\rvert}=K_{\varepsilon}.$$ Finally, by inspection, ${\mathbbm{1}}_{({\varepsilon},1)}{\mathcal{L}}$ is a Hilbert-Schmidt operator and therefore ${\widehat}K_{\varepsilon}$ is trace class.
*Step 2:* We need to study the asymptotics of the traces $\operatorname{Tr}K_{\varepsilon}^m=\operatorname{Tr}{\widehat}K_{\varepsilon}^m$ and to prove that formula holds true for all natural $m$. This has been done in [@FP]; let us briefly recall the key steps of this argument. By and by the cyclicity of trace, we obtain $$\operatorname{Tr}{{\widehat}K_{\varepsilon}}^m
=
\operatorname{Tr}( {\mathbbm{1}}_{({\varepsilon},1)}(\tfrac1\pi{\mathcal{L}}^{2}) {\mathbbm{1}}_{({\varepsilon},1)})^m.$$ Next, we may apply the the result of [@LaSa]: $${\lvert
\operatorname{Tr}g( {\mathbbm{1}}_{({\varepsilon},1)}(\tfrac1\pi {\mathcal{L}}^{2}) {\mathbbm{1}}_{({\varepsilon},1)})
-
\operatorname{Tr}{\mathbbm{1}}_{({\varepsilon},1)}g(\tfrac1\pi{\mathcal{L}}^2){\mathbbm{1}}_{({\varepsilon},1)}\rvert}
\leq
\frac12{\lVertg''\rVert}_{L^\infty(0,1)}{\lVert[\tfrac1\pi{\mathcal{L}}^2,{\mathbbm{1}}_{({\varepsilon},1)}]\rVert}_2^2
\label{i5}$$ for any $g\in C^2$ with $g(0)=0$; this result only uses the fact that ${\mathbbm{1}}_{({\varepsilon},1)}$ is an orthogonal projection and $\tfrac1\pi{\mathcal{L}}^2$ is a bounded operator with the spectrum on the interval $[0,1]$. We use this with $g(t)=t^m$. A direct calculation shows that the r.h.s. in is bounded as ${\varepsilon}\to+0$.
Next, observe that ${\mathcal{L}}^2$ is the Carleman operator, i.e. the integral operator in $L^2({{\mathbb R}}_+)$ with the integral kernel $1/(s+t)$, $s,t\in{{\mathbb R}}$. Using the diagonalisation of the Carleman operator, one can compute the power $(\tfrac1\pi{\mathcal{L}}^2)^m$, which yields [@FP] $$\operatorname{Tr}({\mathbbm{1}}_{({\varepsilon},1)}(\tfrac1\pi{\mathcal{L}}^2)^{m} {\mathbbm{1}}_{({\varepsilon},1)})
=
{\lvert\log {\varepsilon}\rvert}\frac1{2\pi^2}\int_{-\infty}^\infty(\cosh x)^{-m}dx
+
O(1), \quad {\varepsilon}\to+0$$ for all $m\in{{\mathbb N}}$. This gives the required result.
Putting it all together: proof of Theorem \[thm.b1\] {#sec.j}
====================================================
Fix an even integer $q$ such that $q\geq 2p$, $q>1/\gamma$.
*Step 1: assume and let $g(t)=t^m$, $m\in{{\mathbb N}}$, $m\geq q$.*
We follow the strategy outlined in Section \[sec.c1\]. By Lemmas \[lma.c2\], \[lma.c4\], \[lma.c5\] and \[cr.c7\], we have $${\lVertD_{\varepsilon}^{(i)}-D_{\varepsilon}^{(i-1)}\rVert}_q=O(1), \quad {\varepsilon}\to+0
\label{j1}$$ for $i=1,3,4,5$. Recall that our aim is to prove the implication $$P^{(i)} + \eqref{j1} \Rightarrow P^{(i-1)}$$ for $i=1,3,4,5$, where $P^{(i)}$ is the statement $$P^{(i)}\!\!: \qquad
{\lvert\log{\varepsilon}\rvert}^{-1}\operatorname{Tr}(D_{\varepsilon}^{(i)})^m\to\Delta_m,
\quad {\varepsilon}\to+0,
\quad \forall m\geq q.$$ It suffices to prove that $P^{(i)}$ and imply $$\operatorname{Tr}(D_{\varepsilon}^{(i)})^m
-
\operatorname{Tr}(D_{\varepsilon}^{(i-1)})^m
=o({\lvert\log{\varepsilon}\rvert}), \quad {\varepsilon}\to+0.$$ We use the following simple operator theoretic estimates:
\(a) If $X\in{\mathbf{S}}_q$, $q\geq1$, then for all $m\geq q$ $${\lVertX\rVert}_m^m
=
{\lVert{\lvertX\rvert}^m\rVert}_1
\leq
{\lVertX\rVert}^{m-q}{\lVert{\lvertX\rvert}^q\rVert}_1
=
{\lVertX\rVert}^{m-q}{\lVertX\rVert}_q^q.
\label{j3}$$
\(b) If $X,Y\in{\mathbf{S}}_m$, then $${\lvert\operatorname{Tr}X^m-\operatorname{Tr}Y^m\rvert}
\leq
m {\lVertX-Y\rVert}_m\max\{{\lVertX\rVert}_m^{m-1},{\lVertY\rVert}_m^{m-1}\}.
\label{j4}$$ To prove , observe that by cyclicity of the trace one has $$\operatorname{Tr}X^m - \operatorname{Tr}{Y}^m
=
\operatorname{Tr}(X-{Y})(X^{m-1}
+
X^{m-2} {Y} + \cdots + X {Y}^{m-2} + {Y}^{m-1}),$$ and so follows by the application of the Hölder and the triangle inequality for Schatten classes.
We will take $X=D_{\varepsilon}^{(i)}$, $Y=D_{\varepsilon}^{(i-1)}$. By construction, we have a uniform bound of the operator norms: $${\lVertD_{\varepsilon}^{(i)}\rVert}=O(1), \quad {\varepsilon}\to+0
\label{j10}$$ for all $i=1,2,3,4,5$. Next, $P^{(i)}$ with $m=q$ gives (using that $q$ is an even integer) $${\lVertD_{\varepsilon}^{(i)}\rVert}_q^q
=
\operatorname{Tr}(D_{\varepsilon}^{(i)})^q
=
O({\lvert\log{\varepsilon}\rvert}), \quad {\varepsilon}\to+0.$$ From here by we obtain ${\lVertD_{\varepsilon}^{(i)}\rVert}_m^m=O({\lvert\log{\varepsilon}\rvert})$ for all $m\geq q$ and so $${\lVertD_{\varepsilon}^{(i)}\rVert}_m
=
O({\lvert\log{\varepsilon}\rvert}^{1/m}),
\quad {\varepsilon}\to+0,
\quad \forall m\geq q.
\label{j5}$$ By , and we get $${\lVertD_{\varepsilon}^{(i)}-D_{\varepsilon}^{(i-1)}\rVert}_{m}^m
\leq
{\lVertD_{\varepsilon}^{(i)}-D_{\varepsilon}^{(i-1)}\rVert}^{m-q}
{\lVertD_{\varepsilon}^{(i)}-D_{\varepsilon}^{(i-1)}\rVert}_q^{q}
=
O(1), \quad {\varepsilon}\to+0.
\label{j6}$$ Combining with , we get $${\lVertD_{\varepsilon}^{(i-1)}\rVert}_m
\leq
{\lVertD_{\varepsilon}^{(i)}\rVert}_m
+
{\lVertD_{\varepsilon}^{(i-1)}-D_{\varepsilon}^{(i)}\rVert}_{m}
=O({\lvert\log{\varepsilon}\rvert}^{1/m}), \quad {\varepsilon}\to+0.
\label{j7}$$ Substituting , and into , we get $${\lvert\operatorname{Tr}(D_{\varepsilon}^{(i)})^m
-
\operatorname{Tr}(D_{\varepsilon}^{(i-1)})^m\rvert}
\leq
O({\lvert\log{\varepsilon}\rvert}^{(m-1)/m})
=o({\lvert\log{\varepsilon}\rvert}), \quad {\varepsilon}\to+0,$$ as required.
*Step 2: Let $g$ be a polynomial with $g(t)=O(t^q)$ as $t\to0$.* Write $g(t)=\sum_m g_m t^m$; by the previous step, we obtain $$\lim_{{\varepsilon}\to+0}{\lvert\log {\varepsilon}\rvert}^{-1}\operatorname{Tr}g(D_{\varepsilon})
=
\sum_m g_m \Delta_m
=
\frac1{\pi^2}\sum_{m\text{ even}} g_m \sum_{n=1}^N a_n(0)^m \int_{-\infty}^\infty (\cosh x)^{-m}dx.$$ By the change of variable $y=a/\cosh x$, $$\begin{gathered}
a^m
\int_{-\infty}^\infty (\cosh x)^{-m}dx
=
2a^m\int_{0}^\infty (\cosh x)^{-m}dx
\\
=
2\int_0^a \frac{y^m}{y\sqrt{1-y^2/a^2}}dy
=
\int_{-a}^a \frac{y^m}{{\lverty\rvert}\sqrt{1-y^2/a^2}}dy\end{gathered}$$ for even $m$, and so we obtain $$\lim_{{\varepsilon}\to+0}{\lvert\log {\varepsilon}\rvert}^{-1}\operatorname{Tr}g(D_{\varepsilon})
=
\int_{-1}^1 g(y) \mu_0(y)dy,$$ where $\mu_0$ is the weight defined by with $\lambda=0$.
*Step 3: assume and let $g={\mathbbm{1}}_\omega$, where $\omega\subset{{\mathbb R}}$ is an open interval such that $0\notin\overline{\omega}$.* Let $A=2{\lVert\psi\rVert}_{L^\infty}$; then ${\lVertD_{\varepsilon}\rVert}\leq A$. Let $g_\pm$ be polynomials with $g_\pm(t)=O(t^q)$, $t\to0$, and $$g_-(t)\leq {\mathbbm{1}}_\omega(t)\leq g_+(t), \quad {\lvertt\rvert}\leq A.$$ Then $$\operatorname{Tr}g_-(D_{\varepsilon})\leq \operatorname{Tr}{\mathbbm{1}}_\omega(D_{\varepsilon})\leq \operatorname{Tr}g_+(D_{\varepsilon}).$$ By the previous step, it follows that $$\begin{gathered}
\limsup_{{\varepsilon}\to+0}{\lvert\log {\varepsilon}\rvert}^{-1}\operatorname{Tr}{\mathbbm{1}}_\omega(D_{\varepsilon})
\leq
\lim_{{\varepsilon}\to+0}{\lvert\log {\varepsilon}\rvert}^{-1}\operatorname{Tr}g_+(D_{\varepsilon})
=
\int_{-1}^1 g_+(t)\mu_0(t)dt,
\\
\liminf_{{\varepsilon}\to+0}{\lvert\log {\varepsilon}\rvert}^{-1}\operatorname{Tr}{\mathbbm{1}}_\omega(D_{\varepsilon})
\geq
\lim_{{\varepsilon}\to+0}{\lvert\log {\varepsilon}\rvert}^{-1}\operatorname{Tr}g_-(D_{\varepsilon})
=
\int_{-1}^1 g_-(t)\mu_0(t)dt.\end{gathered}$$ Taking infimum over all possible polynomials $g_+$, supremum over all possible polynomials $g_-$ and using the Weierstrass’ approximation theorem, we obtain the required asymptotic relation for $g={\mathbbm{1}}_\omega$.
*Step 4: the general case.*
Let $g={\mathbbm{1}}_\omega$, where $\omega$ is as above, and now we suppose that $\psi$ satisfies instead of the stronger condition . It suffices to consider the cases $\omega=(a,\infty)$ and $\omega=(-\infty,-a)$ with $a>0$. We consider the first case; the second one can be treated in the same way. Given any $d\in(0,a)$, let us represent $\psi=\psi^{(0)}+\psi^{(1)}$, where $\psi^{(0)}$ satisfies the stronger condition , and ${\lVert\psi^{(1)}\rVert}_{L^\infty}<d/2$. Then $$D_{\varepsilon}=
(\psi_{\varepsilon}^{(0)}(H)-\psi_{\varepsilon}^{(0)}(H_0))
-
(\psi_{\varepsilon}^{(1)}(H)-\psi_{\varepsilon}^{(1)}(H_0)),$$ where ${\lVert\psi_{\varepsilon}^{(1)}(H)-\psi_{\varepsilon}^{(1)}(H_0)\rVert}<d$, and therefore $$\psi_{\varepsilon}^{(0)}(H)-\psi_{\varepsilon}^{(0)}(H_0)-d\cdot I
\leq
D_{\varepsilon}\leq
\psi_{\varepsilon}^{(0)}(H)-\psi_{\varepsilon}^{(0)}(H_0)+d\cdot I.$$ By the min-max, we have $$\operatorname{Tr}{\mathbbm{1}}_{(a+d,\infty)}(\psi_{\varepsilon}^{(0)}(H)-\psi_{\varepsilon}^{(0)}(H_0))
\leq
\operatorname{Tr}{\mathbbm{1}}_{(a,\infty)}(D_{\varepsilon})
\leq
\operatorname{Tr}{\mathbbm{1}}_{(a-d,\infty)}(\psi_{\varepsilon}^{(0)}(H)-\psi_{\varepsilon}^{(0)}(H_0)).$$ Applying the previous step of the proof and subsequently letting $d\to0$, we arrive at the required result.
The case of a continuous $g$ which vanishes near the origin follows by approximating $g$ from above and from below by step-functions.
Some applications {#sec.x}
=================
In this section we give some examples of application of Theorem \[thm.b1\] to Schrödinger operators.
Zero background potential {#sec.x1}
-------------------------
Let $$H_0=-\Delta, \quad
H=-\Delta+V
\quad
\text{ in }
{{\mathcal H}}=L^2({{\mathbb R}}^d), \quad d\geq1,$$ where the real-valued potential $V$ satisfies $${\lvertV(x)\rvert}\leq C(1+{\lvertx\rvert})^{-\rho}, \quad \rho>1.
\label{b9}$$ Let ${{\mathcal K}}={{\mathcal H}}$, and let $G$ be the operator of multiplication by $G(x)=(1+{\lvertx\rvert})^{-\rho/2}$; then $V_0=V_0(x)(1+{\lvertx\rvert})^{\rho}$. The statement below is essentially well known.
\[pr.b2\] Assume . Then for any finite interval $\delta=(\lambda_1,\lambda_2)$ with $0<\lambda_1<\lambda_2<\infty$, for any $p>\max\{d/\rho,(d-1)/(\rho-1)\}$ and for any $k\in{{\mathbb N}}$ with $2pk>d$, Assumption \[ass1\] holds true. Moreover, for all $\lambda\in\delta$ the equation has no non-trivial solutions and so $\delta_0=\delta$.
Thus, in this case Theorem \[thm.b1\] holds true with any $\lambda>0$.
The inclusion is well-known. Further, by the Kato–Seiler–Simon bound [@Simon Theorem 4.1], we have for $b>0$ $$(1+{\lvertx\rvert})^{-\rho}(-\Delta+bI)^{-k}\in{\mathbf{S}}_{p}, \quad \rho p>d, \quad 2kp>d.$$ Repeating word for word the argument of [@RS3 Theorem XII.12] for the class ${\mathbf{S}}_p$ instead of ${\mathbf{S}}_1$, we obtain the inclusion .
Let us recall the formulas for the diagonalization ${\mathcal{F}}_0$ of $H_0$. The fiber space ${\mathcal{N}}_d$ is ${\mathcal{N}}_1={{\mathbb C}}^2$ if $d=1$ and ${\mathcal{N}}_d=L^2({{\mathbb S}}^{d-1})$ if $d\geq2$. One defines the operator $${\mathcal{F}}_0: L^2({{\mathbb R}}^d)\to L^2((0,\infty); {\mathcal{N}}_d)$$ by $$\begin{aligned}
({\mathcal{F}}_0u)(\lambda)
&=\frac1{\sqrt{2}}\lambda^{-\frac{1}{4}}
(\hat u(\sqrt{\lambda}),\hat u(-\sqrt{\lambda})),
\quad
\lambda\in(0,\infty), \quad d=1,
\label{x2}
\\
({\mathcal{F}}_0u)(\lambda,\omega)
&=\frac1{\sqrt{2}}\lambda^{\frac{d-2}{4}}\hat u(\sqrt{\lambda}\omega),
\quad
\lambda\in(0,\infty), \quad \omega\in{{\mathbb S}}^{d-1}, \quad d\geq2,
\label{x1}\end{aligned}$$ where $\hat u$ is the standard (unitary) Fourier transform of $u$. It is easy to see that ${\mathcal{F}}_0$ diagonalises $H_0$. From , it is easy to obtain explicit formulas for the operator $Z_0(\lambda):L^2({{\mathbb R}}^d)\to{\mathcal{N}}_d$, defined by . We have for any $\lambda>0$ $$\begin{aligned}
Z_0(\lambda)u
&=\frac1{\sqrt{2}}\lambda^{-\frac{1}{4}}
({\widehat}{Gu}(\sqrt{\lambda}),{\widehat}{Gu}(-\sqrt{\lambda})),
\quad d=1,
\label{x4}
\\
(Z_0(\lambda)u)(\omega)
&=\frac1{\sqrt{2}}\lambda^{\frac{d-2}{4}}{\widehat}{Gu}(\sqrt{\lambda}\omega),
\quad \omega\in{{\mathbb S}}^{d-1}, \quad d\geq2.
\notag\end{aligned}$$ Thus, for $d=1$ the operator $Z_0(\lambda)$ has rank $\leq2$ and it is straightforward to check that $Z_0(\lambda)$ is Hölder continuous in $\lambda>0$. For $d\geq2$, it is also easy to prove the Hölder continuity of $Z_0(\lambda)$ in ${\mathbf{S}}_{2p}$ norm, $p>\frac{d-1}{\rho-1}$; this can be done by interpolating between the cases $p=1$ and $p=\infty$. See e.g. [@Yafaev2 Lemma 8.1.8] for the details.
Finally, the fact that equation has only trivial solutions for $\lambda>0$, follows by the argument involving Agmon’s “bootstrap” and Kato’s theorem on the absence of positive eigenvalues of $H$; see e.g. [@Yafaev2 Section 1.9] for an exposition of this argument.
Constant homogeneous magnetic field in three dimensions {#sec.x2}
-------------------------------------------------------
Let us fix $B=\operatorname{const}>0$ and define the operator $$H_0=(-i\nabla-A(x))^2-B \quad\text{ in ${{\mathcal H}}=L^2({{\mathbb R}}^3)$},
\label{x5}$$ where $$A(x)=(-\tfrac12Bx_2,\tfrac12Bx_1,0)$$ is the vector-potential corresponding to the constant homogeneous magnetic field $(0,0,B)$ in ${{\mathbb R}}^3$. The term $-B$ in is added only for the purposes of normalization; with this normalization, the spectrum of $H_0$ coincides with the interval $[0,\infty)$.
The operator $H_0$ can be written as $H_0=h_0+(-\frac{\partial^2}{\partial x_3^2})$, where $h_0$ is the two-dimensional magnetic operator, $$h_0
=
\biggl(-i\frac{\partial}{\partial x_1}+\frac12 Bx_2\biggr)^2
+
\biggl(-i\frac{\partial}{\partial x_2}-\frac12 Bx_1\biggr)^2
-B
\quad\text{ in $L^2({{\mathbb R}}^2)$.}$$ The spectrum of $h_0$ consists of the Landau levels $\{0,2B,4B,\dots\}$, which are the eigenvalues of $h_0$ of infinite multiplicity. The Landau levels play the role of thresholds in the spectrum of $H_0$. We set $H=H_0+V$, where the real-valued potential $V$ satisfies $${\lvertV(x)\rvert}\leq
C (1+{\lvertx_1\rvert}+{\lvertx_2\rvert})^{-\rho_\perp}
(1+{\lvertx_3\rvert})^{-\rho},
\label{x6}$$ with $$\rho>1 \quad\text{ and }\quad 0<\rho_\perp\leq\rho.
\label{x7}$$ Let $G$ be the operator of multiplication by $G(x)=G_\perp(x_1,x_2)G_3(x_3)$, where $$G_\perp(x_1,x_2)=(1+{\lvertx_1\rvert}+{\lvertx_2\rvert})^{-\rho_\perp/2},
\quad
G_3(x_3)=(1+{\lvertx_3\rvert})^{-\rho/2}.$$
\[prp.x2\] Assume , and let $\delta\subset(0,\infty)$ be an open interval which does not contain any of the Landau levels $\{2nB\}_{n=0}^\infty$. Then Assumption \[ass1\] is satisfied with $k=1$ and with any integer exponent $p>\max\{2,2/\rho_\perp\}$. Let $\delta_*\subset\delta$ be the set of $\lambda\in\delta$ where the equation has a non-trivial solution. Then $\delta_*$ coincides with the set of eigenvalues of $H$ in $\delta$.
We note that the set of the eigenvalues of $H$ in $\delta$ may be non-empty, see examples in [@AHS Section 5]. In any case, we see that Theorem \[thm.b1\] holds true for any $$\lambda>0, \quad \lambda\notin(\{2nB\}_{n=0}^\infty\cup \sigma_p(H)).$$
\(1) The inclusion follows from the diamagnetic inequality, see e.g. [@AHS Section 2].
\(2) By our assumptions on $p$, we have $G\in L^{2p}({{\mathbb R}}^3)$. By the Kato–Seiler–Simon bound [@Simon Theorem 4.1] it follows that $$G(-\Delta+bI)^{-1}\in {\mathbf{S}}_{2p}, \quad b>0.$$ Since $p$ is assumed to be an integer, by the diamagnetic inequality this implies (see [@AHS Section 2]) that $$G(H_0+bI)^{-1}\in {\mathbf{S}}_{2p}, \quad b>0.$$ An application of the resolvent identity (see ) and of the Holder inequality for the Schatten classes (see ) shows that $$(H+bI)^{-1}-(H_0+bI)^{-1}\in {\mathbf{S}}_p$$ for $b>0$, $b>-\inf\sigma(H)$, as required.
\(3) As already mentioned, $H_0$ has a purely a.c. spectrum $[0,\infty)$ with multiplicity $N=\infty$.
\(4) Let us recall the explicit diagonalisation of $H_0$. Let $P_n$ be the orthogonal projection in $L^2({{\mathbb R}}^2)$ onto the eigenspace $\operatorname{Ker}(h_0-2nB)$ of $h_0$. Then $$h_0=\sum_{n=0}^\infty 2nB P_n$$ and so we get $$\varphi(H_0)=\sum_{n=0}^\infty P_n\otimes\varphi(-\tfrac{d^2}{dx_3^2}+2nB)
\quad
\text{ in $L^2({{\mathbb R}}^3)=L^2({{\mathbb R}}^2)\otimes L^2({{\mathbb R}})$}$$ for any bounded function $\varphi$. The operator $-\frac{d^2}{dx_3^2}$ in $L^2({{\mathbb R}})$ is diagonalised by the map ${\mathcal{F}}_0$, see . This allows one to diagonalise $H_0$. In order to display the corresponding formulas, let us assume for simplicity of notation that $\delta\subset(0,2B)$; this corresponds to only one scattering channel being open on $\delta$. Then the diagonalisation operator ${\mathcal{F}}$ for $H_0$ associated with the interval $\delta$ can be written as $${\mathcal{F}}: \operatorname{Ran}{\mathbbm{1}}_\delta(H_0)\mapsto L^2(\delta,{\mathcal{N}}),
\quad
{\mathcal{N}}=\operatorname{Ran}P_0\otimes{{\mathbb C}}^2,$$ $${\mathcal{F}}=P_0\otimes{\mathcal{F}}_0.$$ Further, let $Z_0(\lambda): L^2({{\mathbb R}})\to{{\mathbb C}}^2$ be the operator with $G=G_3$. Then the corresponding operator $Z(\lambda)$ for $H_0$ can be written as $$Z(\lambda): L^2({{\mathbb R}}^3)\to{\mathcal{N}},\quad Z(\lambda)=P_0G_\perp\otimes Z_0(\lambda), \quad \lambda\in \delta.$$ By our assumptions on the exponent $p$, we have $G_\perp\in L^{2p}({{\mathbb R}}^2)$, and therefore $P_0G_\perp\in {\mathbf{S}}_{2p}$ (this can be proven by interpolating between the cases $p=2$ and $p=\infty$; see e.g. [@FR Lemma 3.1]). In the proof of Proposition \[pr.b2\] we have already seen that $Z_0(\lambda)$ is a rank two operator which is Hölder continuous in $\lambda>0$. From here we obtain the strong ${\mathbf{S}}_{2p}$-smoothness assumption for $Z(\lambda)$.
Finally, the required statement about $\delta_*$ was proven in [@AHS Section 4].
Periodic operators {#sec.x3}
------------------
Let $H_0=-\Delta+W$ be the Schrödinger operator in $L^2({{\mathbb R}}^d)$, $d\geq1$, with a periodic background potential $W$, and let $H=H_0+V$, where $V$ satisfies the estimate with $\rho>1$. One can apply Theorem \[thm.b1\] to this pair of operators. In the one-dimensional case $d=1$ this is an easy exercise by using the standard facts about the diagonalisation of $H_0$, see e.g. [@RS4 Section XIII.16]. In the multi-dimensional case, the verification of Assumption \[ass1\] meets the following difficulty: the local structure of the energy band functions of the operator $H_0$ is not fully understood. As a consequence, the required facts about the diagonalisation operator ${\mathcal{F}}$ are not available for all energies. This issue can be overcome by making rather restrictive assumptions on the behaviour of the energy band functions, see e.g. [@BYa].
Acknowledgements {#acknowledgements .unnumbered}
================
The author is grateful to Rupert Frank and Yuri Safarov for useful remarks.
[6]{}
*Functions of operators under perturbations of class ${\mathbf{S}}_p$,* J. Funct. Anal. **258** (2010), no. 11, 3675–3724.
*Schrödinger operators with magnetic fields. I. General interactions,* Duke Math. J. **45** (1978), 847–883.
*Double Stieltjes operator integrals.* Topics in Mathematical Physics, Vol. 1, pp. 25–54, Consultants Bureau Plenum Publishing Corporation, New York 1967.
*Double operator integrals in a Hilbert space,* Integral Equations Operator Theory **47** (2003), 131–168.
*The scattering matrix for a perturbation of a periodic Schršdinger operator by decreasing potential,* St. Petersburg Math. J. **6** (1995), no. 3, 453–474.
*Spectral asymptotics in the semi-classical limit,* Cambridge University Press, 1999.
*On the singularities of the magnetic spectral shift function at the Landau levels,* Ann. H. Poincaré **5** (2004), 381–403.
*A positive density analogue of the Lieb–Thirring inequality*. Duke Math. J. **162** (2013), no. 3, 435–495.
*The spectral density of a product of spectral projections,* to appear in J. Funct. Anal., arXiv:1409.1206.
, *On Krein’s example,* Proc. Amer. Math. Soc. **136** (2008), no. 6, 2067–2071.
*On the trace formula in perturbation theory (Russian)* Mat. Sb. **33 (75)**, no. 3 (1953), 597–626.
*Szegő type limit theorems*, J. Funct. Anal. **138** (1996), no. 2, 544–559.
*The spectral shift function and the Friedel sum rule,* Ann. Henri Poincaré **14** (2013), no. 5, 1413–1424.
*Hankel operators and their applications.* Springer-Verlag, New York, 2003.
*Operator-Lipschitz functions in Schatten-von Neumann classes,* Acta Math. **207** (2011), no. 2, 375–389.
*The scattering matrix and the differences of spectral projections*, Bull. London Math. Soc. **40** (2008), 227–238.
*Scattering matrix and functions of self-adjoint operators*, J. of Spectral Theory **1** (2011), 221–236.
*Spectral theory of discontinuous functions of self-adjoint operators and scattering theory*, J. Funct. Anal. **259** (2010), 1950–1973.
*Spectral theory of piecewise continuous functions of self-adjoint operators*, Proc. Lond. Math. Soc. (3) **108** (2014), no. 5, 1079–1115.
*Methods of modern mathematical physics. III: Scattering theory.* Academic Press, New York, 1979.
*Methods of modern mathematical physics. IV: Analysis of operators.* Academic Press, New York, 1978.
*Trace ideals and their applications*. Second edition. Mathematical Surveys and Monographs **120**. Amer. Math. Soc., Providence, RI, 2005.
*Mathematical scattering theory. General theory.* Amer. Math. Soc., Providence, RI, 1992.
*Mathematical scattering theory. Analytic theory.* Amer. Math. Soc., Providence, RI, 2010.
| {
"pile_set_name": "ArXiv"
} |
epsf
Inflation has become one of the most successful ideas of high energy cosmology today [@Linde]. Actually most of the current work on experimental cosmology are based on ideas, or designed to test inflationary models [@rmp]. No matter which inflationary model we assume, at the end of this period, it must reheat the universe. This period, called reheating, is very important because basically all elementary particles populating the universe were created. Much of the current interest in reheating follows after the discovery of an exponential amplification in the number of the particles produced during preheating. Source of this exponential growth is the presence of the parametric and stochastic resonance [@TB; @KLS; @KLS1; @STB; @BHdV] in the solutions of the equations of motion. Recent investigations address the problem of parametric amplification of super-horizon perturbations [@suphub] during preheating and their consequences in the CMB spectrum [@pertur].
Because the study of models of interacting fields evolving along the expansion of the universe is a highly non-trivial issue, analytical progress in this direction has been difficult to obtain. However, as we describe later, most of the main features of the observed phenomena can be obtained by studying a toy quantum mechanical model of two bi-quadratically coupled oscillators. Furthermore the recent interest in amplification of fluctuations can be described in this context.
In this letter, we describe in a simple way how parametric and stochastic resonance arise from the dynamics of a system of two coupled oscillators and we also identify the main problems that occur with the standard approaches, such as the energy conservation problem. We derive further in a consistent way, a semiclassical equation of motion for the scalar condensate field $R(t) $ (see Eq.(\[leR\]) below). From the analysis of this equation, it is shown that the preheating phase proceeds in a different dynamical way, consistent with the Heisenberg’s principle. This fact also leads to appropriate initial conditions - an initially non-zero value for the fluctuations - to allow an efficient energy transfer between the oscillators. The equations obtained (Eq.(\[leX\]) and Eq.(\[leR\])) can be considered as a classical system, allowing us to perform a complete numerical study [@JC]. A first numerical study shows a novel saturation effect in the small $R$-region of the equation for the fluctuation. This effect holds at the beginning of the energy transfer between the oscillators. We extend our results to quantum field theory, and take into account the expansion of the universe. This is straightforward, and the results obtained in our previous analysis can be extended with slight modifications.
The model we consider is a classical oscillator $X$ interacting with a purely quantum mechanical oscillator $\hat{Y}$ described by the Lagrangian
$$L=\frac{1}{2}\dot{X}^{2}+\frac{1}{2}\dot{\hat{Y}}^{2}-\frac{1}{2}\omega ^{2}X^{2}-\frac{1}{2}\Omega ^{2}\hat{Y}^{2}-\frac{1}{2}g^{2}X^{2}\hat{Y}^{2}, \label{lagra}$$
where $X(t)$ corresponds to the classical inflaton field $\phi (t)$ and $\hat{Y}(t) $ the created scalar field $\hat{\chi}(x)$ of inflationary cosmology respectively. For reasons of simplicity we will first neglect the expansion of the universe, and discuss at the end the effects and possible modifications of including such a friction term. The equations of motion are then corresponding to two coupled harmonic oscillators with time-dependent frequencies
$$\ddot{X}+(\omega ^{2}+g^{2}\hat{Y}^{2})X=0, \label{lagx}$$
$$\ddot{\hat{Y}}+(\Omega ^{2}+g^{2}X^{2})\hat{Y}=0. \label{lagy}$$
Let us now introduce time-independent creation and annihilation Heisenberg operators $a^{\dagger }$ and $a$ respectively using the Ansatz
$$\hat{Y}=f(t)a+f^{\ast }(t)a^{\dagger }, \label{rheis}$$
where, due to the standard commutator $\left[ a,a^{\dagger }\right] =1$ for creation-annihilation operators, $f(t)$ satisfies the Wronskian condition
$$\dot{f}^{\ast }(t)f(t)-f^{\ast }(t)\dot{f}(t)=-i. \label{comf}$$
Inserting the Ansatz of Eq. (\[rheis\]) into Eq. (\[lagy\]), we find that $f(t)$ satisfies the equation of motion
$$\ddot{f}(t)+(\Omega ^{2}+g^{2}X^{2}(t))f(t)=0, \label{mateqf}$$
where the normalization is fixed by the condition (\[comf\]). If we consider $\hat{Y}$ in the frame of quantum field theory, Eq. (\[mateqf\]) would be the equation for the zero-momentum term of its Fourier expansion. Motivated by the plane wave expansion of a quantized-field operator we use the following Ansatz for the function-coefficient $f(t)$
$$f(t)=\frac{1}{\sqrt{2W(t)}}\exp \left( -i\int^{t}W(t^{\prime })dt^{\prime
}\right) ,$$
which satisfies automatically Eq. (\[comf\]). Inserting this Ansatz into Eq. (\[mateqf\]), we obtain the nonlinear differential equation for $W(t)$
$$\frac{1}{2}\frac{\ddot{W}}{W}-\frac{3}{4}\left( \frac{\dot{W}}{W}\right)
^{2}+W^{2}=m^{2}(t), \label{weqt}$$
where $m^{2}(t)=\Omega ^{2}+g^{2}X^{2}(t)$. Eq. (\[weqt\]) describes a non-linear dissipative differential equation for $W(t)$. Instead of solving this rather cumbersome equation numerically, we will consider the Hamiltonian associated to the Lagrangian of Eq. (\[lagra\]) and construct a classical effective Hamiltonian by projecting onto the subspace of the number operator $\hat{n}=a^{\dagger}a$, generated by $\{\left| 4\right\rangle ,\left| 3\right\rangle,\left| 2\right\rangle ,\left| 1\right\rangle ,\left| 0\right\rangle \}$, which we will call $S_{5}$. The vacuum is as usual defined by the condition $a\left| 0\right\rangle =0$. The choice of the non-trivial subspace $S_{5}$ is two fold. First, it attempts a fully quantum-mechanical treatment of this problem, allowing the possibility of quantum interference and other quantum effects in the effective classical theory, and secondly it allows an exact analitical solution of the quantum sector of the model. We first define the classical variable $R^{2}(t)=\left\langle 0\left| \hat{Y}^{2}\right| 0\right\rangle $. The operator $\hat{Y^{2}}$ reads
$$\left[ \hat{Y}^{2}\right] _{nm}=\left[
\begin{array}{ccccc}
\left| f\right| ^{2} & 0 & \sqrt{2}f^{2} & 0 & 0 \\
0 & 3\left| f\right| ^{2} & 0 & \sqrt{6}f^{2} & 0 \\
\sqrt{2}f^{\ast 2} & 0 & 5\left| f\right| ^{2} & 0 & 2\sqrt{3}f^{2} \\
0 & \sqrt{6}f^{\ast 2} & 0 & 7\left| f\right| ^{2} & 0 \\
0 & 0 & 2\sqrt{3}f^{\ast 2} & 0 & 9\left| f\right| ^{2}
\end{array}
\right] .$$ and can be diagonalized in $S_{5}$ explicitly $\left[ \hat{Y}^{2}\right] _{nm}V_{i}=\lambda _{i}V_{i}$, where $V_{i}$ and $\lambda_{i}$ are the eigenvectors and eigenvalues respectively. Because of the similar structure of both operators, $\hat{Y}^2$ and $\hat{P}_{Y}^2=\dot{\hat{Y}}^{2}$, we can use the Ansatz of Eq. (\[rheis\]), and the effective Hamiltonian can be diagonalized exactly as
$$H_{ij}^{eff}=\frac{1}{2}[P_{X}^{2}+P_{R}^{2}+\frac{1}{4R^{2}}+\omega
^{2}X^{2}+m^{2}(t)R^{2}]y_{i}\delta _{ij}, \label{L_eff}$$
where $i,j=0,..4$. Up to irrelevant constants, this Hamiltonian describes the effective dynamics of two classical variables $X$ and $R$, which corresponds to the original classical and quantum oscillators. From Eq. (\[L\_eff\]) the equations of motion are
$$\ddot{X}+(\omega ^{2}+g^{2}R^{2})X=0, \label{leX}$$
$$\ddot{R}+(\Omega ^{2}+g^{2}X^{2}(t))R-\frac{1}{4R^{3}}=0. \label{leR}$$
The above equations actually correspond to two classical coupled particles-like degrees of freedom. Note that the centrifugal term keeps the fluctuation $R$ away from zero, consistent with Heisenberg’s principle and giving us appropriated natural initial conditions for numerical studies [@JC]. These equations are very simple and represent the starting point of our analysis.
A similar derivation of the classical Hamilton dynamics of a classical oscillator interacting with a quantum mechanical oscillator has been made in [@Cooper], in the context of semiquantum chaos. Nevertheless, the chosen classical-variable parametrization leads to cumbersome nonlinear dissipative differential equations.
We note here, that backreaction is taken into account in Eq. (\[leX\]), through the dependence on the classical condensate variable $R$. We start the discussion of the classical dynamics by considering the perturbative regime of the equations of motion, at small coupling constant $g^{2}\ll 1$. For $g^{2}=0$, a solution of Eq. (\[leX\]) is $X(t)=X_{0}\sin (\omega t)$. Inserting this expression into Eq. (\[mateqf\]) we obtain a Mathieu equation
$$f^{\prime \prime }(z)+\left[ A(q)-2q\cos 2z\right] f(z)=0, \label{emath}$$
where $z=\omega t$, $q=g^{2}X_{0}^{2}/4\omega ^{2}$ and $A(q)=\Omega
^{2}/\omega ^{2}+2q$. This equation is well-known and has been used to discuss different mechanisms of particle production during the reheating period of the universe, such as parametric and stochastic resonance [@TB]-[@STB]. In fact, for $2q<A$ this equation leads to narrow parametric resonance [@Landau] and for $2q>A$ leads to broad parametric resonance [@KLS] (called stochastic resonance if expansion of the universe is considered). From the definitions above for $A$ and $q$, we observe that only the first region is allowed. The limit curve $A=2q$ in the space of parameters can be discussed by considering the limit case $\Omega ^{2}=0$ of the equivalent equation
$$\ddot{f(t)}+(\Omega ^{2}+g^{2}X_{0}^{2}\sin ^{2}(\omega t))f(t)=0. \label{trece}$$
This equation leads also to a linear growth of the amplitude $f(t)$, close to the minimum of the potential for $X(t)$. This can be explained intuitively because this equation reads $\ddot{f}(t)=0$ at these points (see Fig. 1), and hence has solutions which raise linearly with time.
=6.0cm
=6.0cm
=6.0cm
=6.0cm
=6.0cm
However, it is natural to ask whether or not a non-exactly oscillatory form of $X$ will lead to an exponential growth of $f(t)$. Moreover, we know from this simple system that as soon as $f(t)$ start to growth, the functional form of $X$ will change, not only in amplitude but also in frequency. Also, attention must be down to solutions of Eq. (\[leR\]) and not of Eq. (\[mateqf\]), because in general its solutions does not satisfy the Wronskian condition Eq. (\[comf\]). The best way to take into account the energy conservation, is by studying a numerical solution [@JC]. Even, the system (\[leX\],\[leR\]) can be seen as a classical system allowing us to perform in a easier way such a study.
So, let us return to equations (\[leX\]) and (\[leR\]). The second equation can be interpreted as a particle moving in the presence of the potential
$$V(R)=\frac{1}{2}m^{2}(t)R^{2}+\frac{1}{8R^{2}}. \label{potR}$$
Figure 2 shows a plot of the potential for a convenient value of $m^{2}$. We see that $R=0$ acts as an infinite wall, which divides the coordinate space into two disconnected regions. Depending on which sign of the root we choose, the particle will remain in one of these two regions.
In order to analyze this potential, there are two scales to be considered. For large amplitudes $R\gg 1/\sqrt{2m}$, the potential looks essentially like a parabola with an infinite wall at $R=0$ (see Fig. 2). The solution of the equation without the non-linear term, which leads to the infinite wall, has resonances at $\Omega _{n}=2\omega /n$, as in the conventional parametric resonance (see [@Landau]). One could naively expect to find resonances of the complete equation at twice this value because of the presence of the wall at $R=0$. But this is not true. The reason for this behavior is as follows. The solution of the problem without the wall-term is invariant under the shift $\omega t\rightarrow \omega t+\pi $. But this corresponds to the same solution of the system in the presence of the wall, when the phase is shifted by $\pi $ as $R$ approaches zero. Figure 3 shows the growth of the amplitude of oscillations for the main resonance band $\Omega =2\omega $. This curve has a similar shape and properties of the one corresponding to narrow parametric resonance, Fig. 2 in [@KLS1].
For small amplitudes $R\sim 1/\sqrt{2m}$, we can expand the potential around the time dependent minimum $R_{m}(t)=\pm 1/\sqrt{2m(t)}$
$$V(R)\simeq \frac{1}{2}m^{2}(t)+2(R-R_{m})^{2}m^{2}(t). \label{potexp}$$
From (\[potexp\]), we see that for amplitude values close to the minimum, the potential develops a linear term, which leads to non-linear oscillations. Such oscillations hold in an almost energy conserved environment, because the amplitude of $R$ is small, so this effect is a genuine one and holds during the beginning of the transfer of energy. Fig. 4 shows the typical behavior of these oscillations. There are two regions separated by the size of the amplitude. When the amplitude is small and a resonance condition is fulfill, the linear term drives the system to larger values of the amplitude, which is characteristic of the broad parametric resonance (see Fig. 3 in [@KLS1]). Nevertheless, when the amplitude grows the non-linear term of the equation becomes important and breaks the resonance tuning and the growth of the amplitude saturates, while the amplitude starts to decrease. Most of the results obtained above can be still applied when we consider the expansion of the universe.
In quantum field theory the process follows similar lines. If we consider the inflaton field $\phi $ and the created field $\chi $ coupled minimally to gravity, the Lagrangian can be written as a sum ${\cal L}={\cal L}_{0}(\phi )+{\cal L}_{0}(\chi )+{\cal L}_{I}(\phi ,\chi ),$ where $${\cal L}_{0}(\phi )=\sqrt{-g}[\frac{1}{2}(\partial \phi )^{2}-\frac{1}{2}m_{\phi }^{2}\phi ^{2}], \label{lagqft}$$ where the interaction is ${\cal L}_{I}=-\frac{1}{2}g^{2}\phi ^{2}\chi ^{2}$. Let us consider $\phi (t)$ and $\chi (t)$ as homogeneous fields, but the former as a [*classical*]{} and the later as a [*quantum*]{} operator field. Assuming a flat FRW metric, the equations of motion obtained from the Lagrangian (\[lagqft\]) are $$\chi ^{\prime \prime }+2{\cal H}\chi ^{\prime }+a^{2}(M_{\chi
}^{2}+g^{2}\phi ^{2})\chi =0, \label{eomxi}$$ $$\phi ^{\prime \prime }+2{\cal H}\phi ^{\prime }+a^{2}(m_{\phi
}^{2}+g^{2}\chi ^{2})\phi =0, \label{eomfi}$$ where $a$ is the scale factor, a prime $^{\prime }$ means a derivative respect to conformal time $\eta =\int dt/a(t)$, and ${\cal H}=a^{\prime }/a$. Defining the conformal fields as $X=a\phi $ and $\hat{Y}=a\chi $, we find the equations $$X^{\prime \prime }+(a^{2}m_{\phi }^{2}+g^{2}\hat{Y}^{2}-\frac{a^{\prime \prime }}{a})X=0, \label{xce}$$ $$\hat{Y}^{\prime \prime }+(a^{2}M_{\chi }^{2}+g^{2}X^{2}-\frac{a^{\prime \prime }}{a})\hat{Y}=0. \label{yce}$$ If we neglect the coupling between $\chi$ with $\phi$ and during the oscillations of the inflaton, the universe expands as a one dominated by matter. Under such circunstances the last term inside the brackets can be dropped out. Moreover, we have checked numerically that the last term $a^{\prime \prime }/a$ can be neglected, even if the coupling is considered.
The semiclassical approximation leads to the equation $$R^{\prime \prime }+(a^{2}M_{\chi }^{2}+g^{2}X^{2}-\frac{a^{\prime \prime }}{a})R-\frac{1}{4R^{3}}=0, \label{eqr}$$ and similarly the corresponding for $X$ (Eq.(\[xce\])) but replacing $\hat{Y}^{2}$ by $R^{2}$. If we consider the corrections beyond the long-wavelenght limit of the field $\hat{Y}$, an infinite set of non-linear coupled differential equations for the modes $R_{k}$ prevents an analogous explicit description as the one given by Eq. (\[eqr\]). Nevertheless, this observation allows a perturbative treatment of this problem by a gradient expansion of the fields, but this is beyond the scope of our analysis. Because the fields are scaled by $a$, the expansion of the universe should damp the oscillations of the created field $R$ (or $\chi $) in Fig. 1 and Fig. 3. In Fig. 5 the combined effects of backreaction and the expansion of the universe in the evolution of the field $R$ is shown. As a consequence, $R$ ceases to grow exponentially, after certain time. The resonance structure of the secondary maxima can be explained by the detuning effect due to the presence of the scale factor in the effective frequency term of Eq. (\[eqr\]).
In this letter we have described a model to study reheating after inflation. We have showed that a simple model of two coupled oscillators can reproduce the main resonances effects, responsible for particle creation during preheating. In addition we have derived a semiclassical equation describing particle production consistent with Heisenberg’s uncertainty principle. From this equation we have also found that at small amplitude of the oscillations, i. e. when the energy transfer starts, there is no resonance amplification. This seems to indicate that there is an energy threshold to get efficient particle production.
Acknowledgments {#acknowledgments .unnumbered}
===============
G. P. would like to thank to R. Labbé for helpful discussions. G. P. was supported in part by the projects FONDECYT 1980608 and DICYT 049631PA. V. H. C. would like to thank CONICYT for financial support.
A. D. Linde, [*Particle Physics and Inflationary Cosmology*]{}, Harwood, NY (1990).
M. S. Turner and J. A. Tyson, Rev. Mod. Phys. [**71**]{}, S145 (1999).
J. Traschen and R. Brandenberger, Phys. Rev. D [**42**]{}, 2491 (1990).
L. A. Kofman, A. D. Linde, and A. A. Starobinsky, Phys. Rev. Lett. [**73**]{}, 3195 (1994); [**76**]{}, 1011 (1996).
L. A. Kofman, A. D. Linde, and A. A. Starobinsky, Phys. Rev. D [**56**]{}, 3258 (1997).
Y. Shtanov, J. Traschen and R. Brandenberger, Phys. Rev. D [**51**]{}, 5438 (1995).
D. Boyanovsky, H. J. de Vega, R. Holman, D. S. Lee, and A. Singh, Phys. Rev. D [**51**]{}, 4419 (1995); D. Boyanovsky, M. D’Attanasio, H. J. de Vega, R. Holman, D. S. Lee, and A. Singh, [*ibid*]{} [**52**]{}, 6805 (1995); D. Boyanovsky, H. J. de Vega, R. Holman, D. S. Lee, A. Singh, and J. F. Salgado, [*ibid*]{} [**54**]{}, 7570 (1996).
B. Basset, D. Kaiser, and R. Maartens, Phys. Lett. [**B455**]{}, 84 (1999); B. Bassett, F. Tamburini, D. Kaiser, and R. Maartens, Nucl. Phys. [**B561**]{}, 188 (1999); B. Bassett, C. Gordon, R. Maartens, and D. Kaiser, Phys. Rev. D [**61**]{}, 061302 (2000); F. Finelli and R. Brandenberger, hep-ph/0003172.
B. A. Bassett and F. Viniegra, Phys. Rev. D [**62**]{}, 043507 (2000); J. P. Zibin, R. Brandenberger, and Douglas Scott, hep-ph/0007219.
S. E. Joras and V. H. Cárdenas, in preparation.
F. Cooper, J. Dawson, D. Meredith, and H. Shepard, Phys. Rev. Lett. [**72**]{}, 1337 (1994); F. Cooper, J. Dawson, S. Habib and R. D. Ryne, Phys. Rev. E [**57**]{}, 1489 (1998).
L. D. Landau y E. M. Lifshitz, [*Mecánica*]{}, Reverté, Barcelona (1970).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In light of the experimental evidence for the existence of a superfluidity region in a Bose–condensed sodium gas a theoretical model is put forward. It will be shown that the predictions of the present work do match with the extant measurement readouts. As a byproduct we also calculate the speed of sound and compare it against the current experimental results.'
author:
- 'A. Camacho'
title: 'Landua Criterion in a Bose–Condensed Sodium Gas'
---
The first experimental evidence concerning the phenomenon of superfluidity can be tracked down to the work of Kamerling Onnes in the year of 1911 with helium when it was found that if cooled below $2.2 ~^{\circ}K$ $He$ did not contract but rather expand [@Onnes]. Since then the amount of theoretical and experimental work has been able to provide a coherent picture to the subjacent Physics [@Khalatnikov; @Nozieres; @Griffin]. Another phenomenon, which also emerged in the last century, is the Bose–Einstein Condensation (BEC) the one, as in the case of superfluidity, is inherently related to the presence of very low temperatures. The need for low temperatures for the appearance of these two effects leads us to the obvious question concerning a possible relation between them. It was Fritz London [@London] who put forward the idea of a connection between these two effects stating that the transition from $He~I$ (the high temperature phase of liquid helium) and $He~II$ (the low temperature phase) should be considered an example of a BEC. Even more, he suggested that in $He~II$ a macroscopic quantum current of matter could be present, i.e., he introduced the idea that BEC and superfluidity could appear, simultaneously, in a system.
It has to be stressed that this understanding cannot be considered a closed issue [@Yukalov], and the answer to this aforementioned interrogant is that, though there is a close relationship, it is not a unique one. Indeed, we may state that BEC is neither necessary nor sufficient for the presence of superfluidity. For instance, and ideal Bose–Einstein condensate shows no superfluidity, and, as a counterpart a two–dimensional superfluid cannot condensate [@Ueda].
Landau (within the two–fluid model proposed by Tisza [@Tisza]) introduced the concept of elementary excitations [@Landau1] as a fundamental element in the description of the behavior of $He~II$. Landau argued [@Landau2] that the normal fluid (the non–superfluid component) could be regarded as a dilute gas whose components are weakly–interacting elementary excitations which move in a background defined by the superfluid component. Accordingly, the phenomenon of superfluidity appears if the velocity of the corresponding flow lies below a certain threshold value given by
$$v_{(crit)}= \min\Bigl(\frac{\epsilon(p)}{p}\Bigr). \label{Landau}$$
Here $\epsilon(p)$ denotes the energy of an elementary excitation and $p$ its corresponding momentum. If the velocity is larger, then the microscopic rugosities of the walls of the container will scatter the particles of the fluid entailing the loss of kinetic energy of the fluid, i.e., viscosity appears.
In the experimental realm the quest for this critical velocity has been carried out in a sodium–BEC, and the results show a possible velocity threshold located around the value of $1.6mm/s$ [@Raman]. The use of sodium–condensed gas in the experimental context is related not only to the aforementioned case but also to the excitation of phonons by light scattering [@Stamper] and the propagation of sound [@Andrews; @Andrews1].
The main purpose of the present work is to obtain a prediction for the critical velocity for a BEC. The deduced model will be compared against the reported measurement readouts [@Raman], and, in addition, the speed of sound for a sodium–condensed gas will be found and compared with the current experimental results [@Andrews; @Andrews1].
From a fundamental point of view our mathematical model can be defined by an $N$–particle Hamiltonian the one in the formalism of second quantization is [@Ueda]
$$\begin{aligned}
\hat{H}= \int d\vec{r}\Bigl[-\hat{\psi}^{\dagger}(\vec{r}, t)
\frac{\hbar^2}{2m}\nabla^2\hat{\psi}(\vec{r}, t)\nonumber\\
+ V(\vec{r})\hat{\psi}^{\dagger}(\vec{r}, t)\hat{\psi}(\vec{r}, t)\nonumber\\
+\frac{U_0}{2}\hat{\psi}^{\dagger}(\vec{r},
t)\hat{\psi}^{\dagger}(\vec{r}, t) \hat{\psi}(\vec{r},
t)\hat{\psi}(\vec{r}, t)\Bigr]. \label{Hamilton1}\end{aligned}$$
In this Hamiltonian $\hat{\psi}^{\dagger}(\vec{r}, t)$ and $\hat{\psi}(\vec{r}, t)$ represent bosonic creation and annihilation operators, respectively. It is restricted to low energies and momenta and implies, as a consequence of the aforementioned conditions, that the interaction among the particles is, as usual, codified by the scattering length parameter $a$, i.e., $U_0=\frac{4\pi a\hbar^2}{m}$. The trapping potential $V(\vec{r})$, for our case, corresponds to an isotropic harmonic oscillator whose frequency reads $\omega$. In addition, there are $N$ particles in the gas, each of them with mass $m$, the volume occupied by the system is $V$.
Our mathematical assumptions are:
\(i) Only the ground and the first excited states are populated. This condition can be justified recalling that for a bosonic system, with chemical potential $\mu$ and energy levels of single–particle $\epsilon$, the occupation number in thermal equilibrium is given by [@Pethick] ($\beta=1/(\kappa T$))
$$\begin{aligned}
<n_{(\epsilon)}>=\frac{1}{e^{(\epsilon-\mu)\beta}-1}.\label{Occunumber}\end{aligned}$$
It is readily seen that we deal with a monotonic decreasing function of $\epsilon$, and this feature justifies the present assumption.
\(ii) The mathematical description of the two occupied states will be done resorting to the Hartree approximation, in which the ground state of the interacting system is deduced by a Ginzburg–Pitaevski–Gross energy functional [@Gross1], and it entails that the ground state wavefunction corresponds to the case of a harmonic oscillator situation but the frequency is modified due to the fact that the system has a non–vanishing scattering length [@Baym], such that the fundamental length parameter reads.
$$R= \Bigl(\frac{2}{\pi}\Bigr)^{1/10}\Bigl(\frac{Na}{l}\Bigr)^{4/5}l.
\label{Radius1}$$
Here $l$ is the radius related to the trap given by the isotropic harmonic oscillator, namely,
$$l= \sqrt{\frac{\hbar}{m\omega}}. \label{Radius2}$$
Clearly this condition implies an effective frequency
$$\tilde{\omega}= \frac{\hbar}{mR^2}. \label{Freq1}$$
Usually the experimental conditions entail $R>l$ [@Raman] and, in consequence, $\tilde{\omega}<\omega$.
In other words, the order parameter related to the particles in the ground state is provided by
$$\begin{aligned}
\psi_{(0)}(\vec{r})=\sqrt{\frac{N_{(0)}}
{(R\sqrt{\pi})^3}}\exp{\Bigl[-\frac{r^2}{2R^2}\Bigr]}.\label{Groustate}\end{aligned}$$
In this last expression $N_{(0)}$ denotes the number of particles in the lowest energy state. The presence of a non–vanishing scattering length entails that in the ground state not all the particles can have zero–momentum, the reason for this lies in the fact that the two–body interaction mixes in components with atoms in other states [@Pethick] and
$$\begin{aligned}
N_{(0)}=N\Bigl[1-\frac{8}{3}\sqrt{\frac{Na}{\pi
V}}\Bigr].\label{Depletion}\end{aligned}$$
Clearly,
$$\begin{aligned}
N_{(0)}= \int\bigl(\psi_{(0)}(\vec{r})\bigr)^2d^3r,\label{Norm1}\end{aligned}$$
$$\begin{aligned}
V=\frac{4\pi}{3}R^3.\label{Volume1}\end{aligned}$$
The wavefunction of the first excited state will be considered as the first excited state of an isotropic oscillator related to frequency given by (\[Freq1\]) and, due to our symmetry, it has three possibilities, all with the same mathematical structure, namely,
$$\begin{aligned}
\psi^{(i)}_{(1)}(\vec{r})=\frac{8}{\sqrt{27\pi}}\sqrt{\frac{N}{V}\sqrt{\frac{Na}{\pi
V}}}\frac{x^{(i)}}{R}\exp{\Bigl[-\frac{r^2}{2R^2}\Bigr]}.\label{Excstate}\end{aligned}$$
Here $x^{(1)}=x$, $x^{(2)}=y$, and $x^{(3)}=z$.
Of course, (\[Excstate\]) must be related to the total number of particles in excited states ($N_{(e)}=\frac{8}{3}\sqrt{\frac{Na}{\pi
V}}$), a condition that becomes [@Pethick]
$$\begin{aligned}
N_{(e)}=\int\Bigl[\sum_{i=1}^{3}
\bigl(\psi^{(i)}_{(1)}(\vec{r})\bigr)^2\Bigr]d^3r.\label{Norm2}\end{aligned}$$
Having stated our assumptions we proceed to compute the speed of sound and the critical velocity. The energy of the ground state in its three possibilities, i.e., kinetic energy, due to the trap, and interaction are [@Baym; @Pethick]
$$\begin{aligned}
\frac{\hbar^2}{2m}\int\Bigl(\nabla\psi_{(0)}(\vec{r})\Bigr)^2d^3r=\frac{3\hbar^2}{4mR^2}N_{(0)}.\label{Kinen}\end{aligned}$$
$$\begin{aligned}
\int V(\vec{r})\psi_{(0)}(\vec{r})d^3r=
\frac{3}{4}m\omega^2R^2N_{(0)}.\label{Trapen}\end{aligned}$$
$$\begin{aligned}
\frac{U_0}{2}\int\Bigl(\psi_{(0)}(\vec{r})\Bigr)^4d^3r=
\frac{U_0}{\sqrt{32\pi^3}R^3}N^2_{(0)}.\label{Inten}\end{aligned}$$
The energy of the ground state (here denoted by $E_{(0)})$, no elementary excitations are present, is the sum of the last three expressions. The corresponding pressure ($P_{(0)}=-\partial
E_{(0)}/\partial V$) is
$$\begin{aligned}
P_{(0)}=\frac{4\pi\hbar^2
N}{mV^{5/3}}\Bigl\{\frac{1}{8\pi}\bigl(\frac{4\pi}{3}\bigr)^{2/3}\Bigl[1+\frac{3}{4}N_{(e)}\Bigr]\nonumber\\
+\Bigl[\frac{1}{\sqrt{18\pi}}\bigl(1-2N_{(e)}\bigr)\bigl(1-N_{(e)}\bigr)\nonumber\\
+\frac{\sqrt{2\pi}}{4\pi}\bigl(1-\frac{3}{2}N_{(e)}\bigr)
\Bigr]\frac{Na}{V^{1/3}} \Bigr\}.\label{Pressure0}\end{aligned}$$
If $v_{(0)}$ denotes the speed of sound related to the last expression ($v^2_{(0)}=-(V^2/mN)\partial P_{(0)}/\partial V$) we are led to
$$\begin{aligned}
v^2_{(0)}=\frac{\hbar^2}{m^2}\frac{4\pi
Na}{V}\Bigl\{\sqrt{\frac{2}{9\pi}}\Bigl[1-\frac{15}{4}N_{(e)}
+4N^2_{(e)} \Bigr]\nonumber\\
+\frac{1}{\sqrt{2\pi}}\Bigl[1-\frac{15}{8}N_{(e)}\Bigr] +
\frac{5}{12\pi}\bigl(\frac{4\pi}{3}\bigr)^{2/3}\frac{V^{1/3}}{Na}\bigl(1\nonumber\\
+\frac{39}{40}N_{(e)}\bigr)\Bigr\}.\label{Sound0}\end{aligned}$$
For the case of sodium [@Andrews; @Andrews1] the physical parameters are: $m=36.8\times 10^{-27}kg$, $a=2.75\times 10^{-9}m$, $N=5\times 10^6$. Considering one of the reported peak densities of the condensate, namely, $1\times 10^{14}cm^{-3}$, we obtain, in the roughest approximation from our calculations, $6.59~mm/s$. Clearly, our prediction is in good agreement with the experiment [@Andrews].
We now proceed to compute the lowest energy and momentum of the elementary excitations, physical parameters required for the deduction of the critical velocity [@Nozieres]. The deduction of the energy of an elementary excitation and of its corresponding momentum requires the knowledge of the energy of a single–particle in the first excited state [@Ueda]. Our assumptions entail that the thermal cloud contains particles subject to an isotropic harmonic oscillator whose frequency is (\[Freq1\]) therefore the energy of an excited particle is given by this assumption and easily calculated as a function of the effective frequency of our variational procedure
$$\begin{aligned}
\tilde{\epsilon}= \frac{5}{2}\hbar\tilde{\omega}.\label{excene1}\end{aligned}$$
According to Bogoliubov [@Ueda; @Bogoliubov] the energy of an elementary excitation, here denoted by $\epsilon$, is a function of the energy of the excited particles of the BEC, namely,
$$\begin{aligned}
\epsilon=\sum\sqrt{(\tilde{\epsilon})^2+\frac{2NU_{(0)}}{V}\tilde{\epsilon}}.\label{excene2}\end{aligned}$$
The energy of all the elementary excitations turns out to be [@Ueda; @Bogoliubov]
$$\begin{aligned}
\tilde{E}=\sum\sqrt{(\tilde{\epsilon})^2+\frac{2NU_{(0)}}{V}\tilde{\epsilon}}<\tilde{n}_{\epsilon}>.\label{Excene2}\end{aligned}$$
Here $<\tilde{n}_{\epsilon}>$ denotes the occupation number of the elementary excitations with energy ${\epsilon}$. The relation between the occupation numbers of particles and elementary excitations is [@Ueda]
$$\begin{aligned}
<\tilde{n}_{\epsilon}>=
\frac{<n_{\epsilon}>}{1+<n_{\epsilon}>}.\label{Occunumber3}\end{aligned}$$
At this point, for the sake of simplicity, we resort to the experimental values related to the detection of a critical velocity in a sodium condensed gas [@Raman] in which the occupation number of the particles in the first excited state fulfill the condition $N_{(e)}\sim 10^2>1$, and, in consequence, $<\tilde{n}_{\epsilon_{(1)}}>=1$. Our assumptions imply $<\tilde{n}_{\epsilon_{(i)}}>=0, ~~\forall i>1$. Indeed, we have considered that the thermal cloud is comprised by particles which occupy only the first excited state, in other words, $<n_{\epsilon_{(i)}}>=0,~~\forall i>1$. Introducing this condition into (\[Occunumber3\]) leads us to the aforementioned result for the occupation number of the elementary excitations.
Casting (\[excene2\]) in terms of the effective volume $V=4\pi
R^3/3$, and using (\[Radius1\]), (\[Radius2\]), and (\[Freq1\]), we have that the energy of our elementary excitation is
$$\begin{aligned}
\epsilon=\Bigl(\frac{4\pi}{3}\Bigr)^{1/3}\frac{\hbar^2}{mV^{2/3}}\sqrt{\frac{25}{4}\Bigl(\frac{4\pi}{3}\Bigr)^{2/3}+
\frac{20\pi Na}{V^{1/3}}}.\label{excene3}\end{aligned}$$
We must now find the momentum of this elementary excitation. Elementary excitations, which define the normal component of the fluid fluid, can be regarded as a bosonic gas whose components are weakly–interacting and moving in a region in which a constant potential exists, and this potential is defined by a mean field approach [@Ueda]. According to this interpretation we may rewrite (\[excene3\]) in the same form as in the case in which our BEC is a homogeneous one [@Ueda]. In other words, we cast our last expression in the following form
$$\begin{aligned}
\epsilon=\frac{\hbar^2k}{2m}\sqrt{k^2+ \frac{16\pi
Na}{V}}.\label{excene4}\end{aligned}$$
Clearly, (\[excene4\]) allows us to deduce the wavenumber related to our elementary excitation and, in consequence, its momentum. Indeed, we have for these two physical variables, respectively, that
$$\begin{aligned}
k=\Bigl(\frac{4\pi}{3}\Bigr)^{1/3}\sqrt{5}\frac{1}{V^{1/3}}
,\label{elemwavenum}\end{aligned}$$
$$\begin{aligned}
p=\Bigl(\frac{4\pi}{3}\Bigr)^{1/3}\sqrt{5}\frac{\hbar}{V^{1/3}}
.\label{elemmom}\end{aligned}$$
Resorting to Landau criterion (\[Landau\]) we obtain that the critical velocity is given by
$$\begin{aligned}
v_{(crit)}=\frac{1}{\sqrt{5}}\frac{\hbar}{mV^{1/3}}\sqrt{\frac{25}{4}\Bigl(\frac{4\pi}{3}\Bigr)^{2/3}+
\frac{20\pi Na}{V^{1/3}}}.\label{crit2}\end{aligned}$$
The experimental parameters [@Raman] are a critical speed of $v^{(e)}_{(crit)}= 1.6~mm/s$. In addition, the number of particles in this experiment has a minimum of $N=3\times 10^6$ and a maximum of $N=12\times 10^6$, and for the evaluation of our expression we will take the arithmetic average, i.e., $N=7.5\times 10^6$. The effective volume is that of an ellipsoid whose axes are $l_1=45\times 10^{-6}m$ and $l_1=150\times 10^{-6}m$ such that $V=\frac{4\pi}{3}l^2_1l_2$.
Introducing these values into (\[crit2\]) entails
$$\begin{aligned}
v^{(m)}=1.95~mm/s.\label{crit3}\end{aligned}$$
The reported critical speed is [@Raman]
$$\begin{aligned}
v^{(e)}= 1.6~mm/s\end{aligned}$$
The ensuing error is less that 18 percent $$\begin{aligned}
\vert v^{(e)}-v^{(m)}\vert/\Bigl(v^{(m)}\Bigr)=0.179.\label{error1}\end{aligned}$$
Since the number of particles in the corresponding experiment varies from $N=3\times 10^6$ to $N=12\times 10^6$ [@Raman] the associated values for the critical speed go from $1.24~mm/s$ to $2.48~mm/s$. If $N=5\times 10^6$, then $v^{(m)}= v^{(e)}$.
In conclusion, we have put forward a theoretical model for the deduction of the critical velocity in a sodium–condensed gas. This threshold speed has been computed and compared against the extant experimental results, having a good agreement between them. A more precise evaluation of the present idea requires a better knowledge of the value of $N$ employed in the experiment. Previous works offer larger critical velocities which have a bigger error than the one here deduced [@Raman; @Frisch], when compared to the experimental result.
H. Kamerling Onnes, Proc. Roy. Acad. Amsterdam. textbf[13]{}, 1903 (1911).
I. M. Khalatnikov, *Introduction to the Theory of Superfluidity* (Benjamin, New York, 1965).
P. Nozièrez and D. Pine, *The Theory of Quantum Liquids.* (Addison–Wesley, Reading MA, 1989 Vol. II).
A. Griffin, T. Nikuni, and E. Zaremba, *Bose–Condensed Gases at Finite Temperatures* (Cambridge University Press, Cambridge, 2009).
F. LOndon, Nature textbf[141]{}, 643 (1938).
V. I. Yukalov, e–print cond–mat/0408599v1.
M. Ueda, *Fundamentals and New Frontiers of Bose–Einstein Condensation* (World Scientific Publishing Co, Singapore, 2010).
L. Tisza, Phys. Rev. **72**, 838 (1947).
L. D. Landau, J. Phys. USSR. **5**, 71 (1941).
L. D. Landau, J. Phys. USSR. **11**, 91 (1947).
C. Raman et al, Phys. Rev. Lett. **83**, 2502 (1999).
D. M. Stamper et al, Phys. Rev. Lett. **83**, 2876 (1999).
M. R. Andrews et al, Phys. Rev. Lett. **79**, 553 (1997).
M. R. Andrews et al, Phys. Rev. Lett. **80**, 2967 (1997).
C. J. Pethick and H. SMith, *Bose–Einstein Condensation in Dilute Gases* (Cambridge University Press, Cambridge, 2008).
E. P. Gross, J. Math. Phys. **4**, 195 (1963).
G. Baym and C. J. Pethick, Phys. Rev. Lett. **76**, 6 (1996).
N. N. Bogoliubov, J. Phys. USSR. **11**, 23 (1947).
T. Frisch, Y. Pomeau, and S. Rica, Phys. Rev. Lett. **69**, 1644 (1992).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We propose a feasible experimental scheme to realize a three-dimensional chiral topological insulator with cold fermionic atoms in an optical lattice, which is characterized by an integer topological invariant distinct from the conventional $\mathbb{Z}_{2}$ topological insulators and has a remarkable macroscopic zero-energy flat band. To probe its property, we show that its characteristic surface states — the Dirac cones — can be probed through time-of-flight imaging or Bragg spectroscopy and the flat band can be detected via measurement of the atomic density profile in a weak global trap. The realization of this novel topological phase with a flat band in an optical lattice will provide a unique experimental platform to study the interplay between interaction and topology and open new avenues for application of topological states.'
author:
- 'S.-T. Wang, D.-L. Deng, and L.-M. Duan'
title: 'Probe of Three-Dimensional Chiral Topological Insulators in an Optical Lattice'
---
The exploration of topological phases of matter has become a major theme at the frontiers of condensed matter physics since the discovery of topological insulators (TIs) [@hasan2010colloquium; @*Qi:2011wt; @*moore2010birth]. The TIs are band insulators with peculiar topological properties that are protected by time reversal symmetry. A recent remarkable theoretical advance is the finding that there are various other kinds of topological phases of free fermions apart from the conventional TIs, which can be classified by a periodic table according to system symmetry and dimensionality [Schnyder:2008ez,\*kitaev2009periodic]{}. An important question then is whether the new topological phases predicted by the periodic table can be physically realized. Several model Hamiltonians have been proposed to have the predicted topological phases as their ground states [@Ryu:2010ko; @neupert2012noncommutative; @qi2009time; @Schnyder:2009he; @moore2008topological; @*Deng:2013fe]. However, these model Hamiltonians typically require complicated spin-orbital couplings that are hard to be realized in real materials. Implementations of these model Hamiltonians still remain very challenging for experiments.
In this Letter, we propose an experimental scheme to realize a three-dimensional (3D) chiral TI with cold fermionic atoms in an optical lattice. The chiral TI is protected by the chiral symmetry, also known as the sublattice symmetry [@Ryu:2010ko; @Hosur:2010ie; @essin2012antiferromagnetic]. Unlike the conventional TIs protected by the time reversal symmetry, which is characterized by a $\mathbb{Z}_{2}$ topological invariant, the chiral TI is characterized by a topological invariant taking arbitrary integer values [@Schnyder:2008ez; @*kitaev2009periodic]. By controlling the spin-orbital coupling of cold fermionic atoms in a tilted optical lattice based on the Raman-assisted hopping [@jaksch2003creation; @Miyake:2013jw; @Aidelsburger:2013ew], we realize a tight-binding model Hamiltonian first proposed in Ref. [@neupert2012noncommutative], which supports a chiral TI with a zero-energy flat band. In such a flat band, the kinetic energy is suppressed and the atomic interaction, which can be tuned by the Feshbach resonance technique [chin2010feshbach]{}, will lead to a novel nonperturbative effect. In a cold atom experiment, flat bands have been studied in a 2D frustrated Kagome lattice [@jo2012ultracold]. Inspired by the discovery of the fractional quantum Hall effect in a topologically nontrivial flat-band Landau level, one expects that the atomic interaction in a flat-band TI may lead to exciting new physics [tang2011high,\*neupert2011fractional,\*sun2011nearly]{}. To probe the properties of the chiral TI in our proposed realization, we show that topological phase transition and the characteristic surface states of the TIs, the Dirac cones, can both be detected by mapping out the Fermi surface structure through time-of-flight imaging [@Spielman2007Mott; @*Kashurnikov2002Revealing; @kohl2005fermionic] or Bragg spectroscopy [@stamper1999excitation]. Furthermore, we show that the flat band can be verified by measurement of the atomic density profile under a weak global harmonic trap [@zhu2007simulation; @Schneider05122008].
We consider realization of the following tight-binding model Hamiltonian in the momentum space [@neupert2012noncommutative]$$\mathcal{H}(\mathbf{k})=\left(
\begin{array}{ccc}
0 & 0 & q_{1}-iq_{2} \\
0 & 0 & q_{3}-iq_{0} \\
q_{1}+iq_{2} & q_{3}+iq_{0} & 0\end{array}\right) , \label{Eqn:Ham}$$with $q_{0}=2t\left( h+\cos k_{x}a+\cos k_{y}a+\cos k_{z}a\right)
$, $q_{1}=2t\sin k_{x}a$, $q_{2}=2t\sin k_{y}a$, $q_{3}=2t\sin k_{z}a$, where $\mathbf{k=}\left( k_{x},k_{y},k_{z}\right) $ denotes the momentum, $a$ is the lattice constant, $t$ is the hopping energy, and $h$ is a dimensionless control parameter. This model Hamiltonian has a chiral symmetry represented by $S\mathcal{H}(\mathbf{k})S^{-1}=-\mathcal{H}(\mathbf{k})$ with the unitary matrix $S \equiv $diag$(1,1,-1)$. It has three bands, with a flat middle band exactly at zero energy protected by the chiral symmetry. The other two bands have energies $E_{\pm }(\mathbf{k})=\pm
2t[\sin ^{2}\left( k_{x}a\right) +\sin ^{2}\left( k_{y}a\right) +\sin
^{2}\left( k_{z}a\right) +(\cos k_{x}a+\cos k_{y}a+\cos k_{z}a+h)^{2}]^{1/2}$. The topological index for this model can be characterized by the integral [@neupert2012noncommutative; @deng2013systematic] $$\Gamma =\frac{1}{12\pi ^{2}}\int_{\text{BZ}}d\mathbf{k}\;\epsilon ^{\alpha
\beta \gamma \rho }\epsilon ^{\mu \nu \tau }\frac{1}{E_{+}^{4}}q_{\alpha
}\partial _{\mu }q_{\beta }\partial _{\nu }q_{\gamma }\partial _{\tau
}q_{\rho },
\label{Eqn:Index}$$where $\epsilon $ is the Levi-Civita symbol with $\left( \alpha ,\beta
,\gamma ,\rho \right) $ and $\left( \mu ,\nu ,\tau \right) $ taking values respectively from $\left\{ 0,1,2,3\right\} $ and $\left\{ k_{x},k_{y},k_{z}\right\} $.
To realize the model Hamiltonian , we consider interaction-free fermionic atoms in an optical lattice and choose three internal atomic states in the ground state manifold to carry three spin states ${| 1 \rangle},{| 2 \rangle},{| 3 \rangle}$. The other levels in the ground state manifold are irrelevant as they are initially depopulated by the optical pumping and transitions to these levels are forbidden during Raman-assisted atomic hopping because of a large energy detuning. The Hamiltonian , expressed in real space, has the following form $$\begin{aligned}
H&= t \sum_{\mathbf{r}}\left[ \left( 2ihc_{3,\mathbf{r}}^{\dag }c_{2,\mathbf{r}}+\text{H.c.}\right) +H_{\mathbf{rx}}+H_{\mathbf{ry}}+H_{\mathbf{rz}}\right] \notag, \\
H_{\mathbf{rx}}& =ic_{3,\mathbf{r-x}}^{\dag }(c_{1,\mathbf{r}}+c_{2,\mathbf{r}})-ic_{3,\mathbf{r+x}}^{\dag }(c_{1,\mathbf{r}}-c_{2,\mathbf{r}})+\text{H.c.,} \notag \\
H_{\mathbf{ry}}& =-c_{3,\mathbf{r-y}}^{\dag }(c_{1,\mathbf{r}}-ic_{2,\mathbf{r}})+c_{3,\mathbf{r+y}}^{\dag }(c_{1,\mathbf{r}}+ic_{2,\mathbf{r}})+\text{H.c.,} \notag \\
H_{\mathbf{rz}}& =2ic_{3,\mathbf{r-z}}^{\dag }c_{2,\mathbf{r}}+\text{H.c.},
\label{Eqn:Ham2}\end{aligned}$$where $\left( \mathbf{x,y,z}\right) $ represents a unit vector along the $\left( x,y,z\right) $-direction of a cubic lattice, and $c_{j,\mathbf{r}}$ $\left( j=1,2,3\right) $ denotes the annihilation operator of the fermionic mode at the lattice site $\mathbf{r}$ with the spin state ${| j \rangle}$. To implement this Hamiltonian, the major difficulty is to realize the spin-transferring hopping terms $H_{\mathbf{rx}},H_{\mathbf{ry}},H_{\mathbf{rz}}$ along each direction [@OpticalLatCTI:Sup]. The hopping terms and the associated spin transformation can be visualized diagrammatically as $$\begin{aligned}
\hspace{-0.2cm}x& \text{-}\text{direction:}\,{| 3 \rangle}\overset{i\sqrt{2}}{\curvearrowleft}{| 1_x \rangle}\overset{\times }{\curvearrowright }\;+\;\overset{\times }{\curvearrowleft }{| 2_x \rangle}\overset{-i\sqrt{2}}{\curvearrowright }{| 3 \rangle}\;+\;\text{H.c.} \notag \\
\hspace{-0.2cm}y& \text{-}\text{direction:}\,{| 3 \rangle}\overset{-\sqrt{2}}{\curvearrowleft }{| 1_y \rangle}\overset{\times }{\curvearrowright }\;+\;\overset{\times }{\curvearrowleft }{| 2_y \rangle}\overset{\sqrt{2}}{\curvearrowright }{| 3 \rangle}\;+\;\text{H.c.} \notag \\
\hspace{-0.2cm}z& \text{-}\text{direction:}\,{| 3 \rangle}\overset{2i}{\curvearrowleft }{| 2 \rangle}\overset{\times }{\curvearrowright }\;+\;\text{H.c.}
\label{Eqn:Hopping}\end{aligned}$$ where $\overset{\times }{\curvearrowright }$ indicates that hopping is forbidden along that direction, and ${| 1_x \rangle}=\left({| 1 \rangle}+{| 2 \rangle}\right)/\sqrt{2},{| 2_x \rangle}=({| 1 \rangle}-{| 2 \rangle})/\sqrt{2},{| 1_y \rangle}=({| 1 \rangle}-i{| 2 \rangle})/\sqrt{2},{| 2_y \rangle}=({| 1 \rangle}+i{| 2 \rangle})/\sqrt{2}$ are superpositions of the original spin-basis vectors ${| 1 \rangle},{| 2 \rangle},{| 3 \rangle}$.
We use Raman-assisted tunneling to achieve the spin-transferring hopping terms depicted in Eq. . Note that the parity (left-right) symmetry is explicitly broken by these hopping terms. To break the parity symmetry, we assume the optical lattice is tilted with a homogeneous energy gradient along the $x$-,$y$-,$z$-directions. This can be achieved, for instance, through the natural gravitational field, the magnetic field gradient, or the gradient of a dc- or ac-Stark shift [@jaksch2003creation; @Miyake:2013jw; @Aidelsburger:2013ew]. Raman-assisted hopping in a tilted optical lattice has been demonstrated in recent experiments [@Miyake:2013jw; @Aidelsburger:2013ew]. In our scheme, we require a different linear energy shift per site $\Delta _{x,y,z}$ along the $x$-,$y$-,$z$-directions. In particular, we take $\Delta _{z}\approx 1.5\Delta _{y}\approx 3\Delta _{x}$ with the energy difference lower bounded by $\Delta _{x}$, and assume the natural tunneling rate $t_{0}\ll \Delta _{x}$ so that the hopping probability $\left(
t_{0}/\Delta _{x}\right) ^{2}$ induced by the natural tunneling is negligible in this tilted lattice.
\(a) laser configuration\
![Schematics of the laser configuration to realize the Hamiltonian in Eq. . Panel (a) shows the propagation direction (big arrows) and the polarization (small arrows) of each laser beam. (b) A linear tilt $\Delta_{x,y,z}$ per site in the lattice along each direction. The detuning in each direction matches the frequency offset of the corresponding Raman beams, which are shown in panels (c), (d), and (e). Polarizations of each beam are shown in brackets. Rabi frequencies for each beam are: $\Omega _{1}^{\pi }=\Omega _{0}e^{ikx}, \Omega _{2}^{\pi}=\Omega _{0}e^{iky}, \Omega _{1}^{x}=i\sqrt{2}\Omega _{0}e^{ikz}, \Omega _{2}^{x}=-i\sqrt{2}\Omega _{0}e^{ikz}, \Omega _{1}^{y} =-\sqrt{2}\Omega _{0}e^{ikz}, \Omega _{2}^{y}=\sqrt{2}\Omega _{0}e^{ikz}, \Omega ^{z} =2i\Omega _{0}e^{ikz}$. [@OpticalLatCTI:Sup][]{data-label="Fig:OpticalLattice"}](Laser "fig:"){width="40.00000%"}\
(b) tilted optical lattice (c) $x$-direction\
![Schematics of the laser configuration to realize the Hamiltonian in Eq. . Panel (a) shows the propagation direction (big arrows) and the polarization (small arrows) of each laser beam. (b) A linear tilt $\Delta_{x,y,z}$ per site in the lattice along each direction. The detuning in each direction matches the frequency offset of the corresponding Raman beams, which are shown in panels (c), (d), and (e). Polarizations of each beam are shown in brackets. Rabi frequencies for each beam are: $\Omega _{1}^{\pi }=\Omega _{0}e^{ikx}, \Omega _{2}^{\pi}=\Omega _{0}e^{iky}, \Omega _{1}^{x}=i\sqrt{2}\Omega _{0}e^{ikz}, \Omega _{2}^{x}=-i\sqrt{2}\Omega _{0}e^{ikz}, \Omega _{1}^{y} =-\sqrt{2}\Omega _{0}e^{ikz}, \Omega _{2}^{y}=\sqrt{2}\Omega _{0}e^{ikz}, \Omega ^{z} =2i\Omega _{0}e^{ikz}$. [@OpticalLatCTI:Sup][]{data-label="Fig:OpticalLattice"}](Potential.eps "fig:"){width="18.00000%"} ![Schematics of the laser configuration to realize the Hamiltonian in Eq. . Panel (a) shows the propagation direction (big arrows) and the polarization (small arrows) of each laser beam. (b) A linear tilt $\Delta_{x,y,z}$ per site in the lattice along each direction. The detuning in each direction matches the frequency offset of the corresponding Raman beams, which are shown in panels (c), (d), and (e). Polarizations of each beam are shown in brackets. Rabi frequencies for each beam are: $\Omega _{1}^{\pi }=\Omega _{0}e^{ikx}, \Omega _{2}^{\pi}=\Omega _{0}e^{iky}, \Omega _{1}^{x}=i\sqrt{2}\Omega _{0}e^{ikz}, \Omega _{2}^{x}=-i\sqrt{2}\Omega _{0}e^{ikz}, \Omega _{1}^{y} =-\sqrt{2}\Omega _{0}e^{ikz}, \Omega _{2}^{y}=\sqrt{2}\Omega _{0}e^{ikz}, \Omega ^{z} =2i\Omega _{0}e^{ikz}$. [@OpticalLatCTI:Sup][]{data-label="Fig:OpticalLattice"}](OpticalLatticeX.eps "fig:"){width="22.00000%"}\
(d) $y$-direction (e) $z$-direction\
![Schematics of the laser configuration to realize the Hamiltonian in Eq. . Panel (a) shows the propagation direction (big arrows) and the polarization (small arrows) of each laser beam. (b) A linear tilt $\Delta_{x,y,z}$ per site in the lattice along each direction. The detuning in each direction matches the frequency offset of the corresponding Raman beams, which are shown in panels (c), (d), and (e). Polarizations of each beam are shown in brackets. Rabi frequencies for each beam are: $\Omega _{1}^{\pi }=\Omega _{0}e^{ikx}, \Omega _{2}^{\pi}=\Omega _{0}e^{iky}, \Omega _{1}^{x}=i\sqrt{2}\Omega _{0}e^{ikz}, \Omega _{2}^{x}=-i\sqrt{2}\Omega _{0}e^{ikz}, \Omega _{1}^{y} =-\sqrt{2}\Omega _{0}e^{ikz}, \Omega _{2}^{y}=\sqrt{2}\Omega _{0}e^{ikz}, \Omega ^{z} =2i\Omega _{0}e^{ikz}$. [@OpticalLatCTI:Sup][]{data-label="Fig:OpticalLattice"}](OpticalLatticeY.eps "fig:"){width="21.00000%"} ![Schematics of the laser configuration to realize the Hamiltonian in Eq. . Panel (a) shows the propagation direction (big arrows) and the polarization (small arrows) of each laser beam. (b) A linear tilt $\Delta_{x,y,z}$ per site in the lattice along each direction. The detuning in each direction matches the frequency offset of the corresponding Raman beams, which are shown in panels (c), (d), and (e). Polarizations of each beam are shown in brackets. Rabi frequencies for each beam are: $\Omega _{1}^{\pi }=\Omega _{0}e^{ikx}, \Omega _{2}^{\pi}=\Omega _{0}e^{iky}, \Omega _{1}^{x}=i\sqrt{2}\Omega _{0}e^{ikz}, \Omega _{2}^{x}=-i\sqrt{2}\Omega _{0}e^{ikz}, \Omega _{1}^{y} =-\sqrt{2}\Omega _{0}e^{ikz}, \Omega _{2}^{y}=\sqrt{2}\Omega _{0}e^{ikz}, \Omega ^{z} =2i\Omega _{0}e^{ikz}$. [@OpticalLatCTI:Sup][]{data-label="Fig:OpticalLattice"}](OpticalLatticeZ.eps "fig:"){width="21.00000%"}
To realize the hopping terms in Eq. , we apply two-photon Raman transitions with the configuration (polarization and propagating direction) of the laser beams depicted in Fig. [Fig:OpticalLattice]{} [@OpticalLatCTI:Sup]. The internal states ${| 1 \rangle},{| 3 \rangle}, {| 2 \rangle}$ differ in the magnetic quantum number $m$ by one successively so that the atomic addressing can be achieved using polarization selection. The $\pi$-polarized lights consist of two laser beams $\Omega _{1}^{\pi }=\Omega _{0}e^{ikx}$ and $\Omega _{2}^{\pi }=\Omega
_{0}e^{iky}$, propagating along the $x$ and $y$ directions respectively, where $k$ is the magnitude of the laser wave vector. The other five beams $\Omega _{1,2}^{x,y,z}$ are all propagating along the $z$ direction and the polarizations are shown in Fig. \[Fig:OpticalLattice\]. The Rabi frequencies $\Omega _{1,2}^{x,y,z}$, expressed in terms of the unit $\Omega
_{0}$, are given in the caption of Fig. \[Fig:OpticalLattice\] to produce the required phase and amplitude relations of the hopping terms in Eq. . Between the sites $\mathbf{r}$ and $\mathbf{r+m}$, the Raman-assisted hopping rate is given by $$t_{\mathbf{r,m}}=\dfrac{\Omega _{\beta \mathbf{m}}^{\ast }\Omega _{\alpha
\mathbf{m}}}{\delta }\int d^{3}\mathbf{r}^{\prime }w^{\ast }(\mathbf{r}^{\prime }-\mathbf{r-m})e^{i\delta \mathbf{k}\cdot \mathbf{r}^{\prime }}w(\mathbf{r}^{\prime }-\mathbf{r}),$$where $\delta $ is a large single-photon detuning to the excited state, $w(\mathbf{r}^{\prime }-\mathbf{r})$ is the Wannier-(Stark) function at the site $\mathbf{r}$ [@OpticalLatCTI:footnote1], and $\delta \mathbf{k=k}_{\alpha}-\mathbf{k}_{\beta}$ is the momentum difference between the relevant Raman beams with the corresponding single-photon Rabi frequencies $\Omega _{\alpha \mathbf{m}}$ and $\Omega_{\beta \mathbf{m}}$. Because of the fast decay of the Wannier function, we consider only the nearest-neighbor Raman-assisted hopping with $\mathbf{m=} \pm \mathbf{x},\pm \mathbf{y},\pm \mathbf{z}$. When $\delta \mathbf{k=0}$, we have $t_{\mathbf{r,m}}=0$ for any $\mathbf{m\neq 0}$ terms because of the orthogonality of Wannier functions. Let us take one of the tunneling terms along the $x$ direction ${| 3 \rangle}\overset{i\sqrt{2}}{\curvearrowleft }{| 1_x \rangle}$ as an example to explain the Raman-assisted hopping rate. The relevant Raman pair is $\Omega _{1}^{x}=i\sqrt{2}\Omega _{0}e^{ikz}$ and $\Omega _{1}^{\pi}=\Omega _{0}e^{ikx}$ in Fig. \[Fig:OpticalLattice\], so $\Omega _{\alpha \mathbf{m} }=i\sqrt{2}\Omega _{0}$ and $\Omega _{\beta \mathbf{m}}=\Omega _{0}$. The laser beam $\Omega _{1}^{x}$ has two frequency components, generated, e.g., by an electric optical modulator (EOM), which are resonant with the levels ${| 1 \rangle},{| 2 \rangle}$ respectively so that in the rotating frame the levels ${| 1 \rangle}$ and ${| 2 \rangle}$ are degenerate in energy. The beam $\Omega _{1}^{x}$ is polarized along the $x$ direction, so, together with $\Omega _{1}^{\pi }$, it couples the state ${| 1_x \rangle}$ to the state ${| 3 \rangle}$ through the two-photon transition. The two-photon detuning $\Delta _{x}$ is in resonance with the potential gradient along the $x$ direction so that the beams only induce the nearest-neighbor hopping from $\mathbf{r}$ to $\mathbf{r-x}$. Using factorization of the Wannier function $w(\mathbf{r}^{\prime
})=w(x^{\prime })w(y^{\prime })w(z^{\prime })$ in a cubic lattice, we find the hopping rate $t_{\mathbf{r,- x}}=i\sqrt{2}\beta \Omega _{\mathbf{R}}e^{i\delta
\mathbf{k}\cdot \mathbf{r}}$, where $\Omega _{\mathbf{R}}\equiv \left\vert
\Omega _{0}\right\vert ^{2}/\delta $ and $\beta \equiv \int dxw^{\ast
}(x+a)e^{-ikx}w(x)\int dyw^{\ast }(y)w(y)\int dzw^{\ast }(z)e^{ikz}w(z)$. For this hopping term, $\delta \mathbf{k}=(-k,0,k)$. Actually, for the beams shown in Fig. \[Fig:OpticalLattice\], any nonzero $\delta \mathbf{k}$ has the form $(\pm k,0,\mp k)$ or $(0,\pm k,\mp k)$, so the site dependent phase term can always be reduced to $e^{i\delta \mathbf{k}\cdot \mathbf{r}}=1$ if we take the lattice constant $a$ to satisfy the condition $ka=2\pi $ by adjusting the interfering angle of the lattice beams. Under this condition, all the hopping terms in Eq. are obtained through the laser beams shown in Fig. \[Fig:OpticalLattice\] with the hopping rate $t=\beta \Omega _{\mathbf{R}}$ [@OpticalLatCTI:Sup]. The on-site spin transferring term $hc_{3,\mathbf{r}}^{\dag }c_{2,\mathbf{r}}$ can be achieved through application of a simple radio-frequency (rf) field (or another copropagating Raman beam). The Raman beams $\Omega _{1,2}^{x,y,z}$ and $\Omega _{1,2}^{\pi}$ may also induce some on-site spin transferring terms, which can be similarly compensated (canceled) with additional rf fields.
Although the laser configuration illustrated in Fig. \[Fig:OpticalLattice\] involves several beams, all of them can be drawn from the same laser, with the small relative frequency shift induced by an acoustic optical modulator (AOM) or EOM. The absolute frequencies of these beams and their fluctuations are not important as long as we can lock the relative frequency differences, which can be well controlled with the driving rf fields of the AOMs and EOMs. To show that the proposed scheme is feasible with current technology, we give a parameter estimation for typical experiments. For instance, with $^{40}$K atoms of mass $m$ in an optical lattice with the lattice constant $a=2\pi
/k=764$ nm [@liu2013realization; @Wang2012Spin], gravity induces a potential gradient (per site) $\Delta =mga/\hbar\approx 2\pi \times 0.75\,$kHz. Gravity provides the gradients for free along three directions with an appropriate choice of the relative axes of the frame to satisfy $\Delta _{x}:\Delta _{y}:\Delta _{z}=1:2:3$ and $\Delta
=\sqrt{\Delta _{x}^{2}+\Delta _{y}^{2}+\Delta _{z}^{2}}$. We then have $\Delta _{x}\approx 2\pi \times 200\,$Hz. For a lattice with depth $V_{0}\approx 2.3E_{r}$, where $E_{r}=\hbar ^{2}k^{2}/2m$ is the recoil energy, the overlap ratio $\beta \approx 0.34$ and the natural tunneling rate $t_{0}/\hbar \approx 2\pi \times 50\,$Hz [@OpticalLatCTI:Sup]. For Raman beams with $\Omega _{0}/2\pi \approx 15\,$MHz and the single-photon detuning $\delta /2\pi \approx 1.7\,$THz [@liu2013realization], we have $\Omega _{\mathbf{R}}=\left\vert \Omega _{0}\right\vert ^{2}/\delta \approx 2\pi \times 120\,$Hz and the Raman-assisted hopping rate $t/\hbar \approx 2\pi \times 40\,$Hz. Apparently, the undesired off-resonant hopping probabilities, upper bounded by $t^{2}/\Delta _{x}^{2}$ or $t_{0}^{2}/\Delta
_{x}^{2}$, are less than $6\%$ and the effective spontaneous emission rate, estimated by $|\Omega _{0}/\delta |^{2}\Gamma _{s}$ ($\Gamma _{s}\approx 2\pi \times 6\,$MHz is the decay rate of the excited state), is negligible during the experimental time of the order of $10/t$.
\
(c)\
![(Color online) (a) The topological index $\Gamma$ as a function of the parameter $h$. (b) Energy dispersion for three bulk bands (surface plot) and surface states (mesh plot) at the boundary along the $z$ direction for $h=2$. (c) Quasimomentum distribution $\rho_{\text{cry}}(\mathbf{k})$ for various $h=0,0.5,1,1.5,2$ at a fixed chemical potential $\protect\mu/2t=-2$ [@OpticalLatCTI:Sup]. One hundreds layers are taken along the $z$ direction with open boundaries in (b) and (c).[]{data-label="fig:Index"}](MomDistPhase "fig:"){width="50.00000%"}
We now proceed to discuss detection methods to probe the exotic phases of the realized Hamiltonian. The topological index $\Gamma$ defined in Eq. \[Eqn:Index\] is shown in Fig. \[fig:Index\](a) under different values of $h$. The system is topologically nontrivial for $|h|<3$, and $\Gamma $ changes at $|h|=1,3$, indicating a topological quantum phase transition. We calculate the band structure numerically for a homogeneous system by keeping $x$ and $y$ directions in momentum space and $z$ direction in real space with open boundaries. Figure \[fig:Index\](b) shows the result, revealing the macroscopic flat band as well as the surface states with Dirac cones. Experimentally, the band structure can be probed by mapping out the crystal quasimomentum distribution $\rho_{\text{cry}}(\mathbf{k})$. By abruptly turning off the lattice potential, one could measure the momentum distribution $\rho(\mathbf{k})$, and the quasimomentum can then be extracted as $\rho_{\text{cry}}(\mathbf{k})=\rho(\mathbf{k})/|w(\mathbf{k})|^{2}$, where $w(\mathbf{k})$ is the Fourier transform of the Wannier function $w(\mathbf{r})$ [@Spielman2007Mott; @*Kashurnikov2002Revealing]. Here, we numerically calculate the crystal quasimomentum distribution, which can be used to track the topological phase transition \[Fig. \[fig:Index\](c)\]. At a fixed chemical potential, as one varies $h$ from $0$ to $2$, the quasimomentum distribution reshapes accordingly when the bulk gap closes and reopens and the number of surface Dirac cones changes from $2$ to $1$, indicated by a change of topology of the Fermi surface [@OpticalLatCTI:Sup].
Bragg spectroscopy is a complementary detection method to reveal the Dirac cone structure [@stamper1999excitation; @zhu2007simulation]. One could shine two laser beams at a certain angle to induce a Raman transition from an occupied spin state to another hyperfine level and focus them near the surface of the 3D atomic gas. The atomic transition rate can be measured, which is peaked when the momentum and energy conservation conditions are satisfied. By scanning the Raman frequency difference, one can map out the surface energy-momentum dispersion relation [@zhu2007simulation]. The surface Dirac cones, with their characteristic linear dispersion, can therefore be probed through Bragg spectroscopy.
So far, we considered a homogeneous system under a box-type trap at zero temperature. In a realistic experiment, finite temperature and a weak confining harmonic trap may introduce noise. To include these effects, an important element to consider is the size of the bulk gap. In our parameter regime, the minimum band gap from the top or bottom bulk band to the middle flat band is $2t=(2\pi \hbar) \times 80\,$Hz at $h=2$ \[Fig. \[fig:Index\](b)\], which corresponds to a temperature around $4 \,$nK. Direct cooling to subnanokelvin temperature is challenging but has been attained experimentally [@Leanhardt12092003; @*Medley2011Spin]. Parametric cooling based on adiabatic preparation can be used to further reduce the effective temperature of the system. With a band gap considerably larger than the probing Raman Rabi frequency, bulk contribution to the Bragg spectroscopy is negligible. In the following, we include the effect of a weak harmonic trap via the local density approximation (LDA) and consider the finite temperature effects to be minimal.
The characteristic flat band can be detected through measurement of the atomic density profile under the global harmonic trap [@bloch2005ultracold; @Schneider05122008]. Under the LDA, the local chemical potential of the system is $\mu (r)=\mu_{0}-m\omega ^{2}r^{2}/2$, where $\mu _{0}$ denotes the chemical potential at the center of a spherically harmonic trap with the potential $V(r)=m\omega^{2}r^{2}/2$. The local atomic density $n(r)$ is uniquely determined by $\mu(r)$, and $\mu _{0}$ is specified by the total atom number $N$ through $\int n(r)4\pi r^{2}\,dr=N$. The atomic density profile $n(r)$, which can be measured *in situ* in experiments [@Schneider05122008], is calculated and shown in Fig. \[fig:LDA\]. A steep fall or rise in $n(r)$ is a clear signature of a macroscopic flat band (horizontal arrows in Fig. \[fig:LDA\]). The plateaus at $1/3$ and $2/3$ fillings \[vertical arrows in Fig. \[fig:LDA\](a)\] reveal the corresponding band gap. At $h=1$, the plateaus vanish \[Fig. \[fig:LDA\](b)\]. The disappearance of the plateaus at this point indicates the phase transition where the band gap closes. In experiments, due to the finite spatial resolution, the detected signal may correspond to a locally averaged $n(r)$. The dashed lines show the local average density $\bar{n}_{i}=\sum_{j=-1}^{1}n_{i+j}/3$, averaged over a spherical shell of $3$ lattice sites. One can see that major features associated with the band gap and the flat band remain clearly visible even when the signal is blurred by the local spatial averaging.
In summary, we have proposed an experimental scheme to realize and probe a 3D chiral TI with a zero-energy flat band. The experimental realization of this model will mark an important advance in the ultracold atom simulation of topological phases.
We thank R. Ma, S.-L. Zhu, C.-J. Wu, K. Sun, and G. Ortiz for discussions. This work was supported by the NBRPC (973 Program) No. 2011CBA00300 (No. 2011CBA00302), the IARPA MUSIQC program, the ARO, and the AFOSR MURI program.
[35]{}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 [****, ()]{} [****, ()](\doibase 10.1103/RevModPhys.83.1057) @noop [****, ()]{} @noop [****, ()]{} in @noop [**]{}, Vol. () @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase
10.1103/PhysRevB.88.201105) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevLett.98.080404) [****, ()](\doibase 10.1103/PhysRevA.66.031601) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase
10.1126/science.1165449) [****, ()](\doibase 10.1103/PhysRevB.89.075126) @noop @noop [****, ()](\doibase 10.1103/PhysRevLett.112.086401) [****, ()](\doibase 10.1103/PhysRevLett.109.095301) [****, ()](\doibase
10.1126/science.1088827) [****, ()](\doibase
10.1103/PhysRevLett.106.195301) @noop [****, ()]{}
Supplemental Material: Probe of Three-Dimensional Chiral Topological Insulators in an Optical Lattice {#supplemental-material-probe-of-three-dimensional-chiral-topological-insulators-in-an-optical-lattice .unnumbered}
=====================================================================================================
> In this supplemental material, we provide more details on the realization scheme of the chiral topological insulator Hamiltonian. Details on the parameter estimation with Wannier functions and additional density of states plots are included.
Realization of the effective Hamiltonian
========================================
Our main result is to use Raman-assisted tunneling to realize the effective Hamiltonian of the chiral topological insulator given by $$\begin{aligned}
H&= t \sum_{\mathbf{r}}\left[ \left( 2ihc_{3,\mathbf{r}}^{\dag }c_{2,\mathbf{r}}+\text{H.c.}\right) +H_{\mathbf{rx}}+H_{\mathbf{ry}}+H_{\mathbf{rz}}\right] \notag, \\
H_{\mathbf{rx}}& =ic_{3,\mathbf{r-x}}^{\dag }(c_{1,\mathbf{r}}+c_{2,\mathbf{r}})-ic_{3,\mathbf{r+x}}^{\dag }(c_{1,\mathbf{r}}-c_{2,\mathbf{r}})+\text{H.c.,} \\
H_{\mathbf{ry}}& =-c_{3,\mathbf{r-y}}^{\dag }(c_{1,\mathbf{r}}-ic_{2,\mathbf{r}})+c_{3,\mathbf{r+y}}^{\dag }(c_{1,\mathbf{r}}+ic_{2,\mathbf{r}})+\text{H.c.,} \notag \\
H_{\mathbf{rz}}& =2ic_{3,\mathbf{r-z}}^{\dag }c_{2,\mathbf{r}}+\text{H.c.} \notag
\label{EqnSup:Ham2}\end{aligned}$$In the following, we provide some complementary details on the realization scheme. The major difficulty is to realize the spin-transferring hopping terms $H_{\mathbf{rx}},H_{\mathbf{ry}},H_{\mathbf{rz}}$ along each direction. Let us focus on a single term first, $H^{(1)}_{\mathbf{rx}}=ic_{3,\mathbf{r-x}}^{\dag }(c_{1,\mathbf{r}}+c_{2,\mathbf{r}})$. This corresponds to an atom in the spin state ${| 1_x \rangle}=\left({| 1 \rangle}+{| 2 \rangle}\right)/\sqrt{2}$ at site $\mathbf{r}$ hopping to site $\mathbf{r-x}$ while changing the spin state to ${| 3 \rangle}$ with hopping strength $i \sqrt{2}$. Diagrammatically, it can be visualized as $$ic_{3,\mathbf{r-x}}^{\dag }(c_{1,\mathbf{r}}+c_{2,\mathbf{r}}) \qquad \Longleftrightarrow \qquad x\text{-}\text{direction:}\quad {| 3 \rangle}\overset{i\sqrt{2}}{\curvearrowleft}{| 1_x \rangle}\overset{\times }{\curvearrowright }\;$$ where $\overset{i\sqrt{2}}{\curvearrowleft}$ means hopping along that direction with strength $i\sqrt{2}$ and $\overset{\times}{\curvearrowright}$ indicates hopping is forbidden. This hopping term can be effected by two Raman beams $\Omega _{1}^{x}=i\sqrt{2}\Omega _{0}e^{ikz}$ and $\Omega _{1}^{\pi}=\Omega _{0}e^{ikx}$ as shown in Fig. \[FigSup:Lattice\].
\(a) (b)\
![(a) A linear tilt $\Delta_{x}$ per site in the lattice along $x$-direction. (b) Two Raman beams $\Omega _{1}^{x}$ and $\Omega _{1}^{\pi}$ used to produce the tunneling in $H^{(1)}_{\mathbf{rx}}$. The unit vectors in brackets show the polarization direction of the corresponding beam. For the complete optical lattice setup, refer to Fig. 1 in the main text.[]{data-label="FigSup:Lattice"}](Sup-Potential.eps "fig:"){width="30.00000%"} ![(a) A linear tilt $\Delta_{x}$ per site in the lattice along $x$-direction. (b) Two Raman beams $\Omega _{1}^{x}$ and $\Omega _{1}^{\pi}$ used to produce the tunneling in $H^{(1)}_{\mathbf{rx}}$. The unit vectors in brackets show the polarization direction of the corresponding beam. For the complete optical lattice setup, refer to Fig. 1 in the main text.[]{data-label="FigSup:Lattice"}](Sup-Lattice.eps "fig:"){width="35.00000%"}
The large single-photon detuning $\delta$ ensures that the population of the excited state, estimated by $|\Omega_{0}/\delta|^{2}$, is negligible. The two-photon detuning $\Delta_{x}$ matches the linear energy shift of the lattice per site, so that it only allows ${| 1_{x} \rangle}$ to hop to the left, and the other direction is forbidden by an energy mismatch $2\Delta_{x}$. The addressing of spin states is done by polarization selection rule. The original spin basis ${| 1 \rangle},{| 3 \rangle}, {| 2 \rangle}$ differ in the magnetic quantum number $m$ by one successively. So a $\pi$-polarized beam $\Omega_{1}^{\pi}$ excites the state ${| 3 \rangle}$ and a linear $\hat{x}$-polarized beam $\Omega_{1}^{x}$ excites the superposition state ${| 1_x \rangle}=\left({| 1 \rangle}+{| 2 \rangle}\right)/\sqrt{2}$ since $\hat{x} \sim (\sigma^{+}+\sigma^{-})$. So together, these two beams induce a Raman-assisted hopping between ${| 1_{x} \rangle}$ and ${| 3 \rangle}$. The hopping amplitude and phase are controlled by the corresponding Raman beam amplitude and phase. In addition, the wave-vector difference of two beams $\delta \mathbf{k}$ ($(-k, 0, k)$ in this case) has to have a component along the hopping direction ($x$-direction) to ensure the hopping strength is non-vanishing.
All the other hopping terms in the Hamiltonian are realized in a similar manner. For example, consider the term $c_{3,\mathbf{r+y}}^{\dag }(c_{1,\mathbf{r}}+ic_{2,\mathbf{r}})$, which can be realized by $\Omega _{2}^{y}=\sqrt{2}\Omega _{0}e^{ikz}$ and $\Omega _{2}^{\pi}=\Omega _{0}e^{iky}$, polarized along $(\hat{x}+\hat{y})$-direction and $\hat{z}$-direction respectively (see Fig. 1 in the main text). Since $(\hat{x}+\hat{y}) \sim (\sigma^{+}+i\sigma^{-})$, it couples the state ${| 2_y \rangle}=({| 1 \rangle}+i{| 2 \rangle})/\sqrt{2}$ and ${| 3 \rangle}$. A wave-vector difference $\delta \mathbf{k} = (0,-k,k)$ and a two-photon energy detuning $\Delta_{y}$ guarantee the desired hopping along $y$-direction.
With a number of laser beams required to realize the full Hamiltonian, it is important to check that undesired tunneling terms are forbidden. To that end, we require a different linear energy shift per site $\Delta _{x,y,z}$ along the $\left(
x,y,z\right) $-direction. The ratio between $\Delta _{x,}\Delta _{y,}\Delta_{z}$ can be adjusted by setting the direction of the gradient field to be in a specific angle with respect to the three axes of the optical lattice. In particular, we set $\Delta _{x}:\Delta _{y}:\Delta _{z}=1:2:3$. The energy difference is lower bounded by $\Delta_{x}$. So if we select a parameter regime such that the Raman-assisted hopping rate $t$ satisfies $t\ll \Delta _{x}$, then the hopping along the $z$ direction induced by $\Omega _{1}^{x}$ and $\Omega _{1}^{\pi }$, for instance, have negligible effects because of the large detuning. Other undesired couplings between different beams are disallowed because the wave-vector difference $\delta \mathbf{k}$ may not have the component along a certain direction to induce a hopping along that direction. For example, $\Omega _{1}^{x}=i\sqrt{2}\Omega _{0}e^{ikz}$ and $\Omega _{2}^{\pi}=\Omega _{0}e^{iky}$ will not induce a hopping along $x$-direction as $\delta \mathbf{k}$ does not include a component along $x$-direction. Moreover, the Raman beams $\Omega _{1,2}^{x,y,z}$ and $\Omega _{1,2}^{\pi}$ may induce some on-site spin transferring terms, which can be compensated with some radio-frequency fields.
Wannier-(Stark) Function Estimation
===================================
In the second quantization representation with the Wannier function basis, the natural tunneling rate can be written as $$t_{0} =\int d^{3}\mathbf{r}^{\prime}
\bar{w}^{\ast }(\mathbf{r}^{\prime }-\mathbf{r_{i}})
\left[ -\dfrac{\hbar^{2}}{2m} \nabla^{2} + V_{0}(\mathbf{r^{\prime}})
\right]
\bar{w}(\mathbf{r}^{\prime }-\mathbf{r_{j}}),$$ where $ \bar{w}(\mathbf{r}^{\prime }-\mathbf{r_{j}})$ is the Wannier function centered at site $\mathbf{r_{j}}$ and $V_{0}(\mathbf{r^{\prime}})$ is the lattice depth at site $\mathbf{r_{j}}$ (We use $\bar{w}(\mathbf{r})$ to denote the Wannier function and $w(\mathbf{r})$ to denote the Wannier-Stark function to be notationally consistent with the main text). With a linear tilt in the optical lattice, translational symmetry is broken and Wannier functions are no longer the proper descriptions of the localized states. Instead, a simple modification with Wannier-Stark functions $w(\mathbf{r})$ will be sufficient [@gluck2002wannier; @miyake2013probing]: $$w_{i}(\mathbf{r^{\prime}}-\mathbf{r}_{l}) = \sum_{m} J_{m-l}
\left( \dfrac{2t_{0}}{\Delta_{i}} \right) \bar{w}(\mathbf{r^{\prime}}-\mathbf{r_{m}}),$$ where $i=x,y,z$, and $w_{i}(\mathbf{r})$ is the Wannier-Stark function, and $\Delta_{i}$ is the linear tilt per site along $i$ direction. $J_{m-l} (x)$ are the bessel functions of the first kind. Fig. \[FigSup:Wannier\] shows the Wannier functions and Wannier-Stark functions with $V_{0} \approx 2.3 E_{r}$. They have close overlaps on the center site, but may differ significantly on neighboring sites. Calculations with the Wannier functions or the Wannier-Stark functions produce the same natural tunneling $t_{0}/ \hbar \approx 2\pi \times 50\,$Hz. The tunneling rate with Raman-assisted hopping can be written as an integral of Wannier-Stark functions (as discussed in the main text): $$t_{\mathbf{r,m}}=\dfrac{\Omega _{\beta \mathbf{m}}^{\ast }\Omega _{\alpha
\mathbf{m}}}{\delta }\int d^{3}\mathbf{r}^{\prime }w^{\ast }(\mathbf{r}^{\prime }-\mathbf{r-m})e^{i\delta \mathbf{k}\cdot \mathbf{r}^{\prime }}w(\mathbf{r}^{\prime }-\mathbf{r}).
\label{EqnSup:hopping}$$ Factorizing the Wannier-Stark functions into each direction, $w(\mathbf{r}^{\prime
})=w(x^{\prime })w(y^{\prime })w(z^{\prime })$ and calculating them along each direction, we can numerically compute the overlap integral $$\beta \equiv \int dxw^{\ast
}(x+a)e^{-ikx}w(x)\int dyw^{\ast }(y)w(y)\int dzw^{\ast }(z)e^{ikz}w(z).$$ With the parameters given in Fig. \[FigSup:Wannier\], we have $\beta \approx 0.34$. For Raman beams with $\Omega_{0}/2\pi \approx 15\,$MHz and the single-photon detuning $\delta/2\pi \approx 1.7\,$THz, we have $\Omega_{\mathbf{R}} = |\Omega_{0}|^{2}/
\delta \approx 2\pi \times 120\,$Hz, and the Raman-assisted hopping rate $t/\hbar \approx 2\pi \times 40\,$Hz. It is worthwhile to point out that the expression given in equation is only valid in the perturbative limit when $t \lesssim t_{0}$. When the Rabi frequency becomes stronger, the Raman-assisted tunneling rate eventually saturates. A more accurate expression may be obtained in the nonperturbative limit with a more accurate analysis [@miyake2013probing; @Miyake:2013jw]. Nevertheless, these numerical calculations only yield rough estimations to experimental parameters, which may need to be fine-tuned in experiments to produce the best result in a topologically nontrivial phase.
Density of States
=================
In Fig. 2(c) of the main text, we plotted the momentum distribution of atoms $\rho_{\text{cry}}(\mathbf{k})$ for various $h$ at a fixed chemical potential $\mu/2t=-2$. It is useful to include the density of states $\rho(E)$ for various values of the parameter $h$. In Fig. \[FigSup:Phase\] here, we show the density of states plots. The macroscopic zero-energy flat band is prominent in each plot. The band gap is also clearly visible for $h=0,2$ (less visible for $h=0.5,1.5$). In Fig. 2(c) of the main text, the figures correspond to a filling up to $\mu/2t=-2$. A change of Fermi surface topology can be observed in those figures.
[3]{}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 [Ph.D. thesis]{}, () @noop [****, ()]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the anomalously large specific heat jump and its systematic change with pressure in CeCoIn$_5$ superconductor. Starting with the general free energy functional of the superconductor for a coupled electron boson system, we derived the analytic result of the specific heat jump of the strong coupling superconductivity occurring in the coupled electron boson system. Then using the two component spin-fermion model we calculate the specific heat coefficient $C(T)/T$ both for the normal and superconducting states and show a good agreement with the experiment of CeCoIn$_5$. Our result also clearly demonstrated that the specific heat coefficient $C(T)/T$ of a coupled electron boson system can be freely interpreted as a renormalization either of the electronic or of the bosonic degrees of freedom.'
address: 'Theoretical Division Los Alamos National Laboratory, Los Alamos, New Mexico 87545 '
author:
- 'Yunkyu Bang$^{*}$, A.V. Balatsky'
title: ' Anomalous specific heat jump in the heavy fermion superconductor CeCoIn$_5$'
---
[2]{}
The superconductivity (SC) in CeMIn$_5$ (M=Co, Rh, Ir) heavy fermion (HF) series compounds [@Hegger; @Petrovic] provides valuable information about the superconductivity nearby a possible quantum critical point (QCP) due to the excellent tunability with pressure and magnetic fields[@Co; @QCP]. Besides a complicated phase diagram due to a competetion/interplay between the magnetism and superconductivity, the superconducting properties as well as the normal state ones around the QCP exhibit various anomalous behaviors such as non-Fermi liquid (NFL) power law of temperature in the resistivity, pseudogap behavior above the superconducting critical temperature, the continuous evolving from the second order to the first order phase transition of the superconductivity with magnetic fields, etc[@Co; @QCP].
In a recent paper[@Bang] we proposed the two-component spin-fermion model as a generic pairing mechanism of this series of compounds CeMIn$_5$. In this work, it is shown that some of the standard features of the d-wave superconductivity mediated by magnetic fluctuations becomes strongly modified in the vicinity of the magnetic critical point. Indeed it is very natural to expect that such a strongly coupled fermion-boson superconductivity shows various anomalies deviating from the standard weak or strong coupling theory of superconductivity[@Abanov].
In this paper we study anomalous property of the specific heat coefficient $C(T)/T$ in CeCoIn$_5$. Sparn et al.[@Sparn] measured $C(T)/T$ of CeCoIn$_5$ with varying pressure and also with applying magnetic fields. The key findings are that (1) suppressing the superconductivity with 8 Tesla fields $C(T)/T$ show $\ln T$ divergence down to 0.4 K; (2) without the magnetic fields the same sample becomes superconducting at 2.3 K with the specific heat jump ratio $\Delta C(T)/C(T) \sim 5$, which is a huge jump considering $C(T_c)/T_c
\sim 400mJ/K^2 mol$; (3) with applying pressure $T_c$ gradually increases to 2.8 K at the pressure of 15 kbar but the jump ratio $\Delta C(T)/C(T)$ drops drastically. We think [*this peculiar behavior is another manifestation of the critical magnetic fluctuations mediated superconductivity.*]{} The huge specific heat jump, indicating the steep drop of the entropy below $T_c$, is actually consistent with the spin-fermion model if we consider that the critical magnetic fluctuations in this model is generated by the fermionic particle-hole excitations, which should be gapped below $T_c$. This means that below $T_c$ not only the fermions suffer the entropy drop by developing pairing correlation but also the spin fluctuations should suffer the entropy drop, which add up to make a larger specific heat jump. As will be shown below, however, this phenomena is actually very general so that it should occur even in the electron-phonon mediated superconductors. The reason why it is phenomenal in CeCoIn$_5$ is simply because the coupling of fermion-boson is strong and in particular the resulting bosonic mode is near critical. There is a recent work [@Kos] addressing the same problem as in this paper, but proposing a different mechanism for the enhanced $\Delta C(T)/C(T)$.
We start with the general free energy functional of the two component spin-fermion model, which is nothing but the same form as the one for electron-phonon model[@Luttinger-Ward; @Eliashberg; @Bardeen-Stephen].
$$\begin{aligned}
\Omega_{total} &=& \Omega_{fer}+ \Omega_{spin} + \Omega_{int} \\
&=& -\frac{T}{V} \sum_P [\ln (-\phi(P)) + 2 \Sigma_{1}(P)G(P) - 2
\Sigma_{2}(P)F(P)] \nonumber \\
&& +(\frac{3}{2} \frac{T}{V}) \sum_q [\ln(-D^{-1}(q)+\pi(q) D(q)] \nonumber \\
&& + (T/V)^2 \sum_{P,P^{'}} \alpha_{P-P^{'}} ^2 [G(P)D(P-P^{'})G(P^{'})
\nonumber \\ && -F(P)D(P-P^{'})F(-P^{'}) ]
\end{aligned}$$
where $P=(\vec{p},i \omega_{n}), q=(\vec{q},i \Omega_{n}); \omega_{n}=\pi T (2n+1),
\Omega_{n}=\pi T (2n)$, and $G(P)=(i \omega_{n}+\epsilon_{p}+\Sigma_{1}(-P))/\phi(P)$, $F(P)=- \Sigma_{2}(-P)/\phi(P)$, $\phi(P)= (i \omega_{n}-\epsilon_{p}-\Sigma_{1}(P))(i
\omega_{n}+\epsilon_{p}+\Sigma_{1}(-P)) - |\Sigma_{2}(P)|^{2}$, respectively. Finally, the dressed spin propagator is written as $D^{-1}(q)=D_{0}
^{-1}(q)-\pi(q)$. $\Omega_{fer}, \Omega_{spin}$, and $\Omega_{int}$ are just apparent separation of the total free energy but the literal meaning of them is not necessarily correct as will be clear later. The notations are standard [@Bardeen-Stephen], and the one important difference is that $D_{0} (q)$ is not explicitly defined in the spin-fermion model in contrast to the electron-phonon case, but only the total spin propagator is defined phenomenologically as follows.
$$D(q)=\frac{D_{0}}{[1-D_{0}\pi(q)]}=\frac{D_{0}}{I(T)+A(\vec{q}-\vec{Q})^2+
|\Omega_n|/\Gamma}.$$
where $I(T)$ defines the distance from the magnetic quantum critical point (QCP) and $A$ and $\Gamma$ are the coefficients determined by the Fermi liquid parameters of the fermion sector [@Hasegawa]. The effective spin correlation length and the relaxational energy scale are $\xi^2=A/I(T)$ and $\omega_{sf}=\Gamma \cdot I(T)$, respectively, and the momenta $\vec{q}$ and $
\vec{Q}$ are assumed to be in the two dimensions[@Rosch]. We take $\Gamma$ as the unit energy scale in our numerical calculations. The bare spin propagator $D_{0}$ is assumed to be trivial regarding the critical low energy behavior and taken to be a constant. As usual, the above free energy functional is constructed in such way that it is stationary with respect to the variations of $\Sigma_{1,2}$ and $\pi$ a la Luttinger-Ward[@Luttinger-Ward] with the following definitions. This property of the functional is very important and utilized critically in our paper.
$$\begin{aligned}
\Sigma_{1}(P) &=&\frac{T}{V} \sum_{P^{'}} \alpha_{P-P^{'}} ^{2} G(P^{'})
D(P-P^{'}),
\\ \Sigma_{2}(P)&=&\frac{T}{V} \sum_{P^{'}} \alpha_{P-P^{'}} ^{2} F(P^{'}) D(P-P^{'}),
\\ \pi(P)& = & - \frac{2}{3} \frac{T}{V} \cdot \nonumber \\
& & \sum_{P} \alpha_{q} ^2 [G(P+q)G(P)-F(P+q)F(-P)].\end{aligned}$$
Now following Ref\[[@Bardeen-Stephen]\], neglecting the momentum dependence of $\Sigma_{1,2}$ and using the definitions, $
%
%\begin{eqnarray}
\Sigma_{1}(P)= i \omega_{n} (1- Z(\omega_{n})),$ and $
\Delta(\omega_{n})=\Sigma_{2}(P)/Z(\omega_{n}),$ we obtain after the momentum integration the following.
$$\begin{aligned}
\Omega_{fer}+ \Omega_{int} = -T \cdot N(0) \cdot \pi \nonumber \\
\sum_{\omega_{n}} [ (Z(\omega_{n})+1) \sqrt{\Delta^2-(i \omega_{n})^2} -
\frac{\Delta^2}{\sqrt{\Delta^2-(i \omega_{n})^2}}]\end{aligned}$$
Note that the above equation didn’t include the $\Omega_{spin}$ term yet [@S_boson]. But due to the stationary property of the total free energy functional as mentioned above, the specific heat jump across the superconducting transition with the changes of $\Sigma_{1} ^{(n)} \rightarrow
\Sigma_{1} ^{(s)}$ and $\pi^{(n)} \rightarrow \pi^{(s)}$ is solely determined by Eq.(7) with $Z(\omega_{n})= Z^{(n)}(\omega_{n})$ at $T=T_{c}$.
The Eq(7) can be written after the Matsubara frequency summation using the contour integration as
$$\begin{aligned}
\Omega_{fer}+ \Omega_{int} = N(0) \int_{0} ^{\infty} d \epsilon [(Z+1) \omega
- Z \frac{\Delta^2}{\omega}] \cdot \tanh(\frac{\omega}{2 T}),\end{aligned}$$
where $\omega^2=\epsilon^2+\Delta^2$ and the specific heat is calculated at $T=T_c$ as
$$C(T)/T=(Z+1) N(0) \frac{\pi^2}{3}+\frac{Z N(0)}{T_c}(-\frac{\partial \Delta^2
}{\partial T})$$
This is our key result in this paper and we can read the specific heat jump ratio as $\frac{\Delta C(T)}{C(T)}=\frac{Z}{Z+1} \cdot \frac{-\frac{\partial
\Delta^2 }{\partial T}/ T_c}{\pi^2/3}$ [@S_boson]. We see from this formula that the jump ratio is modified with the normal self-energy renormalization factor $\frac{Z}{Z+1}$, which is $\frac{1}{2}$ in the BCS limit with $Z=1$ and can have a maximum value $1$ in the limit of $Z(T) \rightarrow \infty$. Apparently $\frac{Z}{Z+1} \rightarrow 1$ is the proper limit for CeCoIn$_5$ and some other heavy fermion compounds exhibiting a large $C(T)/T$ at low temperature. We summarize the consequences of this formula: (1) The jump ratio at most can be doubled to the BCS ratio assuming $\frac{\partial \Delta^2 }{\partial T}$ is the same as the BCS limit, i.e. $9.42 T_c$; (2) Apparently the steeper slope of $\Delta(T)$ can contribute an extra enhancement to the jump ratio. And since our formula is an exact result, this formula can be utilized to estimate the temperature slope of $\Delta(T)$ at $T_c$ from the experimental $\frac{\Delta
C(T)}{C(T)}$ in the limit of $\frac{Z}{Z+1} \rightarrow 1$; (3) Finally, the interpretation of the large $C(T)/T$ and its enhanced jump ratio from the heavily renormalized fermionic quasiparticle due to the scattering from the magnetic fluctuations seems to be legitimate, if we can neglect the contribution from $\Omega_{boson}$ [@S_boson].
Now in order to calculate $C_{total}(T)$ above and below $T_c$, we need to calculate $\Sigma^{(n),(s)} _{1,2}$ with a given $D(q)^{(n),(s)}$. But we found a more convenient way to proceed as follows. Adding $\Omega_{int}$ term to $\Omega_{spin}$ exactly cancels the cumbersome term $\pi(q) D(q)$ as
$$\Omega_{spin} + \Omega_{int} = (\frac{3}{2} T/V) \sum_q [\ln(-D^{-1}(q)].$$
Furthermore, $\Omega_{fer} $ term alone, after the same manipulations used for Eq.(7) and Eq.(8), becomes
$$\begin{aligned}
\Omega_{fer}& = & -T \cdot N(0) \cdot \pi \sum_{\omega_{n}} [ 2
\sqrt{\Delta^2-(i \omega_{n})^2} - 2 \frac{\Delta^2}{\sqrt{\Delta^2-(i
\omega_{n})^2}}] \nonumber \\ & = & N(0) \int_{0} ^{\infty} d \epsilon ~ 2
\omega
\cdot \tanh(\frac{\omega}{2 T}).\end{aligned}$$
A different organization greatly simplifies $\Omega_{fer}$ so that all the renormalization effect due to the interaction miraculously drops and furthermore the explicit $\Delta(T)$ dependent term all cancels out in the final result of Eq.(11). This amazing result is actually the consequence of the Eliashberg free energy functional built with the stationary property with respect to the variations of the self-energies. Considering that the Eliashberg free energy functional is not the exact free energy functional but a consistent approximate functional equivalent to the self-consistent Born approximation for the fermionic part and the random phase approximation (RPA) for the bosonic part, the exactness of Eq.(9) and Eq.(11) has a limited meaning. Nevertheless, besides the accuracy of this functional the insightful reason for this result is the equivalent relation of the following.
$$\frac{T}{V} \sum_P [ \Sigma_{1}(P)G(P) - \Sigma_{2}(P)F(P)]
= - (\frac{3}{2} \frac{T}{V}) \sum_q [\pi(q) D(q)].$$
From the technical viewpoint the above relation is trivial if we consider any linked-cluster diagram obtained from the free energy expansion with $H_{int}$ [@Luttinger-Ward]. However the physical interpretation is rather revealing. Namely, the interacting free energy $\Delta
\Omega=\Omega_{total}-\Omega_{bare}$ and its derived entropy $\Delta S$ all comes from the expansion of $H_{int}$. Then depending on how to view or organize the same diagrams, we can view all interaction effect as the renormalization of either the fermion sector or the bosonic sector [@Kotliar]. While theoretically we have this freedom, in experiments the experimental probe itself determines how to measure the contribution of $H_{int}$ through either fermionic or bosonic degree of freedom. However the thermodynamic measurement like the specific heat measures simultaneously the both degrees of freedom and it is hard to know how much contribution comes from which degrees of freedom. Our analytic result clearly demonstrated that the specific heat coefficient $C(T)/T$ of a coupled fermion-boson system can be freely interpreted as a renormalization either of the electronic or of the bosonic degree of freedom, but not for both. This point should have a far-reaching consequence to the heavy fermion experiments and its related interpretations. This important aspect will be dealt in the separate publication.
Now let us return to the calculation for the total $C(T)_{tot}$. From Eq.(10) we can calculate the entropy $S(T)_{spin+int}$ with the phenomenological $D(q)$ given in Eq.(3) as follows.
$$\begin{aligned}
S_{spin}+S_{int}= \sum_{q} \int_{0} ^{\Lambda_{\omega}} \frac{3 d \omega}{4\pi
} [\frac{1}{\sinh (\frac{\omega}{2 T})} ]^2 \nonumber \\ \cdot
\frac{\omega}{T^2} \cdot \arctan \frac{\omega/\Gamma}{(I_{0}+a T)+ A
(\vec{q}-\vec{Q})^2}.\end{aligned}$$
In real calculation, we need to fix parameters. Our unit energy scale is $\Gamma=1$, and we take $\Lambda_{\omega}=3$, $A
(\vec{q}^{max}-\vec{Q})^2=\frac{\omega_{q} ^{max}}{\Gamma}=1.0$, and $a=1.0$. For the fermionic part $\Omega_{fer}$, it is just the same form as the non-interacting fermion free energy except $\omega =\sqrt{\epsilon_{k}
^2+\Delta^2 (T)}$, and therefore it develops some structure below $T_c$ but no discontinuity at $T_c$. Finally to fix the relative magnitude of $\Omega_{spin+int}$ to $\Omega_{fer}$ we need to determine the spin density of states $N_{s}$ defined as $N_s \int_{0} ^{\Lambda_{\omega}} =\frac{1}{V}
\sum_{q} ^{q_{max}}$. This can be determined using Eq.(12) at normal state, which gives the following relation
$$\begin{aligned}
\frac{3}{2} T N_{s} \sum_{\Omega_{n}} |\Omega_{n}| \ln [ \frac{ I(T)+\omega_{q}
^{max}/\Gamma+|\Omega_{n}|/\Gamma} { I(T)+|\Omega_{n}|/\Gamma}] = \nonumber
\\ T N(0) \pi \sum_{\omega_{n}} (Z^{(n)}-1) |\omega_{n}|.\end{aligned}$$
We assume $\pi_{n}=-\frac{|\Omega_{n}|}{D_0 \Gamma}$ in the derivation, but the above relation is not useful since both side of the equation diverges unless we introduce high frequency cut-off for $\Omega_{n}$ and $\omega_{n}$. Since the above relation should hold for a variation of $\delta \pi_{n}$, we derive another equivalent relation at $T=0$ as
$$\begin{aligned}
\frac{3}{2 \pi} N_{s} \int_{0} ^{\Lambda_{\omega}} d \Omega [
\frac{-\Lambda_{\omega}}{[ I(T)+\Lambda_{\omega}+\Omega] [I(T)+\Omega]}]
\Omega^2 \nonumber \\
= N(0) \int_{0} ^{\infty} d \omega ~~ \omega \cdot \frac{
[ Z^{(n)} (\omega; \pi_{n} +\delta \pi_{n}) - Z^{(n)} (\omega; \pi_{n}) ]}{
\delta \pi_{n}}.\end{aligned}$$
A recent paper by Ref.\[[@Haslinger]\] also use the similar procedure to fix the relation between $N(0)$ and $N_{s}$. Our numerical result gives $N_{s}
\cong 1.5 \frac{D_0 \alpha^2}{\pi} N(0)$. Now we are ready to calculate $C_{tot}(T)$ using Eq.(11) and Eq.(13). For the calculation below $T_c$ we need to know the temperature dependence of the gap function $\Delta(T)$ and $\pi_{s}
(\Delta(T))$. For the gap function we use the generalized BCS form $\Delta(T)=\Delta_0 \tanh (a \sqrt{T_c/T -1})$, where $a=1.764$ and $\Delta_0=
1.74 T_c$ for BCS limit, but for strong coupling superconductor these values can deviate largely from the BCS limit and we take them as parameters. Now much unclear part is the behavior of $\pi_{s} (T)$ below $T_c$. Qualitatively it should be cut off for $\Omega \leq 2 \Delta_0$ at $T=0$ and smoothly recover to the normal state form approaching $T \rightarrow T_c$. The leading expansion of $\pi_{s} (T)$ in $\Delta^2 (T)$ gives $\pi_{s}=\pi_{n}
\exp (- \Delta^2(T)/[\omega^2+T^2])$; but for a larger value of $\Delta(T)$ for $T \rightarrow 0$ this form should not be very correct.
The numerical results are shown in Fig.(1) and Fig.(2). In Fig.(1) the specific heat coefficient $C(T)/T$ calculated from Eq.(13) are shown in unit of $\frac{D_0 \alpha^2}{\pi \Gamma} N(0)$. For our parameter choice producing $Z
>> 1$ (For this we need the critical spin fluctuations ($I_0 << 1$) as well as the effective coupling $\frac{D_0 \alpha^2}{\pi}$ of $O(1)$), the contribution of $C(T)/T$ from $\Omega_{fer}$ Eq.(11) is negligibly small. Therefore we didn’t add it to $C(T)/T$ in Fig(1) but show it separately in Fig.(2) demonstrating its qualitative features. In Fig.(1) the solid line is for the normal state $C(T)/T$ with $I_0=0$ (the spin fluctuations is at QCP), it indeed displays $\ln T$ divergence with decreasing temperature. With SC transition at T=0.2 (open square symbol) it shows the jump ratio $\Delta C(T)/C(T) \sim 5$ with the choice of parameters [@comment2]. Then increasing $I_0$, the $\ln
T$ divergence is quickly suppressed and at the same time the specific heat jump ratio drops with increasing $T_c$ ($T_c$ is increased by hand and the temperature slope of $\Delta(T)$ is reduced accordingly as $\frac{\Delta^{'}(T_{c1})}{\Delta^{'}(T_{c2})} = \frac{T_{c2}}{T_{c1}}$). With this rather crude phenomenological choice of parameters our numerical results reproduce the experimental features of $C(T)/T$ in CeCoIn$_5$ qualitatively as well as quantitatively.
In summary, starting with the general free energy functional for the coupled fermion-boson system, we derived an analytic formula of $\Delta C(T)/C(T)$ for the general strong coupling superconductor. Then we calculate $C(T)/T$ for the spin-fermion model and show that the salient features of $C(T)/T$ of CeCoIn$_5$ for both normal and superconducting states are successfully explained including its anomalous jump ratio $\Delta
C(T)/C(T)$ and the progressive reduction of it with increasing $T_c$. Also with different organization of the Eliashberg free energy functional, which lumps all the interaction effect either into the bosonic or into the fermionic degrees of freedom, we clearly demonstrated that [*the effect of the interaction in the total free energy or entropy of the coupled fermion-boson system can be freely viewed as the renormalization of either the fermion sector or the boson sector.*]{} Finally, our results strongly support the idea that the two dimensional critical magnetic fluctuations plays an essential role in CeMIn$_5$ HF compounds in producing a large and strongly temperature dependent $C(T)/T$ and the SC pairing itself.
We are grateful to J.D. Thompson, J.L. Sarrao, Ar. Abanov, and C. M. Varma for useful discussions. This work was supported by US DoE. Y.B. was also partially supported by the Korean Science and Engineering Foundation (KOSEF) through the Center for Strongly Correlated Materials Research (CSCMR) (2001) and through the Grant No. 1999-2-114-005-5.
On leave from the Department of Physics, Chonnam National University, Kwangju 500-757, Korea.
H. Hegger, C. Petrovic, E. G. Moshopoulou, M. F. Hundley, J. L. Sarrao, Z. Fisk, and J. D. Thompson , Phys. Rev. Lett. [**84**]{}, 4986 (2000).
C. Petrovic, R. Movshovich, M. Jaime, P. G. Pagliuso, M. F. Hundley, J. L. Sarrao, Z. Fisk, J. D. Thompson, Europhys. Lett. [**53**]{}, 354 (2001); C. Petrovic et al., J. Phys. Condens. Matter [**13**]{}, L337 (2001).
V.A. Sidorov, M. Nicklas, P. G. Paguliuso, J. L. Sarrao, Y. Bang, A.V. Balatsky, and J.D. Thompson, Phys. Rev. Lett. [**89**]{}, 157004 (2002); A. Bianchi, R. Movshovich, N. Oeschler, P. Gegenwart, F. Steglich , J.D. Thompson, P.G. Pagliuso, J. L. Sarrao, Phys. Rev. Lett. [**89**]{}, 137002 (2002).
Yunkyu Bang, I. Martin, and A.V. Balatsky, Phys. Rev. B [**66** ]{}, 224501 (2002).
Ar. Abanov, A. V. Chubukov, and A. M. Finkel’stein, Europhys. Lett. [**54**]{}, 488 (2001).
G. Sparn, R. Borth, E. Lengyel, P.G. Pagliuso, J.L. Sarrao, F. Steglich, J.D. Thompson, Physica [**B 312**]{}, 138 (2002)
S. Kos, I. Martin, and C. M. Varma, cond-mat/0302089.
J. M. Luttinger, and J. C. Ward, Phys. Rev. [**118**]{}, 1417 (1960).
G. M. Eliashberg, JETP [**16**]{}, 780 (1963).
J. Bardeen, and M. Stephen, Phys. Rev. [**136**]{}, A1485 (1964).
H. Hasegawa, and T. Moriya, J. Phys. Soc. Japan, [**36**]{}, 1542 (1974).
A. Rosch, A. Schroeder, O. Stockert, and H. v. Loeneysen, Phys. Rev. Lett. [**79**]{}, 159 (1997).
In the electron-phonon case, the term $\Omega_{boson}$ (in this paper $\Omega_{spin}$) is neglected with the argument that the change $(\pi_s
-\pi_n) \sim \Delta^2/E_{f}^2$ is small[@Bardeen-Stephen]. But this argument is wrong because the change of the free energy above and below $T_c$ is always $O(\Delta^2/E_{f}^2)$ and this small change is what makes the superconducting transition. As shown by Chester[@Chester] even before the BCS theory, the actual change of phonon kinetic energy is just the same order of magnitude as the electronic part. The correct reason why still the term $\Omega_{boson}$ can be neglected is that the contribution of the self-energy correction $\pi_{s,n}$ in $\Omega_{boson}$ exactly cancels between $\ln(-D^{-1}(q))$ and $\pi(q) D(q)$ terms for the small frequency limit. Then the actual contribution of $\Omega_{boson}$ is just the one of the unrenormalized bare boson (or phonon) propagator, so it can be dropped.
G. V. Chester, Phys. Rev. [**103**]{}, 1693 (1956).
I. Paul, and G. Kotliar, Phys. Rev. B [**64**]{}, 184414 (2001).
R. Haslinger, and A. V. Chubukov, cond-mat/0209600.
In order to compare with experiment, we should count the gap anisotropy which reduces the specific heat jump by the factor $<\Delta^2(k)>^2/
<\Delta^4(k)>$[@Mishonov] (for d-wave gap, this value is 0.666.). In Fig.(1), we just multiplies 0.666 to the results below $T_c$ for a better comparison with the experiment, but this artificial procedure ruins the entropy conservation.
T.M. Mishonov, E. S. Penev, J. O. Indenkeu, and V. L. Pokrovsky, cond-mat/0209342.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Suppose $2n$ voters vote sequentially for one of two candidates. For how many such sequences does one candidate have strictly more votes than the other at each stage of the voting? The answer is $\binom{2n}{n}$ and, while easy enough to prove using generating functions, for example, only two combinatorial proofs exist, due to Kleitman and Gessel. In this paper we present two new (far simpler) bijective proofs.
[**Key words.**]{} lattice path, bijection, ballot problem
author:
- |
Glenn Hurlbert[^1]\
Vikram Kamat[^2]\
\
Department of Mathematics and Statistics\
Arizona State University\
Tempe, Arizona 85287-1804
title: Two New Bijections on Lattice Paths
---
Ł[[L]{}]{} ¶[[P]{}]{} §[[S]{}]{}
ł([[(]{}]{} )[[)]{}]{} ${{\Biggl(}}
\def$[[)]{}]{} $${{\Biggl[}}
\def$$[[\]]{}]{}
Introduction {#Intro}
============
Suppose $A$ and $B$ are candidates and there are $2n$ voters, voting sequentially. In how many ways can $A$ get $n$ votes and $B$ get $n$ votes such that $A$ is always ahead of or tied with $B$? This is the famous ballot problem, the solution of which is counted by the Catalan number $C_n.$ A more general version of this problem is stated in terms of probability: let candidate $A$ receive $a$ votes and $B$ receive $b$ votes and compute the probability that $A$ never falls behind $B$. This is known to be $\frac{a-b}{a+b}$, first proved by André [@A] using his reflection principle (this also appears in [@F] and again in [@HP]). A $q$-binomial variation of the problem appears in [@KM], and a weighted variation is found in [@EFY]. Some have also considered an $n$-dimensional version by generalizing André’s reflection proof to the many candidate ballot problem – one of these approaches can be found in [@Z]. Others [@FR; @Lo; @Ly] have considered the situation in which the number of votes of the two candidates remains close. In this paper, we concern ourselves with another variation of the two candidate ballot problem discounting all instances when the two candidates are tied. It can be more formally expressed in terms of plus minus sequences.
Let $\S_n=\{s_1\ldots s_{2n}\ |\ s_i\in\{-1,+1\}\}$. For $S\in\S_n$ let $\s(S)=(\s_1,\ldots,\s_{2n})$, where $\s_i=\sum_{j=1}^is_i$. We write that $\s\not=0$ (resp, $\s>0$, $\s<0$) when each $\s_i\not=0$ (resp. $\s_i>0$, $\s_i<0$), and call $S$ [*zero-free*]{} (resp. [*positive*]{}, [*negative*]{}) if $\s(S)\not=0$ (resp. $\s(S)>0$, $\s(S)<0$). The set of zero-free (resp. positive, negative) sequences of $\S_n$ is denoted by $\F_n$ (resp. $\P_n$, $\N_n$). We will find it useful to denote $\s_{2n}$ by $\sum S$. A sequence $S\in\S_n$ is [*balanced*]{} if $\sum S=0$, and we use $\B_n$ to denote all balanced sequences in $\S_n$. We also use the notations $\S_n^+, \F_n^+, \B_n^+$ to denote those sequences that start with $+1$ (with the obvious analogous definitions for $\S_n^-$, etc.). Note that $\F_n^+=\P_n$, $\F_n^-=\N_n$, and $|\P_n|=|\N_n|$.
It is known that $|\F_n|=|\B_n|$ for every $n$ and, to our knowledge, only two bijections have appeared in print, due to Kleitman [@K] and Gessel (see [@S]).[^3] Here we give two new bijections for this result, one indirect (Section \[KP\]) and one direct (Section \[Hurl\]). Of course, it is enough to show these for sequences in $\S_n^+$. That our direct bijection $\sg$ (see Section \[Hurl\]) differs from that of Kleitman’s bijection $\sk$ can be seen for virtually any sequence $P \in P_n$ for any $n$, such as the following. $$\begin{aligned}
P & = & ++-+++---+++--++++-+--+++++- \\
\sg(P) & = & +-+---++++--++++--+-++-----+ \\
\sk(P) & = & ---+++---+++----++-+--+++++- \\ \end{aligned}$$ In addition, ours is considerably simpler to navigate. Like Kleitman’s, Gessel’s bijection does not preserve the first coordinate.
As is well known, one of the interesting applications of this result is the derivation of the generating function $F(x)$ for the sequence $\{\binom{2n}{n}\}$. Indeed, the factoring of a sequence $S\in\S_n$ into its maximum length balanced initial subsequence and corresponding terminal zero-free subsequence results in the relation $|\S_n|=\sum_k|\B_k||\F_{n-k}|$, from which the bijection yields $$\label{sved}
4^n=\sum_k\binom{2k}{k}\binom{2n-2k}{n-k}\ ,$$ one short convolutional step from proving that $F(x)=(1-4x)^{-1/2}$. Marta Sved [@S] recounts the history of Identity (\[sved\]) and notes some combinatorial proofs of it that were submitted by readers.
Most satissfyingly, this paper is the result of the first author challenging the students in his graduate combinatorics course to find such a bijection.
In this article we use the notations $[s]=\{1,2,\ldots,s\}$ and $(r,s)=[r+1,s-1]=\{r+1,\ldots,s-1\}$.
Proofs {#Proofs}
======
Indirect Bijection {#KP}
------------------
Let $\S_{n,k}^+$ be the set of all $S\in\S_n^+$ with $\sum S=2k$. Such an $S$ contains $(n-k)$ ‘$-1$’s and $(n+k)$ ‘$1$’s. Therefore, $$|\S_{n,k}^+|=\binom{2n-1}{n-k}\ ,$$ since each such $S$ begins with a ‘1’. For each $1\leq k\leq n,$ let $\P_{n,k}$ be the set of all $P\in
\P_n$ having $\sum P=2k$. Then $|\P_n|=\sum_{k=1}^n|\P_{n,k}|$. Define $\T_{n,k}^+=S_{n,k}^+-\P_{n,k}$, so that $$|\P_{n,k}|=|\S_{n,k}^+|-|\T_{n,k}^+|=\binom{2n-1}{n-k}-|\T_{n,k}^+|\ .$$ Finally, let $\S_{n,k}^-$ be all sets $S\in S_n^-$ for which $\sum S=2k$. Because such an $S$ begins with a ‘$-1$’, we have $$|\S_{n,k}^-|=\binom{2n-1}{n-(k+1)}\ .$$ If $|\T_{n,k}^+|=|\S_{n,k}^-|$, then $$\begin{aligned}
\sum_{k=1}^{n}|\P_{n,k}|
&=&\sum_{k=1}^{n}\left[\binom{2n-1}{n-k}-\binom{2n-1}{n-(k+1)}\right]\\
&=&\binom{2n-1}{n-1}\ , \end{aligned}$$ by telescoping. Thus, we obtain $|\F_n|=\binom{2n}{n}=|\B_n|$ if $|\T_{n,k}^+|=|\S_{n,k}^-|$. We show that this indeed holds by demonstrating a bijection between the two sets.
[**From $\T_{n,k}^+$ to $\S_{n,k}^-$.**]{}
Let $T\in \T_{n,k}^+$. We can factor $T$ as follows: $$T=T_1T_2$$ where $T_1$ is the smallest balanced initial subsequence, say of length $m$. (In lattice path language, $m$ is the first step that hits the diagonal $y=x$.) By definition every sequence in $\T_{n,k}^+$ can be represented this way. Let $\overline{T_1}$ be the subsequence of length $m$ obtained by negating every term of $T_1.$ Now, define a function $f \colon \T_{n,k}^+ \to \S_{n,k}^-$ as follows: $$f(T)=f(T_1T_2)=\overline{T_1}T_2\ .$$
Consider an example. Let $$\begin{aligned}
T&=&++-++-+--\fbox{$-$}++-++++++-+\ . \end{aligned}$$ Here, we see that $T\in \T_{11,4}^+$ and $m=10$ (with the $10^{\rm th}$ position boxed). Also, $$\begin{aligned}
f(T)&=&--+--+-++\fbox{+}++-++++++-+\ . \end{aligned}$$
Clearly, $f(T)$ belongs to $\S_{11,4}^-.$ The reason why this will hold in general is because only a balanced subsequence is negated and so $\s$ is unchanged. Moreover, the smallest balanced subsequence always contains the first element(which is a $1$) and thus, $f(T)$ will always have $-1$ as its first element.
[**From $\S_{n,k}^-$ to $\T_{n,k}^+$.**]{}
Analogously, define a function $g \colon \S_{n,k}^- \to \T_{n,k}^+$ as follows: for any $S\in \S_{n,k}^-$, $$g(S)=g(S_1S_2)=\overline{S_1}S_2\ ,$$ where $S=S_1S_2$ with $S_1$ its smallest balanced initial subsequence, and $\overline{S_1}$ the negation of $S_1$.
In the example above, we have $$\begin{aligned}
S&=&--+--+-++\fbox{+}++-++++++-+\ \in\ \S_{11,4} \end{aligned}$$ and $$\begin{aligned}
g(S)&=&++-++-+--\fbox{$-$}++-++++++-+\ . \end{aligned}$$
Clearly, $g(S)\in \T_{11,4}^+.$ As above, this holds in general since $\s$ is unchanged and $g(S)$ begins with a ‘1’ because $S$ begins with a ‘$-1$’.
The functions $f \colon \T_{n,k}^+ \to \S_{n,k}^-$ and $g \colon
\S_{n,k}^- \to \T_{n,k}^+$ are inverses of each other.
Let $T\in \T_{n,k}^+$ and let $S=f(T).$ We will show that $g(S)=T.$ We have $T=T_1T_2$ where $|T_1|=m$, as defined above, and $S=\overline{T_1}T_2.$ Now, when we apply function $g$ to $S$, since $T_1$ is the smallest balanced initial subsequence in $T$, $\overline{T_1}$ is the smallest balanced initial subsequence in $S$. This gives $g(S)=T.$
A similar argument proves that, for any $S\in\S_{n,k}^-$, $f(T)=S$ when $g(S)=T$.
Direct Bijection {#Hurl}
----------------
As previously mentioned, we will give a bijection between $\B_n^+$ and $\P_n$. The obvious bijection between $\B_n^-$ and $\N_n$, and hence $\B_n$ and $\F_n$ follows.
[**From $\B_n^+$ to $\P_n$.**]{} For $B\in\B_n^+$ we define a set $\p(B)$ of [*peaks*]{} of $B$ as follows. Set $t=\max\{\s_i(B)\}$ and for $1\le k\le t$ let $\p_k$ be the index of the left-most occurrence of $k$ in $\s=\s(B)$: $\p_k=\min\{i\ |\ \s_i=k\}$. For example, if $$B\ =\ \fbox{+}-+---+++\fbox{+}--++\fbox{+}\fbox{+}--+-++-----+$$ then [$$\s(B)\ =\ (\fbox{1},0,1,0,-1,-2,-1,0,1,\fbox{2},1,0,1,2,\fbox{3},\fbox{4},3,2,3,2,3,4,3,2,1,0,-1,0)$$]{} and $$\p(B)=\{1,10,15,16\}\ .$$ We have boxed in the peak locations as shown.
Next we define the set of intervals $I_k=(\p_k,\p_{k+1})$, with $I=\cup I_k=[2n]-\p(B)$. (Artificially, we set $\p_{t+1}=2n+1$ in order to define $I_t$; in this case we have $\p_5=29$ and $I_4=(16,29)=[17,28]$.) The key property here is that $\s_i\le k$ for every $i\in I_k$.
Finally we define $\ssf=\ssf(B)$ by $\ssf_i=B_i$ for all $i\in\p(B)$ and $\ssf_i=-B_i$ otherwise (for all $i\in I$). For this example, we obtain $$\ssf\ =\ \fbox{+}+-+++---\fbox{+}++--\fbox{+}\fbox{+}++-+--+++++-$$ and [$$\s(\ssf)\ =\ (\fbox{1},2,1,2,3,4,3,2,1,\fbox{2},3,4,3,2,\fbox{3},\fbox{4},5,6,5,6,5,4,5,6,7,8,9,8)\ .$$]{} Note that we have $\s(\ssf)>0$, so that $\ssf(B)\in\P_n$. This holds in general for the following reason. By the definition of $\p_k$ we have $\sum B_{\p_{k+1}-1}=\sum B_{\p_k}$. This means that $B$ is balanced on each interval $I_k$ with $k<t$. Hence each $\sum\ssf_{\p_k}=\sum B_{\p_k}$ and thus $\s(\ssf)_i\ge k$ for every $i\in I_k$. In particular, $\ssf_i\ge 1$ for all $i$.
[**From $\P_n$ to $\B_n^+$.**]{}
For $P\in\P_n$ we define a set $\Pi(P)$ of [*pivots*]{} of $P$ as follows. Set $T=\frac{1}{2}\sum P$, let $\Pi_1=1$, and for $1<k\le T$ let $\Pi_k$ be one more than the index of the right-most occurrence of $k-1$ in $\s=\s(P)$: $\Pi_k=1+\max\{j\ |\ \s_j=k-1\}$. For example, if $$P\ =\ \fbox{+}+-+++---\fbox{+}++--\fbox{+}\fbox{+}++-+--+++++-$$ then [$$\s(P)\ =\ (\fbox{1},2,1,2,3,4,3,2,1,\fbox{2},3,4,3,2,\fbox{3},\fbox{4},5,6,5,6,5,4,5,6,7,8,9,8)$$]{} and $$\Pi(P)=\{1,10,15,16\}\ .$$ We have boxed in the pivot locations as shown.
Next we define the set of intervals $J_k=(\Pi_k,\Pi_{k+1})$, with $J=\cup J_k=[2n]-\Pi(P)$. (Artificially, we set $\Pi_{T+1}=2n+1$ in order to define $J_T$; in this case we have $\Pi_5=29$ and $J_4=(16,29)=[17,28]$.) The key property here is that $\s_j\ge k$ for every $j\in J_k$. Finally we define $\sg=\sg(P)$ by $\sg_j=P_j$ for all $j\in\Pi(P)$ and $\sg_j=-P_j$ otherwise (for all $j\in J$). For this example, we obtain $$\sg\ =\ \fbox{+}-+---+++\fbox{+}--++\fbox{+}\fbox{+}--+-++-----+$$ and [$$\s(\sg)\ =\ (\fbox{1},0,1,0,-1,-2,-1,0,1,\fbox{2},1,0,1,2,\fbox{3},\fbox{4},3,2,3,2,3,4,3,2,1,0,-1,0)\ .$$]{} Note that we have $\sum\sg=0$, so that $\sg(P)\in\B_n^+$. This holds in general for the following reason. By the definition of $\Pi_k$ we have $\sum P_{\Pi_{k+1}-1}=\sum P_{\Pi_k}$. This means that $P$ is balanced on each interval $J_k$ with $k<T$. Hence each $\sum\sg_{\Pi_k}=\sum P_{\Pi_k}$ and thus $\s(\sg)_j\le k$ for every $j\in J_k$. In particular, by the definition of $T$ we have $\sum_{j\in J_T}P_j=T$, so that $\sum_{j\in J_T}\sg_j=-T$, and hence $\sum \sg=0$.
The functions $\ssf:\B_n^+\ra\P_n$ and $\sg:\P_n\ra\B_n^+$ are bijections and, in fact, inverses of each other.
The arguments above show that $\ssf$ and $\sg$ are well-defined. That they are bijections will follow from their inverse relationship.
We suppose first that $\ssf(B)=P$ for $B\in\B_n^+$, and show that $\sg(P)=B$. This will follow from showing inductively that $\Pi(P)=\p(B)$. Of course, $\Pi_1=1=\p_1$, so assume that $\Pi_k=\p_k$. Because we know for all $k$ that $\s(B)_i\le k$ for all $i\in I_k$ and that $\s(B)_{\p_{k+1}}=k+1$, we therefore know for all $k$ that $\s(P)_i\ge k$ for all $i\in I_k$ and that $\s(P)_{\p_{k+1}}=k+1$. In particular, the right-most occurrence of $k$ in $\s(P)$ occurs with index $\p_{k+1}-1$; i.e. $\Pi_{k+1}=\p_{k+1}$.
Next we suppose that $\sg(P)=B$ for $P\in\P_n$, and show that $\ssf(B)=P$. As above, we show that $\p(B)=\Pi(P)$ by induction. Again, $\p_1=1=\Pi_1$, so we assume that $\p_k=\Pi_k$. We know for all $k$ that $\s(P)_j\ge k$ for all $j\in J_k$ and that $\s(P)_{\Pi_{k+1}}=k+1$, and thus we know for all $k$ that $\s(B)_j\le k$ for all $j\in J_k$ and that $\s(B)_{\Pi_{k+1}}=k+1$. In particular, the left-most occurrence of $k+1$ in $\s(B)$ occurs with index $\Pi_{k+1}$; i.e. $\p_{k+1}=\Pi_{k+1}$.
Acknowledgement {#acknowledgement .unnumbered}
===============
We thank Ira Gessel and David Callan for several helpful comments.
[99]{}
D. André, Solution directe du probléme résola par M. Bertnand, C. R. Acad. Sci. Paris 105 (1887), 436–437.
D. Callan, Bijections for the identity $4^n=\sum_{k=0}^n\binom{2k}{k}\binom{2(n-k)}{n-k}$,\
`http://www.stat.wisc.edu/ callan/papersother/`.
S.P. Eu, T.S. Fu and Y.N. Yeh, Refined Chung-Feller theorems for lattice paths, J. Combin. Theory, Ser. A 112 (2005), 143–162.
W. Feller, An introduction to probability theory and its applications, Vol. I, third ed., Wiley, New York, 1968.
R.D. Fray and D.P. Roselle, Weighted lattice paths, Pacific J. Math. 37 (1971), no. 1, 85–96.
P. Hilton and J. Pedersen, The ballot problem and Catalan numbers, Nieuw Arch. Wisk. (4) 8 (1990), no. 2, 209–216.
D. Kleitman, A note on some subset identities, Studies in Appl. Math. 54 (1975), no. 4, 289–292.
C. Krattenthaler and S.G. Mohanty, $q$-generalization of the ballot problem, Discrete Math. 126 (1994), no. 1-3, 195–208.
N. Loehr, Note on André’s reflection principle, Disc. Math. 280 (2004), 233–236.
R.C. Lyness, Al Capone and the death ray, Math. Gaz. 25 (1941), 283–287.
M. Sved, Counting and recounting: the aftermath, Math. Intelligencer 6 (1984), no. 2, 44–45.
D. Zeilberger, André’s reflection proof generalized to the many-candidate ballot problem, Discrete Math. 44 (1983), no. 3, 325–326.
[^1]: `[email protected]`, corresponding author
[^2]: `[email protected]`
[^3]: It has come to our attention recently that a combinatorial proof similar to ours is given by Callan [@C].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present the solution of the problem of the $1/\Box, \Box \to 0,$ asymptotic terms discovered in the one-loop form factors of the gravitational effective action. Owing to certain constraints among their coefficients, which we establish, these terms cancel in the vacuum stress tensor and do not violate the asymptotic flatness of the expectation value of the metric. They reappear, however, in the Riemann tensor of this metric and stand for a new effect: a radiation of gravitational waves induced by the vacuum stress. This coherent radiation caused by the backreaction adds to the noncoherent radiation caused by the pair creation in the case where the initial state provides the vacuum stress tensor with a quadrupole moment.'
title: |
The one-loop form factors in the effective action,\
and\
production of coherent gravitons from the vacuum.
---
[A.G.Mirzabekian]{}\
Lebedev Research Center in Physics, Leninskii Prospect 53,\
Moscow 117924, Russia\
[G.A.Vilkovisky]{}\
Lebedev Research Center in Physics, Leninskii Prospect 53,\
Moscow 117924, Russia\
and\
Istituto Nazionale di Fisica Nucleare, Sezione di Napoli,\
Pad. 20 Mostra d’Oltremare,Napoli 80125, Italia
Introduction
============
$\;\;$
We consider the expectation-value equations for the gravitational field in an in-state [^1]. The model-independent, or phenomenological approach \[4-7\] makes it possible to write down the general form of these equations in terms of the form factors in the vacuum action. The form factors are to be calculated from a given dynamical model. However, for obtaining predictions of various models, the expectation-value equations should be first analysed with arbitrary form factors in order to relate the properties of the form factors to the important properties of the solution \[4\].
This analysis has thus far been limited to the behaviour of the vacuum stress tensor $T^{\mu\nu}_{vac}$ at infinity \[6,7\]. It was shown \[6\] that the requirement of asymptotic flatness of the solution imposes restrictions on the behaviours of the form factors at small values of their arguments. Namely, these behaviours in one (each) of the arguments with the others fixed should be $w(0)\log (-\Box) + O(1), \Box \to -0.$ The coefficients $w(0)$ of the $\log (-\Box), \Box \to -0$ behaviours (the spectral weights at zero spectral mass) determine the rate of the vacuum radiation through the future null infinity $({\cal I}^+).$
However, at each given order in the curvature, only certain combinations of the form factors must behave in this way. The condition of finiteness of $T^{\mu\nu}_{vac}$ at ${\cal I}^+$ which brought to the result above leaves some arbitrariness in the asymptotic behaviours of the individual form factors in the basis decomposition of the action. It is this arbitrariness that allows for the existence of the effect discussed in the present paper.
The present paper deals with the problem which appeared when the field-theoretic form factors were calculated. It is worth noting that, as emphasized in \[6\], the loop expansion of field theory has a domain of validity. It is near null infinity where the results of this expansion are valid and can be used to calculate the energy of the vacuum radiation. The ultraviolet divergent terms which appear in the expectation-value equations when $T^{\mu\nu}_{vac}$ is expanded in loops are local and vanish at infinity; only nonlocal terms survive, and these are unambiguous. Technically, the loop expanded form factors are reliable in the limit of small $\Box$ arguments up to terms $O({\Box}^0), \Box \to -0.$
The field-theoretic form factors have been calculated in the one-loop approximation for a generic quantum field model in refs. \[8-15\]. Their asymptotic behaviours at $\Box \to -0$ up to terms $O({\Box}^0),$ which are of interest for the above-mentioned reason, are presented in ref. \[15\], and these behaviours offer a problem. While the second-order form factors behave as expected: they are $w(0) \log (-\Box) + O(1)$ for $\Box \to -0,$ the third-order form factors contain also the asymptotic terms $1/\Box, \Box \to -0$ which apparently violate the asymptotic flatness of the solution \[6,15\].
Since the third-order form factors are functions of three $\Box$ arguments, their behaviours in one of the arguments with the two others fixed cannot be predicted on dimensional grounds. The $1/\Box$ asymptotic terms in the form factors appear as a result of an explicit calculation of loops \[12,15\]. In the present paper we propose an explanation of this result as well as of the following remarkable fact \[12\] which one can establish by a direct inspection of the expressions in \[15\]. The inspection shows that the alarming $1/\Box$ terms appear only in the curvature invariants containing the gravitational field strength and act [*selectively only*]{} on the Ricci curvature. The matter field strengths contained in the commutator and potential curvatures (see \[15\]) remain unaffected by these terms. Thus the presence of the $1/\Box$ terms breaks the democracy of massless vacuum particles; gravitons appear to be distinguished.
Since the same vacuum action describes also the transition amplitudes between in- and out- states \[8\], the problem of the $1/\Box$ terms appears also in scattering theory where these terms either signal an infrared divergence of the on-shell amplitudes with gravitons or, in the favourable case, stand for some inelastic process. In the language of expectation values, the former case corresponds to a breakdown of the asymptotic flatness. That the situation is not hopeless is seen from the fact that the form factors are not quite the vertices; they are coefficients of the curvatures rather than the field disturbances. As compared to the field disturbances, the curvatures contain extra derivatives which, in the favourable case, may cancel the $1/\Box$ factors in the on-shell amplitudes.
Below we present the solution of the problem as it appears in expectation-value theory. As mentioned above, only certain combinations of the form factors should behave like $\log (-\Box),\Box \to -0$ to ensure finiteness of $T^{\mu\nu}_{vac}$ at ${\cal I}^+.$ We show that the $1/\Box$ terms precisely cancel in these combinations leaving indeed the $\log (-\Box)$ behaviour as the leading one. The cancellation occurs owing to certain constraints among the coefficients of the $1/\Box$ terms which we establish by an analysis of the asymptotic flatness and next check with the explicit expressions in \[15\]. The fulfilment of these constraints is by itself a powerful check on the results in \[15\] apart from the checks that have already been carried out in \[12\]. Thus we prove that the $1/\Box, \Box \to -0$ terms discovered in \[12,15\] do not violate the asymptotic flatness of the solution of the expectation-value equations. The proof is given for a generic quantum field model for which the results in \[15\] are obtained and which is characterized by a set of field strengths consisting of the Ricci, commutator, and potential curvatures. Such a general proof is possible owing to the above-mentioned fact that the $1/\Box$ operators in the asymptotic expressions for the form factors act only on the Ricci curvature. Therefore, for the consideration of the leading asymptotic terms of the equations at ${\cal I}^+,$ one does not need to know the variational derivatives of the commutator and potential curvatures with respect to the metric.
Next, we reveal the significance of the $1/\Box, \Box \to -0$ terms in the vacuum form factors. These terms vanish in the energy-momentum tensor but , as we show, they reappear in the Riemann tensor of the solution and determine its leading $O(1/r)$ behaviour at ${\cal I}^+.$ The coefficient of this behaviour is known to give the energy of the outgoing gravitational waves. Thus the $1/\Box, \Box
\to -0$ terms of the vacuum form factors discovered in \[12,15\] stand for a new effect: a generation of the gravitational waves from the vacuum. This is an effect of the backreaction of the vacuum stress on the metric. All massless particles including gravitons \[16\] contribute to $T^{\mu\nu}_{vac}$ and are radiated through the future null infinity by the quantum mechanism of pair creation. The energy of this component of radiation is determined by the $\log (-\Box), \Box \to -0$ terms of the vacuum form factors. On top of this, $T^{\mu\nu}_{vac}$ as a whole acts as a source of the gravitational field and causes a secondary radiation of gravitons. This component of radiation has the shape of a classical wave but with a quantum amplitude, and its energy is determined by the $1/\Box , \Box \to -0 $ terms of the vacuum form factors.
An important difference between the two cases is that the gravitational-wave component will be nonvanishing only if the initial state has a sufficient asymmetry to provide the vacuum stress with a quadrupole moment whereas the contribution of gravitons in $T^{\mu\nu}_{vac}|_{{\cal I}^+}$ is present even in a spherically symmetric in-state because the out-states in which these gravitons appear at ${\cal I}^+$ are squeezed vacuum states rather than coherent states (see \[17,18\] and references therein).
The paper is organized as follows. In sec. 2 we briefly review the structure of the expectation-value equations and present a new expression for the solution of the Bianchi identities to second order in the curvature. This expression simplifies obtaining the news functions of the gravitational waves. Sec. 3 contains an analysis of contributions to $T^{\mu\nu}_{vac}$ at ${\cal I}^+.$ This analysis is not complete but sufficient for obtaining the asymptotic flatness constraints and for calculating $T^{\mu\nu}_{vac}|_{{\cal I}^+}$ to second order in the curvature. We point out an important distinction of this calculation from the one in two dimensions \[3\], and also make some first step in giving the expectation-value equations a closed form. Sec. 4 contains the derivation of the asymptotic flatness constraints and the proof that they are satisfied by the one-loop form factors. In sec. 5 we propose a new method for calculating the energy of the gravitational waves and obtain the contribution of the third-order vacuum form factors to the news functions. Sec. 6 completes the calculation of the vacuum news functions in the lowest nonvanishing approximation. Appendix A contains reference equations pertaining to the behaviour of the asymptotically flat metric at null infinity. Appendix B summarizes the properties of the retarded Green function used in the text.
The expectation-value equations in an in-state
===============================================
$\mbox{}$
In the framework of quantum field theory one starts with the assumption that there exists a quantum state such that the expectation value of the metric in this state is an asymptotically flat gravitational field. Under this assumption one goes to the past null infinity $({\cal I}^-)$ of the spacetime with the expectation value of the metric and, for all massless fields, builds the Fock space of states (the in-states) based on the standard in-vacuum \[19\]. The assumed state belongs to this space [^2]. The choice of the state determines the initial data at ${\cal I}^-$ for the field’s expectation values and, generally, affects also the form of their dynamical equations since these equations are nonlocal. There exists an action which produces the expectation-value equations although the procedure by which it does so [*is not*]{} the least-action principle (see below).
It makes sense to choose the initial state in which the matter quanta form some heavy classically behaved source of the gravitational field, and gravitons are in a coherent state so that, for the mean metric, they form, generally, a classical incoming gravitational wave. Here we consider the case where such a wave is absent. The action for the expectation value of the metric in such a state can be taken as the sum $$S = S_{vac} + S_{source}$$ where $S_{source}$ is the action of a source which moves along a classical trajectory in the mean metric, and $S_{vac}$ is the action for the gravitational field in the in-vacuum state.
The action $S_{vac}$ is to be calculated from a given quantum field model. Within certain approximations (which are not completely unsatisfactory, see above) this calculation is feasible and, in the one-loop approximation, it can be done for a generic field model \[8-15\]. However, since the gravitational interaction is universal, all particles existing in nature contribute to $S_{vac},$ and, at higher loop orders, all details of their interactions matter. Therefore, if $S_{vac}$ is to be [*ultimately*]{} calculated from a model (of fields or strings or whatever), then this should be the Model and the Calculation.
In the phenomenological approach of refs. \[4-7\], the action (2.1) is viewed as an effective action (in the loose sense) for the observable field which should be a part of predictions of any fundamental dynamical theory. Irrespectively of the nature of this theory, one assumes the existence of a functional, the action $S_{vac},$ which describes the elastic properties of real vacuum i.e. its response to the introduction of a gravitationally charged source. For $S_{vac}$ one writes down the most general expansion in terms of nonlocal invariants of $N$th order in the curvature. One has to go explicitly to $N=3$ because third order in the curvature in the action corresponds to second order in the equations, and it has been shown that, at first order in the curvature, the flux of vacuum energy through ${\cal I}^+$ is pure quantum noise \[7\]. The full bases of nonlocal invariants of second order and third order are built in \[5\] for a set of field strengths consisting of the Ricci, commutator and potential curvatures: $$\Re = \left\{ R_{\mu\nu}, {\hat {\cal R}}_{\mu\nu},
{\hat P} \right\}$$ ( for the definitions see \[15\]). The explanation for the absence of the Riemann tensor from the basis invariants can be found in \[9,4,5\] but it makes sense to repeat it here since the respective equations will be of use below.
By differentiating and contracting the Bianchi identities, one obtains the equation $$\begin{aligned}
\Box R^{\alpha\beta\mu\nu} =
2 \nabla^{[\mu}\nabla^{\alpha} R^{\nu] \beta}
- 2 \nabla^{[\mu}\nabla^{\beta} R^{\nu] \alpha} -
4 R_{\:.\gamma\:.\:\sigma}^{\alpha\:[\mu} R^{\beta\gamma\nu] \sigma}
+ 2 R^{[\mu}_{\;\:\gamma} R^{\alpha\beta\gamma\nu]} - \nonumber \\
{} - R^{\alpha\beta}_{\ \ \gamma\sigma} R^{\mu\nu\gamma\sigma}
\qquad\qquad\end{aligned}$$ which can be solved iteratively with respect to the Riemann tensor. In this equation the Ricci tensor plays the role of a source, and the solution is fixed by the initial data for the gravitational field at ${\cal I}^-.$ The solution with zero data (no incoming gravitational wave) corresponds to the in-vacuum state and is expressed in terms of the retarded Green function (see appendix B):
$$R^{\alpha\beta\mu\nu} = \frac{1}{\Box} \Bigl(
4 \nabla^{[\mu}\nabla^{<\alpha} R^{\nu]\beta>} +
O[R^2_{..}] \Bigr) .$$
Here and below, $1/\Box$ stands for the retarded Green function, and both types of brackets $[\: ]$ and $< >$ denote antisymmetrization of the respective indices. Since the Riemann tensor is expressed in this way through the Ricci tensor, the nonlocal invariants with the Riemann tensor in the vacuum action are redundant.
Below we shall confine ourselves to the case where the flux components of the Ricci tensor at ${\cal I}^-$ vanish [^3]. In this case the derivatives in (2.4a) can be made external by commuting them with the Green function $1/\Box.$ To second order in the Ricci curvature the solution is then of the form
$$\begin{aligned}
R^{\alpha\beta\mu\nu}
&=& \nabla^{[\mu} \nabla^{<\alpha}
\left(\frac{4}{\Box}\right) \Bigl[ R^{\nu] \beta >} +
\Bigl( \nabla^{\nu ]} \frac{1}{\Box} R^{\gamma\delta} \Bigr)
\Bigl( \nabla^{\beta >} \frac{1}{\Box} R_{\gamma\delta} \Bigr) -
\\
& & {} - 2\Bigl( \nabla_{\gamma} \frac{1}{\Box} R^{\nu ]\delta} \Bigr)
\Bigl( \nabla_{\delta} \frac{1}{\Box} R^{\beta >\gamma} \Bigr) \Bigr] +
{} \biggl\{ 8g_{\gamma\delta}\Bigl(\frac{1}{\Box}
\nabla^{[\alpha}R^{<\mu\gamma]}\Bigr)\Bigl(\frac{1}{\Box}
\nabla^{[\beta}R^{\nu>\delta]}\Bigr) + \nonumber \\
& & {} + 8g_{\gamma\delta}\Bigl(\frac{1}{\Box}
\nabla^{[\mu}R^{<\alpha\gamma]}\Bigr)\Bigl(\frac{1}{\Box}
\nabla^{[\nu}R^{\beta>\delta]}\Bigr) -
2\Bigl(\frac{1}{\Box}
\nabla_{\gamma}R^{<\alpha[\mu}\Bigr)\Bigl(\frac{1}{\Box}
\nabla^{\gamma}R^{\beta>\nu]}\Bigr) + \nonumber \\
& & {} + 4\Bigl(\frac{1}{\Box}\nabla^{[\mu}\nabla^{\gamma}R^{
\nu]<\alpha}\Bigr)\Bigl(\frac{1}{\Box}R^{\beta>}_{\:\gamma}
\Bigr) + 4\nabla^{[\mu}\Bigl(\Bigl[\frac{1}{\Box}
\nabla^{<\alpha}R^{\beta>\gamma}\Bigr]\Bigl[
\frac{1}{\Box}R^{\nu]}_{\:\gamma}\Bigr]\Bigr)\biggr\} +
O[R^3_{..}] \nonumber\end{aligned}$$
(cf. the result in \[5,12\]). The advantage of making the derivatives external is in the appearance of the terms in the curly brackets in (2.4b) which have no overall $1/\Box$ factor. These terms [*do not contribute* ]{} to the leading asymptotic behaviour of the Riemann tensor at ${\cal I}^+ .$ On the other hand, the terms which have the overall $1/\Box$ factor have also the overall derivatives. This facilitates solving the equation for the news functions of the gravitational waves ( see sec. 5).
To third order in the curvature the vacuum action is of the form [^4] $$S_{vac} = S(1) + S(2) + S(3) + O[ \Re^4 ] ,$$ $$S(1) = \frac{1}{16\pi} \int dx g^{1/2} R ,$$ $$S(2) = \frac{1}{2{(4\pi)}^2} \int dx g^{1/2} \mbox{tr}
\sum_{i=1}^{5} \gamma_{i} (-\Box_2) \Re_1 \Re_2 (i) ,$$ $$S(3) = \frac{1}{2{(4\pi)}^2} \int dx g^{1/2} \mbox{tr}
\sum_{i=1}^{29} \Gamma_{i} (-\Box_1,-\Box_2,-\Box_3)
\Re_1 \Re_2 \Re_3 (i) ,$$ $$\Box = g^{\mu\nu} \nabla_{\mu} \nabla_{\nu}$$ where $ \Re_1 \Re_2 (i) $ with $ i=1$ to 5, and $\Re_1 \Re_2 \Re_3 (i)$ with $i=1$ to 29 are the quadratic and cubic basis invariants listed in \[15\]. This list is reduced in comparison to the full list in \[5\] because, in the trace of the heat kernel and hence in the one-loop vacuum action, the invariants linear in the commutator curvature all but one prove to be absent \[12,13\]. The only one that is present is number 13 in the list of ref. \[15\] which we use here. In low-dimensional manifolds there exist hidden constraints between nonlocal invariants, reducing the basis. In four dimensions, the second-order basis is unconstrained, and the only constraint which exists among the third-order invariants boils down to the condition that the completely symmetric part of the form factor $\Gamma_{28} $ vanishes identically \[5,12\]. In the field-theoretic form factors of refs. \[12,15\] this condition is explicitly implemented.
The commutator and potential curvatures are functions of the metric and matter fields, different in different models, but in any case their contribution to the purely gravitational sector of the action boils down to a modification of the form factors of the basis invariants with the Ricci tensor only. There are only two such in $S(2)$ and ten in $S(3)$. Below, when referring to the purely gravitational form factors, we shall assume that this reduction has already been made. The full set of invariants for gravity and matter is considered here because it is important that the maintenance of asymptotic flatness be proved for the one-loop action in full generality. Apart from this, our main concern in the discussion below is the vacuum action for the metric.
The functions $\gamma_i $ and $\Gamma_i $ in (2.7) and (2.8) are the second-order and third-order form factors. The principal assumption about these and higher-order form factors made in the axiomatic approach is their analyticity which allows one to put them in the spectral forms \[4\]. For example, the spectral form used in \[6\] for the lowest-order form factors is
$$\gamma (-\Box) = (\Box + \mu^2)^n \int\limits_0^{\infty}
\frac{dm^2}{m^2 - \Box} \frac{w(m^2)}{(m^2 + \mu^2)^n} +
\sum_{k=0}^{n-1} \frac{(-1)^k}{k!} (\Box + \mu^2)^k
{\left( \frac{\partial}{\partial \mu^2}\right)}^k
\gamma(\mu^2)$$
where $w(m^2)$ is the spectral weight $$w(m^2) = \frac{1}{2\pi i} \Bigl[ \gamma(-m^2 - i0)
- \gamma(-m^2 + i0) \Bigr] ,$$ $\mu^2 > 0$ is an arbitrary parameter on which $\gamma(-\Box)$ does not depend, and $n$ is the degree of growth of $\gamma(-\Box)$ at large $\Box$ which will presumably be fixed or bounded by the requirement of regularity of the solution (see \[1-3\]). The requirement of asymptotic flatness of the solution imposes restrictions only on the small-$\Box$ behaviours of the form factors \[6,7\]. In the small-$\Box$ limit, eq. (2.10) reduces to the simple spectral form since the terms modifying this form for $n > 0$ vanish in this limit \[6\]. On the other hand, this form should be generalized to allow for the behaviour $1/\Box, |\Box| \to 0$ of $\gamma(-\Box).$ It remained unnoticed in paper \[6\] that the derivation in this paper for the two purely gravitational second-order form factors brings in fact to the following general result: $$\begin{aligned}
\gamma_1(-\Box)& =& - \frac{2a}{\Box} - w_1(0) \log (-\Box)
+ O(1) ,\quad - \Box \to 0 \\
\gamma_2(-\Box)& =& \frac{a}{\Box} - w_2(0) \log (-\Box)
+ O(1) ,\quad - \Box \to 0\end{aligned}$$ in which there appears an arbitrary constant $a$ [^5]. Only the combination $$\gamma_1(-\Box) + 2\gamma_2(-\Box) =
- \Bigl(w_1(0) + 2w_2(0) \Bigr) \log(-\Box) + O(1), \; -\Box \to 0$$ should behave like $\log (-\Box)$ by the analysis in \[6\]. As will be seen below, a similar situation takes place for the higher-order form factors.
In the form factors $\gamma_i$ calculated from field theory, $a=0$ \[9,15\] since, by dimension, the terms $1/\Box$ cannot appear in the loop expansion of the second-order form factors. They appear, however, already in the third-order form factors $\Gamma_i$ \[15\]. For the generalized spectral forms of these form factors see \[10,12\].
With the form factors in the spectral forms, the only nonlocal structure that remains in $ S_{vac}$ is the inverse operator $1/{(m^2 - \Box)}.$ This simplifies the procedure of obtaining the expectation-value equations. When the action $S_{vac}$ is varied, the inverse operators are regarded as obeying the variational rule $$\delta \frac{1}{m^2 - \Box} = \frac{1}{m^2 - \Box} \delta\Box
\frac{1}{m^2 - \Box} \; ,$$ and, after the variation has been completed, all inverse operators are replaced by the retarded Green functions [^6]. If $${\left.\frac{\delta S_{vac}}{\delta g_{\mu\nu}(x)}\right|}_{\Box
\to \Box_{ret}}$$ is to denote the result of this procedure, and a further notation is introduced to separate the classical term of $S_{vac}:$ $$T^{\mu\nu}_{vac} \equiv \frac{2}{g^{1/2}}{\left. \frac{\delta
S_{vac}}{\delta g_{\mu\nu}}\right|}_{\Box \to \Box_{ret}} -
\frac{2}{g^{1/2}} \frac{\delta S(1)}{\delta g_{\mu\nu}}\; ,$$ then the expectation-value equations corresponding to the action (2.1) are of the form $$R^{\mu\nu} - \frac{1}{2} g^{\mu\nu} R =
8\pi \Bigl( T^{\mu\nu}_{vac} + T^{\mu\nu}_{source}\Bigr)$$ where $$T^{\mu\nu}_{source} = \frac{2}{g^{1/2}}
\frac{\delta S_{source}}{\delta g_{\mu\nu}} \; ,$$ and $T^{\mu\nu}_{vac}$ in (2.17) can be interpreted as the energy-momentum tensor of the in-vacuum. These equations are to be solved with zero initial data for the gravitational field at ${\cal I}^-.$ One arrives at a Cauchy problem \[2,3\] for nonlocal equations with the retarded kernels which are to be integrated from ${\cal I}^-$ to the future until the solution hits a singularity if there remains one. One hopes that it doesn’t.
We do not consider the more general initial data, with a gravitational wave at ${\cal I}^-,$ because in this case the action should also be calculated more generally. Specifically, the solution of eq. (2.3) can no more be taken in the form (2.4).
The structure of $T^{\mu\nu}_{vac}$ at ${\cal I}^+$
===================================================
$\mbox{}$
It is natural to begin the study of the expectation-value equations with the behaviour of $T^{\mu\nu}_{vac}$ at null infinity since the nonlocal terms of the equations should be responsible for the effect of the vacuum radiation.
At the future null infinity one has the Bondi-Sachs equation \[20,21\] (see also appendix A) which is an exact consequence of the expectation-value equations : $$\begin{aligned}
- \frac{dM(u)}{du} = \frac{1}{4\pi} \int d^2 {\cal S}
\left[ \left(\frac{\partial}{\partial u} C_1 \right)^2 +
\left(\frac{\partial}{\partial u} C_2 \right)^2 \right] +
\int d^2 {\cal S} \Bigl( \frac{1}{4} r^2 T^{\mu\nu}_{source}
\nabla_{\mu}v \nabla_{\nu}v + \nonumber \\
{} + \frac{1}{4} r^2 T^{\mu\nu}_{vac}
\nabla_{\mu}v \nabla_{\nu}v \Bigr) \biggl|_{{\cal I}^+} ,\qquad\end{aligned}$$ $$\begin{aligned}
(\nabla u)^2 = 0, {(\nabla u, \nabla r)\Bigr|}_{{\cal I}^+} = -1\:, \\
(\nabla v)^2 = 0, {(\nabla u, \nabla v)\Bigr|}_{{\cal I}^+} = -2\:.\end{aligned}$$ Here $u$ is the retarded time along ${\cal I}^+$ with the natural normalization in (3.2), the integrals are over the 2-sphere ${\cal S}$ (normalized to have the area $4\pi$) at which the null congruence $u=$const. crosses ${\cal I}^+,
r $ is the luminosity distance along the rays of this congruence, $M(u)$ is the Bondi mass, and ${\partial C_1}/{\partial u} ,
{\partial C_2}/{\partial u} $ are the Bondi-Sachs news functions of the gravitational waves.
Eq. (3.1) is the conservation law missing in the theory of quantum fields on a fixed gravitational background. In the collapse problem, this is the backreaction equation relating “the changing mass of the black hole” with the energy of the quanta radiated by this black hole. In full quantum theory, both the “black hole” and the quantum fields “on its background” evolve from one and the same initial state, and one is able to answer the question where does the black-hole radiation take its energy from. It takes it ultimately from the energy of the collapsing source $T^{\mu\nu}_{source}$ which equals the ADM mass of the expectation value of the metric and serves as an initial datum $M(-\infty)$ for eq. (3.1).
The last term of eq. (3.1) is the flux of the vacuum energy through ${\cal I}^+.$ For it to be finite, the flux component of $T^{\mu\nu}_{vac}$ should decrease at ${\cal I}^+$ like $1/r^2.$ This is a necessary condition of asymptotic flatness. Below, terms $O(1/r^3)$ in $T^{\mu\nu}_{vac}$ will be referred to as vanishing at ${\cal I}^+.$
When computing $T^{\mu\nu}_{vac}(x)$ at ${\cal I}^+$ from the action (2.5), all terms in which the Ricci curvature appears at the observation point $x$ can be discarded because $R_{\mu\nu}$ decreases at least like $1/r^2$ and will always be multiplied by a decreasing function. Thus, the covariant derivatives $\nabla$ which appear in the basis invariants $\Re\Re\Re (i),$ etc. need not be varied because the contributions of their variations to $T^{\mu\nu}_{vac}$ contain the curvature at the observation point. For the same reason, in the expression $$\begin{aligned}
\delta R^{\gamma}_{\nu} = \frac{1}{2} g^{\gamma\mu}
\Bigl( \nabla_{\mu}\nabla^{\alpha} \delta g_{\nu\alpha}
+\nabla_{\nu}\nabla^{\alpha} \delta g_{\mu\alpha} -
\nabla_{\mu}\nabla_{\nu} g^{\alpha\beta}
\delta g_{\alpha\beta} - \Box \delta g_{\mu\nu} + \nonumber \\
{} + 2 R_{\mu .\:.\:\nu}^{\;\:\alpha\beta} \delta g_{\alpha\beta}
\Bigr) +
\frac{1}{2} R^{\gamma\alpha} \delta g_{\nu\alpha} -
\frac{1}{2} g^{\gamma\alpha} R^{\beta}_{\nu}
\delta g_{\alpha\beta} \; ,\end{aligned}$$ the terms with the Ricci tensor can be discarded but the term with the Riemann tensor cannot since the Riemann tensor has components decreasing like $1/r.$ In (3.4), the expression for $\delta R^{\gamma}_{\nu}$ has been brought by commutations to the form used below.
Only the variations of the Ricci tensors and the variations of the form factors in $S_{vac}$ can give nonvanishing contributions to $T^{\mu\nu}_{vac}$ at ${\cal I}^+.$ It is easy to see that the variation of a form factor $$\Gamma ( -\Box_1, -\Box_2, ... , -\Box_N )$$ in the argument $\Box_p$ can contribute at ${\cal I}^+$ only if, in this argument, the form factor behaves like $1/\Box_p ,
\Box_p \to -0.$ Thus, assuming $a=0$ in eqs. (2.12), (2.13), as is the case in the field theoretic form factors, one can calculate the variations of $\gamma_i(-\Box)$ in the action (2.7) by using the spectral form (2.10) in which, moreover, the terms appearing at $n > 0 $ can be disregarded when the observation point tends to ${\cal I}^+.$ One finds $$\int dx g^{1/2} \Re_1 \delta \gamma (-\Box) \Re_2 =
\int dx g^{1/2} \int\limits^{\infty}_{0} dm^2 w(m^2)
\Bigl(\frac{1}{m^2 - \Box} \Re_1 \Bigr) \delta \Box
\Bigl(\frac{1}{m^2 - \Box} \Re_2 \Bigr) \: .$$ Since, by the result in \[6\], $$\begin{aligned}
\Bigl(\frac{1}{(m^2 - \Box_{ret})}\: \Re (x)\Bigr)\biggl|_{x
\to {\cal I}^+} \propto
r^{-1}(x) \exp \Bigl( - |\mbox{const.}| m \sqrt{r(x)}\Bigr) (1 +
{\cal O}) , \\
{\cal O} \to 0, r(x) \to \infty, x \to {\cal I}^+,\nonumber\end{aligned}$$ and $$w(0) = \mbox{finite},$$ one concludes that the contribution of $\delta \gamma (-\Box)$ to $T^{\mu\nu}_{vac}$ at ${\cal I}^+$ is $O(1/r^3).$ This result applies to all form factor of the form (3.5) provided that their behaviours in individual arguments are $ O(\log(-\Box)), \Box \to -0.$ On the other hand, with the behaviour $$\Gamma (-\Box_1, -\Box_2, -\Box_3, ... ) =
\frac{1}{\Box_1} F(\Box_2, \Box_3, ... ) (1 + {\cal O}),
\quad {\cal O} \to 0, \Box_1 \to -0$$ the variation of $\Box_1$ in the term $$\int dx g^{1/2} \Gamma (-\Box_1, -\Box_2, -\Box_3, ... )
\Re_1 \Re_2 \Re_3 \;...$$ of the action is asymptotically of the form $$- \int dx g^{1/2} \Bigl(\frac{1}{\Box} \Re_1 \Bigr) \delta \Box
\frac{1}{\Box} \Bigl( F( \Box_2, \Box_3, ... )\Re_2 \Re_3 \; .\:.\:.
\Bigr)$$ and gives a nonvanishing contribution to $T^{\mu\nu}_{vac}$ at ${\cal I}^+$ proportional to $1/r^2.$
The fact that, in four dimensions, the variations of the second-order form factors do not contribute to the energy flux at infinity makes an important distinction of this case from the case in two dimensions where $\gamma (-\Box)
\propto 1/\Box $ and the relevant contribution comes from $\delta \gamma (-\Box)$ \[3\]. The variations of the third-order form factors $\Gamma_i $ in (2.8) can already contribute to $T^{\mu\nu}_{vac}$ at ${\cal I}^+$ owing to the $1/\Box, \Box \to -0 $ terms discovered in \[15\] but, to second order in the curvature (in the equations), only the Ricci tensors in the action $S(3)$ need to be varied.
For all terms of the vacuum action, in which the form factors are of the form (3.5)[^7], it is useful to introduce a quantity, [*the generalized current*]{}, defined by varying the action with respect to the Ricci tensor only. The sum of such terms in $S_{vac}$ can be represented in the form $${\tilde S}_{vac} = \frac{1}{2(4\pi)^2} \sum_{N}
\sum_{i} S_{i}(N) \;,$$ $$S_{i}(N) = \int dx g^{1/2} P_1(\nabla_1) .\:.\:. P_N(\nabla_N)
\Gamma (-\Box_1, ... , -\Box_N) R^{\cdot}_{\cdot}(x_1)
.\:.\:. R^{\cdot}_{\cdot}(x_N)$$ where (3.13) is a contribution of the $N$th order basis invariant number “ $i$ ”, and the sum in (3.12) extends over both $i$ and $N.$ In (3.13),$ R^{\cdot}_{\cdot} $ are the Ricci tensors with mixed indices, and the polynomials in covariant derivatives $P_{n}
(\nabla_n)$ which are generally present in the tensor basis \[15,5\] act on the respective $R^{\cdot}_{\cdot}(x_n)$ [*after*]{} the action of the operator arguments $\Box_n$ of $\Gamma.$ In this representation all operators $\Box_n$ in $\Gamma$ are uniformly defined as applied to a mixed second-rank tensor, and the advantage of taking it mixed is in the absence of $g^{\mu\nu}$ factors contracting indices in (3.13) which otherwise would need to be varied. These factors are generally present in the polynomials $P(\nabla)$ but, in any case, their variations ( as well as the variation of $g^{1/2}$ in the measure) do not contribute to $T^{\mu
\nu}_{vac}$ at ${\cal I}^+.$ The order of operations in (3.13) is different from the one in (2.8) (cf. the explicit expressions in \[15\]) but, at every order in the curvature, the action can be brought to the form (3.13) by commutations.
Since the argument $\Box_n$ of $\Gamma$ is the first operator acting on $R^{\cdot}_{\cdot}(x_n)$ in (3.13), it will be the last in the variational derivative of (3.13) with respect to $R^{\cdot}_{\cdot}(x_n).$ The generalized current $I^{\mu}_{\nu}(\xi, x)$ is then defined by the relation $$\delta_{R}{\tilde S}_{vac} =
\frac{1}{2(4\pi)^2} \int dx g^{1/2}
I^{\mu}_{\nu}(-{\stackrel{\rule{.1mm}{1.95mm}\rule[1.85mm]{17mm}{.1mm}
\hspace{-.6mm}\downarrow}{\Box, x) \delta R^{\nu}_{\mu}(x)}} =
\frac{1}{2(4\pi)^2} \int dx g^{1/2}
I^{\mu}_{\nu}(-{\stackrel{\rule{.1mm}{1.95mm}\rule[1.85mm]{6mm}{.1mm}
\hspace{-.6mm}\downarrow}{\Box\:\strut, x\;}}) \delta R^{\nu}_{\mu}(x)$$ where the notation $\delta_{R}$ points out that only the Ricci tensors in (3.13) are varied, and the argument $\Box$ of $I^{\mu}_{\nu}(-\Box, x)$ is the argument of the form factor $\Gamma$ in (3.13) that acts on the varied Ricci tensor. The $I^{\mu}_{\nu}(\xi, x)$ is a tensor function of the spacetime point $x$ and a function of a parameter $\xi$ which in eq. (3.14) gets replaced by the operator $-\Box.$ This operator next acts in either of the two ways pointed out in (3.14).
Given the action of the form (3.12), it is easy to calculate $I^{\mu}_{\nu}(\xi, x)$ to each given order in the curvature. For the action (2.5) we have $$I^{\alpha}_{\beta}(\xi, x) = I^{\;\alpha}_{2\beta}(\xi, x) +
I^{\;\alpha}_{3\beta}(\xi, x) + O[\Re^3]$$ where $I^{\;\alpha}_{2\beta}(\xi, x)$ and $ I^{\;\alpha}_{3\beta}(\xi, x)$ are the contributions of $S(2)$ and $S(3)$ respectively, and $$I^{\;\alpha}_{2\beta}(\xi, x) = 2 \gamma_1(\xi)
R^{\alpha}_{\beta}(x) + 2 \gamma_2 (\xi) \delta^{\alpha}_{\beta}
R(x) \; ,$$ $$\begin{array}{ll}
I^{\;\alpha}_{3\beta}(\xi, x) = &
\;\;\;3 \Bigl[ \delta^{\mu}_{\nu} \Gamma_{10} ( \xi, -\Box_1, -\Box_2)
R^{\alpha}_{\mu}(x_1) R^{\nu}_{\beta}(x_2) \Bigr]_{x_1=x_2=x}
+ \\
& {} + \Bigl[ \nabla_{1}^{\;\alpha}\nabla_{2\beta} \Gamma_{22}
( \xi, -\Box_1, -\Box_2) R(x_1) R(x_2)\Bigr]_{x_1=x_2=x } - \\
& {} - 2\delta^{\alpha}_{\beta} \nabla_{\mu} \Bigl[ \nabla_{2}^{\;\nu}
\Gamma_{22}( -\Box_1, -\Box_2, \xi ) R^{\mu}_{\nu}(x_1)
R(x_2) \Bigr]_{x_1=x_2=x} + .\:.\:.
\end{array}$$ where only the purely gravitational terms are written down. We do not present the latter expression in full but exemplify it with the contributions of two third-order invariants, number 10 and number 22 (see \[15\]). Eq. (3.17) illustrates the general structure of the current $I^{\mu}_{\nu}(\xi, x).$
A remarkable property of $I^{\mu}_{\nu}(\xi, x)$ is that the variations of both the Ricci tensors and the form factors in (3.12) are expressed entirely through this quantity. Indeed, in addition to (3.14), we have $$\delta_{\Gamma}{\tilde S}_{vac} = \frac{1}{2(4\pi)^2} \int
dx g^{1/2} \delta \Box' \frac{1}{\Box'- \Box''}
\Bigl[ I^{\nu}_{\mu}(-\Box', x'') - I^{\nu}_{\mu}(-\Box'', x'') \Bigr]
R^{\mu}_{\nu}(x')$$ where the notation $\delta_{\Gamma}$ points out that only the form factors $\Gamma$ in (3.13) are varied, and we have used the general formula for a variation of an operator function \[12,14\]: $$\int dx g^{1/2} A \left(\delta f(\Box)\right) B =
\int dx g^{1/2} \delta \Box_{B} \frac{f(\Box_A) -
f(\Box_B)}{\Box_A - \Box_B} A B \; .$$ It is understood that $\Box_A$ (or $\Box_B$) is the operator $\Box$ acting on $A$ (or $B$), and similarly in (3.18) $\Box'$ acts on $x',$ and $\Box''$ on $x''$ with subsequently setting $x'=x''=x.$ The operators $\delta \Box_B$ and $\Box_B,$ and similarly $\delta \Box'$ and $\Box'$ in (3.18), do not commute and act in the indicated order. The identity $$\int dx g^{1/2} ( \delta \Box'' - \delta \Box' ) F(x',x'') =
\int dx g^{1/2} (-\delta \log g^{1/2}) ( \Box'' - \Box' )
F(x',x'')$$ (with an arbitrary two-point tensor $F(x', x'')$ contracting into a scalar at $x'= x''$) serves to check that varying the left-hand side of the equality $$\int dx g^{1/2} f(\Box_B) A B = \int dx g^{1/2} f(\Box_A)
A B$$ with the aid of eq. (3.19) gives the same result as varying its right-hand side.
Although the generalized current $I^{\mu}_{\nu}(\xi, x),$ as calculated from the action (2.5), is given in the form of an expansion, it enters the expectation-value equations as a single whole and determines the vacuum stress at null infinity. Indeed, since (3.14) and (3.18) are the only contributions surviving in (2.16) when the observation point tends to ${\cal I}^+, T^{\mu\nu}_{vac}$ at ${\cal I}^+$ is obtained in a closed form. As seen from (3.16), (3.17), the behaviour of $I^{\mu}_{\nu}(\xi, x)$ in $\xi$ includes all the behaviours of the form factors in individual arguments. The $I^{\mu}_{\nu}(\xi, x)$ may, therefore, be of a significance in axiomatic theory. For this current as a function of its parameter argument one must postulate the existence of a spectral form similar to (2.10) but allowing for the spectral weight to have a $\delta(m^2) $ singularity at $m^2 = 0.$ Assuming for simplicity $n = 0$ in (2.10) (the modification concerning the large $\Box$ is irrelevant to the present discussion) we set $$I^{\mu}_{\nu}(\xi, x) = \int\limits^{\infty}_{-\epsilon}
\frac{dm^2}{m^2 + \xi}\: w^{\mu}_{\nu}(m^2, x) \; ,$$ and then eqs. (3.14) and (3.18) take the form $$\begin{aligned}
\delta_{R}{\tilde S}_{vac} = \frac{1}{2(4\pi)^2}
\int dx g^{1/2} \int\limits^{\infty}_{-\epsilon} dm^2
\delta R^{\nu}_{\mu}(x)\Bigl[ \frac{1}{m^2 - \Box}\:
w^{\mu}_{\nu}(m^2, x)\Bigr] \; , \\
\delta_{\Gamma}{\tilde S}_{vac} = \frac{1}{2(4\pi)^2}
\int dx g^{1/2} \int\limits^{\infty}_{-\epsilon} dm^2
\Bigl[\delta \Box \frac{1}{m^2 - \Box}
R^{\nu}_{\mu}(x) \Bigr] \Bigl[ \frac{1}{m^2 - \Box}
\:w^{\mu}_{\nu}(m^2, x)\Bigr] \; .\end{aligned}$$ Along with expression (3.4) and an easily derivable expression for $\delta \Box $ in (3.24) they determine $T^{\mu\nu}_{vac}$ at ${\cal I}^+.$
The asymptotic flatness constraints
====================================
$\mbox{}$
Even with the $1/\xi, \xi \to 0$ behaviour of $I^{\mu}_{\nu}(\xi, x),$ the contribution to $T^{\mu\nu}_{vac}$ coming from (3.24) is $O(1/r^2)$ at ${\cal I}^+.$ Also the contributions coming from the term with the Riemann tensor and the term with the $\Box$ operator in (3.4) are $O(1/r^2).$ We have, therefore, from (3.14) and (3.4) $$\begin{aligned}
T^{\mu\nu}_{vac}(x) = \frac{1}{2(4\pi)^2} \Bigl(
\nabla^{\mu}\nabla_{\alpha}I^{\alpha\nu}
(-{\stackrel{\rule{.1mm}{1.95mm}\rule[1.85mm]{9mm}{.1mm}
\hspace{-.6mm}\downarrow}{\Box_{ret},\strut x\;}}) +
\nabla^{\nu}\nabla_{\alpha}I^{\alpha\mu}
(-{\stackrel{\rule{.1mm}{1.95mm}\rule[1.85mm]{9mm}{.1mm}
\hspace{-.6mm}\downarrow}{\Box_{ret},\strut x\;}}) - \qquad\qquad \\
{} - g^{\mu\nu} \nabla_{\beta}\nabla_{\alpha}I^{\alpha\beta}
(-{\stackrel{\rule{.1mm}{1.95mm}\rule[1.85mm]{9mm}{.1mm}
\hspace{-.6mm}\downarrow}{\Box_{ret},\strut x\;}}) \Bigr) +
\; O(r^{-2}(x)),\;\; x \to {\cal I}^{+} \quad \nonumber\end{aligned}$$ $$I^{\alpha\beta}(\xi, x) = g^{(\alpha \gamma}(x) I^{\beta)}_{
\gamma}(\xi, x)\; .$$ The terms of order $r^{-2}(x)$ in (4.1) are the ones to be retained but it is not our purpose here to calculate $T^{\mu\nu}_{vac}$ [^8]. The point is that the remaining terms in (4.1) should also behave like $O(r^{-2}(x))$ for the asymptotic flatness to be maintained. This behaviour should , moreover, hold at every order in the curvature because even small disturbances of the metric can violate the asymptotic flatness. This does not mean, however, that the function $$I^{\alpha\mu}(-{\stackrel{\rule{.1mm}{1.95mm}\rule[1.85mm]{9mm}{.1mm}
\hspace{-.6mm}\downarrow}{\Box_{ret},\strut x\;}})$$ should behave like $O(r^{-2}(x));$ it suffices that $$\nabla_{\alpha}I^{\alpha\mu}(-{\stackrel{\rule{.1mm}{1.95mm}\rule[1.85mm]{9mm}{.1mm}
\hspace{-.6mm}\downarrow}{\Box_{ret},\strut x\;}})
= O(r^{-2}(x))\;, \;\; x \to {\cal I}^+.$$ Owing to this fact, the $1/\xi, \xi \to 0$ behaviour of $I^{\alpha
\mu}(\xi, x)$ is not completely ruled out but condition (4.4) imposes a constraint on the coefficient of this behaviour.
For implementing this constraint the derivative $\nabla_{\alpha}$ in (4.4) should be commuted with the operator $\Box_{ret}:$ $$\begin{aligned}
\nabla_{\alpha}I^{\alpha\mu}(-{\stackrel{\rule{.1mm}{1.95mm}\rule[1.85mm]{9mm}{.1mm}
\hspace{-.6mm}\downarrow}{\Box_{ret},\strut x\;}}) =
\int\limits^{\infty}_{-\epsilon} dm^2
\Bigl( \frac{1}{m^2 - \Box_{ret}} \nabla_{\alpha}
w^{\alpha\mu}(m^2, x) + \qquad\quad \\
\qquad\qquad{} + \frac{1}{m^2 - \Box_{ret}}
[\nabla_{\alpha}, \Box] \frac{1}{m^2 - \Box_{ret}}
w^{\alpha\mu}(m^2, x) \Bigr)\; . \nonumber\end{aligned}$$ Here all operators act to the right on $x,$ and $$w^{\mu\nu}(m^2, x) = g^{(\mu\gamma}(x)\;
w^{\nu)}_{\gamma}(m^2, x)\; .$$ Both terms in (4.5) are generally $O(r^{-1}(x))$ but the commutator term contains an extra power of the curvature. If we denote $$w^{\mu}(m^2, x) \equiv \nabla_{\alpha}w^{\alpha\mu}(m^2, x)
+ [\nabla_{\alpha}, \Box]\: \frac{1}{m^2 - \Box_{ret}}\:
w^{\alpha\mu}(m^2, x)$$ and $$I^{\mu}(\xi, x) \equiv \int\limits^{\infty}_{0}
\frac{dm^2}{m^2 + \xi}\: w^{\mu}(m^2, x)\; ,$$ so that $$\nabla_{\alpha}I^{\alpha\mu}
(-{\stackrel{\rule{.1mm}{1.95mm}\rule[1.85mm]{9mm}{.1mm}
\hspace{-.6mm}\downarrow}{\Box_{ret},\strut x\;}}) =
I^{\mu}(-{\stackrel{\rule{.1mm}{1.95mm}\rule[1.85mm]{9mm}{.1mm}
\hspace{-.6mm}\downarrow}{\Box_{ret},\strut x\;}})\; ,$$ then, by (4.4) and the result in \[6\], the behaviour of the function (4.8) should already be $$I^{\mu}(\xi, x) = -\; w^{\mu}(0, x) \log \xi + O(1)\;, \; \xi \to 0$$ $$w^{\mu}(0, x) = \mbox{finite.}$$
The vector current (4.8) is obtained from (3.15) as an expansion: $$I^{\mu}(\xi, x) = I_2^{\mu}(\xi, x) + I^{\mu}_3(\xi, x) +
O[\Re^3]$$ where $I^{\mu}_2$ and $I^{\mu}_3$ are the contributions of the actions $S(2)$ and $S(3).$ For the contribution of the action $S(2)$ one finds by using eq. (3.16): $$\begin{aligned}
I^{\mu}_2(-{\stackrel{\rule{.1mm}{1.95mm}\rule[1.85mm]{9mm}{.1mm}
\hspace{-.6mm}\downarrow}{\Box_{ret},\strut x\;}})
= \Bigl(\gamma_1(-\Box_{ret}) +
2\gamma_2(-\Box_{ret})\Bigr)\nabla^{\mu}R +
2\:[\nabla_{\alpha}, \gamma_1(-\Box_{ret})] R^{\alpha\mu} +
\quad \\
{} + 2\:[\nabla^{\mu}, \gamma_2(-\Box_{ret})] R\; , \nonumber\end{aligned}$$ and $$I^{\mu}_2(\xi, x) = \Bigl( \gamma_1(\xi) + 2\gamma_2(\xi)
\Bigr) \nabla^{\mu}R(x) + O[\Re^2]$$ whence the constraint (2.14) for the second-order form factors immediately follows. Should the constant $a$ in (2.12),(2.13) be nonvanishing, the commutator terms in (4.13) would contribute to the constraint condition for the third-order form factors. Since, however, $a=0$ as discussed above, these commutator terms are $O(1/r^2)$ as seen from their spectral forms.
The constraint condition for the third-order form factors is thus of the form $$I_3^{\mu}(-{\stackrel{\rule{.1mm}{1.95mm}\rule[1.85mm]{9mm}{.1mm}
\hspace{-.6mm}\downarrow}{\Box_{ret},\strut x\;}})
+ O[\Re^3] = O\Bigl( r^{-2}(x)\Bigr)
\; ,\;\; x \to {\cal I}^+$$ and $$I_3^{\mu}(\xi, x) = \nabla_{\alpha}I_3^{\alpha\mu}(\xi, x) +
O[\Re^3]$$ since, in this case, the commutator of $\nabla_{\alpha}$ with $\xi = -\Box$ contributes to $O[\Re^3]$ already. On the other hand, with the field-theoretic form factors \[15\] one obtains $$\begin{aligned}
I_3^{\alpha\mu}(\xi, x)& =& \frac{1}{\xi} A^{\alpha\mu}(x) +
(\log \xi) B^{\alpha\mu}(x) + O(\xi^0)\;, \\
I_3^{\mu}(\xi, x)& =& \frac{1}{\xi} A^{\mu}(x) + O(\log \xi) \; ,
\xi \to 0\; ,\\
A^{\mu}(x)& =& \nabla_{\alpha}A^{\alpha\mu}(x) + O[\Re^3]\end{aligned}$$ where the coefficients $A^{\mu}(x),$ etc. can be expanded in some vector basis second-order in the curvature: $$A^{\mu}(x) = {\mbox{tr}}\sum_{p}A_p(\Box_1, \Box_2)
{\Re_1\Re_2}^{\: \mu}(p) + O[\Re^3]\;.$$ Examples of the basis structures in (4.20) are $$\begin{aligned}
{\Re_1\Re_2}^{\:\mu}(1)& =& \nabla^{\mu}R_1\cdot R_2 {\hat 1}\;, \\
{\Re_1\Re_2}^{\:\mu}(2)& =& \nabla^{\mu}R_1^{\alpha
\beta}\cdot\nabla_{\alpha}\nabla_{\beta}R_2 {\hat 1}\;, \\
{\Re_1\Re_2}^{\:\mu}(3)& =& {\hat {\cal R}}_1^{\alpha\beta}\nabla^{\mu}
{\hat {\cal R}}_{2\alpha\beta} \;, \\
{\Re_1\Re_2}^{\:\mu}(4)& =& R_1^{\mu\alpha}
\nabla_{\alpha}{\hat P}_2 \;,\end{aligned}$$ etc. where all curvatures of the set (2.2) participate, and the trace in (4.20) refers to the matrices ${\hat 1},{\hat P},$ etc. (cf. a similar construction of the basis of invariants in \[15\]). We do not present the basis in (4.20) in full although it is important to have it in full for obtaining the results below. In order that (4.15) hold for any choice of the in-state [^9], there should be $$A^{\mu}(x) + O[\Re^3] = 0$$ and hence $$A_p(\Box_1, \Box_2) = 0\; .$$
Eqs. (4.26) are the constraints to be satisfied by the coefficients of the $1/\Box$ asymptotic behaviours of the third-order form factors. Let us introduce a notation for these coefficients: $$\begin{aligned}
\Gamma_{i}(-\Box_1,-\Box_2, -\Box_3) =
\frac{1}{\Box_1}F^1_i(\Box_2, \Box_3) + O(\log(-\Box_1))\;,
\Box_1 \to -0 \\
\Gamma_{i}(-\Box_1,-\Box_2, -\Box_3) =
\frac{1}{\Box_2}F^2_i(\Box_3, \Box_1) + O(\log(-\Box_2))\;,
\Box_2 \to -0 \\
\Gamma_{i}(-\Box_1,-\Box_2, -\Box_3) =
\frac{1}{\Box_3}F^3_i(\Box_1, \Box_2) + O(\log(-\Box_3))\;,
\Box_3 \to -0\end{aligned}$$ where there appear functions of two variables $F^m_i$ with $m=1$ to 3 and $i=1$ to 29. The functions $A_p$ in (4.20) are certain linear combinations of the $F^m_i.$ By an explicit calculation with the action $S(3)$ in (2.8) and \[15\] one can work up these combinations to see if they vanish. The commutator and potential curvatures appear in the basis in (4.20) but the contributions of $\delta
{\hat {\cal R}}_{\mu\nu}$ and $\delta {\hat P}$ to (3.14) may, in this calculation, be omitted since all $F^m_i$ with $\Box_m$ acting on ${\hat {\cal R}}_{\mu\nu} $ or ${\hat P}$ vanish \[12,15\]. This makes it possible to carry out the check of asymptotic flatness for a generic quantum field model.
The results are as follows. Of $3\times 29$ functions $F^m_i$ only 21 in the table of ref. \[15\] do not vanish and are not related to each other by the symmetries of the form factors. With the 21 nonvanishing $F^m_i$ the expansion (4.20) gives rise to 14 constraints (4.26) which, by linearly combining them, can be brought to the following form: $$\begin{aligned}
F^3_{25}(\Box_1, \Box_2)& =& \frac{1}{2}(\Box_1 - \Box_2)
F^3_{28}(\Box_1, \Box_2) \; ,\\
F^1_{10}(\Box_1, \Box_2)& =& -\frac{1}{12}(\Box_2 - \Box_1)^2
F^3_{28}(\Box_1, \Box_2) \; , \\
F^1_{28}(\Box_1, \Box_2)& =& 2 F^3_{27}(\Box_1, \Box_2) +
\frac{3}{2}(\Box_2 - \Box_1)F^1_{29}(\Box_1, \Box_2) \; ,\\
F^3_{24}(\Box_1, \Box_2)& =&\frac{1}{2}F^1_{25}(\Box_1, \Box_2)
- \frac{1}{4}(\Box_1 - \Box_2)F^1_{28}(\Box_2, \Box_1) \; , \\
F^3_{22}(\Box_1, \Box_2)& =& - \frac{1}{2}F^3_{24}(\Box_1, \Box_2)
+ \frac{1}{8}(\Box_1 + \Box_2)F^1_{28}(\Box_1, \Box_2) \; , \\
F^1_{11}(\Box_1, \Box_2)& =& - \frac{1}{4}(\Box_2 - \Box_1)
F^1_{23}(\Box_1, \Box_2)
\;, \\
F^1_{23}(\Box_1, \Box_2)& =& -2 F^3_{22}(\Box_1, \Box_2) +
(\Box_1 - \Box_2)F^1_{27}(\Box_1, \Box_2) + \\
& {} +{} & \frac{1}{4}(\Box_1 + \Box_2)F^1_{28}(\Box_2, \Box_1) +
\frac{3}{4}(\Box_2 - \Box_1)\Box_1 F^1_{29}(\Box_1, \Box_2)
\; , \nonumber\\
F^1_9(\Box_1, \Box_2)& =& \frac{1}{32}(\Box_1 + \Box_2)
(\Box_1 - \Box_2)^2 F^1_{29}(\Box_1, \Box_2) - \\
& & {} - \quad\frac{1}{24}
(\Box_1^2 + \Box_2^2) F^3_{27}(\Box_1, \Box_2) \; ,\nonumber \\
F^1_{22}(\Box_1, \Box_2)& =& - \frac{1}{4}F^1_{25}(\Box_1, \Box_2)
+ \frac{3}{8}(\Box_1 - \Box_2)^2 F^1_{29}(\Box_1, \Box_2) \; ,\\
F^1_5(\Box_1, \Box_2)& =& \frac{1}{4}(\Box_1 - \Box_2)^2
F^1_{26}(\Box_1, \Box_2) \; , \\
F^1_{16}(\Box_1, \Box_2)& =& \quad(\Box_1 - \Box_2)
F^1_{26}(\Box_1, \Box_2) \; , \\
F^1_8(\Box_1, \Box_2)& =& \frac{1}{4}(\Box_1 + \Box_2)
F^1_{21}(\Box_1, \Box_2) \; , \\
F^1_{18}(\Box_1, \Box_2)& =& \frac{1}{2}
F^1_{21}(\Box_1, \Box_2) \; , \\
F^1_{19}(\Box_1, \Box_2)& =& - \frac{1}{4}
F^1_{21}(\Box_1, \Box_2) \; .\end{aligned}$$ It is now a matter of a direct inspection to check if the $F^m_i$ calculated in \[15\] satisfy these constraints. They do!
Relations (4.30) - (4.43) leave only 7 independent nonvanishing $F^m_i$ for which one can take the functions $ F^1_{21}, F^1_{25}, F^1_{26}, F^1_{27},
F^3_{27}, F^3_{28},
F^1_{29}.$ With the exception of $F^1_{26}$ and $F^1_{27},$ these functions are symmetric in their $\Box$ arguments: $F^1_{25},
F^3_{27}$ and $F^3_{28}$ are symmetric owing to the respective symmetries of the form factors $\Gamma_{25}, \Gamma_{27}$ and $\Gamma_{28},$ and the symmetry of $F^1_{21}, F^1_{29}$ is a property of the explicit expressions in \[15\].
By expressing all $F^m_i$ through the 7 independent ones, one can bring the coefficient $A^{\alpha\mu}(x)$ in (4.17) to the form $$A^{\mu\nu}(x) = - \nabla_{\alpha} \nabla_{\beta}
K^{[\mu\alpha][\nu\beta]}(x) + O[\Re^3]$$ in which the fulfilment of condition (4.25) is manifest, and the function $K^{\mu\alpha\nu\beta}(x)$ which appears antisymmetrized in (4.44) is of the following form:
$$\begin{aligned}
\; K^{\mu\alpha\nu\beta}(x)& =& 3 {\mbox{tr}}F^1_{29}(\Box_1, \Box_2)
\Bigl\{ 4 \nabla^{\mu}\nabla^{\nu} R^{\gamma\sigma}_1
\cdot \nabla_{\gamma}\nabla_{\sigma}R^{\alpha\beta}_2 +
2 ( \Box_2 - \Box_1) \Bigl[ \nabla^{(\mu}R_1 \cdot
\nabla^{\nu)}R^{\alpha\beta}_2 \Bigr] - \nonumber\\
& & {} - \Box_2 (\Box_2 - \Box_1) g^{\mu\nu} \Bigl[ R_1^{\alpha\beta}
\cdot R_2 \Bigr] \Bigr\}{\hat 1} + {}
{\mbox{tr}}F^3_{28}(\Box_1, \Box_2)\Bigl\{ 4\nabla^{\mu}R^{\alpha
\gamma}_1 \cdot \nabla^{\nu}R^{\:\:\beta}_{2\gamma}\Bigr\}
{\hat 1} + \nonumber \\
& & {}+ {\mbox{tr}}F^1_{27}(\Box_1, \Box_2)\Bigl\{ 8\nabla^{\mu}
\nabla^{\nu}R_1^{\alpha\beta} \cdot R_2 \Bigr\}
{\hat 1} + 8\;{\mbox{tr}}F^1_{26}(\Box_1, \Box_2)\Bigl\{
\nabla^{\mu}\nabla^{\nu}R_1^{\alpha\beta} \cdot
{\hat P}_2 \Bigr\} + \nonumber\\
& & {} + {\mbox{tr}}F^1_{25}(\Box_1, \Box_2)\Bigl\{
4 R_1^{\mu\beta}\cdot R^{\nu\alpha}_2 + 2g^{\mu\nu}
R_1^{\alpha\beta}\cdot R_2 \Bigr\} {\hat 1} + \nonumber \\
& & {}+ \frac{1}{2}{\mbox{tr}}F^1_{21}(\Box_1, \Box_2)\Bigl\{
{\hat {\cal R}}_1^{\mu\alpha}\cdot {\hat {\cal R}}_2^{\nu\beta}
\Bigr\} +{} {\mbox{tr}}F^3_{27}(\Box_1, \Box_2)\Bigl\{
8 \nabla^{(\mu}R_1^{\nu)\gamma}\cdot \nabla_{\gamma}
R_2^{\alpha\beta} + \nonumber \\
& & {} + 4\:g^{\mu\nu}\nabla_{\gamma}R_1^{
\sigma\alpha}\cdot \nabla_{\sigma}R_2^{\gamma\beta}
+ 4 \nabla^{(\mu}R_1 \cdot \nabla^{\nu)}R_2^{\alpha
\beta} + \\
& & {} + 2 (\Box_1 - \Box_2) g^{\mu\nu}\Bigl[ R_1 \cdot R_2^{\alpha
\beta}\Bigr] + g^{\mu\nu} \nabla^{\alpha}\nabla^{\beta}R_1
\cdot R_2 \Bigr\}{\hat 1} \; .\nonumber\end{aligned}$$
The one-loop expressions for the functions $F$ entering (4.45) are given in \[15\] but one may conjecture that the 7 independent structures in (4.45) and (4.44) (5 in the case of pure gravity) is the general result independent of models and approximations (although this remains to be checked by repeating the analysis above for the general form of the action). As will be seen below, expression (4.45) gets considerably simplified when inserted in the formula for the news functions.
The news functions. Contribution of the
========================================
third-order form factors {#third-order-form-factors .unnumbered}
========================
$\mbox{} $
A significance of the $1/\Box$ asymptotic terms in the vacuum form factors is that they contribute to the energy of the outgoing gravitational waves. This energy is the term with the news functions ${\partial C_1}/{\partial u}, {\partial C_2}/{\partial u}$ in the Bondi-Sachs equation (3.1).
Obtaining the news functions requires solving the dynamical equations already. However, there is a short cut : eq. (2.4). We may use the fact that the news functions appear as a coefficient of the $1/r$ behaviour of the Riemann tensor at null infinity. Indeed, we have (see appendix A) $$\nabla_{\alpha}v \nabla_{\mu}v\: m_{\beta} m_{\nu}
R^{\alpha\beta\mu\nu}\biggl|_{{\cal I}^+} =
-\frac{8}{r}\: \frac{\partial^2}{{\partial u}^2}\:{\bf C}
+ O\left( \frac{1}{r^2}\right)$$ where $${\bf C} = C_1 + i C_2 \; ,$$ and $m_\beta$ is a complex null vector tangent to the 2-sphere ${\cal S}$: $$(m, \nabla u) = (m, \nabla v) = (m, m) = 0 ,\quad (m, m^*) = -2$$ with $ \nabla u, \nabla v$ and ${\cal S}$ in (3.1)-(3.3), and $m^*$ complex conjugate to $m.$ The contribution of the outgoing gravitational waves to the mass loss, eq. (3.1), is then $$\frac{1}{4\pi} \int d^2 {\cal S} \left|\: \frac{\partial}{
\partial u}\: {\bf C}\:\right|^2 \; ,$$ and $$-\frac{\partial^2}{{\partial u}^2}\: {\bf C}
= \left\{\frac{r}{8}\:\nabla_{\alpha}v \nabla_{\mu}v \:
m_{\beta} m_{\nu}R^{\alpha\beta\mu\nu}\right\}
\Biggl|_{{\cal I}^+}\;.$$
On the other hand, the Riemann tensor can be calculated with the aid of eq. (2.4). Its $1/r$ behaviour is then obtained as [*the leading asymptotic behaviour of the retarded Green function*]{}. Using (2.4b) we have $$-\frac{\partial^2}{{\partial u}^2}\: {\bf C}
= \left\{\frac{r}{2}\nabla_{\alpha}v \nabla_{\mu}v \:
m_{\beta} m_{\nu} \nabla^{[\mu}\nabla^{<\alpha}
\frac{1}{\Box}\Bigl( R^{\nu]\beta>} + O[R^2_{..}]
\Bigr)\right\}\biggl|_{{\cal I}^+}$$ where the quadratic terms are to be copied from (2.4b), and the quadratic terms with no overall $1/\Box$ factor - all terms in the curly brackets in (2.4b) - do not contribute.
Expression (5.6) can next be simplified as follows. The derivatives in (5.6) appear projected either as $\nabla^{\nu}v \nabla_{\nu}$ or as $m^{\nu} \nabla_{\nu}.$ In both cases the projected derivatives can be commuted with the remaining factors of $\nabla v$ and $m$ since the components of $\nabla \nabla
v$ and $\nabla m$ in the null tetrad basis are $O(1/r)$ at ${\cal I}^+.$ The projected derivatives become then acting on some scalar $X$ which behaves like $1/r$ at ${\cal I}^+,$ and in this case $$\begin{aligned}
m^{\nu} \nabla_{\nu} X\biggl|_{{\cal I}^+} & =
& O\left( \frac{1}{r^2} \right) \; , \\
\nabla^{\nu}v \nabla_{\nu}X\biggl|_{{\cal I}^+} &
= & -2\: \frac{\partial}{\partial u}\: X + O\left(
\frac{1}{r^2} \right)\end{aligned}$$ (see appendix A). In this way one obtains $$\frac{\partial^2}{{\partial u}^2}\: {\bf C}
= -\: \frac{\partial}{\partial u}\:
\left\{\:\frac{\partial}{\partial u}\; \frac{r}{2} \:
m_{\beta} m_{\nu}\: \frac{1}{\Box}\:\Bigl( R^{\nu\beta}
+ O[R^2_{..}] \Bigr)\right\}\Biggl|_{{\cal I}^+} \; .$$
Eq. (5.9) can now be integrated over $u$ from $-\infty$ to a given point of ${\cal I}^+$ to obtain the news functions. Since, at $u = -\infty, {\partial {\bf C}}/{\partial u} = 0,$ there remains to be shown that the expression in the curly brackets in (5.9) also vanishes at $u = -\infty.$ This is shown in appendix B. As a result one obtains $$\begin{aligned}
\frac{\partial}{\partial u}\: {\bf C}& = &
- \frac{\partial}{\partial u}\: \Bigl\{\: \frac{r}{2}\:m_{\beta}
m_{\nu} \frac{1}{\Box}\Bigl[ R^{\nu\beta} +
(\nabla^{\nu} \frac{1}{\Box}R^{\gamma\delta})(\nabla^{\beta}
\frac{1}{\Box}R_{\gamma\delta}) - \qquad\\
& &\qquad\qquad\quad\qquad {} - 2(\nabla_{\gamma} \frac{1}{\Box}R^{\nu\delta}
)(\nabla_{\delta} \frac{1}{\Box}R^{\beta\gamma})
+ O[R^3_{..}] \Bigr] \Bigr\} \Biggl|_{{\cal I}^+}\nonumber\end{aligned}$$ where for the Ricci tensor one can use the dynamical equations [^10].
There is, of course, a classical gravitational radiation or, more generally, a radiation induced by the classical source $T^{\mu\nu}_{source}$ in eq. (2.18). The respective contribution[^11] to the news functions (call it ${\partial {\bf C}_{source
}}/{\partial u}$) is obtained by substituting for $R^{\nu\beta}$ its Einstein value $$R^{\nu\beta}_{cl} = 8\pi \left( T^{\nu\beta}_{source} -\:
\frac{1}{2}\: g^{\nu\beta} T_{source} \right)$$ in all terms of (5.10) including $ O[R^3_{..}].$ The remaining contributions in (5.10) stand for the gravitational radiation induced by the vacuum stress, $$\frac{\partial}{\partial u}\: {\bf C} =
\frac{\partial}{\partial u}\: {\bf C}_{source} + \:
\frac{\partial}{\partial u}\: {\bf C}_{vac}\; .$$ By (5.10) and the dynamical equations (2.18)[^12], $$\frac{\partial}{\partial u}\: {\bf C}_{vac} =
- \frac{\partial}{\partial u}\: \Bigl\{ 4\pi r\: m_{\beta} m_{\nu}
\frac{1}{\Box}\Bigl(\: T^{\nu\beta}_{vac} +\:
O[R^2_{..}] \Bigr) \Bigr\}\biggl|_{{\cal I}^+}$$ and $$T^{\mu\nu}_{vac} = T^{\mu\nu}_{vac}(2) +
T^{\mu\nu}_{vac}(3) + O[R^3_{..}]$$ where $T^{\mu\nu}_{vac}(2)$ and $T^{\mu\nu}_{vac}(3)$ are the contributions of the actions $S(2)$ and $S(3)$ in (2.5).
In (5.10) and (5.13) there appears an expression of the form $$\Bigl\{ r\:m_{\beta} m_{\nu} \frac{1}{\Box} X^{\nu\beta}
\Bigr\} \biggl|_{{\cal I}^+}$$ with some tensor $X^{\mu\nu},$ and we have used the fact that a contribution of this form with $$X^{\mu\nu} = g^{\mu\nu} X$$ vanishes by the orthogonality relation in (5.3). Another property of expression (5.13) which will be used below is that a contribution of the form (5.15) with $$X^{\mu\nu} = \nabla^{\mu} X^{\nu} +
\nabla^{\nu} X^{\mu}$$ is of a higher order in the curvature : $$O[ X^{\mu}\times R_{..} ] \; .$$ Indeed, in this case, $$m_{\beta}m_{\nu}\: \frac{1}{\Box} X^{\nu\beta} =
2 m_{\beta}m_{\nu} \nabla^{\nu} \frac{1}{\Box} X^{\beta} +
2 m_{\beta}m_{\nu} \frac{1}{\Box} \Bigl(
[\nabla^{\nu}, \Box] \frac{1}{\Box} X^{\beta}\Bigr) \; .$$ By (5.7), the first term of this expression is $O(1/r^2),$ and the remaining term contains a commutator.
To lowest order in the curvature, only the contribution of the action $S(2)$ is to be considered. By (3.4) and (3.14), $$\begin{aligned}
T^{\mu\nu}_{vac}(2) & = & \frac{1}{2(4\pi)^2}
\Bigl( \nabla^{\mu}\nabla_{\alpha}I^{\alpha\nu}_2 +
\nabla^{\nu}\nabla_{\alpha}I^{\alpha\mu}_2 -
g^{\mu\nu}\nabla_{\beta}\nabla_{\alpha}I^{\alpha\beta}_2
- \qquad \\
& & \qquad\qquad\qquad\qquad\qquad\qquad\qquad\quad
{} - \Box I^{\mu\nu}_2 \Bigr) + O[ R^2_{..} ]\nonumber\end{aligned}$$ where $$I^{\mu\nu}_{2} = I_{2}^{\mu\nu}(-
{\stackrel{\rule{.1mm}{1.95mm}\rule[1.85mm]{9mm}{.1mm}
\hspace{-.6mm}\downarrow}{\Box_{ret},\strut x\;}}) \; ,$$ and the explicit form of $I_2(\xi, x)$ is given in (3.16). The first two terms in (5.20) are of the form (5.17), and, therefore, their contribution to (5.13) is $O[R^2_{..}].$ The third term in (5.20) is of the form (5.16), and, therefore, its contribution to (5.13) vanishes. In the remaining term of (5.20),the $\Box$ operator kills the Green function: $$\frac{\partial}{\partial u}\: {\bf C}_{vac} =
\:\frac{\partial}{\partial u}\: \Bigl\{\: \frac{r}{8\pi}\:
m_{\beta}m_{\nu}I_2^{\nu\beta}(
-{\stackrel{\rule{.1mm}{1.95mm}\rule[1.85mm]{9mm}{.1mm}
\hspace{-.6mm}\downarrow}{\Box_{ret},\strut x\;}})
\Bigr\}\biggl|_{{\cal I}^+} + O[R^2_{..}] \; ,$$ [*and this is the reason why at all the asymptotic behaviours of the form factors at small*]{} $\Box$ [*are relevant to the gravitational waves.*]{} Since the second-order form factors $\gamma_1, \gamma_2$ behave like $\log (-\Box)$ at small $\Box,$ we have $$I_2^{\nu\beta}( -
{\stackrel{\rule{.1mm}{1.95mm}\rule[1.85mm]{9mm}{.1mm}
\hspace{-.6mm}\downarrow}{\Box_{ret},\strut x\;}}) = O\left( \frac{1}{r^2}\right)
\;,\qquad x \to {\cal I}^+$$ and the contribution (5.22) vanishes. Thus the vacuum contribution to the news functions begins with second order in the curvature: $$\frac{\partial}{\partial u}\: {\bf C}_{vac} = O[R^2_{..}]\; .$$
At second order in the curvature both $S(2)$ and $S(3)$ contribute to ${\bf C}_{vac}.$ We are presently interested in the contribution of $S(3)$ which we shall denote ${\bf C}_{vac}(3).$ Since the action $S(3)$ is cubic in the curvature, the accuracy in (5.13) is sufficient for calculating this contribution : $$\frac{\partial}{\partial u}\: {\bf C}_{vac}(3) =
\:-\: \frac{\partial}{\partial u}\: \Bigl\{ 4\pi r\: m_{\beta} m_{\nu}
\frac{1}{\Box}\Bigl( T^{\nu\beta}_{vac}(3) +
O[R^3_{..}] \Bigr) \Bigr\}\biggl|_{{\cal I}^+}~~~.$$ To second order in the curvature, $T^{\mu\nu}_{vac}(3)$ is of the form similar to (5.20) : $$\begin{aligned}
T^{\mu\nu}_{vac}(3) & = & \frac{1}{2(4\pi)^2}
\Bigl( \nabla^{\mu}\nabla_{\alpha}I^{\alpha\nu}_3 +
\nabla^{\nu}\nabla_{\alpha}I^{\alpha\mu}_3 -
g^{\mu\nu}\nabla_{\beta}\nabla_{\alpha}I^{\alpha\beta}_3
- \qquad \\
& &\qquad\qquad\qquad\qquad\qquad\qquad\qquad\quad
{} - \Box I^{\mu\nu}_3 \Bigr) + O[ R^3_{..} ] \nonumber\end{aligned}$$ with $$I^{\mu\nu}_3 = I^{\mu\nu}_3(-{\stackrel{\rule{.1mm}{1.95mm}\rule[1.85mm]{9mm}{.1mm}
\hspace{-.6mm}\downarrow}{\Box_{ret},\strut x\;}})\; .$$ Again, the first two terms in (5.26) are of the form (5.17), and their contribution to (5.25) is $O[R^3_{..}].$ Again the third term is proportional to the metric, and its contribution vanishes. Again the $\Box$ operator in the remaining term kills the Green function: $$\frac{\partial}{\partial u}\: {\bf C}_{vac}(3) =
\:\frac{\partial}{\partial u}\: \Bigl\{\: \frac{r}{8\pi}\:
m_{\beta}m_{\nu}I_3^{\nu\beta}(-
{\stackrel{\rule{.1mm}{1.95mm}\rule[1.85mm]{9mm}{.1mm}
\hspace{-.6mm}\downarrow}{\Box_{ret},\strut x\;}})
\Bigr\}\biggl|_{{\cal I}^+} +\; O[R^3_{..}]$$ but this time the contribution (5.28) does not vanish because the third-order form factors behave like $1/\Box$ at small $\Box.$ By (4.17), $$I^{\nu\beta}_{3}(-{\stackrel{\rule{.1mm}{1.95mm}\rule[1.85mm]{5mm}{.1mm}
\hspace{-.6mm}\downarrow}{\Box,\strut x\;}}) =
- \Bigl({\stackrel{\rule{.1mm}{1.95mm}\rule[1.85mm]{14mm}{.1mm}
\hspace{-.6mm}\downarrow}{\frac{1}{\Box}\Bigr) A^{\nu\beta}(x)}}
+ \log(-{\stackrel{\rule{.1mm}{1.95mm}\rule[1.85mm]{13mm}{.1mm}
\hspace{-.6mm}\downarrow}{\Box)\strut B^{\nu\beta}(x)}}
+\; O\left(\frac{1}{r^3}\right)\; ,$$ and only the term with $1/\Box$ survives in (5.28). With the expression for $A^{\nu\beta}(x)$ given in (4.44) the result is $$\frac{\partial}{\partial u}\: {\bf C}_{vac}(3) =
\:\frac{\partial}{\partial u}\: \Bigl\{\: \frac{r}{8\pi}
\:m_{\mu}m_{\nu}\:\frac{1}{\Box} \nabla_{\alpha}
\nabla_{\beta}K^{[\mu\alpha][\nu\beta]}(x) +
O[R^3_{..}] \Bigr\}\biggl|_{{\cal I}^+} \; .$$
Expression (4.45) for $K^{\mu\alpha\nu\beta}$ can now be simplified by using that (i) all terms containing the metric with the indices of $K^{\mu\alpha\nu\beta}$ can be discarded since any contraction among $m_{\mu}m_{\nu}
\nabla_{\alpha} \nabla_{\beta}$ in (5.30) results in a vanishing contribution, and (ii) any derivative $\nabla$ with the indices of $K^{\mu\alpha\nu\beta}$ can be treated as in integration by parts because the respective total derivative contracts with either $m$ or $\nabla,$ and its contribution vanishes. As a result, $K^{\mu\alpha
\nu\beta}$ in (5.30) can be replaced by the following expression : $$\begin{aligned}
& &{\tilde K}^{\mu\alpha\nu\beta}(x) = 12 {\mbox{tr}}F^1_{29}(
\Box_1, \Box_2) \Bigl[\nabla^{\mu}\nabla^{\nu}R^{\gamma
\sigma}_1 \cdot \nabla_{\gamma}\nabla_{\sigma} R^{\alpha
\beta}_2 \Bigr] {\hat 1} + \\
& & {} + 4 {\mbox{tr}}F^3_{28}(\Box_1, \Box_2)
\Bigl[\nabla^{\mu}R^{\alpha\gamma}_1 \cdot\nabla^{\nu}
R^{\;\beta}_{2\gamma} \Bigr] {\hat 1} +
8 {\mbox{tr}}F^3_{27}(\Box_1, \Box_2)
\Bigl[\nabla^{(\mu}R^{\nu)\gamma}_1
\cdot \nabla_{\gamma}R^{\alpha\beta}_2 \Bigr] {\hat 1} +
\nonumber \\
& & {} + 4 {\mbox{tr}}F^1_{25}(\Box_1, \Box_2) \Bigl[ R_1^{\mu\beta}\cdot
R_2^{\nu\alpha} \Bigr] {\hat 1}
+ {\mbox{tr}}\Bigl\{ 6(\Box_2 - \Box_1) F^1_{29}(\Box_1, \Box_2) +
8 F^1_{27} (\Box_1, \Box_2) - \nonumber \\
& & {} - 4 F^3_{27}(\Box_1, \Box_2)\Bigr\}
\Bigl[ \nabla^{\mu}\nabla^{\nu} R^{\alpha\beta}_1 \cdot R_2
\Bigr] {\hat 1} + \;
8 {\mbox{tr}}F^1_{26}(\Box_1, \Box_2) \Bigl[ \nabla^{\mu}
\nabla^{\nu} R^{\alpha\beta}_1 \cdot {\hat P}_2 \Bigr] +\nonumber \\
& & \qquad\qquad\qquad {} + \frac{1}{2} {\mbox{tr}}F^1_{21}(\Box_1, \Box_2)
\Bigl[ {\hat {\cal R}}^{\mu\alpha}_1 \cdot
{\hat {\cal R}}^{\nu\beta}_2 \Bigr]\; .\qquad\qquad\nonumber\end{aligned}$$
The vacuum news functions in the lowest
========================================
nonvanishing approximation {#nonvanishing-approximation .unnumbered}
==========================
$\mbox{} $
To complete the calculation of the vacuum news functions in the lowest nonvanishing approximation we must consider the contribution of the action $S(2):$ $$\begin{aligned}
\frac{\partial}{\partial u}\: {\bf C}_{vac}(2)& =&
- \frac{\partial}{\partial u}\Bigl\{ 4 \pi r m_{\beta}
m_{\gamma} \frac{1}{\Box} \Bigl[ T^{\beta\gamma}_{vac}(2)\;
+ 2 \Bigl( \nabla^{\beta} \frac{1}{\Box}R^{cl}_{\mu\nu}\Bigr)
\Bigl(\nabla^{\gamma}\frac{1}{\Box} T^{\mu\nu}_{vac}(2)\Bigr) -
\nonumber\\
& & {} - \Bigl( \nabla^{\beta} \frac{1}{\Box}R^{cl}\Bigr)
\Bigl(\nabla^{\gamma}\frac{1}{\Box} T_{vac}(2)\Bigr) -
\: 4 \Bigl( \nabla_{\sigma} \frac{1}{\Box}R_{cl}^{\beta\mu}\Bigr)
\Bigl(\nabla_{\mu}\frac{1}{\Box} T^{\gamma\sigma}_{vac}(2)\Bigr) +
\nonumber \\
& & {} + 2 \Bigl( \nabla^{\gamma} \frac{1}{\Box}R_{cl}^{\beta\mu}\Bigr)
\Bigl(\nabla_{\mu}\frac{1}{\Box} T_{vac}(2)\Bigr)\Bigr] +
O( \hbar^2 ) + O[ R^3_{..} ] \Bigr\}\biggl|_{{\cal I}^+}\; .\end{aligned}$$ Here $R^{\mu\nu}_{cl}$ is the notation in (5.11), and in the nonlinear terms of eq. (5.10) we omitted the contributions $T_{vac}\times
T_{vac} = O(\hbar^2)$ of second order in the Planck constant. As shown above, expression (6.1) is of second order in the curvature; the contribution of first order in this expression vanishes.
The expression (5.20) for $T^{\mu\nu}_{vac}(2)$ completed with terms $O[ R^2_{..} ]$ is of the form $$\begin{aligned}
T^{\mu\nu}_{vac}(2)& = & \frac{1}{2(4\pi)^2} \Bigl(
\nabla^{\mu}\nabla_{\alpha} I_2^{\alpha\nu} +
\nabla^{\nu}\nabla_{\alpha} I_2^{\alpha\mu} -
g^{\mu\nu} \nabla_{\beta}\nabla_{\alpha}I_2^{\alpha
\beta} - \Box I_2^{\mu\nu} + \\
& &\qquad\qquad\qquad{} + 2 R_{\alpha .\:.\:\beta}^{\:\:\mu\nu}
I_2^{\alpha\beta}
{} + \frac{1}{2}g^{\mu\nu} R_{\alpha\beta}
I_2^{\alpha\beta}\Bigr) + \Pi^{\mu\nu} \nonumber\end{aligned}$$ where $$\Pi^{\mu\nu} = \frac{2}{g^{1/2}}\frac{\delta_{\gamma}S(2)}{
\delta g_{\mu\nu}}\biggl|_{\Box \to \Box_{ret}}$$ with $$\delta_{\gamma}S(2) = \frac{1}{2(4\pi)^2} \int dx g^{1/2}
\Bigl( R^\mu_\nu\: \delta \gamma_1( -\Box) R^\nu_\mu +
R\: \delta \gamma_2( -\Box) R \Bigr)$$ is the contribution of the variations of the lowest-order form factors (see eqs. (3.19), (3.24)). When inserting expression (6.2) in the linear term of (6.1), all terms in (6.2) proportional to $g^{\mu\nu}$ or to the $\Box$ operator can be omitted, and for the contribution of the first two terms in (6.2) one can use eq. (5.19).
In the nonlinear terms of (6.1), the expression (6.2) is needed only up to $O[ R^2_{..}].$ By using (3.16), we obtain $$\begin{aligned}
T^{\mu\nu}_{vac}(2) = \frac{1}{(4\pi)^2} \Bigl[
\nabla^{\mu}\nabla^{\nu} \Bigl(\gamma_1(-\Box) +
2\gamma_2(-\Box)\Bigr) R - \Box \gamma_1(-\Box)
R^{\mu\nu} - \nonumber \\
{} - \frac{1}{2}\: g^{\mu\nu} \Box \Bigl( \gamma_1(-\Box) +
4 \gamma_2(-\Box) \Bigr) R \Bigr] + O[ R^2_{..} ] \; .\end{aligned}$$ This makes it possible to calculate also the Riemann tensor with accuracy $O[R^2_{..}] :$ $$\begin{aligned}
R^{\alpha\beta\mu\nu} = R^{\alpha\beta\mu\nu}_{cl}
- \frac{2}{\pi} \gamma_1(-\Box) \nabla^{[\mu}\nabla^{
<\alpha}R^{\nu] \beta>} + \frac{1}{4\pi}\Bigl(
\gamma_1(-\Box) + 2\gamma_2(-\Box)\Bigr) \times \nonumber \\
\times \Bigl( g^{\nu\beta} \nabla^{\mu}\nabla^{\alpha} -
g^{\nu\alpha} \nabla^{\mu}\nabla^{\beta} -
g^{\mu\beta} \nabla^{\nu}\nabla^{\alpha} +
g^{\mu\alpha} \nabla^{\nu}\nabla^{\beta}\Bigr) R
+ O[ R^2_{..} ] \; .\end{aligned}$$ Here we used eq. (2.4a) which defines also $$R^{\alpha\beta\mu\nu}_{cl} = \frac{1}{\Box} \Bigl(
2 \nabla^{\mu}\nabla^{<\alpha}R^{\nu\beta>}_{cl} -
2 \nabla^{\nu}\nabla^{<\alpha}R^{\mu\beta>}_{cl} +
O[ R^2_{..} ] \Bigr) \; .$$
Finally, by combining the results above, the following expression is obtained for the contribution ${\partial {\bf C}_{vac}(2)}/{\partial u}$ to the news functions: $$\begin{aligned}
\frac{\partial}{\partial u}\: {\bf C}_{vac}(2)& = &
- \frac{\partial}{\partial u} \biggl\{ \frac{r}{4\pi}
m_{\mu}m_{\nu} \frac{1}{\Box}\Bigl(
\gamma_1(-\Box_2) \Bigl[ 4\nabla_{\alpha}\frac{1}{\Box}
R^{\mu\beta}_{1\: cl}\cdot \nabla_{\beta}R^{\nu\alpha}_{
2\: cl} + \\
& & {} + 4 \nabla^{\alpha}\nabla^{\mu} \frac{1}{\Box} R^{\beta
\nu}_{1\: cl}\cdot R^{cl}_{2\alpha\beta} -
2\nabla^{\alpha}\nabla^{\beta}\frac{1}{\Box}R^{\mu\nu}_{
1\: cl}\cdot R^{cl}_{2\alpha\beta} +
R^{\mu\nu}_{1\: cl}\cdot R_{2\: cl} \Bigr] + \nonumber\\
& &{} + \Bigl( \gamma_1(-\Box_2) + 2 \gamma_2(-\Box_2)
\Bigr)\Bigl[ \nabla^{\mu}\frac{1}{\Box}R_{1\: cl}\cdot
\nabla^{\nu}R_{2\: cl} - 2 \nabla^{\nu}\frac{1}{\Box}
R^{\mu\alpha}_{1\: cl}\cdot \nabla_{\alpha}R_{2\: cl}+ \nonumber \\
& &{} + \frac{1}{2}\frac{1}{\Box}R^{\mu\nu}_{1\: cl}\cdot
\Box R_{2\: cl}
{} - \frac{1}{2}R^{\mu\nu}_{1\: cl}\cdot R_{2\: cl} \Bigr] +
(4\pi)^2 \Pi^{\mu\nu}\Bigr) + O(\hbar^2) +
O[ R^3_{..} ] \biggr\} \Biggl|_{{\cal I}^+} \; .\nonumber\end{aligned}$$ This contribution involves the lowest-order vacuum form factors in the whole range of their dependence on the $\Box$ argument.
The total result is $$\frac{\partial}{\partial u}\:{\bf C} = \frac{\partial}{\partial u}
\:{\bf C}_{source} +
\frac{\partial}{\partial u}\:{\bf C}_{vac}(2) +
\frac{\partial}{\partial u}\:{\bf C}_{vac}(3) + O(\hbar^2) +
O[ R^3_{..} ]$$ with ${\partial {\bf C}_{vac}(2)}/{\partial u}$ in (6.8) and ${\partial {\bf C}_{vac}(3)}/{\partial u}$ in (5.30). Note that the news functions appear squared in the mass-loss formula (3.1), and the vacuum contribution to ${\partial {\bf C}}/{\partial u}$ begins with second order in the curvature, eq. (5.24). Therefore, in the absence of a classical radiation, the energy of the vacuum gravitational waves is of order $O[ \hbar^2 R^4_{..} ]$ which makes this effect difficult to be noticed in perturbation theory.
The authors are grateful to B.S.DeWitt and L.P.Grishchuk for helpful discussions.
The present work was supported in part by the Russian Foundation for Fundamental Research (Grant 93-02-15594), the International (Soros) Science Foundation (Grant MQY000), the EC grant INTAS-93-493, and the NATO travel grant CRG 920991.
Appendix A. The asymptotically flat metric at null infinity {#appendix-a.-the-asymptotically-flat-metric-at-null-infinity .unnumbered}
===========================================================
$\mbox{} $
The general asymptotically flat metric in a chart covering ${\cal I}^+$ is built by considering a congruence of null hypersurfaces $u =$ const. generated by the light rays reaching ${\cal I}^+.$ The generators are labeled by two parameters $\theta, \varphi$ taking values on a 2 - sphere ${\cal S},$ and the luminosity distance $r$ is used as a parameter along the generators. The metric is then of the form \[21\] $$ds^2 = - V du^2 + 2 \Psi du dr + r^2 f_{ab}(dx^{a} -
U^a du)(dx^b - U^b du)$$ where $$\begin{aligned}
a, b = 1,2 ;& &\quad\quad x^1 = \theta,\; x^2 = \varphi\; ,
\qquad\qquad \nonumber \\
f_{ab}dx^a dx^b& = &\frac{1}{2}(e^{2\gamma} + e^{2\delta})
d\theta^2 + (e^{\gamma - \delta} - e^{\delta - \gamma})
\sin \theta d\theta d\varphi + \\
& &\qquad\qquad\quad \qquad\qquad\qquad\qquad {}
+\: \frac{1}{2}(e^{-2\gamma} + e^{-2\delta})\sin^2\theta
d\varphi^2 \nonumber\end{aligned}$$ and $$\begin{aligned}
& & \frac{1}{\Psi} = (\nabla u, \nabla r) < 0 \; , \\
& & \Psi \biggl|_{{\cal I}^+} = -1\; .\end{aligned}$$ In this metric, $(\nabla u)^2 = 0$ is to ensure that the hypersurfaces $u =$ const. are null, $(\nabla x^a, \nabla u
) = 0 $ is to ensure that the lines $u =$ const.,$x^a =$ const. are null geodesics, det$ f_{ab} = \sin^2 \theta$ is to ensure that $r$ is the luminosity parameter along these geodesics, condition (A.3) is to ensure that this parameter is monotonic, and condition (A.4) is to choose the retarded time $u$ coincident with the proper time of an observer at large and constant $r.$
At the limit of ${\cal I}^+ (r \to \infty, u =$ const., $
\theta = $const., $\varphi = $const. ) the metric behaves as follows \[21\]: $$\begin{aligned}
V &=& 1 - \frac{2{\cal M}}{r} + O\left(\frac{1}{r^2}\right)\; , \\
\frac{\gamma + \delta}{2}& =& \frac{C_1}{r} +
O\left(\frac{1}{r^2}\right)\; , \\
\frac{\gamma - \delta}{2}& =& \frac{C_2}{r} +
O\left(\frac{1}{r^2}\right)\; , \\
U^a& =& \frac{2N^a}{r^2} + O\left(\frac{1}{r^3}\right) \; ,\\
\Psi& =& -1 - \frac{2B}{r^2} + O\left(\frac{1}{r^3}\right)\end{aligned}$$ where ${\cal M}, C_1, C_2, N^a, B$ are functions of $\theta, \varphi,
u.$
The $C_1, C_2 $ differentiated with respect to $u$ are the Bondi-Sachs news functions \[20, 21\]. In the gauge (A.1) they stand for the radiation degrees of freedom of the gravitational field. The Bondi mass ${M}(u)$ is obtained by averaging the coefficient ${\cal M}$ in (A.5) over the unit 2-sphere : $$M(u) = \frac{1}{4\pi} \int d^2{\cal S} {\cal M} =
\frac{1}{4\pi} \int\limits_0^{2\pi}d\varphi \int\limits_0^{\pi}
d\theta \sin \theta {\cal M}(\theta, \varphi, u)\; .$$ Its limiting value at $u \to - \infty$ is the ADM mass, and the difference $$M(-\infty) - M(u) = \int\limits_{-\infty}^{u} du \Bigl(
- \frac{dM}{du}\Bigr)$$ is the energy radiated away through ${\cal I}^+$ by the instant $u$ of retarded time ( see \[22\] and references therein).
Some components of the Riemann and Ricci tensors calculated in the metric (A.1) - (A.9) at ${\cal I}^+$ are as follows [^13]: $$\begin{aligned}
R_{u\theta u\theta}\biggl|_{{\cal I}^+}& =&
{} - r\: \frac{\partial^2}{{\partial u}^2}\:C_1 + O(1)\; ,\\
R_{u\theta u\varphi}\biggl|_{{\cal I}^+}& = &
{} - r \sin \theta\: \frac{\partial^2}{{\partial u}^2}\:C_2 + O(1)\; ,\\
R_{u\varphi u\varphi}\biggl|_{{\cal I}^+}& =&
r \sin^2 \theta\:\frac{\partial^2}{{\partial u}^2}\:C_1 + O(1)\; ,\\
R_{rr}\biggl|_{{\cal I}^+}& =&
{} - \frac{2}{r^4}\Bigl(C_1^2 + C_2^2 + 4B\Bigr) +
O\left(\frac{1}{r^5}\right)\; ,\\
R_{r\theta}\biggl|_{{\cal I}^+}& = &
\: - \:\frac{1}{r^2}\:\Bigl[ 2N^{\theta} +
\:\frac{1}{\sin^2 \theta}\:\frac{\partial}{\partial \theta}
\Bigl(C_1 \sin^2 \theta\Bigr) + \frac{\partial}{\partial
\varphi}\:\Bigl(\frac{C_2}{\sin \theta}\Bigr) \Bigr]
{} + O\Bigl(\frac{1}{r^3}\Bigr) \; , \\
R_{r\varphi}\biggl|_{{\cal I}^+}& = &
\: - \:\frac{1}{r^2}\:\Bigl[ 2N^{\varphi}\sin^2 \theta +
\:\frac{1}{\sin \theta}\:\frac{\partial}{\partial \theta}
\Bigl(C_2 \sin^2 \theta\Bigr) - \frac{\partial}{\partial
\varphi}\: C_1 \Bigr] + O\Bigl(\frac{1}{r^3}\Bigr) \; , \\
R_{ur}\biggl|_{{\cal I}^+}& =&
{} \:\frac{1}{r^3}\:\frac{\partial}{\partial u}\:
\Bigl(C_1^2 + C_2^2 + 4B\Bigr) + O\Bigl(\frac{1}{r^4}\Bigr)\;,\\
R_{u\theta}\biggl|_{{\cal I}^+}& = &
\: \:\frac{1}{r}\:\frac{\partial}{\partial u}\:\Bigl[ 2N^{\theta} +
\:\frac{1}{\sin^2 \theta}\:\frac{\partial}{\partial \theta}
\Bigl(C_1 \sin^2 \theta\Bigr) + \frac{\partial}{\partial
\varphi}\:\Bigl(\frac{C_2}{\sin \theta}\Bigr) \Bigr]
{} + O\Bigl(\frac{1}{r^2}\Bigr) \; , \\
R_{u\varphi}\biggl|_{{\cal I}^+}& = &
\: \:\frac{1}{r}\:\frac{\partial}{\partial u}\:
\Bigl[ 2N^{\varphi}\sin^2 \theta +
\:\frac{1}{\sin \theta}\:\frac{\partial}{\partial \theta}
\Bigl(C_2 \sin^2 \theta\Bigr) - \frac{\partial}{\partial
\varphi}\: C_1 \Bigr] + O\Bigl(\frac{1}{r^2}\Bigr) \; , \\
R_{uu}\biggl|_{{\cal I}^+}& =&
{} - \frac{2}{r^2}\:\Bigl[\frac{\partial}{\partial u}\:
{\cal M} + \Bigl( \frac{\partial}{\partial u}\:C_1\Bigr)^2 +
{} \Bigl( \frac{\partial}{\partial u}\:C_2\Bigr)^2 +
{} \frac{1}{\sin \theta}\:\partial_{a} \Bigl(
\sin \theta\:\frac{\partial}{\partial u}\: N^{a}
\Bigr) \Bigr] + \nonumber \\
& & \qquad\qquad\quad\qquad\qquad\qquad\qquad
{} + O\Bigl(\frac{1}{r^3}\Bigr) \; .\end{aligned}$$
When referring to tensors at ${\cal I}^+$ we always mean their projections on the null tetrad $\nabla u, \nabla v, m, m^{*}$ introduced in (3.2), (3.3) and (5.3) where $v$ and $m_{\alpha}$ are asymptotically of the form $$\begin{aligned}
v\Bigl|_{{\cal I}^+}& =& 2r + u + O\left(\frac{1}{r}\right)\; ,\\
m_{\alpha}\Bigl|_{{\cal I}^+}& =& r(\nabla_{\alpha}\theta +
i\sin \theta \nabla_{\alpha}\varphi) + O(1)\; .\end{aligned}$$ The null-tetrad components of physical quantities are regular at ${\cal I}^+$ i.e. are either finite or decreasing like inverse powers of $r.$ This is true specifically of tensors obtained by the action of the retarded form factors (see below). The null-tetrad vectors may be regarded as covariantly constant at ${\cal I}^+$ since the null-tetrad components of their derivatives are $O(1/r).$ Thus, up to curvature terms, $$\nabla_{\mu}\nabla_{\alpha}v\biggl|_{{\cal I}^+} =
\frac{1}{2r} (m_{\mu}m_{\alpha}^{*} +
m^{*}_{\mu}m_{\alpha}) + O[R_{..}]\;,$$ and the curvature terms are $O(1/r).$ Eqs. (5.7), (5.8) are obtained by calculating the derivatives projected on the null tetrad in terms of the Bondi-Sachs coordinates. Specifically, eq. (5.7) owes to the fact that $ m^{\alpha}\Bigl|_{{\cal I}^+} = O(1/r)$ as seen from (A.23) and (A.1).
It follows from the asymptotic expressions above that the null-tetrad components of the curvature tensor decrease at ${\cal I}^+$ like $1/r$ or faster. Specifically, $$\nabla_{\alpha}v \nabla_{\mu}v\: m_{\beta}m_{\nu}
R^{\alpha\beta\mu\nu}\biggl|_{{\cal I}^+} =
-\: \frac{8}{r}\:\Bigl(\frac{\partial^2}{{\partial u}^2}C_1 +
i\frac{\partial^2}{{\partial u}^2}C_2 \Bigr) +
O\left(\frac{1}{r^2}\right)$$ which is eq. (5.1). The energy flux component of the Ricci tensor at ${\cal I}^+$ $$R_{\mu\nu}\nabla^{\mu}v \nabla^{\nu}v\biggl|_{
{\cal I}^+} = 4R_{uu}\biggl|_{{\cal I}^+}$$ is given in eq. (A.21). By averaging (A.21) over the unit 2-sphere one obtains the relation $$- \frac{dM(u)}{du} = \frac{1}{4\pi} \int d^2{\cal S} \Bigl[
\Bigl( \frac{\partial}{\partial u}C_1 \Bigr)^2 +
\Bigl( \frac{\partial}{\partial u}C_2 \Bigr)^2 \Bigr] +
\frac{1}{8\pi} \int d^2{\cal S}\:\frac{1}{4}\: r^2
R^{\mu\nu}\nabla_{\mu}v \nabla_{\nu}v\biggl|_{{\cal I}^+}$$ which, after using the dynamical equations $$R^{\mu\nu} - \frac{1}{2}g^{\mu\nu}R = 8\pi T^{\mu\nu}_{total}
\; ,$$ becomes the conservation law (3.1). Here $T^{\mu\nu}_{total}$ is the total energy-momentum tensor which in eq. (2.18) is $$T^{\mu\nu}_{total} = T^{\mu\nu}_{source} +
T^{\mu\nu}_{vac}\; .$$
Appendix B. The retarded Green function in the past of ${\cal I}^+$ {#appendix-b.-the-retarded-green-function-in-the-past-of-cal-i .unnumbered}
===================================================================
$\mbox{}$
To lowest order in the curvature, the retarded operator $1/\Box$ acting on an arbitrary tensor source $X^{\alpha_1 \ldots \alpha_k}$ is of the following form: $$\begin{aligned}
\lefteqn{
- \frac{1}{\Box}X^{\alpha_1 \ldots \alpha_k}(x) = } & \\
&= \frac{1}{4\pi} \int\limits_{{past\: of}\; x}
d{\bar x} {\bar g}^{1/2}
\delta ( \sigma( x, {\bar x} )) g^{\alpha_1}_{\;\;{\bar \alpha}_1}
( x, {\bar x}) \ldots g^{\alpha_k}_{\;\;{\bar \alpha}_k}( x,
{\bar x}) X^{{\bar \alpha}_1 \ldots {\bar \alpha}_k}({\bar x})
+ O[X \times \Re] \nonumber\end{aligned}$$ where $\sigma(x, {\bar x})$ is the world function \[23\], or geodetic interval biscalar \[24\], $ g^{\alpha}_{\;\;{\bar
\alpha}}( x, {\bar x}) $ is the geodetic parallel displacement bivector \[24\], and the integration point ${\bar x}$ is in the past of the observation point $x.$ Here and below, the bar over a symbol means that this symbol refers to the point ${\bar x}.$
It follows from a comparison of the equations defining $ g^{\alpha}_{\;\;{\bar \alpha}}( x, {\bar x}) $ with the ones defining $\sigma(x, {\bar x})$ that, up to the curvature terms, the parallel displacement bivector can be calculated as follows: $$g^{\alpha}_{\;\;{\bar \alpha}}( x, {\bar x}) =
- \nabla^{\alpha} {\bar \nabla}_{\bar \alpha}
\sigma (x, {\bar x}) + O[R_{..}]$$ whence it also follows that $$\nabla_{\mu}
g^{\alpha}_{\;\;{\bar \alpha}}( x, {\bar x}) =
O[R_{..}]\; .$$ If $\ell^{\mu}_{i}$ with $i=1$ to 4 are the vectors of the null tetrad, then the null-tetrad components of the tensor (B.1) are obtained by calculating the contractions $$\ell^{\mu}_{i}(x) g_{\mu{\bar \mu}}( x, {\bar x}) \; .$$ By using eq. (B.2) and a perturbative expression for the world function, it is easy to see that, when $x$ tends to ${\cal I}^+,$ and ${\bar x}$ is in a compact domain, the contractions (B.4) remain finite. Hence the null-tetrad components of the tensor (B.1) decrease at ${\cal I}^+$ like $O(1/r)$ - the fact assumed in the main text. The expression for the massive retarded Green function is similar to (B.1) \[6\]. Therefore, generally, the null-tetrad components of tensors obtained by the action of the retarded form factors are regular at ${\cal I}^+.$
As seen from (B.1), we are always dealing with some scalar source $${\cal Y}({\bar x}) =
g^{\alpha_1}_{\;\;{\bar \alpha}_1}
( x, {\bar x}) \ldots g^{\alpha_k}_{\;\;{\bar \alpha}_k}( x,
{\bar x}) X^{{\bar \alpha}_1 \ldots {\bar \alpha}_k}({\bar x})$$ which may depend parametrically on the observation point but it suffices to consider the action of the Green function on a scalar : $$- \frac{1}{\Box}X^{\alpha_1 \ldots \alpha_k}(x) =
\frac{1}{4\pi} \int\limits_{{past\: of}\; x} d{\bar x} {\bar g}^{1/2}
\delta ( \sigma( x, {\bar x} )){\cal Y}({\bar x}) + O[X \times \Re] .$$ The integration over the light cone in (B.6) includes subintegrations along the light rays coming from ${\cal I}^-$ to the observation point $x$ : $$\int\limits_0^{\infty} d\mu (\rho) {\cal Y}\Bigl|_{L}\; .$$ Here $L$ is a generator of the past light cone of $x, \rho$ is the luminosity parameter along $L,$ and the measure in (B.7) is asymptotically of the form $$d \mu (\rho)\Bigl|_{\rho \to \infty} = d \rho \cdot \rho\; .$$ We shall, therefore, assume that the source decreases at ${\cal I}^-$ like $$X^{\alpha_1 \ldots \alpha_k}\Bigl|_{{\cal I}^-} =
O\left( \frac{1}{r^3}\right)\; .$$ Another important assumption \[6\] is analyticity of the source in time including the past timelike infinity $(i^-).$ The real sources appearing in the expectation-value equations are built out of the curvature, and the condition of analyticity implies in particular that, at $i^-,$ the metric becomes asymptotically static. This should be provided by imposing the respective requirement on $T^{\mu\nu}_{source}.$ By analyticity, the limit $r \to \infty$ of the source at $i^-$ coincides with its limit in the past of ${\cal I}^-:$ $$\Bigl(X^{\alpha_1 \ldots \alpha_k}
\Bigl|_{i^-}\Bigr)\biggr|_{r \to \infty} =
\Bigl(X^{\alpha_1 \ldots \alpha_k}
\Bigl|_{{\cal I}^-}\Bigr)\biggr|_{v \to -\infty}$$ where $v$ is the advanced time along ${\cal I}^-.$ Hence, by (B.9), $$\Bigl(X^{\alpha_1 \ldots \alpha_k}
\Bigl|_{i^-}\Bigr)\biggr|_{r \to \infty} = O\left(
\frac{1}{r^3}\right)\; .$$
For the calculation of the integral (B.6) at $x \to {\cal I}^+$ we may use the Bondi-Sachs frame (A.1). For the past of ${\cal I}^+$ this is safe even if the metric has closed apparent horizons \[2,3\] since no one of these will be encountered by the light rays emitted sufficiently early. To lowest order in the curvature, the world function is then of the form $$\sigma( x, {\bar x}) = - (u - {\bar u})
( r - {\bar r} + \frac{u - {\bar u}}{2}) +
r{\bar r}(1 - \cos \omega) + O[R_{..}]$$ where $\frac{1}{2}\omega^2 $ is the world function on the 2-sphere : $$\cos \omega = \cos \theta \cos {\bar \theta} +
\sin \theta \sin {\bar \theta} \cos (\varphi -
{\bar \varphi}) \; .$$ By solving the equation $ \sigma (x, {\bar x}) = 0$ with respect to ${\bar u}$ and choosing the solution which corresponds to the past light cone of $x, {\bar u} = f,$ we obtain $$\begin{aligned}
-\frac{1}{\Box} X^{\alpha_1 \ldots \alpha_k}
( r, \theta, \varphi, u ) = \frac{1}{4\pi} \int
d^2 {\bar {\cal S}} \int\limits^{\infty}_{0}
d{\bar r} {\bar r}^2 \Bigl| ({\bar \nabla}
{\bar u}, {\bar \nabla}{\bar r}) \cdot
\frac{\partial \sigma}{\partial {\bar u}}
\Bigr|^{-1} {\cal Y}({\bar r}, {\bar \theta}, {\bar \varphi},
{\bar u})\biggl|_{{\bar u} = f} + \nonumber \\
+\; O[X\times\Re] \qquad\end{aligned}$$ and $$-\frac{1}{\Box} X^{\alpha_1 \ldots \alpha_k}
( r, \theta, \varphi, u )\biggr|_{{\cal I}^+} =
\frac{1}{r}\: Q(\theta, \varphi, u) + O\left(\frac{1}{r^2}\right)$$ where $$Q( \theta, \varphi, u) = - \frac{1}{4\pi} \int d^2
{\bar {\cal S}} \int\limits_{0}^{\infty} d {\bar r}
(\log{\bar r}) \frac{\partial}{\partial {\bar r}}
\Bigl[ {\bar r}^3 {\cal Y}({\bar r}, {\bar \theta},
{\bar \varphi}, {\bar u})\Bigl|_{{\bar u} = f^{*}}
\Bigr]
+\; O[ X\times\Re ] \; ,\quad$$ $$f^{*} = u - {\bar r}(1 - \cos \omega) + O[R_{..}]\;,$$ and the equation ${\bar u}=f^{*}$ is the equation of the limiting light cone of the point $x$ at ${\cal I}^+.$ Here we wrote $ {\bar r}^2 = {\bar r}^3 {\partial \log {\bar
r}}/{\partial {\bar r}}$ and integrated by parts for being able to consider sources decreasing at ${\cal I}^-$ like $1/r^3.$ At the limit $u \to -\infty,$ the source in (B.16) turns out to be at ${\bar u} \to -\infty.$ By the assumption of analyticity, we then have $$Q( \theta, \varphi, u )\Bigl|_{u \to -\infty} =
Q_0 + \frac{1}{u}Q_1 + \cdots$$ with $$Q_0 = - \frac{1}{4\pi} \int d^2
{\bar {\cal S}} \int\limits_{0}^{\infty} d {\bar r}
(\log{\bar r}) \frac{\partial}{\partial {\bar r}}
\Bigl[ {\bar r}^3 {\cal Y}({\bar r}, {\bar \theta},
{\bar \varphi}, -\infty)\Bigr] + O[ X\times\Re ] \; ,$$ etc., and the convergence of the integral in (B.19) is now owing to (B.11).
Eqs. (B.15)-(B.19) make manifest the fact which is more general than the approximations made. Namely, the integral (B.1) in the past of ${\cal I}^+$ involves only the source $X^{\alpha_1 \ldots \alpha_k}$ at $i^-.$ Under the assumption of analyticity at $i^-,$ this integral becomes a static Coulomb potential. One can show that it preserves this form also at two other limits : at the past null infinity and spatial infinity ($i^0$), $$- \frac{1}{\Box} X^{\alpha_1 \ldots \alpha_k}\biggl|_{{\cal
I}^- or\: i^0\: or {\cal I}^+, u \to -\infty} =
\frac{Q_0}{r} (1 + {\cal O})\;,\quad {\cal O} \to 0 \;.$$
Consider now equation (5.9) for the news functions which is valid under the assumption made in sec. 2 that the flux components of the Ricci tensor at ${\cal I}^-$ vanish. Since, in this case, $R^{\mu\nu}$ at ${\cal I}^-$ is $O(1/r^3),$ the integral implied in $(1/\Box) R^{\mu\nu}$ converges. Moreover, the derivatives of this integral in the null-tetrad basis behave at ${\cal I}^-$ like $$\Bigl( \nabla_{\gamma} \frac{1}{\Box} R^{\mu\nu}
\Bigr) \biggr|_{{\cal I}^-} = O\left( \frac{1}{r^2}\right)$$ (see below). Therefore, the nonlinear additions to $R^{\nu\beta}$ in eq. (5.10) are $O(1/r^4)$ at ${\cal I}^-.$ We then have [^14] $$\begin{aligned}
m_{\beta}m_{\nu} \frac{1}{\Box} \Bigl[ R^{\nu\beta}
+ (\nabla^{\nu} \frac{1}{\Box} R^{\gamma\delta})(
\nabla^{\beta} \frac{1}{\Box} R_{\gamma\delta}) -
2 (\nabla_{\gamma} \frac{1}{\Box} R^{\nu\delta})(
\nabla_{\delta} \frac{1}{\Box} R^{\beta\gamma}) +
O[ R^3_{..}] \Bigr]\biggr|_{{\cal I}^+} =
\nonumber \\
{} = \frac{1}{r} q( \theta, \varphi, u) + O\left( \frac{1}{r^2}\right)
\; .\qquad\quad\end{aligned}$$ We need to calculate the limit $$\lim_{u \to -\infty} \frac{\partial}{\partial u}\; q( \theta, \varphi,
u)$$ which serves as an initial datum for eq. (5.9). By analyticity of the metric at $i^-,$ $$q( \theta, \varphi, u)\Bigl|_{u \to -\infty} = q_0 +
\frac{1}{u}\: q_1 + \cdots \;,$$ and the limit (B.23) vanishes owing to the presence of the derivative ${\partial}/{\partial u}.$
Eq. (B.20) can also be used to prove that the solution of the Bianchi identities (2.3) with zero initial data for the gravitational field at ${\cal I}^-$ is expressed indeed in terms of the retarded Green function. This is easily seen in expression (2.4b). Since, by (B.20), $$- \frac{1}{\Box}\Bigl( R^{\nu\beta} + O[R^2_{..}]
\Bigr) \biggr|_{{\cal I}^-} =
\frac{Q^{\nu\beta}_0}{r} + O\left(\frac{1}{r^2}\right) \; ,$$ and $Q^{\nu\beta}_0$ [*does not depend on the advanced time along*]{} ${\cal I}^-,$ the derivatives in (2.4b) enhance the power of $1/r.$ Eq. (B.21) is valid for the same reason.
The proof is more involved in the general case, where the fluxes of $R^{\mu\nu}$ through ${\cal I}^-$ are nonvanishing, since one has to address eq. (2.4a). In terms of the null tetrad at ${\cal I}^-$ $$\nabla^{[\mu} \nabla^{<\alpha} R^{\nu]\beta>}\biggl|_{
{\cal I}^-} = \frac{\nabla^{[\mu}v \nabla^{<\alpha}v}{
(\nabla u, \nabla v)^2} (\nabla_{\gamma}u) ( \nabla_{\sigma}
u) \nabla^{\gamma}\nabla^{\sigma} R^{\nu]\beta>} +
O\left( \frac{1}{r^3}\right) \; .$$ Since the flux components of $R^{\nu\beta}$ at ${\cal I}^-$ are proportional either to $\nabla^{\nu}v$ or to $\nabla^{\beta}v,$ they cancel in (B.26) owing to the antisymmetrizations : $$\nabla^{[\mu} \nabla^{<\alpha} R^{\nu]\beta>}\Bigl|_{
{\cal I}^-} = O\left( \frac{1}{r^3}\right)\; .$$ Thus the source in (2.4a) satisfies condition (B.9), and we have $$R^{\alpha\beta\mu\nu}\biggl|_{{\cal I}^-} =
- \frac{Q_0^{\alpha\beta\mu\nu}}{r} + O\left(\frac{1}{r^2}\right)$$ with $Q_0^{\alpha\beta\mu\nu}$ constant along ${\cal I}^-.$ By (B.19) and (B.5), $$\begin{aligned}
Q_0^{\alpha\beta\mu\nu}& =& - \frac{1}{4\pi} \int d^2 {\bar {
\cal S}} \int\limits_0^{\infty} d{\bar r} (\log {\bar r})
\frac{\partial}{\partial {\bar r}} \left( {\bar r}^3
{\bar {\cal Y}}\Bigl|_{i^-}\right) + O[R^2_{..}] \; , \\
{\bar {\cal Y}}& =& 4 g^{[\mu}_{\;\:{\bar \mu}}
g^{\nu]}_{\;\:{\bar \nu}} g^{[\alpha}_{\;\:{\bar \alpha}}
g^{\beta]}_{\;\:{\bar \beta}} {\bar \nabla}^{\bar \mu}
{\bar \nabla}^{\bar \alpha} {\bar R}^{{\bar \nu}{\bar \beta}}\; ,\end{aligned}$$ and, up to $O[R^2_{..}], {\bar {\cal Y}}$ is a total derivative: $$\begin{aligned}
{\bar {\cal Y}}\biggl|_{i^-}& =& {\bar g}^{-1/2}
{\bar \partial}_{\bar \mu} \left( {\bar g}^{1/2} {\bar Z}^{\bar \mu}
\right) \biggl|_{i^-} +\;\: O[R^2_{..}]\; , \\
{\bar Z}^{\bar \mu}& =& 4 g^{[\mu{\bar \mu}}g^{\nu]{\bar \nu}}
g^{[\alpha{\bar \alpha}}g^{\beta]{\bar \beta}}
{\bar \nabla}_{\bar \alpha} {\bar R}_{{\bar \nu}{\bar \beta}} \; ,\end{aligned}$$ where use is made of eq. (B.3). Since, in this total derivative, the time derivative vanishes owing to the asymptotic stationarity of the metric at $i^-,$ and the angle derivatives vanish in the integral $\int d^2{\bar {\cal S}},$ we obtain $$Q_0^{\alpha\beta\mu\nu} = - \frac{1}{4\pi} \int d^2 {\bar {
\cal S}} \int\limits_0^{\infty} d{\bar r} (\log {\bar r})
\frac{\partial}{\partial {\bar r}} \left[ {\bar r}
\frac{\partial}{\partial {\bar r}} \Bigl( {\bar r}^2
{\bar \nabla}_{\bar \mu} {\bar r}\; {\bar Z}^{\bar \mu}
\biggl|_{i^-}\Bigr)\right] + O[R^2_{..}]\; .$$ Since the power of decrease of ${\bar Z}^{\bar \mu}\Bigl|_{i^-}$ at ${\bar r} \to \infty $ is at least $O(1/{{\bar r}^2}),$ we may integrate by parts in (B.33) [*to remove the integration*]{} over ${\bar r}$ completely : $$Q_0^{\alpha\beta\mu\nu} = \lim_{{\bar r} \to \infty}
\frac{1}{4\pi} \int d^2 {\bar {
\cal S}}\Bigl( 4 g^{[\mu{\bar \mu}}g^{\nu]{\bar \nu}}
g^{[\alpha{\bar \alpha}}g^{\beta]{\bar \beta}}
{\bar \nabla}_{\bar \mu}{\bar r} \Bigr)
\Bigl({\bar r}^2 {\bar \nabla}_{\bar \alpha}
{\bar R}_{{\bar \nu}{\bar \beta}}\biggl|_{i^-}\Bigr)
+ O[R^2_{..}] \; .$$ Only the terms in ${\bar R}_{{\bar \nu}{\bar \beta}}$ that decrease like $1/{{\bar r}^2}$ can survive in (B.34). However, the integrand in (B.34) contains one more derivative of ${\bar R}_{{\bar \nu}{\bar \beta}},$ and, again, the respective time derivative vanishes at $i^-.$ Therefore, this integrand is $O(1/{\bar r})$ (actually $O(1/{{\bar r}^2}),$ see below), and the limit (B.34) vanishes. The presence of an extra derivative is in fact not essential. By (B.10), the sequence of limits in (B.34) can be replaced by the limit of the same quantity in the past of ${\cal I}^-.$ It is then seen that the constant (B.34) hangs solely on the flux components of the Ricci tensor [*in the past of*]{} ${\cal I}^-.$ However, the metric in the past becomes asymptotically static, and for a static metric the fluxes of $R_{\nu\beta}$ through ${\cal I}^-$ are absent as seen from the counterparts at ${\cal I}^-$ of eqs. (A.18) - (A.21). We have, therefore, $$( R_{\nu\beta}\Bigl|_{i^-} )\biggl|_{r \to \infty} =
( R_{\nu\beta}\Bigl|_{{\cal I}^-} )\biggl|_{v \to -\infty} =
O\left(\frac{1}{r^3}\right)\; ,$$ and $$Q_0^{\alpha\beta\mu\nu} = 0\; .$$ Hence $$R^{\alpha\beta\mu\nu}\Bigl|_{{\cal I}^-} = O\left(\frac{1}{r^2}
\right)$$ which proves that, in the retarded solution, there are no incoming gravitational waves.
References. {#references. .unnumbered}
===========
1. V.P.Frolov and G.A.Vilkovisky,\
, ed. R.Ruffini (Amsterdam: North-Holland 1982) p.455\
2. V.P.Frolov and G.A.Vilkovisky, Phys. Lett. B 106 (1981) 307\
3. V.P.Frolov and G.A.Vilkovisky,\
, ed. M. A. Markov and P.C.West (London: Plenum, 1983) p. 267\
4. G.A.Vilkovisky, Class. Quantum Grav. 9 (1992) 895\
5. A.O.Barvinsky, Yu.V.Gusev, G.A.Vilkovisky, and V.V.Zhytnikov, J.Math. Phys. 35 (1994) 3525\
6. A.G.Mirzabekian and G.A.Vilkovisky, Phys. Lett. B 317 (1993) 517\
7. A.G.Mirzabekian, Zh. Eksp. Teor. Fiz. 106 (1994) 5 \[ JETP 79 (1994) 1 \]\
8. A.O.Barvinsky and G.A.Vilkovisky, Nucl. Phys. B 282 (1987) 163\
9. A.O.Barvinsky and G.A.Vilkovisky, Nucl. Phys. B 333 (1990) 471\
10. A.O.Barvinsky and G.A.Vilkovisky, Nucl. Phys. B 333 (1990) 512\
11. G.A.Vilkovisky, Preprint CERN-TH. 6392/92;\
, vol. 43 (Strasbourg, 1992) p.203\
12. A.O.Barvinsky, Yu.V.Gusev, G.A.Vilkovisky,and V.V.Zhytnikov,\
, Report of the University of Manitoba ( Winnipeg : U. of Manitoba 1993) pp. 1-192\
13. A.O.Barvinsky, Yu.V.Gusev, G.A.Vilkovisky, and V.V.Zhytnikov, J.Math. Phys. 35 (1994) 3543\
14. A.O.Barvinsky, Yu.V.Gusev, G.A.Vilkovisky, and V.V.Zhytnikov, Nucl. Phys. B, to appear\
15. A.O.Barvinsky, Yu.V.Gusev, V.V.Zhytnikov, and G.A.Vilkovisky, (to be published)\
16. L.P.Grishchuk, Zh. Eksp. Teor. Fiz. 67 (1974) 825 \[ Sov. Phys. JETP 40 (1975) 409 \]\
17. L.P.Grishchuk and Yu.V.Sidorov, Phys. Rev. D 42 (1990) 3413\
18. L.P.Grishchuk, Class. Quantum Grav. 10 (1993) 2449\
19. B.S.DeWitt, , ed. B.S.DeWitt and R.Stora (Amsterdam:North-Holland, 1984) p. 221\
20. H.Bondi, M.G.J. van der Burg, and A.W.K.Metzner, Proc. R. Soc. London A269 (1962) 21\
21. R.Sachs, , ed. C.DeWitt and B.S.DeWitt ( New York: Gordon and Breach 1964)\
22. R.M.Wald, (Chicago: The University of Chicago Press, 1984)\
23. J.L.Synge, (Amsterdam: North-Holland 1960 )\
24. B.S.DeWitt, (New York: Gordon and Breach, 1965)
[^1]: The gravitational collapse problem was first considered in this setting in refs. \[1-3\].
[^2]: After the choice of the state has been made, one checks the original assumption. Since the causality relationship is set by the expectation value of the metric, this is a self-consistent problem even in its original setting. That the consistency check is nontrivial is seen, for example, from the fact that with the massive quantum fields one generally arrives at a contradiction with the asymptotic flatness.
[^3]: For the expectation value of the metric this condition restricts the choice of the classical source which can be taken, e.g., as having a compact spatial support. Then, since the energy-momentum tensor of the in-vacuum has no incoming fluxes at ${\cal I}^-,$ the same will be true also of $R^{\mu\nu}$ (see eq. (2.18) below). In the general case, the flux components of $R^{\mu\nu}$ at ${\cal I}^-$ cancel in the combination $\nabla^{[\mu} \nabla^{<\alpha} R^{\nu ]\beta >}$ appearing in (2.4a) but the derivatives cannot be commuted with $1/\Box$ for otherwise the action of the retarded Green function will become ill-defined.
[^4]: We follow the notation of ref. \[15\] but change the overall sign of the action as appropriate for the lorentzian signature of the metric. We use the signature $(- + + +)$ and the conventions $ R^{\mu}_{\;\alpha\nu\beta} = \partial_{\nu}
\Gamma^{\mu}_{\alpha\beta} - ... , R_{\alpha\beta} =
R^{\mu}_{\;\alpha\mu\beta} , R = g^{\alpha\beta} R_{\alpha\beta}.$
[^5]: Eq. (39) of ref. \[6\] admits one more solution: $ \Box \Bigl[ \gamma_1 (-\Box) + 3 \gamma_2(-\Box) \Bigr] \to a,
- \Box \to 0 $ which had been overlooked.
[^6]: For the derivation of this procedure in QFT see \[8\] and references therein. In the phenomenological approach this set of rules is taken for granted \[4\].
[^7]: We do not consider here the more general form factors $ \Gamma (-\Box_1, ... , -\Box_N, -\Box_{1+2},
-\Box_{1+3}, ... ) $ in which the operator arguments $\Box_{n+m}$ act on products of two curvatures since such form factors appear in $S_{vac}$ only beginning with $N=4$ \[4,5\].
[^8]: For this calculation to lowest order in the curvature see \[6,7\]. The cancellation of the $1/\Box$ terms being established in the present paper, this calculation can now be done to second order in the curvature on the basis of the results in \[15\].
[^9]: With the specifications made in sec. 2, this choice boils down to the choice of $T^{\mu\nu}_{source}$ in eq. (2.18). Variations in $T^{\mu\nu}_{source}$ induce variations in the curvature of the solution. Eq. (4.15) should, therefore, hold for any configuration of the curvature. It is also important for inferring (4.25) that, by the construction of the curvature basis in the action \[5\], $I_3^{\alpha\mu}(\xi, x)$ can have no total derivative terms of the form $\Box X^{\alpha\mu}(\xi, x)$ or $\nabla^{(\alpha}X^{\mu)}(\xi, x).$ Hence $A^{\alpha\mu}(x)$ can have no such terms.
[^10]: This method can also be applied to the classical problems in the gravitational radiation.
[^11]: Even this contribution is not purely classical since the metric to be used is the solution of the expectation-value equations. The same concerns the flux of $T^{\mu\nu}_{source}$ at ${\cal I}^+$ in eq. (3.1). It would be interesting to consider a case where $T^{\mu
\nu}_{source}$ does not induce the gravitational waves but $T^{\mu\nu}_{vac}$ does.
[^12]: Since, at ${\cal I}^-,$ the flux components of $T^{\mu\nu}_{vac}$ vanish, eq. (5.13) is valid even without the limitation implied in (2.4b) (see sec. 2).
[^13]: By our calculations, the equation $R_{rr} =
(2/r)(\Psi^{-1}\partial_{r}\Psi - (\partial_{r}\gamma)^2)$ ( in the present notation) given in \[21\] for the axisymmetric case is in error. We obtain $ R_{rr} =
(2/r)\Psi^{-1}\partial_{r}\Psi - 2(\partial_{r}\gamma)^2.$ It follows from the former equation that, for $R_{\mu\nu}$ of a compact spatial support, $\Psi = -1 + O(1/r^3).$ With our result, $\Psi = -1 + O(1/r^2).$ The significance of the behaviour of $\Psi$ is seen from eqs. (A.15), (A.18).
[^14]: Since, at ${\cal I}^+$, the flux components of $R^{\mu\nu}$ do not vanish, the integral $(1/\Box)
R^{\mu\nu}$ behaves at ${\cal I}^+$ like $(\log r)/r .$ The $(\log r)/r$ asymptotic terms vanish in expression (2.4b) owing to the presence of the antisymmetrized derivatives, and in expressions (5.10), (5.13), (B.22) owing to the contraction with $m_{\beta}m_{\nu}.$
| {
"pile_set_name": "ArXiv"
} |
0.7cm
UT-754\
TU-505\
hep-ph/9607234\
July, 1996
0.35cm
[**Is the ee$\gamma \gamma+/\!\!\!\! E_T$ Event an Evidence of the Light Axino?** ]{} 1.2cm J. Hisano$^{(a)}$[Fellows of the Japan Society for the Promotion of Science.]{}[The address after the next September is, School of Physics and Astronomy, University of Minnesota, Minneapolis, MN 55455, USA.]{} , K. Tobe$^{(a)(b)*}$, and T. Yanagida$^{(a)}$
0.4cm
[*(a) Department of Physics, University of Tokyo, Tokyo 113, Japan*]{}\
[*(b) Department of Physics, Tohoku University, Sendai 980-77, Japan*]{}
1.5cm
The no-scale supergravity [@no-scale] has attracted many physicists in particle physics, since it may arise from a class of space-time compactifications in superstring theories [@witten]. It is also interesting in cosmology, since it can naturally accommodate the chaotic inflation [@MSYY], but it also provides a consistent solution [@MYY] to the serious cosmological problem in supergravity; [*i.e.*]{}, the Polonyi problem [@Polonyi]. However, it has been recently pointed out [@KMY] that one needs a new fine tuning to solve the Polonyi problem if the usual lightest supersymmetric particle (LSP) (which is perhaps a bino-dominated neutralino) is stable [@MYY]. Therefore, we are led to consider the unstable “LSP”. A possible way to have the unstable “LSP” is to break R parity. In this case, however, we must invoke the other mechanism to avoid a rapid proton decay. In Ref. [@KMY] it has been suggested that the bino-dominated “LSP” (we call it, “bino”, hereafter) decays into the axino (a fermionic superpartner of the axion) [@RTW], since there is a possibility that the axino remains light even after supersymmetry (SUSY) breaking and becomes the true LSP in the no-scale supergravity model [@GY].
In this letter we point out that if the Peccei-Quinn (PQ) symmetry breaking scale $F_a$ lies in a range of the hadronic axion window ($F_a/N=(0.7-2)
\times 10^6$GeV with $N$ being the QCD anomaly factor of the PQ symmetry) [@Ressell], the “bino” decay into the axino can explain the ee$\gamma\gamma+/\!\!\!\! E_T $ event recently observed in the CDF experiment [@CDF]. We also stress that the constraint on the hadronic axion window derived from the intergalactic photon search [@MBRT][@Ressell] is irrelevant in our no-scale supergravity model, since the decay of the Polonyi field produces a large amount of entropy at the late epoch of the universe evolution and dilutes the abundance of the cosmic axion density substantially. (The dilution factor for relativistic particles is about $10^{-13}$.)[^1]
We consider, in this letter, an example of SUSY hadronic axion model [@hadronicaxion]. The extension of our analysis based on more general models is straightforward. We assume $N$ pairs of massless new chiral superfields $\Psi_A=(Q, L)_A$ and $\bar{\Psi}_A=(\bar{Q}, \bar{L})_A$ ($A=1-N$) which transform as $\bf 5$ and $\bf 5^*$ under the grand unified gauge group $\rm SU(5)_{GUT}$, respectively, in addition to the SUSY standard model (SM) sector. They are assumed to have the PQ charge $+1$ and hence there are massless as far as the PQ symmetry is unbroken. All fields in the SUSY SM sector have no PQ charge. In order to break the PQ symmetry we introduce a superfield $\Phi$ whose PQ charge is set as $-2$ so that the $\Phi$ can couple to the $N$ pairs of $\Psi$ and $\bar{\Psi}$ as $$\begin{aligned}
W=\lambda_A \bar{\Psi}_A \Psi_A \Phi.
$$ We assume that some physics involving the $\Phi$ field gives a nonzero vacuum expectation value to the $\Phi$ and then the PQ symmetry is spontaneously broken by the vacuum condensation $\langle\Phi \rangle
\neq 0$.
The Nambu-Goldstone chiral multiplet arising from the PQ symmetry breaking contains pseudo-scalar field axion $a(x)$, real-scalar field saxion $s(x)$, and their fermionic partner called axino $\tilde{a}(x)$. The $N$ pairs of $\bar{\Psi}$ and $\Psi$ have masses of $\lambda_A \langle \Phi \rangle$. The axion acquires a mass of order of $\Lambda_{\rm QCD}^2/\langle \Phi \rangle$ through QCD instanton effects, where $\Lambda_{\rm QCD}$ is the QCD scale $\sim 100 \rm{MeV}$. On the other hand, the axino gets a mass[^2] $$\begin{aligned}
m_{\tilde a} \simeq \sum_A\frac{1}{16 \pi^2} \lambda_A^2 m_{\rm SUSY}
$$ through one loop diagrams in the no-scale supergravity model as shown in Ref. [@MY]. Here, the $m_{\rm SUSY}$ is an induced SUSY breaking soft mass of $\Psi$ and $\bar{\Psi}$. If one takes $m_{\rm SUSY} \sim O(100) {\rm GeV}$ and $\lambda_A \sim O(0.1)$, one gets $m_{\tilde a} \sim O(10) {\rm MeV}$. Notice that this axino $\tilde{a}$ is harmless in cosmology since the large entropy production from the Polonyi field decay dilutes the axino density substantially. This large entropy production is also very important to dilute the cosmic axion density as stressed in the introduction.
A crucial point in this letter is that the axion superfield $\Phi_a(x,\theta)$ couples to the gauge superfields through anomalies of the PQ current as $$\begin{aligned}
{\cal{L}}= -\sqrt{2}\frac{\alpha_i}{8 \pi}
\int d^2 \theta \frac{\Phi_a}{F_a/N} W_\alpha^i W_\alpha^i,
\label{anomaly}\end{aligned}$$ where $F_a = \langle \Phi \rangle$, and $W_\alpha^i$ ($i=1-3$) are gauge superfields of the SM gauge groups U(1)$_Y$, SU(2)$_L$, and SU(3)$_C$, and $\alpha_i$ are corresponding gauge coupling constants ($\alpha_1= 5/3 \alpha_Y= 0.017 $, $\alpha_2 = 0.034 $, and $\alpha_3 = 0.12$ at the electroweak scale). This induced interactions in Eq. (\[anomaly\]) contain a bino-axino-photon coupling as $${\cal L}
=
- \frac{5\alpha_{em}}{24\pi} \frac{1}{\cos\theta_W} \frac{1}{F_a/N}
\overline{\tilde{a}}\gamma_5\sigma_{\mu\nu} \tilde{B} F^{\mu\nu},
\label{interaction}$$ from which we can estimate the decay width of $\tilde{B}\rightarrow \tilde{a}+ \gamma$ as $$\Gamma(\tilde{B}\rightarrow \tilde{a}+ \gamma)
=\frac{25\alpha_{em}^2}{1152\pi^3}\frac{1}{\cos^2\theta_W} \frac{M_{\tilde{B}}^3}{({F_a}/N)^2}$$ where $M_{\tilde{B}}$ is the bino mass, and then,[^3] $$c\tau_{\tilde{B}} = 0.36~
\left(\frac{100{ {\rm GeV} }}{M_{\tilde{B}}}\right)^3
\left(\frac{F_a/N}{10^6{ {\rm GeV} }}\right)^2
~{\rm c.m.}.$$ We are now at the point of this letter. If the $F_a$ lies in the range of hadronic axion window [@Ressell], [*i.e.*]{}, $$F_a/N \simeq (0.7-2) \times 10^6 ~~{\rm GeV},
$$ we obtain $$c\tau_{\tilde{B}}= (0.18-1.44) ~{\rm c.m.},
$$ for $M_{\tilde{B}}=100 {\rm GeV}$. It is now clear that the $\tilde{B}\rightarrow \tilde{a} +\gamma$ decay can be a source of the hard photon in the ee$\gamma \gamma + /\!\!\!\!E_T$ event observed in the CDF experiment [@CDF].
It has been already shown in recent papers [@gravitino] that masses of a selectron $\tilde e$ and the “bino” $\tilde B$ must be $m_{\tilde e}= (80-130) {\rm GeV}$, and $M_{\tilde B}=(38-100) {\rm GeV}$ to explain the ee$\gamma \gamma + /\!\!\!\!E_T$ event by sequent decays; $\tilde{e}^-( \tilde{e}^+) \rightarrow e^- (e^+)+\tilde{B}$ and $\tilde{B}
\rightarrow \gamma+ {\rm LSP}$. Let us now discuss the low-energy mass spectrum of SUSY particles in our model, and show that it is very much welcome to this event. In the no-scale supergravity model sfermion masses are induced by the SM gauge interactions with non-vanishing gaugino masses. Then, the right-handed selectron and the “bino” are naturally expected to be the lightest two among SUSY particles except for the axino, since they have only the U(1)$_Y$ gauge interaction. (See Ref. [@IKYY] for a detailed calculation.) We use renormalization group (RG) equations to evaluate a ratio of the right-handed selectron to the bino masses. The RG equations of the right-handed selectron mass ($m_{\tilde{e}_R}$) and the bino mass ($M_{\tilde{B}}$) above the mass scale of $\Psi$ and $\bar{\Psi}$ are $$\begin{aligned}
\mu \frac{\partial m_{\tilde{e}_R}^2}{\partial\mu} &=& -8 \frac{\alpha_Y}{4\pi} M_{\tilde{B}}^2,
\nonumber\\
\mu \frac{\partial M_{\tilde{B}}}{\partial\mu} &=& 2 b_Y \frac{\alpha_Y}{4\pi} M_{\tilde{B}},
$$ where $b_Y(=11+5N/3)$ is a coefficient of beta function of the U(1)$_Y$ gauge coupling constant and $\mu$ the renormalization point. These equations become those of the SUSY SM below the mass scale of $\Psi$ and $\bar{\Psi}$. If the no-scale boundary conditions for these masses such as $ m_{\tilde{e}_R}=0$ is imposed at $\mu = 10^{16}$GeV, the mass ratio between the right-handed selectron and the “bino” at the electroweak scale is given by $$\begin{aligned}
\frac{m_{\tilde{e}_R}}{M_{\tilde{B}}}&=&
\left\{
\begin{array}{cc}
1.1 & (N=1) \\
1.3 & (N=2) \\
1.7 & (N=3) \\
2.6 & (N=4)
\end{array}
\right. .\end{aligned}$$ Here, the masses of $\Psi$ and $\bar{\Psi}$ are taken at $10^5$GeV. The mass ratios for $N=2,3$ are suitable to the ee$\gamma
\gamma + /\!\!\!\!E_T$ event in the CDF experiment as shown in Refs. [@gravitino][@BKW][@DTW]. The masses of $m_{\tilde{e}_R}$ and $M_{\tilde{B}}$ themselves take the values around $100$ GeV to cause the correct electroweak symmetry breaking as shown in Ref. [@IKYY], which are also desirable for the explanation of the event. Notice that the right-handed selectron is heavier than wino ($\tilde{W}$) for $N=4$ if the gaugino masses satisfy the GUT relation ($M_{\tilde{W}}\simeq 2 M_{\tilde{B}}$). In this case, the CDF event can have another interpretation that it is a wino pair production accompanied with the sequent decays as $\tilde{W}^-(\tilde{W}^+)\rightarrow e^-(e^+)+ \bar{\nu}(\nu)+ \tilde{B}$ and $\tilde{W}^0 \rightarrow e^-(\nu)+ e^+ (\bar{\nu}) +\tilde{B}$, assuming two body decays into $W^{\pm}(Z^0) +\tilde{B}$ are suppressed by phase space (that is, $M_{\tilde{W}}-M_{\tilde{B}}<
m_W(m_Z)$) [@BKW][@AKKMM]. If this interpretation is right, it is expected to observe multi-leptons and 2 photons events with missing energy.[^4] If the boundary condition is given at the gravitational scale ($\mu = 10^{18}$GeV), the mass ratio of the right-handed selectron to the “bino” becomes larger, and we can get the suitable mass spectrum even for $N=1$ case ($m_{\tilde{e}_R}/M_{\tilde{B}}= 1.6$).
So far we do not have taken constraints from the effects of axion emission upon the life cycle of red-giant (RG) and horizontal-branch (HB) stars in our analysis, since these constraints are based on the statistics of small number [@Ressell]. If one takes the constraints seriously, one obtains [@Raffelt] $$\begin{aligned}
F_a/N &>& 3 \times 10^6 ~{ {\rm GeV} }~~~~~(\rm{RG}),
\\
F_a/N &>& 9 \times 10^6 ~{ {\rm GeV} }~~~~~(\rm{HB}),
$$ which is already outside the axion window. However, this problem can be easily solved, since the axion-photon-photon ($a \gamma \gamma$) coupling depends on the details of models [@Choi]. For example, we assign the different PQ charges $Q_L$ and $Q_Q$ to the doublet $L$ and triplet $Q$, respectively. [^5] Then, we obtain the $a \gamma \gamma$ coupling as $$\begin{aligned}
{\cal{L}}_{a \gamma \gamma}&=&\frac{\kappa}{4}
a F_{\mu \nu} \tilde{F}^{\mu \nu},
\\
\kappa&=&\frac{\alpha_{em}}{2\pi}\frac{1}{F_a/N}
\left[ \frac{2}{3}(1+3 \gamma)-\frac{2(4+z)}{3(1+z)}\right],
\label{agammagamma}\end{aligned}$$ where $\gamma$ is the ratio of the PQ charge, $\gamma=Q_L/Q_Q$ and $z$ the mass ratio of up- and down-quarks, $z=m_u/m_d$. Notice that the second term in the bracket of Eq.(\[agammagamma\]) denotes the contribution from the long-distance effect. We can see that $a \gamma \gamma$ coupling is almost vanishing for $z \simeq 0.56$ and $\gamma \simeq 2/3$. In this case, the above constraints from RG and HB stars become weaker [@Ressell] $$\begin{aligned}
F_a/N &>& 0.2 \times 10^6 ~{ {\rm GeV} }~~~~~(\rm{RG}),
\\
F_a/N &>& 0.6 \times 10^6 ~{ {\rm GeV} }~~~~~(\rm{HB}),
$$ since subject to the constraints is the $a \gamma \gamma$ coupling strength but not $F_a$ itself. On the other hand, the bino-axino-photon ($\tilde{B}\tilde{a} \gamma$) coupling does not have the contribution from the long-distance effect and hence there is not such a cancellation in the $\tilde{B}\tilde{a} \gamma$ coupling. Thus, the analysis in this paper is unchanged.
In summary, we argue that the hadronic axion window is not yet excluded by the intergalactic photon search, since there is a substantial dilution of the cosmic axion density in the no-scale supergravity model. If the PQ symmetry breaking scale is in the hadronic axion window, the ee$\gamma\gamma+ /\!\!\!\! E_T$ event in the CDF experiment can be naturally interpreted as a result of the cascade decays; $\tilde{e}_R^-(\tilde{e}_R^+)\rightarrow
e^-(e^+) + \tilde{B}$ and $\tilde{B}\rightarrow {\rm axino} + \gamma$. We hope that this hadronic axion window will be tested by future axion searches [@moriyama].
1.5cm [**Acknowledgment**]{}
We thank K. Choi for informing the presence of Ref. [@Choi].
[99]{} J. Ellis, A.B. Lahanas, D.V. Nanopoulos, and K. Tamvakis, [[*Phys. Lett.*]{}]{} [B134]{} (1984) 429;\
See also, A.B. Lahanas, and D.V. Nanopoulos, [*Phys. Rep.*]{} [*145*]{} (1987) 1. E. Witten, [[*Phys. Lett.*]{}]{} [B155]{} (1985) 151. H. Murayama, H. Suzuki, T. Yanagida, and J. Yokoyama, [[*Phys. Rev.*]{}]{} [D50]{} (1994) 2356. T. Moroi, M. Yamaguchi, and T. Yanagida, [[*Phys. Lett.*]{}]{} [B342]{} (1995) 105. G.D. Coughlan, N. Fischler, E.W. Kolb, S. Raby and G.G. Ross, [[*Phys. Lett.*]{}]{} [B131]{} (1983) 59;\
T. Banks, D.B. Kaplan, and A.E. Nelson, [[*Phys. Rev.*]{}]{} [D49]{} (1994) 779;\
B. de Carlos, J.A. Casas, F. Quevedo, and E. Roulet, [[*Phys. Lett.*]{}]{} [B318]{} (1993) 447. M. Kawasaki, T. Moroi, and T. Yanagida, [[*Phys. Lett.*]{}]{} [B370]{} (1996) 52. K. Rajagopal, M.S. Turner, and F. Wilczek, [[*Nucl. Phys.*]{}]{} [B358]{} (1991) 447. T. Goto and M. Yamaguchi, [[*Phys. Lett.*]{}]{} [B276]{} (1992) 103. See references in M.T. Ressell, [[*Phys. Rev.*]{}]{} [D44]{} (1991) 3001. S. Park, “Search for New Phenomena in CDF”, $10^{\rm th}$ Topical Workshop on Proton–Anti-proton Collider Physics, edited by Rajendran Raja and John Yoh, AIP Press, 1995. M.A. Bershady, M.T. Ressell, and M.S. Turner, [[*Phys. Rev. Lett.*]{}]{} [D66]{} (1991) 1398. J.E. Kim, [[*Phys. Rev. Lett.*]{}]{} [43]{} (1979) 103;\
M.A. Shifman, A.I. Vainshtein, and V.I. Zakharov, [[*Nucl. Phys.*]{}]{} [B166]{} (1980) 493. S. Chang and H.B. Kim, FTUAM-96-15, hep-ph/9604222. P. Moxhay and K. Yamamoto, [[*Phys. Lett.*]{}]{} [B151]{} (1985) 363. S. Ambrosanio, G.L. Kane, G.D. Kribs, P. Stephen, and S. Mrenna, [[*Phys. Rev. Lett.*]{}]{} [76]{} (1996) 3498;\
S. Dimopoulos, M. Dine, S. Raby, and S. Thomas, [[*Phys. Rev. Lett.*]{}]{} [76]{} (1996) 3494. K. Inoue, M. Kawasaki, M. Yamaguchi, and T. Yanagida, [[*Phys. Rev.*]{}]{} [D45]{} (1992) 328. K.S. Babu, C. Kolda, and F. Wilczek, IASSNS-HEP-96-55, hep-ph/9605408. S. Dimopoulos, S. Thomas, and J.D. Wells, SLAC-PUB-7148, hep-ph/9604452. S. Ambrosanio, G.L. Kane, G.D. Kribs, S.P. Martin, and S. Mrenna, hep-ph/9605398. G.G. Raffelt, and D.S.P. Dearborn, [[*Phys. Rev.*]{}]{} [D36]{} (1987) 2211. S. Chang, and K. Choi, [[*Phys. Lett.*]{}]{} [B316]{} (1993) 51. S. Moriyama, [[*Phys. Rev. Lett.*]{}]{} [75]{} (1995) 3222.
[^1]: In Ref. [@MYY] it is shown that enough baryon asymmetry is created in spite of the large entropy production if we assume the Affleck-Dine mechanism for baryogenesis.
[^2]: The saxion mass is given by $m_{s}^2= \sum_A (c/16\pi^2) \lambda_A^2 m_{SUSY}^2$ with $c \sim O(1)$. If one takes $m_{\rm SUSY} \sim O(100) {\rm GeV}$ and $\lambda_A \sim O(0.1)$, one gets $m_{s} \sim O(1) {\rm GeV}$. This is cosmologically harmless since the lifetime $\tau_{s\rightarrow gg}$ is $10^{-7}$ sec. for $F_a=10^6{ {\rm GeV} }$ [@saxion].
[^3]: The decay $\tilde{B}\rightarrow \tilde{a}+ \gamma$ is the main decay mode. Thus, the branching ratio ${\rm {Br}}(\tilde{B}\rightarrow \tilde{a}+ \gamma)
\simeq 100 \%$ which is also a favorable point for explaining the ee$\gamma \gamma+/\!\!\!\! E_T$ event.
[^4]: The decay into jets is suppressed since squarks are heavier than sleptons due to the SU(3)$_C$ interactions.
[^5]: Since $L$ and $Q$ have different PQ charges, the multiplets $\Psi_A(Q,L)$ and $\bar{\Psi}_A(\bar{Q},\bar{L})$ do not form $\rm{SU(5)}_{\rm{GUT}}$ multiplets.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a system of session types based on *adjoint logic* which generalize standard binary session types [@Honda93concur]. Our system allows us to uniformly capture several new behaviors in the space of asynchronous message-passing communication, including *multicast*, where a process sends a single message to multiple clients, *replicable services*, which have multiple clients and replicate themselves on-demand to handle requests from those clients, and *cancellation*, where a process discards a channel without communicating along it. We provide session fidelity and deadlock-freedom results for this system, from which we then derive a logically justified form of garbage collection.'
author:
- Klaas Pruiksma
- Frank Pfenning
bibliography:
- 'fp.bib'
title: 'A Message-Passing Interpretation of Adjoint Logic'
---
Introduction
============
*Binary session types* [@Honda93concur] were designed to specify the communication behavior between two message-passing processes. But there are patterns of communication that do not fall into this category. One example is one provider of a *replicable service* with multiple clients. Another is a *multicast*, that is, a process sending one message to multiple recipients. A third one is a client that no longer wishes to use a service, a form of *cancellation*. In this paper we provide a uniform language and operational semantics rooted in logic that captures such patterns of asynchronous communication. It generalizes the usual binary session types by supporting multiple *modes of communication*. In each of these modes every channel has a unique *provider* (which may send or receive), and possibly multiple clients. We identify the following modes: *linear* (a unique client that must communicate, as with the usual binary session types), *affine* (a unique client that may communicate or cancel), *strict* (multiple clients, each of which must communicate), and *unrestricted* (multiple clients, each of which may or may not communicate, which captures both replicable services and multicast).
A type system that uniformly integrates all of these patterns is not obvious if we want to preserve the desirable properties of session fidelity and deadlock freedom that we obtain from binary session types. Underlying our approach is *adjoint logic* [@Reed09un; @Licata16lfcs; @Pruiksma18un], which generalizes *intuitionistic linear logic* [@Girard87tcs; @Girard87tapsoft] and LNL [@Benton94csl] by decomposing the usual exponential modality $!A$ into two adjoint modal operators and also affords individual control over the structural rules of weakening and contraction. We provide a formulation of adjoint logic in which cut reduction corresponds to asynchronous communication, and from which session fidelity and deadlock freedom derive. Moreover, our formulation uses a form of explicit structural rules embedded in a multicut, where weakening corresponds to cancellation and contraction corresponds to duplication of a message or service.
Some of these patterns have been previously addressed with varying degrees of proximity to an underlying logic. A replicable service with multiple clients can be achieved with *access points* [@Gay10jfp] or *persistent services* of type $!A$ [@Caires10concur]. Cancellation can be addressed with affine types [@Mostrous14coordination; @Scalas16ecoop; @Padovani17icfp] further developed for asynchronous communication and general handling of failure [@Fowler19popl]. Cancellation can also be handled with modalities used to label cancellable types [@Caires2017linearity]. This approach differs from ours in a few respects — first, Caires and P[é]{}rez work in a purely synchronous setting, without multicast, and second, they focus heavily on introducing nondeterminism, which we believe to be orthogonal to (our form of) cancellation. Closest to the present proposal is a polarized formulation of asynchronous communication in adjoint logic [@Pfenning15fossacs] which had several shortcomings that are addressed here. Specifically, the mode hierarchy was fixed to have only three modes (linear, affine, and unrestricted), and the unrestricted mode only allowed a single kind of proposition ${\up}^\mU_m A_m$. This meant that, for example, multicast was not representable. Also, the rules left weakening and contraction implicit, which means that there is no explicit cancellation or distributed garbage collection, which is only briefly hinted at as a possibility [@Griffith16phd].
The Curry-Howard correspondence relates propositions to types, proofs to programs, and proof reduction to computation. Cut reductions in a pure sequent calculcus for linear logic [@Caires10concur; @Wadler12icfp] naturally correspond to synchronous communication because both premises of the cut are reduced at the same time. We reformulate adjoint logic with a nonstandard sequent calculus in which noninvertible rules are presented as axioms, that is, rules with no premises. As our operational interpretation shows, an axiom can be seen as a message and cut reduction in this sequent calculus corresponds to asynchronous communication. Another unusual aspect of our sequent calculus is that we generalize cut to a sound rule of multicut [@Gentzen35; @Negri01book], which operationally allows one provider to connect with multiple clients. Two further consequences of this reformulation are that (a) no explicit rules are needed for weakening and contraction, and yet (b) channels and resources are tracked with sufficient precision that computation in a network of processes “leaves no garbage” (see \[sec:metatheory\]). This is the concurrent realization of the early observation by Girard and Lafont [@Girard87tapsoft] that functional computation based on intuitionistic linear logic does not require a garbage collector. Cancellation [@Mostrous14coordination; @Fowler19popl] is a natural consequence, without requiring any special mechanism, but our system goes beyond it in the sense that processes with multiple clients will also terminate once no clients are left.
We begin with a brief discussion of our type system (\[sec:typing\]), deferring discussion of the underlying logic to \[app:adjoint-logic\], in order to focus on the programming system. We then present an operational semantics (\[sec:operational\]): our first major contribution. It models a variety of asynchronous communication behaviors, uniformly generalizing previous systems. We close by briefly presenting our results on session fidelity and deadlock-freedom, along with a brief discussion of the “garbage-collection” result that follows from them (\[sec:metatheory\]).
Language and Typing {#sec:typing}
===================
$\displaystyle
\begin{array}{c}
\infer[\id]
{(a : A_m) \seq {{\color{red} c \leftarrow a}} :: (c : A_m)}
{\mathstrut}
\qquad
\infer[\cut(S)]
{\Psi \ \Psi' \seq {{\color{red} S \leftarrow (\nu x)P; Q}} :: (c : C_k)}
{\Psi \geq m \geq k & |S| \sim m
&\Psi \seq {{\color{red} P}} :: (x : A_m)
&(S : A_m) \ \Psi' \seq {{\color{red} Q}} :: (c : C_k)
}
\\[1em]
\infer[{\oplus} R_\ell^0]
{(a : A_m^\ell) \seq {{\color{red} c.\m{\ell}(a)}} :: (c : \mathop{\oplus}\limits_{i \in I} A_m^i)}
{\ell \in I}
\qquad
\infer[{\oplus} L]
{\Psi \ (a : \mathop{\oplus}\limits_{i \in I} A_m^i) \seq {{\color{red} \m{case}\; a\; (i(x_i) \Rightarrow P_i)_{i \in I}}} :: (c : C_k)}
{\Psi \ (x_i : A_m^i) \seq {{\color{red} P_i}} :: (c : C_k) \text{ for each $i \in I$}}
\\[1em]
\infer[{\with} R]
{\Psi \seq {{\color{red} \m{case}\; c\; (i(x_i) \Rightarrow P_i)_{i \in I}}} :: (c : \mathop{\with}\limits_{i \in I} A_m^j)}
{\Psi \seq {{\color{red} P_i}} :: (x_i : A_m^i) \text{ for each $i \in I$}}
\qquad
\infer[{\with} L_\ell^0]
{(a :\mathop{\with}\limits_{i \in I} A_m^i) \seq {{\color{red} a.\ell(c)}} :: (c : A_m^\ell)}
{\ell \in I}
\\[1em]
\infer[{\otimes} R^0]
{(a : A_m) \ (b : B_m) \seq {{\color{red} c.\langle a, b \rangle}} :: (c : A_m \otimes B_m)}
{\mathstrut}
\qquad
\infer[{\otimes} L]
{\Psi \ (a : A_m \otimes B_m) \seq {{\color{red} \m{case}\; a(\langle x, y \rangle \Rightarrow P)}} :: (c : C_k)}
{\Psi \ (x : A_m) \ (y : B_m) \seq {{\color{red} P}} :: (c : C_k)}
\\[1em]
\infer[{\one} R]
{\cdot \seq {{\color{red} c.\langle \rangle}} :: (c : \one_m)}
{\mathstrut}
\qquad
\infer[{\one} L]
{\Psi \ (a : \one_m) \seq {{\color{red} \m{case}\; a(\langle \rangle \Rightarrow P)}} :: (c : C_k)}
{\Psi \seq {{\color{red} P}} :: (c : C_k)}
\\[1em]
\infer[{\lolli} R]
{\Psi \seq {{\color{red} \m{case}\; c(\langle x, y \rangle \Rightarrow P)}} :: (c : A_m \lolli B_m)}
{(x : A_m) \ \Psi \seq {{\color{red} P}} :: (y : B_m)}
\qquad
\infer[{\lolli} L^0]
{(a : A_m) \ (c : A_m \lolli B_m) \seq {{\color{red} c.\langle a, b \rangle}} :: (b : B_m)}
{\mathstrut}
\\[1em]
\infer[{\upshift} R]
{\Psi \seq {{\color{red} \m{case}\; c(\m{shift}(x) \Rightarrow P)}} :: (c : \upshift_k^m A_k)}
{\Psi \seq {{\color{red} P}} :: (x :A_k)}
\qquad
\infer[{\upshift} L^0]
{(a : \upshift_k^m A_k) \seq {{\color{red} a.\m{shift}(c)}} :: (c : A_k)}
{\mathstrut}
\\[1em]
\infer[{\downshift} R^0]
{(a : A_m) \seq {{\color{red} c.\m{shift}(a)}} :: (c : \downshift_k^m A_m)}
{\mathstrut}
\qquad
\infer[{\downshift} L]
{\Psi \ (a : \downshift_k^m A_m) \seq {{\color{red} \m{case}\; a(\m{shift}(c) \Rightarrow P)}} ::(c : C_\ell)}
{\Psi \ (x : A_m) \seq {{\color{red} P}} :: (c : C_\ell)}
\end{array}
$
Our typing judgment for processes $P$ is based on *intuitionistic sequents* of the form $$(x^1 : A^1) \cdots (x^n : A^n) \vdash P :: (x : A)$$ where each of the $x^i$ are *channels* that $P$ *uses* and $x$ is a channel that $P$ *provides*. All of these channels must be distinct and we abbreviate the collection of antecedents as $\Psi$. The *session types* $A^i$ and $A$ specify the communication behavior that the process $P$ must follow along each of the channels.
Such sequents are standard for the intuitionistic approach to understanding binary session types (e.g., [@Caires10concur]) where the channels are *linear* in that every channel in a network of processes has exactly one provider and exactly one client. In the closely related formulation based on classical linear logic [@Wadler12icfp] all channels are on the right-hand side of the turnstile, but each linear channel still has exactly two endpoints.
We generalize this significantly by assigning to each channel an intrinsic *mode* $m$. Each mode $m$ is assigned a set of structural properties $\sigma(m)$ among W (for weakening) and C (for contraction). Separating $m$ from $\sigma(m)$ allows us to have multiple modes with the same set of structural properties.[^1] No matter which structural properties are available for a channel, each active channel will still have *exactly one provider*. Beyond that, a channel $x_m$ with ${\rm W} \in \sigma(m)$ may not have any clients. Furthermore, a channel $x_m$ with ${\rm C} \in \sigma(m)$ may have multiple clients. All other properties of our system of session types for processes derive systematically from these simple principles.
The modes are organized into a preorder where $m \geq k$ requires that $\sigma(m) \supseteq \sigma(k)$, that is, $m$ must allow more structural properties than $k$. In order to guarantee session fidelity and deadlock freedom, for any sequent $\Psi \vdash P :: (x_m : A_m)$ is must be the case that for every $y_k : B_k \in \Psi$ we have $k \geq m$. For example, if $m$ permits contraction and therefore $P$ may have multiple clients, then for any $y_k$ in $\Psi$, mode $k$ must also permit contraction because (intuitively) if $x_m$ is referenced multiple times then, indirectly, so is $y_k$. If $k \geq m$ then this is ensured. We express this with the *presupposition* that $$\Psi \vdash P :: (x_m : A_m)\quad \mbox{requires} \quad \Psi \geq m$$ where $\Psi \geq m$ simply means $k \geq m$ for every $y_k : A_k \in \Psi$. We will only consider sequents satisfying this presupposition, so our rules, when they are used to break down a conclusion into the premises, must preserve this fundamental property which we call *the declaration of independence*.
In our formulation, channels $x_m$ as well as types $A_m$ are endowed with modes which must always be consistent between a channel and its type ($x_m : A_m$). We therefore often omit redundant mode annotations on channels.
The complete set of rules for the typing judgment are given in Fig. \[fig:message-passing-typing-rules\]. We first examine the judgmental rules that explain the meaning of identity and composition. Identity (rule $\id$) is straightforward: a process $c \leftarrow a$ providing $c$ defers to the provider of $a$, which is possible as long as $a$ and $c$ have the same type and mode. This is usually called *forwarding* or *identification* of the channels $a$ and $c$.
The usual logical rule of cut corresponds to the parallel composition of two processes with a single private channel for communication between them. However, ordinary cut is insufficiently general to describe the situation where a single provider of a channel $x_m$ may have multiple clients ($C \in \sigma(m)$) or no clients ($W \in \sigma(m)$). We therefore generalize it to a form of multicut, [^2] where the channel $x_m$ provided by $P$ is known by multiple aliases in the set of channels $S$ in $Q$ as long as the multiplicity of the aliases is permitted by the mode. This is expressed as $|S| \sim m$ and is sufficient for static typing. Formally, we define this condition by $0 \sim m$ if $\m{W} \in \sigma(m)$, $1 \sim m$ always, and $k \sim m$ for $k \geq 2$ if $\m{C} \in \sigma(m)$. When processes execute we will have an even more general situation where one provider has multiple separate client processes, which is captured in the typing judgment for process configurations (\[sec:operational\]).
Next we come to the various session types. From the logical perspective, these are the propositions of adjoint logic. $$\begin{array}{lcl}
A_m, B_m & \Coloneqq & p_m \mid A_m \lolli_m B_m \mid A_m \tensor_m B_m \mid \one_m \mid \mathop{\oplus_m}\limits_{i \in I} A_m^i
\mid \mathop{\with_m}\limits_{i \in I} A_m^i \mid \up_k^m A_k \mid \down_m^\ell A_\ell
\end{array}$$ Here, $p_m$ stands for atomic propositions at mode $m$. The other connectives, other than $\up_k^m$ and $\down_m^\ell$, are standard linear logic connectives, except that they are only allowed to combine types (propositions) at the same mode. Since the mode of a connective can be inferred from the modes of the types it connects (other than for shifts), we omit subscripts on connectives. Note also that $\with$ and $\oplus$ have been generalized to $n$-ary forms from the usual binary forms. This is convenient for programming. We will use a label set $I = \{\pi_1, \pi_2\}$ when working with the binary forms $A_m \with B_m$ and $A_m \oplus B_m$, where $\pi_1$ selects the left-hand type and $\pi_2$ selects the right-hand type. The operational meaning of these connectives (as discussed further in \[sec:operational\]) is largely similar to that in past work (e.g. [@Caires10concur]), with $\lolli_m$ and $\tensor_m$ sending channels along other channels, $\one_m$ sending an end-of-communication message, and $\oplus_m$ and $\with_m$ sending labels. The shifts send a simple $\m{shift}$ message to signal a transition between modes, either *up* ($\up_k^m$) from $k$ to some $m \geq k$ or *down* ($\down_m^\ell$) from $\ell$ to some $m \leq \ell$.
We provide proof terms for the rules in our sequent calculus, as shown in Figure \[fig:message-passing-typing-rules\]. We can then interpret the proof terms as process expressions, and these rules are used to give the typing judgment for such processes. gives the informal meaning of each such process term.
In general, our process syntax represents an intermediate point between a programmer-friendly syntax and a notation in which it is easy to describe the operational semantics and prove progress and preservation. When compared to, for instance, SILL [@Toninho13esop], the main revisions are that (1) we make channel continuations explicit in order to facilitate asynchronous communication while preserving message order [@DeYoung12csl], and (2) we distinguish between an *internal name* for the channel provided by a process and *external names* connecting it to multiple clients.
Process term Meaning
----------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
$a \leftarrow c$ Identify channels $a$ and $c$.
$S \leftarrow (\nu x)P \semi Q$ Spawn a new process $P$ providing channel $x$ with aliases $S$ to be used by $Q$. Here, $x$ is the *internal name* in $P$ for the channel offered by $P$, and $S$ is the set of *external names* of the same channel as used in $Q$.
\[1.2ex\] $c.\ell(a)$ Send the label $\ell$ and the channel $a$ along $c$.
$\m{case}\; c (i(x_i) \Rightarrow P_i\}_{i \in I})$ Receive a label $i$ and a channel $x_i$ from $c$, continue as $P_i$.
\[1.2ex\] $c.\langle a, b \rangle $ Send the channels $a$ and $b$ along $c$.
$\m{case}\; c(\langle x, y \rangle \Rightarrow P)$ Receive channels $x$ and $y$ from $c$ to be used in $P$.
\[1.2ex\] $c.\langle \rangle$ End communication over $c$ by sending a terminal message.
$\m{case}\; c(\langle \rangle \Rightarrow P)$ Wait for $c$ to be closed, continue as $P$.
\[1.2ex\] $c_m.\m{shift}(a_k)$ Send a shift, from mode $m$ to mode $k$
$\m{case}\; c_m\; (\m{shift}(x_k) \Rightarrow P)$ Receive a shift from mode $m$ to mode $k$
\[0.2ex\]
: Informal Meanings of Process Terms[]{data-label="tbl:proc-meanings"}
[**Some simple examples.**]{} We provide here some small examples with their types; additional examples which highlight more interesting behavior can be found in \[sec:example\].
First, we have a process that can be written at any mode $m$, which witnesses that $\otimes_m$ is commutative.
$(x : A_m \otimes B_m) \vdash \m{case}\; x\,(\langle y, x' \rangle \Rightarrow z.\langle x', y \rangle) :: (z : B_m \otimes A_m)$
If $m$ is a mode that admits contraction, we can write the following process, which witnesses that $A_m \with B_m$ proves $A_m \otimes B_m$ in the presence of contraction. ‘$\%$’ starts a comment.
$(p : A_m \with B_m) \vdash$ = $\{p_1, p_2\} \leftarrow (\nu q)\, (q \leftarrow p); \qquad$ = $\%\ \{p_1, p_2\} \leftarrow \m{copy}\; p$\
$x \leftarrow (\nu a)\, p_1.\pi_1(a);$\
$y \leftarrow (\nu b)\, p_2.\pi_2(b);$\
$z.\langle x, y \rangle$ $:: (z : A_m \otimes B_m)$
If $m$ is a mode that admits weakening, we can write the following process, which witnesses that $A_m \otimes B_m$ proves $A_m \with B_m$ in the presence of weakening.
$(x : A \otimes B) \vdash$ = $\m{case}\; p\;$ = $($ = $\pi_1(p_1) \Rightarrow$ = $\m{case}\; x\,(\langle y, z \rangle \Rightarrow$\
$\{\,\} \leftarrow (\nu a)\, (a \leftarrow z); \qquad \quad$ = $\%\ \m{drop}\; z$\
$p_1 \leftarrow y)$\
$|\,\pi_2(p_2) \Rightarrow$ = $\m{case}\; x\,(\langle y, z \rangle \Rightarrow$\
$\{\,\} \leftarrow (\nu a)\, (a \leftarrow y); \qquad \quad$ = $\%\ \m{drop}\; y$\
$p_2 \leftarrow z \,))$\
$:: (p : A \with B)$
Operational Semantics {#sec:operational}
=====================
In order to describe the computational behavior of process expressions, we need to first give some syntax for the computational artifacts, which are running processes ${\m{proc}}(S, \Delta, a, P)$. Such a process executes $P$ and provides a channel $a$ while using the channels in the channel set $\Delta$. $S$ is a set of aliases for the channel $a$, which can be referred to by one or more clients. Each alias $c \in S$ is used by at most one client, but one client may use multiple such aliases. Note that as the aliases in $S$ are the only way to interact with the channel $a$ from an external process, the objects ${\m{proc}}(S, \Delta, a, P)$ and ${\m{proc}}(S, \Delta, b, P[b/a])$ are equivalent — changing the internal name of a process has no effect on its interactions with other processes.
A *process configuration* is a multiset of processes: $$\begin{array}{lcl}
\CC & \Coloneqq & {\m{proc}}(S, \Delta, a, P) \mid (\cdot) \mid \CC\ \CC'
\end{array}$$ where we require that all the aliases or names provided by the processes ${\m{proc}}(S, \Delta, a, P)$ are distinct, i.e., given objects ${\m{proc}}(S, \Delta_1, a, P)$ and ${\m{proc}}(T, \Delta_2, b, Q)$ in the same process configuration, $S$ and $T$ are disjoint. We will specify the operational semantics in the form of *multiset rewriting rules* [@Cervesato09ic]. That means we show how to rewrite some subset of the configuration while leaving the remainder untouched. This form provides some assurance of the locality of the rules.
It simplifies the description of the operational semantics if for any process ${\m{proc}}(S, \Delta, a, P)$, $\Delta$ consists of exactly the free channels (other than $a$) in $P$. This requires that we restrict the labeled internal and external choices, $\mathop{\oplus}\limits_{i \in I} A^i_m$ and $\mathop{\with}\limits_{i \in I} A^i_m$ to the case where $I \neq \emptyset$. Since a channel of empty choice type can never carry any messages, this is not a significant restriction in practice.
In order to understand the rules of the operational semantics, it will be helpful to understand the typing of configurations. The judgment has the form $\Psi \vDash \CC :: \Psi'$ which expresses that using the channels in $\Psi$, configuration $\CC$ provides the channels in $\Psi'$. This allows a channel that is not mentioned at all in $\CC$ to appear in both $\Psi$ and $\Psi'$—we think of such a channel as being “passed through” the configuration.
Note that while the configuration typing rules induce an ordering on a configuration, the configuration itself is not inherently ordered. The key rule is the first: for any object ${\m{proc}}(S, \Delta, a, P)$ we require that $P$ is well-typed on some subset of the available channels while the others are passed through. Here we write $\overline{\Psi}$ for the set of channels declared in $\Psi$, which must be exactly those used in the typing of $P$. Moreover, *externally* such a process provides the channels $S = \{a^1_m, \ldots, a^n_m\}$, all of the same type $A_m$. We use the abbreviation $(S : A_m)$ for $a^1_m : A_m, \ldots, a^n_m : A_m$. Finally, we enforce that the number of clients must be compatible with the mode $m$ of the offered channel, which is exactly that $|S| \sim m$, as defined in \[sec:typing\]. $$\infer[\m{Proc}]
{\Psi\ \Psi' \vDash {\m{proc}}(S, \overline{\Psi'}, a, P) :: \Psi\ (S : A_m)}
{|S| \sim m
&\Psi' \seq P :: (a : A_m)
}
\qquad
\infer[\m{Id}]
{\Psi \vDash (\cdot) :: \Psi}
{\mathstrut}
\qquad
\infer[\m{Comp}]
{\Psi \vDash \mathcal C\ \mathcal C' :: \Psi''}
{\Psi \vDash \mathcal C :: \Psi'
&\Psi' \vDash \mathcal C' :: \Psi''
}$$ The identity and composition rules are straightforward. The empty context $(\cdot)$ provides $\Psi$ if given $\Psi$, since it does not use any channels in $\Psi$ or provide any additional channels. Composition just connects configurations with compatible interfaces: what is provided by $\CC$ is used by $\CC'$.
$\displaystyle
\begin{array}{rcl}
\begin{array}{r}
{\m{proc}}(T \cup \{c\}, \Delta, x, P)\\
{\m{proc}}(S, \{c\}, y, y \leftarrow c)\\
\end{array}
&\overset{\id}{\Longrightarrow}&
\begin{array}{l}
{\m{proc}}(T \cup S, \Delta, x, P)\\
\end{array}
\\[1.2em]
\begin{array}{r}
{\m{proc}}(T, \Delta_P \cup \Delta_Q, y, S \leftarrow (\nu x)P; Q)\\
(\text{$S'$ a fresh set of channels matching $S$})
\end{array}
&\overset{\cut(S)}{\Longrightarrow}&
\begin{array}{l}
{\m{proc}}(S', \Delta_P, x, P) \\
{\m{proc}}(T, \Delta_Q \cup \{S'\}, y, Q[S'/S])\\
\end{array}
\\[1.2em]
\begin{array}{r}
(P \text{ not an identity}) \qquad {\m{proc}}(\emptyset, \Delta, x, P)
\end{array}
&\overset{\m{drop}}{\Longrightarrow}&
\begin{array}{l}
{\m{proc}}(\emptyset, \{b\}, y, y \leftarrow b)_{b \in \Delta} \\
\end{array}
\\[1.2em]
\begin{array}{r}
{\m{proc}}(S \cup T, \Delta, x, P)\\
(\text{$P$ not an identity and $S, T$ non-empty})
\end{array}
&\overset{\m{copy}}{\Longrightarrow}&
\begin{array}{l}
{\m{proc}}(\{b', b''\}, \{b\}, y, y \leftarrow b)_{b \in \Delta}\\
{\m{proc}}(S, \{b'\}_{b \in \Delta}, x, P[b'/b])\\
{\m{proc}}(T, \{b''\}_{b \in \Delta}, x, P[b''/b])\\
\end{array}
\\[1.2em]
\begin{array}{r}
{\m{proc}}(\{b\}, \{c\}, x, x.\ell(c))\\
{\m{proc}}(S, \Delta \cup \{b\}, z, \m{case}\; b(i(y_i) \Rightarrow P_i)_{i \in I})\\
\end{array}
&\overset{{\oplus}\; C}{\Longrightarrow}&
\begin{array}{l}
{\m{proc}}(S, \Delta \cup \{c\}, z, P_\ell[c/y_\ell])\\
\end{array}
\\[1.2em]
\begin{array}{r}
{\m{proc}}(\{b\}, \Delta, x, \case\; x(i(y_i) \Rightarrow P_i)_{i \in I})\\
{\m{proc}}(\{c\}, \{b\}, z, b.\ell(z))\\
\end{array}
&\overset{{\with}\; C}{\Longrightarrow}&
\begin{array}{l}
{\m{proc}}(\{c\}, \Delta, z, P_\ell[z/y_\ell])\\
\end{array}
\\[1.2em]
\begin{array}{r}
{\m{proc}}(\{b\}, \{c, d\}, w, w.\langle c, d \rangle)\\
{\m{proc}}(S, \Delta \cup \{b\}, z, \m{case}\; b(\langle x, y \rangle \Rightarrow P)\\
\end{array}
&\overset{{\otimes}\; C}{\Longrightarrow}&
\begin{array}{l}
{\m{proc}}(S, \Delta \cup \{c, d\}, z, P[c/x, d/y])\\
\end{array}
\\[1.2em]
\begin{array}{r}
{\m{proc}}(\{b\}, \Delta, w, \m{case}\; w(\langle x, y \rangle \Rightarrow P)\\
{\m{proc}}(\{c\}, \{b, d\}, z, b.\langle d, z \rangle)\\
\end{array}
&\overset{{\lolli}\; C}{\Longrightarrow}&
\begin{array}{l}
{\m{proc}}(\{c\}, \Delta \cup \{d\}, z, P[d/x, z/y])\\
\end{array}
\\[1.2em]
\begin{array}{r}
{\m{proc}}(\{b\}, \emptyset, x, x.\langle \rangle)\\
{\m{proc}}(S, \Delta \cup \{b\}, y, \m{case}\; b (\langle \rangle \Rightarrow P))\\
\end{array}
&\overset{{\one}\; C}{\Longrightarrow}&
\begin{array}{l}
{\m{proc}}(S, \Delta, y, P)\\
\end{array}
\\[1.2em]
\begin{array}{r}
{\m{proc}}(\{b_k\}, \{c_m\}, x_k, x_k.\m{shift}(c_m))\\
{\m{proc}}(S, \Delta \cup \{b_k\}, y, \m{case}\; b_k(\m{shift}(z_m) \Rightarrow P))\\
\end{array}
&\overset{{\down_k^m}\; C}{\Longrightarrow}&
\begin{array}{l}
{\m{proc}}(S, \Delta \cup \{c_m\}, y, P[c_m/z_m])\\
\end{array}
\\[1.2em]
\begin{array}{r}
{\m{proc}}(\{b_m\}, \Delta, x_m, \case\; x_m(\m{shift}(z_k) \Rightarrow P))\\
{\m{proc}}(\{c_k\}, \{b_m\}, y_k, b_m.\m{shift}(y_k))\\
\end{array}
&\overset{{\up_k^m}\; C}{\Longrightarrow}&
\begin{array}{l}
{\m{proc}}(\{c_k\}, \Delta, y_k, P[y_k/z_k])\\
\end{array}
\\[1.2em]
\end{array}
$
The computation rules we discuss in this section can be found in Figure \[fig:dynamics1\]. Remarkably, the computation rules do not depend on the modes, although some of the rules will naturally only apply at modes satisfying certain structural properties.
[**Judgmental rules.**]{} The identity rule (written as $\overset{\id}{\Longrightarrow}$) describes how an identity process (for instance, ${{\m{proc}}(S, \{c\}, a, a \leftarrow c)}$) may interact with other processes. We think of such a process as connecting the provider of $c$ to clients in $S$, and therefore sometimes call it a *forwarding process*. A forwarding process interacts with the provider of $c$, telling it to replace $c$ with $S$ in its set of clients. In adding $S$ to the set of clients, the forwarding process accomplishes its goal of connecting the provider of $c$ to $S$, and so it can terminate.
The cut rule steps by spawning a new process which offers along a fresh set of channels $S'$, all of which are used in $Q$, the continuation of the original process. Here we write $\Delta_P$ and $\Delta_Q$ for the set of free channels in $P$ and $Q$, respectively.
[**Structural rules.**]{} A process with no clients can terminate (rule $\overset{\m{drop}}{\Longrightarrow}$), but must notify all of the processes it uses that they should also terminate. It does so by sending each one a forwarding message, effectively embodying a cancellation. In concert with the identity rule this accomplishes cascading cancellation in the distributed setting. Note that the mode $m$ of channel $a$ must admit weakening in order for the process on the left-hand side of the rule to be well-typed.
Similarly, a process with multiple clients can spawn a copy of itself, each with a strictly smaller set of clients (rule $\overset{\m{copy}}{\Longrightarrow}$). If the process $P$ is a replicable service, that is, if it has a negative type $\with$, $\lolli$, ${\up}^m_k$, then this corresponds to actual process replication. If it has a positive type $\oplus$, $\otimes$, $\one$, ${\down}^m_k$, this corresponds to duplicating a multicast message into copies for different subsets of recipients. The mode $m$ of the channel $a$ must admit contraction in order for the process on the left-hand side of the rule to be well-typed.
While both the $\m{drop}$ and $\m{copy}$ rules can be applied to any process with $0$ or multiple clients, respectively, this does not cause any problems as long as we forbid them from executing on identity processes. If we apply drop or copy to an identity process, we end up with another process of the same form on the right-hand side of the rule, and so we could repeatedly apply drop or copy and not make any progress. As such, we forbid this use of the drop and copy rules.
For any other type of process, regardless of whether we drop/copy first or execute another communication rule first, we can eventually reach the same state, and so we do not need to make additional restrictions (though an actual implementation would likely pick either a maximally eager or a maximally lazy strategy for applying these rules).
[**Additive and multiplicative connectives.**]{} In the rule for $\oplus$, the process $\m{proc}(\{b\}, \{c\}, a, a.\ell(c))$ represents the message ‘label $\ell$ with continuation $c$’. After this message has been received, the process terminates since $b$ was its only client. The recipient selects the appropriate branch of the $\m{case}$ construct and also substitutes the continuation channel $c$ for the continuation variable $d_\ell$.
The $\with$ computation rule is largely similar to that for $\oplus$, except that communication proceeds in the opposite direction—messages are sent *to* providers *from* clients, rather than from providers to clients as in the case of $\oplus$.
The multiplicative connectives $\otimes$ and $\lolli$ behave similarly to their additive counterparts, except that rather than sending and receiving labels, they send and receive channels together with a continuation, and so an extra substitution is required when receiving messages, while the $\one$ behaves as a nullary $\otimes$, allowing us to signal that no more communication is forthcoming along a channel, and to wait for such a signal before continuing to compute.
[**Shifts.**]{} We present the computation rules for shifts with modes marked explicitly on the relevant channels. Channels whose modes are unmarked may be at any mode (provided, of course, that the declaration of independence is respected).
Operationally, $\upshift$ behaves essentially the same as unary $\with$, while $\downshift$ behaves as unary $\oplus$. Their significance lies in the *mode shift* of the continuation channel that is transmitted, which is required for the configuration to remain well-typed.
The messages $\m{shift}(a_k)$ or $\m{shift}(c_m)$ should be thought of as signaling a transition between modes — to mode $k$ for the former, and to mode $m$ for the latter. Whether the transition is up or down depends on which direction the message is being sent in. As with other messages (in particular, the messages for $\oplus$ and $\with$), the continuation channels are made explicit.
Session Fidelity, Deadlock-Freedom, and Garbage Collection {#sec:metatheory}
==========================================================
While we can prove cut elimination for the form of adjoint logic presented in \[app:adjoint-logic\], from a programmer’s perspective we are not interested in eliminating all cuts (which would correspond to reducing under $\lambda$-abstractions in a functional language) but rather we block when waiting to receive a message, analogous to a $\lambda$-abstraction waiting for input before it can reduce. What we prove instead are session fidelity and deadlock-freedom.
[**Session fidelity.**]{} The session fidelity theorem follows from a case analysis on the computation rule used to get that $\mathcal C \Rightarrow \mathcal C'$. In each case, we break $\mathcal C$ down to find the processes on which the computation rule acts, along with some collections of processes which are unaffected by the computation. From these pieces, we build a proof that $\Psi \vDash \mathcal C' :: \Psi'$.
If $\Psi \vDash \mathcal C :: \Psi'$ and $\mathcal C \Rightarrow \mathcal C'$, then $\Psi \vDash \mathcal C' :: \Psi'$.
[**Deadlock-freedom.**]{} The progress theorem for a functional language states that an expression is either a value or it can take a step. Here we do not have values, but there is nevertheless a clear analogue between, say, a value $\lambda x.e$ that waits for an argument, and a process $\m{case}\; x\, (\langle y, z \rangle \Rightarrow P)$ that waits for an input. We formalize this in the definition below.
We say that a process ${\m{proc}}(S, \Delta, a, P)$ is *poised* on $a$ if:
1. it is a process ${\m{proc}}(S, \Delta, a, P)$ that sends on $a$ — that is, $P$ is of the form $(a.\uscore)$, or
2. it is a process ${\m{proc}}(S, \Delta, a, P)$ that receives on $a$ — that is, $P$ is of the form $(\m{case}\; a\; (\uscore))$.
Intuitively, ${\m{proc}}(S, \Delta, a, P)$ is poised on $a$ if it is blocked trying to communicate along $a$. This definition allows us to state the following progress theorem:
If $(\cdot) \vDash \mathcal C :: \Psi$, then exactly one of the following holds:
1. There is a $\mathcal C'$ such that $\mathcal C \Rightarrow \mathcal C'$.
2. Every ${\m{proc}}(S, \Delta, a, P)$ in $\mathcal C$ is poised on $a$.
In order to prove this theorem, we first prove a lemma allowing us to take advantage of the ordering induced by configuration typing. We note that if object $\psi$ is a client of object $\phi$, $\psi$ must occur to the right of $\phi$ in the ordering, and so if we can analyze a configuration from right to left, we consider each process before (or after, depending on your view of induction) all of its dependencies. To formalize this, we present a second set of rules defining another form of configuration typing (which will turn out to prove the same judgments as the original form). $$\infer[\m{Empty}]
{\Psi \vDash' (\cdot) :: \Psi}
{\mathstrut}
\qquad
\infer[\m{Extend}]
{\Psi \vDash' \mathcal C\ {\m{proc}}(S, \overline{\Psi'}, a, P) :: \Psi''\ (S : A_m)}
{|S| \sim m
&\Psi \vDash' \mathcal C :: \Psi'\ \Psi''
&\Psi' \seq P :: (a : A_m)
}$$ It is clear that if $\vDash$ and $\vDash'$ are the same, then we can perform induction using the $\m{Empty}$ and $\m{Extend}$ rules rather than the $\m{Id}$, $\m{Comp}$, and $\m{Proc}$ rules, allowing us to analyze a configuration from right to left. We formalize this as \[lem:config-order\].
\[lem:config-order\] $\Psi \vDash \mathcal C :: \Psi'$ if and only if $\Psi \vDash' \mathcal C :: \Psi'$.
This lemma is nearly immediate — all of the rules for $\vDash'$ are derivable from the rules of $\vDash$, and all rules of $\vDash$ but $\m{Comp}$ are derivable from the rules of $\vDash'$. We therefore need only show (by an induction over the right-hand premise) that the version of the $\m{Comp}$ rule with $\vDash$ replaced by $\vDash'$ is admissible.
The proof of deadlock-freedom then proceeds by an induction on the derivation of $(\cdot) \vDash \mathcal C :: \Psi$, using \[lem:config-order\] to work right to left. Writing $\mathcal C = \mathcal C'\; {\m{proc}}(S, \overline{\Psi'}, a.P)$, we see that either $\mathcal C'$ can step, in which case so can $\mathcal C$, or every process in $\mathcal C'$ is poised. Now we carefully distinguish cases on $S$ (empty, singleton, or greater) and apply inversion to the typing of $P$ to see that in each case the process either is poised, can take a step independently, or can interact with provider of a channel in $\overline{\Psi'}$.
[**Garbage collection.**]{} As we can see from the preservation theorem, the interface to a configuration never changes. While new processes may be spawned, they will have clients and are therefore not visible at the interface. This is in contrast to the semantics of shared channels in prior work (for example, in [@Caires10concur; @Pfenning15fossacs]) where shared channels may show up as newly provided channels. Therefore they may be left over at the end of a computation without any clients.
This cannot happen here. Initially, at the top level, we envision starting with the configuration below on the left. Assuming this computation completes, by the progress property and the definition of *poised*, computation could only halt with the configuration on the right. In other words: no garbage! $$\cdot \vDash {\m{proc}}(\{c_0\}, \cdot, c, P_0) :: (c_0 : \one) \qquad \qquad
\cdot \vDash {\m{proc}}(\{c_0\}, \cdot, c, c.\langle \rangle) :: (c_0 : \one)$$
One can generalize this to allow nontrivial output by allowing any purely positive type (that is, one which only uses the fragment of the logic with connectives $\oplus$, $\otimes$, $\one$, and $\downshift$), such as $\oplus\{\m{false} : \one, \m{true} : \one\}$.
We can formalize this intuition by defining an *observable* configuration $\mathcal C$ which corresponds to our intuitive notion of garbage-free. We only define what it means for a configuration with purely positive type to be observable. It is likely that this definition can be extended to encompass negative types as well, but it is not nearly as natural to do so.
A configuration $\mathcal C$ for which there is $\Psi$ composed entirely of purely positive types such that $\cdot \vDash \mathcal C :: \Psi$ is *observable* at $\Psi$ if, when we repeatedly receive messages from all channels we know about, starting from a state where we only know about $\Psi$, we eventually receive a message from every object in $\mathcal C$. If we do not care about the particular channels in $\Psi$, we may say simply that $\mathcal C$ is *observable*.
We define what it means for a configuration $\mathcal C$ to be observable at $\Psi$ (written $\mathcal C \rhd \Psi$) inductively over the structure of $\mathcal C$.
1. ${\m{proc}}(\{c\}, \cdot, x, x.\langle \rangle) \rhd (c : \one)$.
2. If $\mathcal C \rhd \Psi\; (d : A_m^\ell)$, then $\mathcal C\; {\m{proc}}(\{c\}, \{d\}, x, x.\ell(d)) \rhd \Psi\; (c : \mathop{\oplus}\limits_{i \in I} A_m^i)$.
3. If $\mathcal C \rhd \Psi\; (d : A_m)$, then $\mathcal C\; {\m{proc}}(\{c\}, \{d\}, x, x.\m{shift}(d)) \rhd \Psi\; (c : \down_k^m A_m)$.
4. If $\mathcal C \rhd \Psi\; (d : A_m)\; (e : B_m)$, then $\mathcal C\; {\m{proc}}(\{c\}, \{d, e\}, x, x.\langle d, e \rangle) \rhd \Psi\; (c : A_m \otimes B_m)$.
We can then give the following corollary of our deadlock-freedom theorem:
If $\cdot \vDash \mathcal C :: \Psi$ for some $\Psi$ consisting entirely of purely positive types and $\mathcal C$ cannot take any steps, then $\mathcal C \rhd \Psi$.
This proof proceeds by a simple induction on the derivation of $\cdot \vDash \mathcal C :: \Psi$, using () to work from right to left. At each step, we note that the rightmost process is poised. Because $\Psi$ consists only of purely positive types, the rightmost process must therefore be sending a positive message. Moreover, it can only use channels of purely positive type. Well-typedness of the configuration then lets us apply the inductive hypothesis to the remainder of the configuration, at which point we can simply apply the definition of observability.
Perhaps most interesting here is an extension, following SILL [@Toninho13esop; @Pfenning15fossacs] in a straightforward way, where we allow recursive types and recursive definition beyond the pure logic. For example, we can define binary numbers in “little endian” representation (least significant bit first) as
$\m{bits}_m = \oplus\{\m{b0} : \m{bits}_m, \m{b1} : \m{bits}_m, \m{e} : \one_m\}$
We do not specify here the mode $m$ because the examples below will work for any mode, regardless of its structural properties! Then the number $5 = (101)_2$ is represented by the following process:
$\cdot \vdash \m{five}(x_0) :: (x_0 : \m{bits})$\
$\m{five}(x_0) =$ = $x_4 \leftarrow (\nu a)a.\langle \rangle;$\
$x_3 \leftarrow (\nu a)a.\m{e}(x_4);$\
$x_2 \leftarrow (\nu a)a.\m{b1}(x_3);$\
$x_1 \leftarrow (\nu a)a.\m{b0}(x_2);$\
$\m{send}\; x_0\; \m{b1}(x_1)$
which evolves into five processes: three representing the bits $1$, $0$, $1$; one ($\m{e}$) for the end of the number; and one to close the channel. In fact, by the progress and preservation theorems and inversion on typing, we know that if any process $\cdot \vdash P :: (x_0 : \m{bits}_m)$ terminates, then $P$ will represent a binary number with one process for each of its bits and two more to mark the end of the number and close the channel.
As a last example, the recursive process definition $\m{inc}(x,y)$ reads a stream of bits along channel $x$ and sends an incremented stream along $y$.
$(x : \m{bits}) \vdash \m{inc}(x, y) :: (y : \m{bits})$\
$\m{inc}(x, y) =$ = $\m{case}\; x$ = $\{$ = $\m{b0}(x') \Rightarrow$ = $\m{send}\; y\; \m{b1}(x')$\
$|\,\m{b1}(x') \Rightarrow$ = $y' \leftarrow \nu a.(\m{inc}(x', a));$\
$\m{send}\; y\; \m{b0}(y')$\
$|\,\m{e}(x') \Rightarrow$ = $y' \leftarrow \nu a.(\m{send}\; a\; \m{e}(x'))$\
$\m{send}\; y\; \m{b1}(y') \,\}$
We can obtain the representation of 6 by incrementing 5.
$\cdot \vdash \m{six}(x_0) :: (x_0 : \m{bits})$\
$\m{six}(x_0) = $ = $x_1 \leftarrow \nu a.(\m{five}(a));$\
$\m{inc}(x_1, x_0)$
Related Work {#sec:related}
============
[^3]
Various items of related work have already been mentioned in the preceding sections either in examples or technical cross-references.
The most closely related work and immediate inspiration comes from the unpublished [@Reed09un] which introduces an arbitrary preorder on modes with a uniform logical language and sequent calculus rules. It uses the notation $F_{q \geq p}A_q = \down^q_pA_q$ and $U_{q \geq p}A_q = \up^p_q A_q$. It mostly stays in the realm of structural logics, but Section 4.4 sketches intuitionistic linear logic and LNL as examples. It does not use explicit weakening and contraction rules (which are incorporated into the other rules as is frequently done for sequent calculi), and, while it uniformly proves cut elimination in the case of structural modes, it does not provide an operational semantics.
Recently,Licata et al. [@Licata17fscd] have further generalized Reed’s adjoint logic by uniquely labeling antecedents and then controlling their use through a resource annotation of a sequent. Their resource annotations are made in an expressive *mode theory* which allows a richer set of logics to be represented than in our system here, including non-associative, ordered, and bunched logics. In addition, Licata et al. view multiplicative connectives such as tensor ($\otimes$) or linear implication ($\lolli$) as instances of a new generalized form of the adjoint modal operators, which paves the way for yet additional operators to be represented. In particular, their system allows $?_a A$ of intuitionistic subexponential logic (see Section \[ssec:isell-appendix\]) and $\diamondsuit A$ of judgmental modal logic (see Example \[ex:js4\]) to be encoded directly, which, as far as we can tell, requires at least a 4-point lattice and an additional distinguished atom $a_r$ [@Reed09un Section 4.5] in our setting.
This generality also comes at a price. The declaration of independence is no longer a fundamental notion, but a roughly corresponding *strengthening lemma* has to be proved and requires some complex conditions on the mode theory [@Licata17fscd Lemma 7.1]. Adequacy of encodings also becomes more complicated. Furthermore, Licata et al. do not provide an operational semantics; their interest (like Reed’s) lies on the logical and categorical side.
In a different direction, there is some related work on the use of linearity for garbage collection. The work of Wadler [@wadler1990linear] presents a simple language which uses linearity for state change without a need for garbage collection. Roughly contemporaneous with this is work by Chirimar et al. [@chirimar1992proving], which focuses more on the idea of implementing garbage collection with linear logic, allowing the programmer to “dispose” of variables which are not needed in a type-safe fashion. Again, however, this focuses on a functional language, with the intent to allow for easier reasoning about memory optimizations in functional programming. More recently, [@igarashi2002resource; @igarashi2000garbage] deal with garbage collection (and the more general problem of resource management) using linearity. Yet again, this deals only with the case of sequential functional programs. A major difference between the prior work and our work, then, is that we work in a concurrent setting, and indeed our garbage collection is concurrent as well.
Conclusion
==========
At this point, our formulation of adjoint logic and its operational semantics seem to provide a good explanation for a variety of patterns of asynchronous communication. The key behaviors which we can model (and importantly, model in a uniform fashion) are cancellation, replication, and multicast. We also obtain a foundation for a system of distributed garbage collection. Moreover, if used linearly, our semantics coincides with the purely linear semantics developed in prior work.
In parallel work we have also provided a shared memory semantics for a closely related formulation of adjoint logic with implicit structural rules [@Pfenning18tlla]. In future work, we plan to investigate if the declaration of independence is sufficient to allow a *modular* combination of different operational interpretations for different modes. Of particular interest here would be the semantics with manifest sharing [@Balzer17icfp].
Acknowledgments {#acknowledgments .unnumbered}
===============
Supported by NSF Grant No. CCF-1718267: “Enriching Session Types for Practical Concurrent Programming”
Adjoint Logic {#app:adjoint-logic}
=============
We present here a brief overview of the formulation of adjoint logic that we take as a basis for the semantics presented in the main body of the paper. Adjoint logic can be thought of as a schema to define particular logics. The schema is parameterized by a set of modes of truth $m$, where each proposition and logical connective is indexed by its mode. Furthermore, each mode intrinsically carries a set of structural properties $\sigma(m) \subseteq \{\m{W},\m{C}\}$ where $\m{W}$ stands for *weakening* and $\m{C}$ stands for *contraction*. As a concession to simplicity of the presentation, in this paper we always allow exchange, although nothing stands in the way of an even more general framework [@Pfenning16course]. In addition, an instance requires a preorder between modes, where $m \geq k$ expresses that the proof of a proposition of mode $k$ may depend on a hypotheses of mode $m$. This preorder embodies the *declaration of independence*:
> *A proof of $A_k$ may only depend on hypotheses $B_m$ for $m \geq k$.*
The form of a sequent is $$\Psi \vdash A_k \quad \mbox{where $\Psi \geq k$}$$ where $\Psi$ is a collection of *antecedents* of the form $(x_i : B^i_{m_i})$ with each $m_i \geq k$, where all the variables $x_i$ are distinct. This critical presupposition is abbreviated as $\Psi \geq k$. Furthermore, the order of the antecedents does not matter since we always allow exchange.
In addition, we require the preorder between modes to be compatible with their structural properties: that is, $m \geq k$ implies $\sigma(m) \supseteq \sigma(k)$. This is necessary to guarantee cut elimination.
Finally, we may define fragments by restricting the set of propositions we consider for a given mode.
The propositions at each mode are constructed uniformly, remaining within the same mode, except for the *shift operators* that move between modes. They are $\up_k^m A_k$ (pronounced *up*), which is a proposition at mode $m$ and requires $m \geq k$; and $\down^\ell_m A_\ell$ (*down*), which is also a proposition at mode $m$, and which requires $\ell \geq m$.
At this point we can already write out the syntax of propositions. $$\begin{array}{lcl}
A_m, B_m & \Coloneqq & p_m \mid A_m \lolli_m B_m \mid A_m \tensor_m B_m \mid \one_m \mid \mathop{\oplus_m}\limits_{i \in I} A_m^i
\mid \mathop{\with_m}\limits_{i \in I} A_m^i \mid \up_k^m A_k \mid \down_m^\ell A_\ell
\end{array}$$ Here $p_m$ stands for atomic propositions at mode $m$. Due to the needs of our operational interpretation, we generalize internal and external choice to $n$-ary constructors parameterized by an index set $I$. So we write $A_m^1 \oplus A_m^2 = \mathop{\oplus}\limits_{i \in \{1,2\}} A_m^i$. Remarkably, the right and left rules in the sequent calculus defining the logical connectives are the same for each mode and are complemented by the permissible structural rules.
$\displaystyle
\begin{array}{c}
\infer[\id]
{(x : A_m) \seq A_m}
{\mathstrut}
\qquad
\infer[\cut]
{\Psi \ \Psi' \seq C_k}
{\Psi \geq m \geq k
& \Psi \vdash A_m
& (x : A_m)\ \Psi' \vdash C_k
}
\\[1em]
\infer[\weaken]
{\Psi \ (x : A_m) \seq C_k}
{\m{\m{W}} \in \sigma(m)
&\Psi \seq C_k
}
\qquad
\infer[\contract]
{\Psi \ (x : A_m) \seq C_k}
{\m{\m{C}} \in \sigma(m)
&\Psi \ (y : A_m) \ (z : A_m) \seq C_k
}
\\[1em]
\infer[{\oplus} R_\ell]
{\Psi \seq \mathop{\oplus}\limits_{i \in I} A_m^i}
{\ell \in I
& \Psi \seq A_m^\ell}
\qquad
\infer[{\oplus} L]
{\Psi \ (x : \mathop{\oplus}\limits_{i \in I} A_m^i) \seq C_k}
{\Psi \ (y : A_m^i) \seq C_k \text{ for each $i \in I$}}
\\[1em]
\infer[{\with} R]
{\Psi \seq \mathop{\with}\limits_{i \in I} A_m^i}
{\Psi \seq A_m^i \text{ for each $i \in I$}}
\qquad
\infer[{\with} L_\ell]
{\Psi\ (x :\mathop{\with}\limits_{i \in I} A_m^i) \seq C_k}
{\ell \in I
& \Psi\ (y : A_m^\ell) \seq C_k
}
\\[1em]
\infer[{\otimes} R]
{\Psi \ \Psi' \seq A_m \otimes B_m}
{\Psi \seq A_m
&\Psi' \seq B_m
}
\qquad
\infer[{\otimes} L]
{\Psi \ (x : A_m \otimes B_m) \seq C_k}
{\Psi \ (y : A_m) \ (z : B_m) \seq C_k}
\qquad\qquad
\infer[{\one} R]
{\cdot \seq \one_m}
{\mathstrut}
\qquad
\infer[{\one} L]
{\Psi \ (x : \one_m) \seq C_k}
{\Psi \seq C_k}
\\[1em]
\infer[{\lolli} R]
{\Psi \seq A_m \lolli B_m}
{(x : A_m) \ \Psi \seq B_m}
\qquad
\infer[{\lolli} L]
{\Psi \ \Psi' \ (x : A_m \lolli B_m) \seq C_k}
{\Psi' \geq m
&\Psi' \seq A_m
&\Psi \ (y : B_m) \seq C_k
}
\\[1em]
\infer[{\upshift} R]
{\Psi \seq \upshift_k^m A_k}
{\Psi \seq A_k}
\qquad
\infer[{\upshift} L]
{\Psi \ (x : \upshift_k^m A_k) \seq C_\ell}
{k \geq \ell
&\Psi \ (y : A_k) \seq C_\ell
}
\qquad\qquad
\infer[{\downshift} R]
{\Psi \seq \downshift_k^m A_m}
{\Psi \geq m
&\Psi \seq A_m
}
\qquad
\infer[{\downshift} L]
{\Psi \ (x : \downshift_k^m A_m) \seq C_\ell}
{\Psi \ (y : A_m) \seq C_\ell}
\end{array}
$
Judgmental and structural rules
-------------------------------
The rules for adjoint logic can be found in \[fig:basic-logic-rules\], in which we give a more standard presentation of the logic than that used by the operational semantics (\[fig:message-passing-typing-rules\]). We begin with the judgmental rules of identity and cut, which express the connection between antecedents and succedents. Identity says that if we assume $A_m$ we are allowed to conclude $A_m$. Cut says the opposite: if we can conclude $A_m$ we are allowed to assume $A_m$ *as long as the declaration of independence is respected*.
As is common for the sequent calculus, we read the rules in the direction of bottom-up proof construction. For the cut rule, this means we should assume that the conclusion $\Psi\ \Psi' \vdash C_k$ is well-formed and, in particular, that $\Psi \geq k$ and $\Psi' \geq k$. Therefore, if we check that $m \geq k$, then we know that the second premise, $(x : A_m)\ \Psi' \vdash C_k$, will also be well-formed. For the first premise to be well-formed, we need to check outright that $\Psi \geq m$.
In LNL [@Benton94csl] we have two modes, $\mU > \mL$ with $\sigma(\mU) = \{\m{W},\m{C}\}$ and $\sigma(\mL) = \{\,\}$. This one cut rule encompasses exactly the three cut rules in LNL and their restrictions via the declaration of independence:
$m = k = \mL$
: corresponds to $\mathcal{L}\mathcal{L}\mbox{\it-cut}$ with no further restriction.
$m = \mU$ and $k = \mL$
: corresponds to $\mathcal{C}\mathcal{L}\mbox{\it-cut}$ where $\Psi \geq \mU$ enforces that the first premise depends only on structural antecedents $B_\mU$.
$m = \mU$ and $k = \mU$
: corresponds to $\mathcal{C}\mathcal{C}\mbox{\it-cut}$ where $\Psi \geq \mU$ is already known by the well-formedness of the conclusion: both $\Psi$ and $\Psi'$ consist only of structural propositions $B_\mU$.
The structural rules of weakening and contraction just need to verify that the mode of the principal formula permits the rule.
Additive and multiplicative connectives
---------------------------------------
The logical rules defining the additive and multiplicative connectives are simply the linear rules for all modes, since we have separated out the structural rules. Except in one case, ${\lolli}L$, the well-formedness of the conclusion implies the well-formedness of all premises.
As for ${\lolli}L$, we know from the well-formedness of the conclusion that $\Psi \geq k$, $\Psi' \geq k$, and $m \geq k$. These facts by themselves already imply the well-formedness of the second premise, but we need to check that $\Psi' \geq m$ in order for the first premise to be well-formed.
Shifts
------
The shifts represent the most interesting aspects of the rules. Recall that in $\up^m_k A_k$ and $\down^m_k A_m$ we require that $m \geq k$. We first consider the two rules for $\up$. We know from the conclusion of the right rule that $\Psi \geq m$ and from the requirement of the shift that $m \geq k$. Therefore, as $\geq$ is transitive, $\Psi \geq k$ and the premise is always well-formed. This also means (although we do not prove it here) that this rule is *invertible*.
From the conclusion of the left rule, we know $\Psi \geq \ell$, $m \geq \ell$, and $m \geq k$. This does not imply that $k \geq \ell$, which we need for the premise to be well-formed and thus needs to be checked. Therefore, this rule is non-invertible.
The downshift rules are constructed analogously, taking only the declaration of independence and properties of the preorder $\leq$ as guidance. Note that in this case the left rule is always applicable (that is, invertible), while the right rule is non-invertible.
Logic Examples
--------------
We now describe how adjoint logic can be used to embed various other logics, and provide some examples to justify our presentation.
We obtain intuitionistic linear logic [@Girard87tcs; @Barber96] by using two modes, $\mU$ (for *structural*) and $\mL$ (for *linear*) with $\mU > \mL$. Moreover, $\sigma(\mU) = \{\m{W},\m{C}\}$ and $\sigma(\mL) = \{\,\}$, and the structural layer contains only the shifted proposition. $$\begin{array}{lcl}
A_\mU & \Coloneqq & \up_\mL^\mU A_\mL \\
A_\mL, B_\mL & \Coloneqq & p_\mL \mid A_\mL \lolli B_\mL \mid A_\mL \tensor B_\mL \mid \one
\mid \mathop{\oplus}\limits_{i \in I} A_\mL^i
\mid \mathop{\with}\limits_{i \in I} A_\mL^i
\mid \down_\mL^\mU A_\mU
\end{array}$$ In this representation the exponential modality is decomposed into shift modalities $\bang A_\mL = \down^\mU_\mL\, \up^\mU_\mL A_\mL$. Unlike Chang et al. [@Chang03tr], our sequent calculus employs explicit structural rules of weakening and contraction on unrestricted propositions $A_\mU$. We do not state an explicit correctness theorem because it follows from the embedding of LNL (Theorem \[thm:lnl\]) and Benton’s results [@Benton94csl].
\[ex:LNL\] We obtain LNL [@Benton94csl] just like linear logic with two modes $\mU > \mL$, but we populate the unrestricted layer with additional propositions, where we write ${\times} = {\tensor_\mU}$ and ${\rightarrow} = {\lolli_\mU}$. $$ \begin{array}{lcc@{{}\mid{}}c@{{}\mid{}}c@{{}\mid{}}c@{{}\mid{}}c}
A_\mU, B_\mU & \Coloneqq & p_\mU & A_\mU \rightarrow B_\mU & A_\mU \times B_\mU & 1_\mU & \up_\mL^\mU A_\mL \\
A_\mL, B_\mL & \Coloneqq & p_\mL & A_\mL \lolli B_\mL & A_\mL \tensor B_\mL & \one_\mL & \down_\mL^\mU A_\mU
\end{array}$$ Benton’s notation for shifts is $F = \down_\mL^\mU$ and $G = \up_\mL^\mU$. Our formulation then combines the various versions of the rules by combining the two contexts, using the declaration of independence instead to force that unrestricted succedents depend only on unrestricted antecedents. A small difference arises only in the ${\times}\textit{-left}$ rules where our version has both components in the premise, which is of course logically equivalent to LNL in the presence of weakening and contraction.
\[thm:lnl\] If we let $\tau$ embed propositions of LNL into the instance of adjoint logic described above, then
(a) $\Theta \seq_{\mathcal C} X$ in LNL iff $\tau(\Theta) \seq \tau(X)$ in adjoint logic.
(b) $\Theta; \Gamma \seq_{\!\mathcal L} A$ in LNL iff $\tau(\Theta), \tau(\Gamma) \seq \tau(A)$ in adjoint logic.
\[ex:counter-independence\] Consider linear logic or LNL and consider the following **faulty(!)** “proof” showing that contraction for linear propositions is derivable: $$\infer[\m{cut}]
{(x:A_\mL) \vdash C_\mL}
{\infer[{\up}R]
{(x:A_\mL) \vdash^{??} \up_\mL^\mU A_\mL}
{\infer[\m{id}]{(x:A_\mL) \vdash A_\mL}{\mathstrut}}
&
\infer[\m{contract}]
{(u:\up^\mU_\mL A_\mL) \vdash C_\mL}
{\infer[{\up}L]
{(v:\up^\mU_\mL A_\mL)\ (w:\up^\mU_\mL A_\mL) \vdash C_\mL}
{\infer[{\up}L]
{(y:A_\mL)\ (w:\up^\mU_\mL A_\mL) \vdash C_\mL}
{(y:A_\mL)\ (z:A_\mL) \vdash C_\mL}}}}$$ The fallacy lies with the sequent marked $\vdash^{??}$ because it violates our declaration of independence: the succedent $\up^\mU_\mL A_\mL$ of mode $\mU$ depends on an antecedent of mode $\mL$, and $\mL \not \geq \mU$.
If we wanted to blame a particular inference, it would be either $\m{cut}$, viewed bottom-up, or ${\up}R$, viewed top-down. In our case, the bottom-up construction of this proof would fail because the condition $(x:A_\mL) \geq \mU$ of the cut rule is violated.
It is an immediate corollary that cut elimination fails if the declaration of independence is not enforced. For example, using the above faulty reasoning, we could prove $A_\mL \vdash A_\mL \tensor A_\mL$, which in general has no cut-free proof.
Multicut {#ssec:multicut}
--------
Because we have an explicit rule of contraction, cut elimination does not follow by a simple structural induction. However, we can follow Gentzen [@Gentzen35] and allow multiple copies of the same proposition to be removed by the cut, which then allows a structural induction argument. In anticipation of the operational interpretation, we have labeled our antecedents with unique variables, so the generalized form of cut called *multicut* (see, for example, [@Negri01book]) can remove $n \geq 0$ copies. Of course, such cuts are only legal if the propositions that are removed satisfy the necessary structural rules. For $n = 0$, we require that the mode $m$ support weakening. $$\infer[\cut(\emptyset)]
{\Psi \ \Psi' \seq C_k}
{\Psi \geq m \geq k
& \m{W} \in \sigma(m)
&\Psi \seq A_m
&\Psi' \seq C_k
}$$ For $n = 1$, we obtain the usual cut rule and no special requirements are needed. $$\infer[\cut(\{x\})]
{\Psi \ \Psi' \seq C_k}
{\Psi \geq m \geq k
& \Psi \seq A_m
& (x : A_m) \ \Psi' \seq C_k
}$$ For $n \geq 2$, the mode of the cut formula must admit contraction. $$\infer[\cut(S \cup \{x, y\})]
{\Psi \ \Psi' \seq C_k}
{\begin{array}[b]{l}
\m{C} \in \sigma(m) \\
\Psi \geq m \geq k
\end{array}
\; \Psi \seq A_m
& (S \cup \{x,y\} : A_m) \ \Psi' \seq C_k
}$$ Here, we have used the abbreviation $(\{x_1, \dotsc, x_n\} : A_m)$ to stand for $(x_1 : A_m) \ldots (x_n : A_m)$.
Note that each of these rules has a side condition that can be interpreted informally as stating that the number of antecedents cut must be compatible with the mode $m$: if there are no antecedents removed, $m$ must admit weakening, and if we remove two or more, $m$ must admit contraction. This is exactly $|S| \sim m$ as defined in \[sec:typing\].
This allows us to write down a single rule encompassing all three of the above cases for multicut: $$\infer[\cut(S)]
{\Psi \ \Psi' \seq C_k}
{\Psi \geq m \geq k
&|S| \sim m
&\Psi \seq A_m
& (S : A_m) \ \Psi' \seq C_k
}$$ Note that the standard cut rule is the instance of the multicut rule where $|S| = 1$, and so proving multicut elimination for adjoint logic also yields cut elimination for the standard cut rule.
Identity Expansion and Cut Elimination
--------------------------------------
We present standard identity expansion and cut elimination results as evidence for the correctness of the sequent calculus as capturing the meaning of the logical connectives via their inference rules. Cut-free proofs will always decompose propositions when read from conclusion to premise and thus yield a conservative extension result. Finally, the fine detail of the proof is significant because the cut reductions, which constitute the essence of the proof, are the basis for the operational semantics.
If $\Psi \seq A_m$, then there exists a proof that $\Psi \seq A_m$ using identity rules only at atomic propositions, which is cut-free if the original proof is.
We begin by proving that for any formula $A_m$, there is a cut-free proof that $(x : A_m) \seq A_m$ using identity rules only at atomic propositions. This follows easily from an induction on $A_m$.
Now, we arrive at the theorem by induction over the structure of the given proof that $\Psi \seq A_m$.
\[thm:cut-elim\] If $\Psi \seq A_m$, then there is a cut-free proof of $\Psi \seq A_m$.
This proof follows the structure of many cut-elimination results. First we prove admissibility of multicut in the cut-free system. This is established by a straightforward nested induction, first on the proposition $A_m$ and then simultaneously on the structure of the deductions $\DD$ and $\EE$. This is followed by a simple structural induction to prove cut elimination, using the admissibility of (multi)cut when it is encountered. If we ignore the modes, this proof is very similar to the original proof of Gentzen [@Gentzen35].
Adjoint logic is a conservative extension of each of the logics at a fixed mode. That is, if $\Psi \seq A_m$ is a sequent purely at mode $m$ (in that every type in $\Psi$ is at mode $m$ and neither $A_m$ nor the types in $\Psi$ make use of shifts), then $\Psi \seq A_m$ is provable using the rules of adjoint logic iff it is provable using the rules which define the logic at mode $m$.
$\displaystyle
\begin{array}{rl}
&\infer-[\cut(\emptyset)]
{(x : A_m) \ \Psi' \seq C_k}
{
(x : A_m) \ge m \ge k
&|\emptyset| \sim m
&\infer[\id_{A_m}]{(x : A_m) \sseq A_m}{}
&\deduce
{\Psi' \sseq C_k}
{\EE}
}
\\[1em]
\Longrightarrow \qquad
&\infer[\weaken]
{(x : A_m) \ \Psi' \seq C_k}
{\m{W} \in \sigma(m)
&\deduce
{\Psi' \seq C_k}
{\EE}
}
\label{eq:cut-id-EE}
\\[2em]
&\infer-[\cut(\{y\})]
{\Psi \ \Psi' \seq C_k}
{
\Psi \ge m \ge k
&|\{y\}| \sim m
&\deduce{\Psi \sseq A_m}{\DD}
&\infer[\contract]
{\Psi' \ (y : A_m) \sseq C_k}
{\m{C} \in \sigma(m)
&\deduce{\Psi' \ (z : A_m) \ (w : A_m) \sseq C_k}{\EE'}
}
}
\\[1em]
\Longrightarrow \qquad
&\infer[\ih{A_m, \DD, \EE'}(\{z, w\})]
{\Psi \ \Psi' \seq C_k}
{\Psi' \ (z : A_m) \ (w : A_m) \ge m \ge k
&|\{z, w\}| \sim m
&\deduce{\Psi \sseq A_m}{\DD}
&\deduce{\Psi' \ (z : A_m) \ (w : A_m) \sseq C_k}{\EE'}
}
\label{eq:cut-DD-contract}
\\[2em]
&\infer-[\cut(T \cup \{x, y\})]
{\Psi \Psi' \sseq C_k}
{
\Psi \ge m \ge k
&|T \cup \{x, y\}| \sim m
&\inferlabel{\DD = {}}
\infer[{\oplus} R_\ell]
{\Psi \sseq \mathop{\oplus}\limits_{i \in I} A_m^i}
{\ell \in I
&\deduce{\Psi \sseq A_m^\ell}{\DD_1}
}
&\infer[{\oplus} L]
{\Psi' \ (T \cup \{x, y\} : \mathop{\oplus}\limits_{i \in I} A_m^i) \sseq C_k}
{
\deduce{\Psi' \ (T \cup \{x\} : \mathop{\oplus}\limits_{i \in I} A_m^i) \ (z : A_m^i) \sseq C_k}{\EE_i}
\text{ for each $i \in I$}
}
}
\\[1em]
\Longrightarrow \qquad
&\infer[\contract^*]
{\Psi \ \Psi' \sseq C_k}
{\m{C} \in \sigma(m)
&\infer[\ih{A_m^\ell, \DD_1, ...}(\{z\})]
{\Psi \ \Psi \ \Psi' \sseq C_k}
{\Psi \ge m \ge k
&\deduce{\Psi \sseq A_m^\ell}{\DD_1}
&\infer[\ih{...}]
{\Psi \ \Psi' \ (z : A_m^\ell) \sseq C_k}
{\Psi \ge m \ge k
&|T \cup \{x\}| \sim m
&\deduce{\Psi \sseq \mathop{\oplus}\limits_{i \in I} A_m^i}{\DD}
&\deduce{\Psi' \ (T \cup \{x\} : \mathop{\oplus}\limits_{i \in I} A_m^i) (z : A_m^\ell) \sseq C_k}{\EE_\ell}
}
}
}
\\[2em]
&\infer-[\cut(S)]
{\Psi \ \Psi' \sseq A_m \lolli B_m}
{
\Psi \ge k \ge m
&|S| \sim k
&\deduce
{\Psi \sseq C_k}
{\DD}
&\infer[{\lolli} R]
{(S : C_k) \ \Psi' \sseq A_m \lolli B_m}
{
\deduce
{(x : A_m) \ (S : C_k) \ \Psi' \sseq B_m}
{\EE_1}
}
}
\\[1em]
\Longrightarrow \qquad
&\infer[{\lolli} R]
{\Psi \ \Psi' \sseq A_m \lolli B_m}
{\infer[\ih{C_k, \DD, \EE_1}(S)]
{(x : A_m) \ \Psi \ \Psi' \sseq B_m}
{\Psi \ge k \ge m
&|S| \sim k
&\deduce{\Psi \sseq C_k}{\DD}
&\deduce{(x : A_m) \ (S : C_k) \ \Psi' \sseq B_m}{\EE_1}
}
}
\end{array}
$
Adjunction properties {#ssec:adjunction}
---------------------
As yet, we have not discussed the meaning of the name “-0.1em*adjoint logic*”. This can be justified by showing that for fixed $k \le m$, $\downshift_k^m$ and $\upshift_k^m$ yield an adjoint pair of functors $\downshift_k^m \dashv \upshift_k^m$. Since prior results (see [@Benton94csl] and [@Licata17fscd]) already establish this property and we have little new to contribute here, we omit the details here.
Asynchronous Adjoint Logic {#app:async-adjoint-logic}
==========================
As has been observed before, intuitionistic and classical linear logics can be put into a Curry–Howard correspondence with session-typed communicating processes [@Caires10concur; @Wadler12icfp; @Caires16mscs]. A linear logical proposition corresponds to a session type, and a sequent proof to a process expression. The transition rules of the operational semantics derive from the cut reductions.
Under the intuitionistic interpretation a sequent proof[^4] of $$(x_1 : A_\mL^1)\cdots(x_n : A_\mL^n) \vdash (x : A_\mL)$$ corresponds to a process $P$ that *provides* channel $x$ and uses channels $x_i$. The types of the channels prescribe the pattern of communication: in the succedent, positive types (${\oplus}, {\tensor}, {\one}$) will send and negative types (${\with}, {\lolli}$) will receive. In the antecedent, the roles are reversed. Cut corresponds to parallel composition of two processes, with a private channel between them, while identity simply equates two channels.
Enforcing Asynchronous Communication {#ssec:async}
------------------------------------
Under this interpretation, a cut of a right rule against a matching left rule allows computation to proceed by mimicking the cut reduction from the proof of Theorem \[thm:cut-elim\]. For example, a cut at type $\mathop{\oplus}\limits_{i \in I} A_\mL^i$ is replaced by a cut at type $A_\mL^\ell$ for some $\ell \in I$. This corresponds to passing a message (‘$\ell$’) from the process *providing* $x : \mathop{\oplus}\limits_{i \in I} A_\mL^i$ to the process *using* $x$. By its very nature, this form of cut reduction is *synchronous*: both provider and client proceed simultaneously because the channel $x : A_\ell$ connects the two process continuations.
For realistic languages, and also for the paradigm to smoothly extend to the case of adjoint logic where some modes permit weakening and contraction, we would like to prescribe *asynchronous communication* instead.
We observe that the *asynchronous $\pi$-calculus* replaces the usual action prefix for output $x\langle y\rangle. P$ by a process expression $x\langle y\rangle$ *without a continuation*, thereby ensuring that communication is asynchronous. Such a process represents the message $y$ sent along channel $x$. Under our interpretation, the continuation process corresponds to the proof of the premise of a rule. Therefore, if we can restructure the sequent calculus so that the rules that send (${\oplus}R$, ${\one}R$, ${\tensor}R$, ${\down}R$, ${\with}L$, ${\lolli}L$, ${\up}L$) have zero premises, then we may achieve a similar effect.
As an example, we consider the two right rules for ${\oplus}$. Reformulated as axioms, they become $$\infer[{\oplus}R^0_1]
{A \vdash A \oplus B}
{} \hspace{3em}
\infer[{\oplus}R^0_2]
{B \vdash A \oplus B}
{} $$ In the presence of cut, these two rules together produce the same theorems as the usual two right rules. In one direction, we use cut $$\infer[\m{cut}_A]
{\Delta \vdash A \oplus B}
{\Delta \vdash A
& \infer[{\oplus}R^0_1]
{A \vdash A \oplus B}
{}}
\hspace{3em}
\infer[\m{cut}_B]
{\Delta \vdash A \oplus B}
{\Delta \vdash B
& \infer[{\oplus}R^0_2]
{B \vdash A \oplus B}
{}}$$ and in the other direction we use identity $$\infer[{\oplus}R_1]
{A \vdash A \oplus B}
{\infer[\m{id}_A]{A \vdash A}{}}
\hspace{3em}
\infer[{\oplus}R_2]
{B \vdash A \oplus B}
{\infer[\m{id}_B]{B \vdash B}{}}$$ to derive the other rules.
Returning to the $\pi$-calculus, instead of explicitly *sending* a message $a\langle b\rangle.\, P$ we *spawn* a new process in parallel $a\langle b\rangle \mid P$. This use of parallel composition corresponds to a cut; receiving a message is achieved by cut reduction: $$\infer[\m{cut}_{A \oplus B}]
{\Delta', A \vdash C}
{\infer[{\oplus}R^0_1]{A \vdash A \oplus B}{\mathstrut}
& \infer[{\oplus}L]
{\Delta', A \oplus B \vdash C}
{\deduce[Q_1]{\Delta', A \vdash C}{}
& \deduce[Q_2]{\Delta', B \vdash C}{}}}
\quad \Longrightarrow \quad
\deduce[Q_1]{\Delta', A \vdash C}{}$$ We see the cut reduction completely eliminates the cut in one step, which corresponds precisely to receiving a message. In this example the message would be $\pi_1$ since the axiom ${\oplus}R^0_1$ was used; for ${\oplus}R^0_2$ it would be $\pi_2$.
In summary, if we restructure the sequent calculus so that the non-invertible rules (those that send) have zero premises, then (1) messages are proofs of axioms, (2) message sends are modeled by cut, and (3) message receives are a new form of cut reduction with a single continuation.
In the process we give something up, namely the traditional cut elimination theorem. For example, the sequent $\cdot \vdash \one \oplus \one$ has no cut-free proof since no rule matches this conclusion. The saving grace is that we can reach a normal form where each cut just simulates the usual rules of the sequent calculus. This can be shown by translation to the ordinary sequent calculus, applying cut elimination, and translating the result back. Proofs in this normal form have the subformula property. Perhaps more importantly, we have session fidelity and deadlock freedom (\[sec:metatheory\]) for the corresponding process calculus even in the presence of recursive types and processes, which is ultimately what we care about for the resulting concurrent programming language.
Eliminating Weakening and Contraction {#ssec:structural-elimination}
-------------------------------------
We have introduced multicut entirely with the standard motivation of providing a simple proof of the admissibility of cut using structural induction. Surprisingly, we can streamline the system further by using multicut to eliminate weakening and contraction from the logic altogether, as in the system we use as the basis for our typing rules (\[fig:message-passing-typing-rules\]).
Consider a mode $m$ with $\m{C} \in \sigma(m)$. Then contraction is a simple instance of multicut with an instance of the identity rule. $$\infer[\cut(\{y, z\})]
{\Psi \ (x : A_m) \seq C_k}
{\infer[\id]
{(x : A_m) \seq A_m}
{\mathstrut}
& \Psi \ (y : A_m) \ (z : A_m) \seq C_k
}$$ Similarly, for a mode $m$ with $\m{W} \in \sigma(m)$, weakening is also an instance of multicut. $$\infer[\cut(\emptyset)]
{\Psi \ (x : A_m) \seq C_k}
{\infer[\id]
{(x : A_m) \seq A_m}
{\mathstrut}
& \Psi \seq C_k}$$ Cut reductions in the presence of contraction entail many residual contractions, as is evident already from Gentzen’s original proof. Under our interpretation of contraction above, these residual contractions simply become multicuts with the identity. The operational interpretation of identities then plays three related roles: with one client, an identity achieves a renaming, redirecting communication; with two or more clients, an identity implements copying; with zero clients, its effect is cancellation or garbage collection. The central role of identities can be seen in full detail in Figure \[fig:dynamics1\], once we have introduced our notation for processes and process configurations.
Program Examples {#sec:example}
================
In the examples that follow, we will work with two modes, $\mL$ and $\mU$, with $\mL < \mU$, $\sigma(\mL) = \emptyset$, and $\sigma(\mU)~=~\{W, C\}$. In these examples we also use recursively defined types and processes without formally defining these constructs, since they are well-known from the literature and orthogonal to our concerns (see, for example, [@Toninho13esop]).
Example: Circuits {#sec:circuits}
-----------------
We call channels $c_\mU$ that are subject to weakening and contraction *shared channels*. As an example that requires shared channels we use circuits. We start by programming a nor gate that processes infinite streams of zeros and ones.
$\m{bits}^\infty_\mU = {\oplus}\{\m{b0} : \m{bits}^\infty_\mU, \m{b1} : \m{bits}^\infty_\mU\}$\
$x : \m{bits}^\infty_\mU, y : \m{bits}^\infty_\mU \vdash \mi{nor} :: (z : \m{bits}^\infty_\mU)$\
$z \leftarrow \mi{nor} \leftarrow x,y =$\
$\m{case}\, x\,$ = $(\,\m{b0}(x') \Rightarrow \m{case}\, y\,$ = $(\, \m{b0}(y') \Rightarrow$ = $z' \leftarrow z.\m{b1}(z') \semi$\
$z' \leftarrow \mi{nor} \leftarrow x',y'$\
$\mid \m{b1}(y') \Rightarrow z' \leftarrow z.\m{b0}(z') \semi$\
$z' \leftarrow \mi{nor} \leftarrow x',y'\,)$\
$\mid \m{b1}(x') \Rightarrow \m{case}\, y\,$ = $(\, \m{b0}(y') \Rightarrow$ = $z' \leftarrow z.\m{b0}(z') \semi$\
$z' \leftarrow \mi{nor} \leftarrow x',y'$\
$\mid \m{b1}(y') \Rightarrow z' \leftarrow z.\m{b0}(z') \semi$\
$z' \leftarrow \mi{nor} \leftarrow x',y'\,)\,)$
This is somewhat verbose, but note that all channels here are shared. For this particular gate they could also be linear because they are neither reused nor canceled. This illustrates that programming can be uniform at different modes, which is a significant advantage of our system over systems of session types based on linear logic with an exponential ${!}A$. Our implementation of $\mi{nor}$ has the property that for bits $A$, $B$, and $C$ with $C = \lnot (A \lor B)$, the following transitions are possible and characterize $\mi{nor}$:
${\m{proc}}(\{a\}, \{a'\}, a, a.A(a')), {\m{proc}}(\{b\}, \{b'\}, b, b.B(b')),
{\m{proc}}(S, \{a, b\}, c, c \leftarrow \mi{nor} \leftarrow a,b)$\
$\null \longrightarrow^*
{\m{proc}}(c', \{a', b'\}, c', c' \leftarrow \mi{nor} \leftarrow a',b'),
{\m{proc}}(S, \{c'\}, c, c.C(c'))$ ($c'$ fresh)
This multi-step reduction is shown in full (one step at a time) below. We only show the initial portion of each process term, which is enough to disambiguate where in the program we are, as otherwise process terms become unwieldy and reduce clarity. We also assume the existence of a rule $\m{call}$ that lets us invoke a defined process, replacing the call with the process definition, after appropriate substitution. At each step, we have highlighted in red the process(es) that are about to transition.
$$\begin{array}{rc}
{\m{proc}}(\{a\}, \{a'\}, a, a.A(a')), {\m{proc}}(\{b\}, \{b'\}, b, b.B(b')),
{\color{red} {\m{proc}}(S, \{a, b\}, c, c \leftarrow \mi{nor} \leftarrow a,b)}
&\overset{\m{call}}{\Longrightarrow} \\[.5em]
{\color{red} {\m{proc}}(\{a\}, \{a'\}, a, a.A(a'))}, {\m{proc}}(\{b\}, \{b'\}, b, b.B(b')),
{\color{red} {\m{proc}}(S, \{a, b\}, c, \m{case}\; a \ldots)}
&\overset{{\oplus}\; C}{\Longrightarrow} \\[.5em]
{\color{red} {\m{proc}}(\{b\}, \{b'\}, b, b.B(b'))},
{\color{red} {\m{proc}}(S, \{a', b\}, c, \m{case}\; b \ldots)}
&\overset{{\oplus}\; C}{\Longrightarrow} \\[.5em]
{\color{red} {\m{proc}}(S, \{a', b'\}, c, z' \leftarrow \ldots)}
&\overset{\cut(\{z'\})}{\Longrightarrow} \\[.5em]
{\m{proc}}(\{c'\}, \{a', b'\}, z', z' \leftarrow \mi{nor} \leftarrow a', b'), {\m{proc}}(S, \{c'\}, c, c.C(c'))
\end{array}$$
When we build an or-gate out of a nor-gate we need to exploit sharing to implement simple negation. In the example below, $u$ and $u'$ are both names for the same shared channel. The process invoked as $\mi{nor} \leftarrow x, y$ will multicast a message to the clients of $u$ and $u'$.
$x : \m{bits}^\infty_\mU, y : \m{bits}^\infty_\mU \vdash \mi{or} :: (z : \m{bits}^\infty_\mU)$\
$z \leftarrow \mi{or} \leftarrow x,y =$\
$\{u, u'\} \leftarrow \mi{nor} \leftarrow x,y$\
$z \leftarrow \mi{nor} \leftarrow u,u'$
An analogous computation to the above is possible, except that at an intermediate stage of the computation, we will also have a shared channel $d$ carrying the (multicast) message ${\m{proc}}(\{u, u'\}, \{d'\}, d, d.D(d'))$ with $D = \lnot(A \lor B)$.
Example: Map {#sec:map}
------------
Mapping a process over a list allows us to demonstrate the use of replicable services, as well as cancellation. We define a whole family of types indexed by a type $A$, which is not formally part of the language but is expressed at the metalevel.
$\mi{list}_A = {\oplus}\{\m{cons} : A \tensor \mi{list}_A, \m{nil} : \one\}$
Such a list should not be viewed as a data structure in memory. Instead, it is a behavioral description of a stream of messages. A process that maps a channel of type $A$ to one of type $B$ will itself have type $A \lolli B$. However, this process must be shared since it needs to be applied to every element. We therefore obtain the following type and definition, where all channels not annotated with a mode subscript are at mode $\mL$.
$\m{f}_\mU : \up_\mL^\mU(A_\mL \lolli B_\mL), l : \mi{list}_A \vdash \mi{map} :: (k : \mi{list}_B)$\
$k \leftarrow \mi{map} \leftarrow f_\mU, l =$\
$\m{case}\, l\,$ = $(\, \m{cons}(l') \Rightarrow$ = $\m{case}\; l' (\langle x, l''\rangle \Rightarrow$ = % receive element $x:A$ with continuation $l''$\
$\{f_\mU', f_\mU''\} \leftarrow (\nu a)a \leftarrow f_\mU$ % duplicate the channel $f_\mU$\
$f' \leftarrow f_\mU'.\m{shift}(f') \semi$ % obtain a fresh linear instance $f'$ of $f_\mU'$\
$y \leftarrow f'.\langle x,y\rangle \semi $ % send $x$ to $f'$, response will be along fresh $y$\
$k' \leftarrow k.\m{cons}(k') \semi$ % select $\m{cons}$\
$k'' \leftarrow k'.\langle y,k''\rangle \semi$ % send $y$ with continuation $k''$\
$k'' \leftarrow \mi{map} \leftarrow f_\mU'', l'')$ % recurse with continuation channels\
$\mid \m{nil}(l') \Rightarrow$ $\emptyset \leftarrow (\nu a)a \leftarrow f_\mU$ % Cancel the channel $f_\mU$\
$k' \leftarrow k.\m{nil}(k') \semi$ % select $\m{nil}$\
$\m{case}\; l' (\langle \rangle \Rightarrow$ % wait for $l'$ to close\
$k'.\langle\,\rangle \,))$ % close $k'$ and terminate
In this example, $f_\mU$ is a replicable and cancelable service. In the case of a nonempty list, we create two names for the channel $f_\mU$ — one to use immediately and one to pass to the recursive call. Note that the service itself remains a single service with two clients until the message $\m{shift}(f')$ is sent to it, at which point it replicates itself, creating one copy to handle this request and leaving another to deal with future requests. In the case of an empty list, we have no elements to map over, and so we do not need to use $f_\mU$. As such, we cancel it before continuing.
Typing and Reduction Examples {#sec:typing-reduction-examples}
=============================
We present here several example typing derivations for the processes shown in \[sec:typing\], as well as several examples of reduction sequences to illustrate the dynamic semantics presented in \[sec:operational\]. First, we provide the following derivation for typing a process witnessing commutativity of $\otimes$: $$\infer[{\otimes L}]
{(x : A_m \otimes B_m) \vdash \m{case}\; x\,(\langle y, x' \rangle \Rightarrow z.\langle x', y \rangle) :: (z : B_m \otimes A_m)}
{\infer[{\otimes R}^0]
{(y : A_m) (x' : B_m) \vdash z.\langle x', y \rangle :: (z : B_m \otimes A_m)}
{\mathstrut}
}$$
We now provide the derivation for typing the following process, which witnesses that (in the presence of contraction at mode $m$), $A_m \with B_m$ proves $A_m \otimes B_m$.
$(p : A_m \with B_m) \vdash$ = $\{p_1, p_2\} \leftarrow (\nu q)\, (q \leftarrow p); \qquad$ = $\%\ \{p_1, p_2\} \leftarrow \m{copy}\; p$\
$x \leftarrow (\nu a)\, p_1.\pi_1(a);$\
$y \leftarrow (\nu b)\, p_2.\pi_2(b);$\
$z.\langle x, y \rangle$ $:: (z : A_m \otimes B_m)$
We omit the process terms and types from the derivation for space reasons, but they can easily be filled in — this mostly serves to show the structure of the typing derivation, rather than being a full example.
$$\infer[\cut(\{p_1, p_2\})]
{p \vdash z}
{\infer[\id]
{p \vdash q}
{\mathstrut}
&\infer[\cut(\{x\})]
{p_1, p_2 \vdash z}
{\infer[{\with L}^0_{\pi_1}]
{p_1 \vdash x}
{\mathstrut}
&\infer[\cut(\{y\})]
{x, p_2 \vdash z}
{\infer[{\with L}^0_{\pi_2}]
{p_2 \vdash y}
{\mathstrut}
&\infer[{\otimes R}^0]
{x, y \vdash z}
{\mathstrut}
}
}
}$$ [^5]
We now move on to demonstrate a possible reduction sequence for the process $\m{nor}$ defined in \[sec:circuits\], showing the full steps of the following transition (where $C$ is the nor of $A$ and $B$):
${\m{proc}}(\{a\}, \{a'\}, a, a.A(a')), {\m{proc}}(\{b\}, \{b'\}, b, b.B(b')),
{\m{proc}}(S, \{a, b\}, c, c \leftarrow \mi{nor} \leftarrow a,b)$\
$\null \longrightarrow^*
{\m{proc}}(c', \{a', b'\}, c', c' \leftarrow \mi{nor} \leftarrow a',b'),
{\m{proc}}(S, \{c'\}, c, c.C(c'))$ ($c'$ fresh)
We only show the initial portion of each process term, which is enough to disambiguate where in the program we are, as otherwise process terms become unwieldy and reduce clarity. We also assume the existence of a rule $\m{call}$ that lets us invoke a defined process, replacing the call with the process definition, after appropriate substitution.
$$\begin{array}{rc}
{\m{proc}}(\{a\}, \{a'\}, a, a.A(a')), {\m{proc}}(\{b\}, \{b'\}, b, b.B(b')),
{\m{proc}}(S, \{a, b\}, c, c \leftarrow \mi{nor} \leftarrow a,b)
&\overset{\m{call}}{\Longrightarrow} \\[.5em]
{\m{proc}}(\{a\}, \{a'\}, a, a.A(a')), {\m{proc}}(\{b\}, \{b'\}, b, b.B(b')),
{\m{proc}}(S, \{a, b\}, c, \m{case}\; a \ldots)
&\overset{{\oplus}\; C}{\Longrightarrow} \\[.5em]
{\m{proc}}(\{b\}, \{b'\}, b, b.B(b')),
{\m{proc}}(S, \{a', b\}, c, \m{case}\; b \ldots)
&\overset{{\oplus}\; C}{\Longrightarrow} \\[.5em]
{\m{proc}}(S, \{a', b'\}, c, z' \leftarrow \ldots)
&\overset{\cut(\{z'\})}{\Longrightarrow} \\[.5em]
{\m{proc}}(\{c'\}, \{a', b'\}, z', z' \leftarrow \mi{nor} \leftarrow a', b'), {\m{proc}}(S, \{c'\}, c, c.C(c'))
\end{array}$$
Further Logic Examples {#app:further-logic-examples}
======================
\[ex:js4\] The ($\diamondsuit$-free portion of) judgmental modal logic S4 [@Pfenning01mscs] arises from two modes $\mV$ (validity) and $\mU$ (truth) with $\mV > \mU$. The declaration of independence here expresses that *validity is categorical with respect to truth*—that is, a proof of $A_\mV$ may not depend on any hypotheses of the form $B_\mU$. Previously, this had been enforced by segregating the antecedents into two zones and managing their dependence accordingly. $$\begin{array}{lcl}
A_\mV & \Coloneqq & \up_\mU^\mV A_\mU \\
A_\mU, B_\mU & \Coloneqq & p_\mU \mid A_\mU \lolli B_\mU \mid A_\mU \tensor B_\mU \mid \one
\mid \mathop{\oplus}\limits_{i \in I} A_\mU^i
\mid \mathop{\with}\limits_{i \in I} A_\mU^i
\mid \down_\mU^\mV A_\mV
\end{array}$$ Analogous to the encoding of linear logic, we only need to allow $\up^\mV_\mU A_\mU$ in the validity layer. Under that interpretation, we encode $\Box A_\mU = \down^\mV_\mU \,\up^\mV_\mU A_\mU$, which is entirely analogous to the representation of $\bang A$ in linear logic.
The adjoint reconstruction now gives rise to a richer logic where additional connectives speaking about validity can be decomposed directly via their left and right rules.
\[thm:js4\] If we let $\tau$ embed propositions of judgmental S4 into the instance of adjoint logic described above, then
(a) $\Delta; \Gamma \seq A$ in judgmental S4 iff0.1em $\up_\mU^\mV \tau(\Delta), \tau(\Gamma) \seq \tau(A)$ in adjoint logic.
(b) $\Delta; \cdot \seq A$ in judgmental S4 iff0.1em $\up_\mU^\mV \tau(\Delta) \seq \up_\mU^\mV \tau(A)$ in adjoint logic.
\[ex:lax\] Lax logic [@Fairtlough97ic; @Pfenning01mscs] encodes a weaker form of truth called *lax truth*. We can represent it as a substructural adjoint logic with two modes, $\mU > \mX$, where both modes satisfy weakening and contraction. We restrict the lax layer to a single connective and omit additive connectives for simplicity. $$\begin{array}{lcl}
A_\mU, B_\mU & \Coloneqq & p_\mU \mid A_\mU \rightarrow B_\mU \mid A_\mU \times B_\mU \mid 1_\mU \mid \up_\mX^\mU A_\mX \\
A_\mX & \Coloneqq & \down_\mX^\mU A_\mU
\end{array}$$ Now the lax modality is defined as $\bigcirc A_\mU = \up^\mU_\mX\, \down^\mU_\mX A_\mU$.
We can now add further connectives directly operating on the lax layer and obtain consistent left and right rules for them.
If we let $\tau$ embed propositions of lax logic into the instance of adjoint logic described above, then
(a) $\Gamma \seq A\; \mi{true}$ in lax logic iff $\tau(\Gamma) \seq \tau(A)$ in adjoint logic.
(b) $\Gamma, \Gamma' \seq A\; \mi{lax}$ in lax logic iff $\tau(\Gamma), \down_\mX^\mU \tau(\Gamma') \seq \down_\mX^\mU \tau(A)$ in adjoint logic.
[^1]: This allows us, for example, to model the modal logic S4 or lax logic (the logical origins of comonadic and monadic programming), each with two modes both satisfying weakening and contraction, as well as linear analogues of these constructions.
[^2]: The term “multicut” has been used in the literature for different rules. We follow here the proof theory literature [@Negri01book Section 5.1], where it refers to a rule that cuts out some number of copies of the *same* proposition A, as in Gentzen’s original proof of cut elimination [@Gentzen35], where he calls it “Mischung”.
[^3]: TODO: This section needs to be revised to account for the changed goal of the paper. -kmp
[^4]: for now on the linear fragment, and also labeling the succedent with a fresh variable
[^5]: Is this example worth including with how much needs to be stripped out? -kp
| {
"pile_set_name": "ArXiv"
} |
---
address: 'Rutherford Appleton Laboratory, Didcot, Chilton, Oxford, OX11 0QX, UK'
author:
- 'F.F. WILSON'
title: CHARMLESS HADRONIC $B$ DECAYS AT BELLE and
---
Introduction.
=============
The study of the branching fractions and angular distributions of decays to hadronic final states without a charm quark probes the dynamics of both the weak and strong interactions, and plays an important role in understanding Violation (CPV) in the quark sector. Violation at the $B$ factories is described graphically by a triangle with sides formed from the CKM matrix elements $V_{qd}V^{\ast}_{qb}$ ($q=u,c,t$) and internal angles $\alpha,\beta,\gamma$ (or $\phi_2,\phi_1,\phi_3$). Discrepancies in the measured values of the sides and angles could be an indication of New Physics beyond the Standard Model (SM) due to enhanced branching fractions or modified asymmetries. The experimental measurements of branching fractions, asymmetries, polarization and phases (both weak and strong) can be compared to theoretical models based on, for example, QCD factorization, SU(3) symmetry and Lattice QCD.
The results presented below assume charge-conjugate states and all branching fraction upper limits (UL) are at the 90% confidence level (C.L.). The time-integrated asymmetry is defined as $\ACP =
(N_b - N_{\overline{b}}) / (N_b + N_{\overline{b}})$ where $N_b$ ($N_{\overline{b}}$) is the number of containing a $b$($\overline{b}$) quark. The latest results are based on a total dataset of $467\times 10^{6}$ pairs for and $657\times
10^{6}$ pairs for Belle, unless indicated.
Decays involving two-body final states. {#subsec:twobody}
=======================================
The last few years have seen considerable advancement in the prediction of the branching fractions and polarizations of decays to Vector-Vector (), Vector-Scalar () and Vector-Tensor () final states. In general, there has been good agreement between theory and experiment on branching fractions (with some notable exceptions) but the polarization measurements have presented a challenge. The states are expected to be almost fully longitudinally polarized ($\fl \sim 1$) and this should remain true in the presence of penguin loop decays. However, penguin-dominated decays seem to have a smaller (e.g. $\fl \sim 0.5$ for $B\to\phi K^{\ast}$) [@bib:p4a].
Belle has recently measured the decay $\Bm \to K^{\ast0}K^-$ which is dominated by $b\to d s\bar{s}$ gluonic penguin diagrams. They measure a yield of $47.7\pm11.1$ events, corresponding to a branching fraction $\calB (\Bm \to K^{\ast0}K^-) = (0.68\pm0.16\pm0.10) \times 10^{-6}$ with a 4.4$\sigma$ significance [@bib:p4b]. The event yield for $\Bm \to K_2^{\ast0}(1430)\Km$ is measured to be $23.4\pm12.1$ with an upper limit on the branching fraction of $\calB (\Bm \to
K_2^{\ast0}(1430)K^-) < 1.1\times 10^{-6}$. A similar analysis has been done for decays to the final states $\rho^0 K^{\ast0}$ and $f_0 K^{\ast0}$ [@bib:p7]. Unlike an earlier analysis [@bib:p7b], Belle sees no evidence for $\rho^0 K^{\ast0}$ and $f_0 K^{\ast0}$ (and, consequently, do not measure ) but observes $\Bz\to \rho^0\Kp\pim$ and sees first evidence for $\Bz\to
f_0\Kp\pim$ and $\Bz\to \pip\pim K^{\ast0}$, with branching fractions (significance) of $(2.8 \pm 0.5 \pm 0.5)\times 10^{-6}$ ($5.0\sigma$), $(1.4 \pm 0.4^{+0.3}_{-0.4})\times 10^{-6}$ ($3.5\sigma$), and $(4.5^{+1.1+0.9}_{-1.0-1.6})\times 10^{-6}$ ($4.5\sigma$), respectively. has measured decay to an $\omega$ accompanied by a $K^{\ast}$, $\rho$ or $f_0$. Five measurements have a significance above 5$\sigma$, with another two above 3$\sigma$. This has allowed to measure both and . The branching fractions agree with theory predictions and the asymmetries are consistent with zero, as expected, while $\fl \sim 0.5$ except for $\omega\rho^+ \sim 0.9$. The results [@bib:p8] are summarized in Table \[tab:vv\].
Mode Decay S($\sigma$) UL A$_{\CP}$
------ ----------------------------- ------------- -------------------- ----- ---------------------- -----------------------
$\omega K^{\ast0}$ $4.1$ $2.2\pm0.6\pm0.2$ - $0.72\pm0.14\pm0.02$ $+0.45\pm0.25\pm0.02$
$\omega K^{\ast+}$ $2.5$ $2.4\pm1.0\pm0.2$ 7.4 $0.41\pm0.18\pm0.05$ $+0.29\pm0.35\pm0.02$
$\omega (K\pi)^{\ast0}_0$ $9.8$ $18.4\pm1.8\pm1.7$ - - $-0.07\pm0.09\pm0.02$
$\omega (K\pi)^{\ast+}_0$ $9.2$ $27.5\pm3.0\pm2.6$ - - $-0.10\pm0.09\pm0.02$
$\omega K^{\ast}_2(1430)^0$ $5.0$ $10.1\pm2.0\pm1.1$ - $0.45\pm0.12\pm0.02$ $-0.37\pm0.17\pm0.02$
$\omega K^{\ast}_2(1430)^+$ $6.1$ $21.5\pm3.6\pm2.4$ - $0.56\pm0.10\pm0.04$ $+0.14\pm0.15\pm0.02$
$\omega \rho^0$ $1.9$ $0.8\pm0.5\pm0.2$ 1.6 - -
$\omega f_0$ $4.5$ $1.0\pm0.3\pm0.1$ 1.5 - -
$\omega \rho^+$ $9.8$ $15.9\pm1.6\pm1.4$ - $0.90\pm0.05\pm0.03$ $-0.20\pm0.09\pm0.02$
: Branching fraction central value () and upper limit (UL) in units of $10^{-6}$, significance S in standard deviations, longitudinal polarization () and asymmetry A$_{\CP}$ for the Vector-Vector (), Vector-Scalar () and Vector-Tensor () decays of $B\to\omega K^{\ast}$, $\omega
f_0$ and $\omega\rho$. \[tab:vv\]
Decays involving three-body final states. {#subsec:threebody}
=========================================
An interesting use of the decay to final states with three particles is the search by Belle for the exotic state X(1812) in the decay $\Bp\to \Kp\ X(1812), X(1812)\to \omega\phi$. This is similar to the observation by Belle of the Y(3940) resonance in $\Bp\to\Kp\omega\psi$ [@bib:p5a]. Belle observe $N_{\Kp\omega\phi}
= 22.1^{+8.3}_{-7.2}$ events leading to a branching fraction for the Dalitz plot of $\calB(\Bp\to\Kp\omega\phi) = (
1.15^{+0.43+0.14}_{-0.38-0.13})\times 10^{-6}$ (2.8$\sigma$) and an upper limit $< 1.9\times 10^{-6}$. Assuming the X(1812) masses and width from BES [@bib:bes], Belle searches for a near-threshold enhancement in the $M_{\pip\pim\piz\Kp\Km}$ mass spectrum. No significant yield is seen and an upper limit of $3.2\times 10^{-7}$ is placed on the product branching fraction $\calB(\Bp\to\Kp X(1812),
X(1812)\to \omega\phi)$ [@bib:p5b].
has also looked at rare processes in Dalitz plots. Previous measurements have shown that almost 50% of the events in $\Bz\to\Kp\Km\pip$ can be assigned to an ill-defined resonance, called $f_X(1500)$ by . If this is an even-spin resonance, the rate for $f_X(1500)\to \KS\KS$ would be expected to be half the rate for $f_X(1500)\to \Kp\Km$. They see $15\pm15$ events in the whole Dalitz plot placing an upper limit on the total branching fraction of $\calB(\Bp\to \KS\KS\pip) < 5.1\times 10^{-7}$. This makes the even-spin hypothesis unlikely but interpretation is difficult as the exact quantum numbers of the $f_X(1500)$ are unknown [@bib:p6a].
Some MSSM models could enhance the branching fractions of SM-suppressed decays from the SM values of $\sim10^{-16}$ to $\sim10^{-6}$. has searched for $\Bm\to \Kp\pim\pim$ and $\Bm\to \Km\Km\pip$ and placed upper limits of $9.5\times10^{-7}$ and $1.6\times10^{-7}$, respectively, on the branching fractions [@bib:p6b].
The decay $\Bp\to\pip\pip\pim$ can in principle be used to extract the CKM angle $\gamma$ by measuring the interference between $\pip\pim$ resonances and the $\chi_{c0}$ resonance which has no violating phase. It can also be helpful in understanding broad resonances and nonresonant backgrounds that are present in $\Bz\to\pip\pim\piz$ and so improve our measurement of the CKM angle $\alpha$. ’s results [@bib:p9] for $\Bp\to\pip\pip\pim$ are summarized in Table \[tab:pipipi\]. No significant direct asymmetry is measured and, although some resonances are significant, no evidence is found for $\chi_{c0}$ and $\chi_{c2}$, leading to branching fraction upper limits for $\Bp\to\chi_{c0}\pip<1.5\times 10^{-5}$ and $\Bp\to\chi_{c2}\pip<2.0\times 10^{-5}$, making the measurement of $\gamma$ in this mode unlikely at Belle or .
Decay Fit Fraction (%) ($\times 10^{-6}$) A$_{CP}$ (%)
--------------------------------------- ---------------------------------- --------------------------------------- ---------------------------------
Total - $15.2\pm0.6\pm1.2^{+0.4}_{-0.3}$ $3.2\pm4.4\pm3.1^{+2.5}_{-2.0}$
nonresonant $34.9\pm4.2\pm2.9^{+7.5}_{-3.4}$ $5.3\pm0.7\pm0.6^{+1.1}_{-0.5}$ $-14\pm14\pm7^{+17}_{-3}$
$\rho^0(770)\pipm;\rho^0\to\pip\pim$ $53.2\pm3.7\pm2.5^{+1.5}_{-7.4}$ $8.1\pm0.7\pm1.2^{+0.4}_{-1.1}$ $18\pm7\pm5^{+2}_{-14}$
$\rho^0(1450)\pipm;\rho^0\to\pip\pim$ $9.1\pm2.3\pm2.4^{+1.9}_{-4.5}$ $1.4\pm0.4\pm0.4^{+0.3}_{-0.7}$ $-6\pm28\pm20^{+12}_{-35}$
$f_2(1270)\pipm;f_2\to\pip\pim$ $5.9\pm1.6\pm0.4^{+2.0}_{-0.7}$ $0.9\pm0.2\pm0.1^{+0.3}_{-0.1}$ $41\pm25\pm13^{+12}_{-8}$
$f_0(1370)\pipm;f_0\to\pip\pim$ $18.0\pm3.3\pm2.6^{+4.3}_{-3.5}$ $2.9\pm0.5\pm0.5^{+0.7}_{-0.5}(<4.0)$ $72\pm15\pm14^{+7}_{-8}$
$f_0(980)\pipm;f_0\to\pip\pim$ - $<1.5$ -
$\chi_{c0}\pipm;\chi_{c0}\to\pip\pim$ - $<0.1$ -
$\chi_{c2}\pipm;\chi_{c2}\to\pip\pim$ - $<0.1$ -
: Branching fraction (), asymmetry A$_{CP}$, and Fit Fraction for the decay $\Bp\to\pip\pip\pim$ with the resonance decaying to $\pip\pim$\[tab:pipipi\]. The errors are statistical, systematic and model-dependent, respectively.
Violation and the CKM angle $\alpha(\phi_2)$. {#subsec:cpv}
==============================================
The precision of the measurement of the CKM angle $\alpha(\phi_2)$ continues to improve. In the absence of penguin loops in the decays, the angle $\alpha$ can be measured in the time-dependent decay of $\Bz\to\rho\rho$ and $\Bz\to\pi\pi$. However the penguin contribution, particularly in $\piz\piz$, is not small and so the measured $\alpha_{eff}$ differs from the true $\alpha$ by $\Delta\alpha=\alpha-\alpha_{eff}$. $\Delta\alpha$ can be constrained by performing an Isospin analysis on the decays $\Bz\to\rho^0\rho^0$, $\Bpm\to\rho^{\pm}\rho^0$ and $\Bz\to\rho^+\rho^-$. Table \[tab:rhorho\] summarizes the measurements from [@bib:p10], where the parameters are quoted for the longitudinally polarized (-even) component of the decays. When combined, $\Delta\alpha$ is constrained to be between $-1.8^{o}$ and $6.7^{o}$ (68% C.L.). The angle $\alpha$ is measured to be $(92.4^{+6.0}_{-6.5})^{o}$ and can be compared to the recent result from Belle [@bib:p10a] of $\alpha =
(91.7\pm14.9)^{o}$. A similar analysis using $B\to\pi\pi$ decays produces a looser constraint $\mid\Delta\alpha\mid < 43^{o}$, which results in an exclusion range for $\alpha$ between $23^{o}$ and $43^{o}$ at the 90% C.L. The result of combining these measurements using the CKMfitter programme [@bib:ckmfitter] with earlier measurements of $B\to\pi\rho$ are shown in Fig. \[fig:ckmfitter\].
$\Bz\to\rho^+\rho^-$ $\Bz\to\rho^0\rho^0$ $\Bp\to\rho^+\rho^0$
-------------------------- ----------------------------------- ---------------------- -------------------------
$\calB (\times 10^{-6})$ $25.5\pm2.1^{+3.6}_{-3.9}$ $0.92\pm0.32\pm0.14$ $23.7\pm1.4\pm1.4$
$0.992\pm0.024^{+0.026}_{-0.013}$ $0.75\pm0.14\pm0.04$ $0.950\pm0.015\pm0.006$
C$_L$ $0.01\pm0.15\pm0.06$ $0.2\pm0.8\pm0.3$ -
S$_L$ $-0.17\pm0.20^{+0.05}_{-0.06}$ $0.3\pm0.7\pm0.2$ -
A$_{CP}$ - - $-0.054\pm0.055\pm0.01$
: Branching fraction (), longitudinal polarization (), direct asymmetry (C$_L$), asymmetry in the interference between mixing and decay (S$_L$) and asymmetry A$_{CP}$ for the decays $\Bz\to\rho^+\rho^-$, $\Bz\to\rho^0\rho^0$ and $\Bp\to\rho^+\rho^0$ measured by .\[tab:rhorho\]
[c]{}
Belle has seen direct in $\Bz\to\pip\pim$ but does not, reporting only that $C_{\pip\pim} = -0.25\pm0.08\pm0.02$ with a significance of just 2.2$\sigma$. However, both experiments see significant direct in $\Bz\to\Kp\pim$ with reporting $\ACP = -0.107\pm0.016^{+0.006}_{-0.004}$ with 6.1$\sigma$ significance, to be compared to $-0.094\pm0.018\pm0.008$ from Belle. Both experiments also measure for $\Bpm\to\Kpm\piz$ to be slightly positive but consistent with zero. should be similar for both $K\pi$ modes but Belle reports a 4.4$\sigma$ difference and sees a similar discrepancy [@bib:p11].
References {#references .unnumbered}
==========
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper presents a geometric description on Lie algebroids of Lagrangian systems subject to nonholonomic constraints. The Lie algebroid framework provides a natural generalization of classical tangent bundle geometry. We define the notion of nonholonomically constrained system, and characterize regularity conditions that guarantee that the dynamics of the system can be obtained as a suitable projection of the unconstrained dynamics. The proposed novel formalism provides new insights into the geometry of nonholonomic systems, and allows us to treat in a unified way a variety of situations, including systems with symmetry, morphisms, reduction, and nonlinearly constrained systems. Various examples illustrate the results.'
address:
- 'Jorge Cortés: Department of Mechanical and Aerospace Engineering, University of California at San Diego, La Jolla, California 92093, USA'
- 'Manuel de León: Instituto de Matemáticas y F[í]{}sica Fundamental, Consejo Superior de Investigaciones Científicas, Serrano 123, 28006 Madrid, Spain'
- 'Juan C. Marrero: Departamento de Matemática Fundamental y Unidad Asociada ULL-CSIC Geometría Diferencial y Mecánica Geométrica, Facultad de Matemáticas, Universidad de la Laguna, La Laguna, Tenerife, Canary Islands, Spain '
- 'Eduardo Martínez: Departamento de Matemática Aplicada, Facultad de Ciencias, Universidad de Zaragoza, 50009 Zaragoza, Spain'
author:
- Jorge Cortés
- Manuel de León
- 'Juan C. Marrero'
- Eduardo Martínez
title:
- Nonholonomic Lagrangian systems on Lie algebroids
- Foundations of Mechanics
- Nonholonomic reduction
- Nonholonomic Mechanics and Control
- Nonholonomic mechanical systems with symmetry
- 'Lie algebroid morphisms, Poisson sigma models, and off-shell closed gauge symmetries'
- Geometric Models for Noncommutative Algebras
- 'On the geometry of generalized [C]{}haplygin systems'
- Reduction of nonholonomic mechanical systems with symmetries
- Reduction of constrained systems with symmetries
- 'On almost-Poisson structures in nonholonomic mechanics'
- 'On almost-Poisson structures in nonholonomic mechanics II. The time-dependent framework'
- 'Lagrangian systems with constraints: a geometrical approach to the method of Lagrange multipliers'
- Reduction of Lie algebroid structures
- Reduction of Lagrangian mechanics on Lie algebroids
- Lagrangian Reduction by Stages
- 'On the theory of the motion of nonholonomic systems. Theorem on the reducing multiplier'
- 'Geometric, Control and Numerical Aspects of Nonholonomic Systems'
- Reduction and reconstruction of the dynamics of nonholonomic systems
- A survey of Lagrangian mechanics and control on Lie algebroids and groupoids
- Mechanical control systems on Lie algebroids
- Nonholonomic integrators
- Tangent bundle geometry for Lagrangian dynamics
- 'Geometry of non-holonomic constraints'
- Nonholonomic reduction for free and proper actions
- 'Nonholonomic systems via moving frames: Cartan equivalence and Chaplygin Hamiltonization'
- Nonholonomic LR systems as generalized Chaplygin systems with an invariant measure and flows on homogeneous spaces
- 'Various aspects of $n$-dimensional rigid body dynamics'
- Discrete nonholonomic LL systems on Lie groups
- 'Lie algebroids, holonomy and characteristic classes'
- 'On the geometry of Lagrangian mechanics with non-holonomic constraints'
- 'Theoretische Mechanik: Eine einheitliche Einfhrung in die gesamte Mechanik (1949)'
- Algebraic constructions in the category of Lie algebroids
- Dirac brackets in constrained dynamics
- 'Nonholonomic geodesic flows on Lie groups and the integrable Suslov problem on $SO(4)$'
- Nonholonomic left and right flows on Lie groups
- Espaces variationnels et mécanique
- Reduction of some classical nonholonomic systems with symmetry
- The Hamiltonian and Lagrangian approaches to the dynamics of nonholonomic systems
- Poisson reduction of nonholonomic mechanical systems with symmetry
- 'On the geometry of non-holonomic Lagrangian systems'
- Mechanical systems with nonlinear constraints
- Geometric integrators and nonholonomic mechanics
- Lagrangian submanifolds and dynamics on Lie algebroids
- Methods of Differential Geometry in Analytical Mechanics
- Affine connections and distributions with applications to nonholonomic mechanics
- Lie algebroids and mechanics
- Reduction of constrained mechanical systems and stability of relative equlibria
- Lie Groupoids and Lie Algebroids in Differential Geometry
- Integrability of the problem of rolling of a sphere with a multiply connected cavity filled with an ideal fluid
- 'Introduction to Mechanics and Symmetry. A Basic Exposition of Classical Mechanical Systems'
- Lagrangian mechanics on Lie algebroids
- Lie algebroid structures and Lagrangian systems on affine bundles
- Reduction in optimal control theory
- 'Classical field theory on Lie algebroids: multisymplectic formalism'
- Lagrangian reduction by stages for nonholonomic systems in a Lie algebroid framework
- A Lie algebroid framework for nonholonomic systems
- Dynamics of Nonholonomic Systems
- Vector form brackets in Lie algebroids
- Geometric objects defined by almost Lie structures
- Prolongations of Lie groupoids and Lie algebroids
- A geometrical framework for the study of nonholonomic Lagrangian systems
- Nonholonomic Noether theorem and reduction of symmetries
- Theoretical Mechanics
- Model reduction method for nonholonomic mechanical systems with semidirect product symmetry
- 'On the Hamiltonian formulation of non-holonomic mechanical systems'
- Integrable nonholonomic systems on Lie groups
- The structure of Hamiltonian mechanics
- Lagrangian Mechanics and groupoids
- 'Dynamics of the $n$-dimensional Suslov problem'
- Invariant measures of nonholonomic flows with internal degrees of freedom
---
[^1]
Introduction
============
The category of Lie algebroids has proved useful to formulate problems in applied mathematics, algebraic topology, and differential geometry. In the context of Mechanics, an ambitious program was proposed in [@Weinstein] in order to develop formulations of the dynamical behavior of Lagrangian and Hamiltonian systems on Lie algebroids and discrete mechanics on Lie groupoids. In the last years, this program has been actively developed by many authors, and as a result, a powerful mathematical structure is emerging.
The main feature of the Lie algebroid framework is its inclusive nature. Under the same umbrella, one can consider such disparate situations as systems with symmetry, systems evolving on semidirect products, Lagrangian and Hamiltonian systems on Lie algebras, and field theory equations (see [@CoLeMaMaMa; @LeMaMa] for recent topical reviews illustrating this). The Lie algebroid approach to Mechanics builds on the particular structure of the tangent bundle to develop a geometric treatment of Lagrangian systems parallel to Klein’s formalism [@Cr; @Klein]. At the same time, the attention devoted to Lie algebroids from a purely geometrical viewpoint has led to an spectacular development of the field, e.g., see [@BoKoSt; @CaNuSaI; @Mackenzie; @Sau] and references therein. The merging of both perspectives has already provided mutual benefit, and will undoubtedly lead to important developments in the future.
The other main theme of this paper are nonholonomic Lagrangian systems, i.e., systems subject to constraints involving the velocities. This topic is a classic subject in Mathematics and Mechanics, dating back to the early times of Lagrange; a comprehensive list of classical references can be found in [@NF]. At the beginning of the nineties, the work [@Ko] sparked a renewed interest in the geometric study of nonholonomic mechanical systems, with a special emphasis on symmetry aspects. In the last years, several authors have extended the ideas and techniques of the geometrical treatment of unconstrained systems to the study of nonholonomic mechanical systems, see the recent monographs [@Bl; @cortes]. These include symplectic [@CaRa; @LeMa; @LeMa2], Hamiltonian [@VaMa], and Lagrangian approaches [@CuKeSnBa; @KoMa1], the study of almost Poisson brackets [@CaLeMa; @IbLeMaMa; @KoMa2], and symmetry and reduction of the dynamics [@BaSn; @BlKrMaMu; @CaCoLeMa; @CaLeMaMa; @CaLeMaMa2; @CoLe; @marle].
In this paper we develop a comprehensive treatment of nonholonomic systems on Lie algebroids. This class of systems was introduced in [@CoMa] when studying mechanical control systems (see also [@MeLa] for a recent approach to mechanical systems on Lie algebroids subject to linear constraints). Here, we build on the geometry of Lie algebroids to identify suitable regularity conditions guaranteeing that the nonholonomic system admits a unique solution. We develop a projection procedures to obtain the constrained dynamics as a modification of the unconstrained one, and define an almost-Poisson nonholonomic bracket. We show that many of the properties that standard nonholonomic systems enjoy have their counterpart in the proposed setup. As important examples, we highlight that the analysis here provides a natural interpretation for the use of pseudo-coordinates techniques and lends itself to the treatment of constrained systems with symmetry, following the ideas developed in [@CoMa; @MaROMP]. We carefully examine the reduction procedure for this class of systems, paying special attention to the evolution of the momentum map.
From a methodological point of view, the approach taken in the paper has enormous advantages. This fact must mainly be attributed to the inclusive nature of Lie algebroids. Usually, the results on nonholonomic systems available in the literature are restricted to a particular class of nonholonomic systems, or to a specific context. However, as illustrated in Table \[tab:examples\], many different nonholonomic systems fit under the Lie algebroid framework, and this has the important consequence of making the results proved here widely applicable. With the aim of illustrating this breadth, we consider various examples throughout the paper, including the Suslov problem, the Chaplygin sleigh, the Veselova system, Chaplygin Gyro-type systems, the two-wheeled planar mobile robot, and a ball rolling on a rotating table. We envision that future developments within the proposed framework will have a broad impact in nonholonomic mechanics. In the course of the preparation of this manuscript, the recent research efforts [@CaNuSaII; @Me] were brought to our attention. These references, similar in spirit to the present work, deal with nonholonomic Lagrangian systems and focus on the reduction of Lie algebroid structures under symmetry.
Lie algebroid Dynamics Example
-------------------- ------------------ ----------------------- --------------------
Standard Tangent bundle Lagrande-d’Alembert Rolling disk [@NF]
On a Lie algebra Lie algebra Euler-Poincaré-Suslov
Symmetry-invariant Atiyah algebroid
: The Lie algebroid framework embraces different classes of nonholonomic systems.[]{data-label="tab:examples"}
The paper is organized as follows. In Section \[preliminaries\] we collect some preliminary notions and geometric objects on Lie algebroids, including differential calculus, morphisms and prolongations. We also describe classical Lagrangian systems within the formalism of Lie algebroids. In Section \[linear\], we introduce the class of nonholonomic Lagrangian systems subject to linear constraints, given by a regular Lagrangian $L : E
\longrightarrow \R$ on the Lie algebroid $\tau : E \longrightarrow
M$ and a constraint subbundle $D$ of $E$. We show that the known results in Mechanics for these systems also hold in the context of Lie algebroids. In particular, drawing analogies with d’Alembert principle, we derive the Lagrange-d’Alembert equations of motion, prove the conservation of energy and state a Noether’s theorem. We also derive local expressions for the dynamics of nonholonomic Lagrangian systems, which are further simplified by the choice of a convenient basis of $D$. As an illustration, we consider the class of nonholonomic mechanical systems. For such systems, the Lagrangian $L$ is the polar form of a bundle metric on $E$ minus a potential function on $M$. In Section \[sec:regularity\], we perform the analysis of the existence and uniqueness of solutions of constrained systems on general Lie algebroids, and extend the results in [@BaSn; @CaLeMaMa; @CaLeMa; @CoLe; @LeMa] for constrained systems evolving on tangent bundles. We obtain several characterizations for the regularity of a nonholonomic system, and prove that a nonholonomic system of mechanical type is always regular. The constrained dynamics can be obtained by projecting the unconstrained dynamics in two different ways. Under the first projection, we develop a distributional approach analogous to that in [@BaSn], see also [@MeLa]. Using the second projection, we introduce the nonholonomic bracket. The evolution of any observable can be measured by computing its bracket with the energy of the system. Section \[sec:reduction\] is devoted to studying the reduction of the dynamics under symmetry. Our approach follows the ideas developed in [@CeMaRa], who defined a minimal subcategory of the category of Lie algebroids which is stable under Lagrangian reduction. We study the behavior of the different geometric objects introduced under morphisms of Lie algebroids, and show that fiberwise surjective morphisms induce consistent reductions of the dynamics. This result covers, but does not reduce to, the usual case of reduction of the dynamics by a symmetry group. In accordance with the philosophy of the paper, we study first the unconstrained dynamics case, and obtain later the results for the constrained dynamics using projections. A (Poisson) reduction by stages procedure can also be developed within this formalism. It should be noticed that the reduction under the presence of a Lie group of symmetries $G$ is performed in two steps: first we reduce by a normal subgroup $N$ of $G$, and then by the residual group. In Section \[momentum-equation\], we prove a general version of the momentum equation introduced in [@BlKrMaMu]. In Section \[examples\], we show some interesting examples and in Section \[nonlinear\], we extend some of the results previously obtained for linear constraints to the case of nonlinear constraints. The paper ends with our conclusions and a description of future research directions.
Preliminaries
=============
In this section we recall some well-known facts concerning the geometry of Lie algebroids. We refer the reader to [@CaWe; @HiMa; @Mackenzie] for details about Lie groupoids, Lie algebroids and their role in differential geometry.
Lie algebroids
--------------
Let $M$ be an $n$-dimensional manifold and let ${\tau\colonE\toM}$ be a vector bundle. A vector bundle map ${\rho\colonE\toTM}$ over the identity is called an *anchor map*. The vector bundle $E$ together with an anchor map $\rho$ is said to be an *anchored vector bundle* (see [@PoPo]). A structure of *Lie algebroid* on $E$ is given by a Lie algebra structure on the ${C^\infty(M)}$-module of sections of the bundle, $({\operatorname{Sec}_{}(E)},[\cdot\,,\cdot])$, together with an anchor map, satisfying the compatibility condition $$[\sigma,f\eta] = f[\sigma,\eta] + \bigl( \rho(\sigma)f \bigr) \eta
.$$ Here $f$ is a smooth function on $M$, $\sigma$, $\eta$ are sections of $E$ and $\rho(\sigma)$ denotes the vector field on $M$ given by $\rho(\sigma)(m)=\rho(\sigma(m))$. From the compatibility condition and the Jacobi identity, it follows that the map $\sigma\mapsto\rho(\sigma)$ is a Lie algebra homomorphism from the set of sections of $E$, ${\operatorname{Sec}_{}(E)}$, to the set of vector fields on $M$, $\mathfrak{X}(M)$.
In what concerns Mechanics, it is convenient to think of a Lie algebroid ${\rho\colonE\toTM}$, and more generally an anchored vector bundle, as a substitute of the tangent bundle of $M$. In this way, one regards an element $a$ of $E$ as a generalized velocity, and the actual velocity $v$ is obtained when applying the anchor to $a$, i.e., $v=\rho(a)$. A curve ${a\colon[t_0,t_1]\toE}$ is said to be *admissible* if $\dot{m}(t)=\rho(a(t))$, where $m(t)=\tau(a(t))$ is the base curve. We will denote by ${\operatorname{Adm}(E)}$ the space of admissible curves on $E$.
Given local coordinates $(x^i)$ in the base manifold $M$ and a local basis $\{e_\alpha\}$ of sections of $E$, we have local coordinates $(x^i,y^\alpha)$ in $E$. If $a\in E$ is an element in the fiber over $m\in M$, then we can write $a=y^\alpha
e_\alpha(m)$ and thus the coordinates of $a$ are $(m^i,y^\alpha)$, where $m^i$ are the coordinates of the point $m$. The anchor map is locally determined by the local functions $\rho^i_\alpha$ on $M$ defined by $\rho(e_\alpha)=\rho^i_\alpha(\partial/\partial
x^i)$. In addition, for a Lie algebroid, the Lie bracket is determined by the functions $C^\gamma_{\alpha\beta}$ defined by $[e_\alpha,e_\beta]=C^\gamma_{\alpha\beta}e_\gamma$. The functions $\rho^i_\alpha$ and $C^\gamma_{\alpha\beta}$ are called *the structure functions* of the Lie algebroid in this coordinate system. They satisfy the following relations $$\begin{aligned}
\rho^j_\alpha{\frac{\partial \rho^i_\beta}{\partial x^j}} -
\rho^j_\beta{\frac{\partial \rho^i_\alpha}{\partial x^j}} = \rho^i_\gamma
C^\gamma_{\alpha\beta} \quand
\sum_{\mathrm{cyclic}(\alpha,\beta,\gamma)} \left[\rho^i_\alpha{\frac{\partial
C^\nu_{\beta\gamma}}{\partial x^i}} + C^\mu_{\beta\gamma}
C^\nu_{\alpha\mu}\right]=0,\end{aligned}$$ which are called *the structure equations* of the Lie algebroid.
Exterior differential
---------------------
The anchor $\rho$ allows to define the differential of a function on the base manifold with respect to an element $a\in E$. It is given by $$df(a)=\rho(a)f.$$ It follows that the differential of $f$ at the point $m\in M$ is an element of $E_m^*$. Moreover, a structure of Lie algebroid on $E$ allows to extend the differential to sections of the bundle ${\bigwedge\nolimits^{p}{E}}$, which will be called $p$-sections or just $p$-forms. If $\omega\in{\operatorname{Sec}_{}({\bigwedge\nolimits^{p}{E}})}$, then $d\omega\in{\operatorname{Sec}_{}({\bigwedge\nolimits^{p+1}{E}})}$ is defined by $$\begin{aligned}
d\omega(\sigma_0,\sigma_1,\ldots,\sigma_p) &=
\sum_i(-1)^i\rho(\sigma_i)(
\omega(\sigma_0,\ldots,\widehat{\sigma_i},\ldots,\sigma_p))\\
&\qquad{}+ \sum_{i<j}(-1)^{i+j}
\omega([\sigma_i,\sigma_j],\sigma_0,\ldots,
\widehat{\sigma_i},\ldots,\widehat{\sigma_j},\ldots,\sigma_p).\end{aligned}$$ It follows that $d$ is a cohomology operator, that is, $d^2=0$. Locally the exterior differential is determined by $$dx^i=\rho^i_\alpha e^\alpha \qquand
de^\gamma=-\frac{1}{2}C^\gamma_{\alpha\beta}e^\alpha\wedge e^\beta.$$ Throughout this paper, the symbol $d$ will refer to the exterior differential on the Lie algebroid $E$ and not to the ordinary exterior differential on a manifold. Of course, if $E=TM$, then both exterior differentials coincide.
The usual Cartan calculus extends to the case of Lie algebroids (see [@Mackenzie; @Nijenhuis]). For every section $\sigma$ of $E$ we have a derivation $i_\sigma$ (contraction) of degree $-1$ and a derivation $d_\sigma=i_\sigma\circ d+d\circ i_\sigma$ (Lie derivative) of degree $0$. Since $d^2=0$, we have that $d_\sigma\circ d=d\circ
d_\sigma$.
Morphisms
---------
Let ${\tau\colonE\toM}$ and ${\tau'\colonE'\toM'}$ be two anchored vector bundles, with anchor maps ${\rho\colonE\toTM}$ and ${\rho'\colonE'\toTM'}$. A vector bundle map ${\Phi\colonE\toE'}$ over a map ${\varphi\colonM\toM'}$ is said to be *admissible* if it maps admissible curves onto admissible curves, or equivalently $T\varphi\circ\rho = \rho'\circ\Phi$. If $E$ and $E'$ are Lie algebroids, then we say that $\Phi$ is a *morphism* if $\Phi{^\star}d\theta=d\Phi{^\star}\theta$ for every $\theta\in{\operatorname{Sec}_{}({\bigwedge\nolimits^{}{E'}})}$. It is easy to see that morphisms are admissible maps.
In the above expression, the pullback $\Phi{^\star}\beta$ of a $p$-form $\beta$ is defined by $$(\Phi{^\star}\beta)_m(a_1,a_2,\ldots,a_p)=
\beta_{\varphi(m)}\bigl(\Phi(a_1),\Phi(a_2),\ldots,\Phi(a_p)\bigr),$$ for every $a_1,\ldots,a_p\in E_m$. For a function $f\in{C^\infty(M')}$ (i.e., for $p=0$), we just set $\Phi{^\star}f=f\circ\varphi$.
Let $(x^i)$ and $(x'{}^i)$ be local coordinate systems on $M$ and $M'$, respectively. Let $\{e_\alpha\}$ and $\{e'_\alpha\}$ be local basis of sections of $E$ and $E'$, respectively, and $\{e^\alpha\}$ and $\{e'{}^\alpha\}$ the corresponding dual basis. The bundle map $\Phi$ is determined by the relations $\Phi{^\star}x'{}^i = \phi^i(x)$ and $\Phi{^\star}e'{}^\alpha = \phi^\alpha_\beta e^\beta$ for certain local functions $\phi^i$ and $\phi^\alpha_\beta$ on $M$. Then, $\Phi$ is admissible if and only if $$\rho^j_\alpha{\frac{\partial \phi^i}{\partial x^j}}=\rho'{}^i_\beta\phi^\beta_\alpha.$$ The map $\Phi$ is a morphism of Lie algebroids if and only if, in addition to the admissibility condition above, one has $$\phi^\beta_\gamma C^\gamma_{\alpha\delta} =
\left(\rho^i_\alpha{\frac{\partial \phi^\beta_\delta}{\partial x^i}} -
\rho^i_\delta{\frac{\partial \phi^\beta_\alpha}{\partial x^i}}\right) +
C'{}^\beta_{\theta\sigma}\phi^\theta_\alpha\phi^\sigma_\delta.$$ In these expressions, $\rho^i_\alpha$, $C^\alpha_{\beta\gamma}$are the local structure functions on $E$ and $\rho'{}^i_\alpha$, $C'{}^\alpha_{\beta\gamma}$ are the local structure functions on $E'$.
Prolongation of a fibered manifold with respect to a Lie algebroid
------------------------------------------------------------------
Let ${\pi\colonP\toM}$ be a fibered manifold with base manifold $M$. Thinking of $E$ as a substitute of the tangent bundle of $M$, the tangent bundle of $P$ is not the appropriate space to describe dynamical systems on $P$. This is clear if we note that the projection to $M$ of a vector tangent to $P$ is a vector tangent to $M$, and what one would like instead is an element of $E$, the ‘new’ tangent bundle of $M$.
A space which takes into account this restriction is the *$E$-tangent bundle* of $P$, also called the *prolongation* of $P$ with respect to $E$, which we denote by ${\@ifstar{\@proldf}{\@prolpf}}[E]{P}$ (see [@LeMaMa; @LMLA; @MaMeSa; @PoPo]). It is defined as the vector bundle ${\tau^E_P\colon{\@ifstar{\@proldf}{\@prolpf}}[E]{P}\toP}$ whose fiber at a point $p\in
P_m$ is the vector space $${\@ifstar{\@proldf}{\@prolpf}}[E]{P}[p] ={\{\,(b,v)\in E_m\times T_pP\,|\,\rho(b)=T_p\pi(v)\,\}}.$$ We will frequently use the redundant notation $(p,b,v)$ to denote the element $(b,v)\in{\@ifstar{\@proldf}{\@prolpf}}[E]{P}[p]$. In this way, the map $\tau^E_P$ is just the projection onto the first factor. The anchor of ${\@ifstar{\@proldf}{\@prolpf}}[E]{P}$ is the projection onto the third factor, that is, the map ${\rho^1\colon{\@ifstar{\@proldf}{\@prolpf}}[E]{P}\toTP}$ given by $\rho^1(p,b,v)=v$. The projection onto the second factor will be denoted by ${{\@ifstar{\@proldf}{\@prolpf}}{\pi}\colon{\@ifstar{\@proldf}{\@prolpf}}[E]{P}\toE}$, and it is a vector bundle map over $\pi$. Explicitly ${\@ifstar{\@proldf}{\@prolpf}}{\pi}(p,b,v)=b$.
An element $z\in{\@ifstar{\@proldf}{\@prolpf}}[E]{P}$ is said to be vertical if it projects to zero, that is ${\@ifstar{\@proldf}{\@prolpf}}{\pi}(z)=0$. Therefore it is of the form $(p,0,v)$, with $v$ a vertical vector tangent to $P$ at $p$.
Given local coordinates $(x^i,u^A)$ on $P$ and a local basis $\{e_\alpha\}$ of sections of $E$, we can define a local basis $\{{\mathcal{X}}_\alpha,{\mathcal{V}}_A\}$ of sections of ${\@ifstar{\@proldf}{\@prolpf}}[E]{P}$ by $${\mathcal{X}}_\alpha(p)
=\Bigl(p,e_\alpha(\pi(p)),\rho^i_\alpha{\frac{\partial }{\partial x^i}}{\Big|_{p}}\Bigr) \qquand
{\mathcal{V}}_A(p) = \Bigl(p,0,{\frac{\partial }{\partial u^A}}{\Big|_{p}}\Bigr).$$ If $z=(p,b,v)$ is an element of ${\@ifstar{\@proldf}{\@prolpf}}[E]{P}$, with $b=z^\alpha
e_\alpha$, then $v$ is of the form $v=\rho^i_\alpha
z^\alpha{\frac{\partial }{\partial x^i}}+v^A{\frac{\partial }{\partial u^A}}$, and we can write $$z=z^\alpha{\mathcal{X}}_\alpha(p)+v^A{\mathcal{V}}_A(p).$$ Vertical elements are linear combinations of $\{{\mathcal{V}}_A\}$.
The anchor map $\rho^1$ applied to a section $Z$ of ${\@ifstar{\@proldf}{\@prolpf}}[E]{P}$ with local expression $Z = Z^\alpha{\mathcal{X}}_\alpha+V^A{\mathcal{V}}_A$ is the vector field on $P$ whose coordinate expression is $$\rho^1(Z) = \rho^i_\alpha Z^\alpha {\frac{\partial }{\partial x^i}} + V^A{\frac{\partial }{\partial u^A}}.$$
If $E$ carries a Lie algebroid structure, then so does ${\@ifstar{\@proldf}{\@prolpf}}[E]{P}$. The associated Lie bracket can be easily defined in terms of projectable sections, so that ${\@ifstar{\@proldf}{\@prolpf}}{\pi}$ is a morphism of Lie algebroids. A section $Z$ of ${\@ifstar{\@proldf}{\@prolpf}}[E]{P}$ is said to be projectable if there exists a section $\sigma$ of $E$ such that ${\@ifstar{\@proldf}{\@prolpf}}{\pi}\circ
Z=\sigma\circ\pi$. Equivalently, a section $Z$ is projectable if and only if it is of the form $Z(p)=(p,\sigma(\pi (p)),X(p))$, for some section $\sigma$ of $E$ and some vector field $X$ on $E$ (which projects to $\rho(\sigma)$). The Lie bracket of two projectable sections $Z_1$ and $Z_2$ is then given by $$[Z_1,Z_2](p)=(p,[\sigma_1,\sigma_2](m),[X_1,X_2](p)), \qquad p \in
P,\,\;\;\; m=\pi(p).$$ It is easy to see that $[Z_1,Z_2](p)$ is an element of ${\@ifstar{\@proldf}{\@prolpf}}[E]{P}[p]$ for every $p\in P$. Since any section of ${\@ifstar{\@proldf}{\@prolpf}}[E]{P}$ can be locally written as a linear combination of projectable sections, the definition of the Lie bracket for arbitrary sections of ${\@ifstar{\@proldf}{\@prolpf}}[E]{P}$ follows.
The Lie brackets of the elements of the basis are $$[{\mathcal{X}}_\alpha,{\mathcal{X}}_\beta]= C^\gamma_{\alpha\beta}\:{\mathcal{X}}_\gamma, \qquad
[{\mathcal{X}}_\alpha,{\mathcal{V}}_B]=0 \qquand [{\mathcal{V}}_A,{\mathcal{V}}_B]=0,$$ and the exterior differential is determined by $$\begin{aligned}
&dx^i=\rho^i_\alpha {\mathcal{X}}^\alpha,
&&du^A={\mathcal{V}}^A,\\
&d{\mathcal{X}}^\gamma=-\frac{1}{2}C^\gamma_{\alpha\beta}{\mathcal{X}}^\alpha\wedge{\mathcal{X}}^\beta,
&&d{\mathcal{V}}^A=0,\end{aligned}$$ where $\{{\mathcal{X}}^\alpha,{\mathcal{V}}^A\}$ is the dual basis corresponding to $\{{\mathcal{X}}_\alpha,{\mathcal{V}}_A\}$.
Prolongation of a map
---------------------
Let ${\Psi\colonP\toP'}$ be a fibered map from the fibered manifold ${\pi\colonP\toM}$ to the fibered manifold ${\pi'\colonP'\toM'}$ over a map ${\varphi\colonM\toM'}$. Let ${\Phi\colonE\toE'}$ be an admissible map from ${\tau\colonE\toM}$ to ${\tau'\colonE'\toM'}$ over the same map $\varphi$. The prolongation of $\Phi$ with respect to $\Psi$ is the mapping ${{\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Psi}\colon{\@ifstar{\@proldf}{\@prolpf}}[E]{P}\to{\@ifstar{\@proldf}{\@prolpf}}[E']{P'}}$ defined by $${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Psi}(p,b,v) =(\Psi(p),\Phi(b),(T_p\Psi)(v)).$$ It is clear from the definition that ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Psi}$ is a vector bundle map from ${\tau^E_P\colon{\@ifstar{\@proldf}{\@prolpf}}[E]{P}\toP}$ to ${\tau^{E'}_{P'}\colon{\@ifstar{\@proldf}{\@prolpf}}[E']{P'}\toP'}$ over $\Psi$. Moreover, in [@CFTLAMF] it is proved the following result.
The map ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Psi}$ is an admissible map. Moreover, ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Psi}$ is a morphism of Lie algebroids if and only if $\Phi$ is a morphism of Lie algebroids.
Given local coordinate systems $(x^i)$ on $M$ and $(x'{}^i)$ on $M'$, local adapted coordinates $(x^i,u^A)$ on $P$ and $(x'{}^i,u'{}^A)$ on $P'$ and a local basis of sections $\{e_\alpha\}$ of $E$ and $\{e'_\alpha\}$ of $E'$, the maps $\Phi$ and $\Psi$ are determined by $\Phi{^\star}e'{}^\alpha=\Phi^\alpha_\beta e^\beta$ and $\Psi(x,u)=(\phi^i(x),\psi^A(x,u))$. Then the action of ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Psi}$ is given by $$\begin{aligned}
({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Psi}){^\star}{\mathcal{X}}'{}^\alpha
&= \Phi_\beta^\alpha{\mathcal{X}}^\beta,\\
({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Psi}){^\star}{\mathcal{V}}'{}^A
&=\rho^i_\alpha{\frac{\partial \psi^A}{\partial x^i}}{\mathcal{X}}^\alpha+{\frac{\partial \psi^A}{\partial u^B}}{\mathcal{V}}^B.\end{aligned}$$
We finally mention that the composition of prolongation maps is the prolongation of the composition. Indeed, let $\Psi'$ be another bundle map from ${\pi'\colonP'\toM'}$ to another bundle ${\pi''\colonP''\toM''}$ and $\Phi'$ be another admissible map from ${\tau'\colonE'\toM'}$ to ${\tau''\colonE''\toM''}$ both over the same base map. Since $\Phi$ and $\Phi'$ are admissible maps then so is $\Phi'\circ\Phi$, and thus we can define the prolongation of $\Psi'\circ\Psi$ with respect to $\Phi'\circ\Phi$. We have that ${\@ifstar{\@proldf}{\@prolpf}}[\Phi'\circ\Phi]{(\Psi'\circ\Psi)}
=({\@ifstar{\@proldf}{\@prolpf}}[\Phi']{\Psi'})\circ({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Psi})$.
In the particular case when the bundles $P$ and $P'$ are just $P=E$ and $P'=E'$, whenever we have an admissible map ${\Phi\colonE\toE'}$ we can define the prolongation of $\Phi$ along $\Phi$ itself, by ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(a,b,v)=(\Phi(a),\Phi(b),T\Phi(v))$. From the result above, we have that ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}$ is a Lie algebroid morphism if and only if $\Phi$ is a Lie algebroid morphism. In coordinates we obtain $$\begin{aligned}
({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}){^\star}{\mathcal{X}}'{}^\alpha
&= \Phi_\beta^\alpha{\mathcal{X}}^\beta,\\
({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}){^\star}{\mathcal{V}}'{}^\alpha
&=\rho^i_\beta{\frac{\partial \Phi^\alpha_\gamma}{\partial x^i}}y^\gamma{\mathcal{X}}^\beta +
\Phi^\alpha_\beta{\mathcal{V}}^\beta,\end{aligned}$$ where $(x^i,y^\gamma)$ are the corresponding fibred coordinates on $E$. From this expression it is clear that ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}$ is fiberwise surjective if and only if $\Phi$ is fiberwise surjective.
Lagrangian Mechanics
--------------------
In [@LMLA] (see also [@PoPo]) a geometric formalism for Lagrangian Mechanics on Lie algebroids was defined. Such a formalism is similar to Klein’s formalism [@Klein] in standard Lagrangian mechanics and it is developed in the prolongation ${\mathcal{T}^E_{}E}$ of a Lie algebroid $E$ over itself. The canonical geometrical structures defined on ${\mathcal{T}^E_{}E}$ are the following:
- The *vertical lift* ${\xi{{^{\scriptscriptstyleV}}}\colon\tau^*E\to{\mathcal{T}^E_{}E}}$ given by $\xi{{^{\scriptscriptstyleV}}}(a,b)=(a,0,b{{^{\scriptscriptstyleV}}}_a)$, where $b{{^{\scriptscriptstyleV}}}_a$ is the vector tangent to the curve $a+tb$ at $t=0$,
- The *vertical endomorphism* ${S\colon{\mathcal{T}^E_{}E}\to{\mathcal{T}^E_{}E}}$ defined as follows: $$S(a,b,v)=\xi{{^{\scriptscriptstyleV}}}(a,b)=(a,0,b_a{{^{\scriptscriptstyleV}}}),$$
- The *Liouville section* which is the vertical section corresponding to the Liouville dilation vector field: $$\Delta(a)=\xi{{^{\scriptscriptstyleV}}}(a,a)=(a,0,a_a{{^{\scriptscriptstyleV}}}).$$
A section $\Gamma$ of ${\mathcal{T}^E_{}E}$ is said to be a [<span style="font-variant:small-caps;">sode</span>]{} section if $S\Gamma = \Delta$.
Given a Lagrangian function $L\in{C^\infty(E)}$ we define the *Cartan 1-form* $\theta_L$ and the *Cartan 2-form* $\omega_L$ as the forms on ${\@ifstar{\@proldf}{\@prolpf}}[E]{E}$ given by $$\label{Cartan-forms}
\theta_L=S^*(dL)\qquand \omega_L=-d\theta_L.$$ The real function $E_{L}$ on $E$ defined by $E_{L} = d_{\Delta}L -
L$ is the *energy function* of the Lagrangian system.
By a solution of the Lagrangian system (a solution of the *Euler-Lagrange equations*) we mean a [<span style="font-variant:small-caps;">sode</span>]{} section $\Gamma$ of ${\mathcal{T}^E_{}E}$ such that $$\label{Euler-Lagrange}
i_\Gamma\omega_L-dE_L=0.$$
The local expressions for the vertical endomorphism, the Liouville section, the Cartan $2$-form and the Lagrangian energy are $$\label{endverlo}
S{\mathcal{X}}_{\alpha} = {\mathcal{V}}_{\alpha}, \makebox[.3cm]{} S{\mathcal{V}}_{\alpha} = 0,
\makebox[.3cm]{} \mbox{ for all } \alpha,$$ $$\label{Lioulo} \Delta = y^{\alpha}{\mathcal{V}}_{\alpha},$$ $$\label{omegaL}
\omega_L
={\frac{\partial ^2L}{\partial y^\alpha\partial y^\beta}}{\mathcal{X}}^\alpha\wedge {\mathcal{V}}^\beta
+\frac{1}{2}\left(
{\frac{\partial ^2L}{\partial x^i\partial y^\alpha}}\rho^i_\beta-{\frac{\partial ^2L}{\partial x^i\partial
y^\beta}}\rho^i_\alpha+{\frac{\partial L}{\partial y^\gamma}}C^\gamma_{\alpha\beta}
\right){\mathcal{X}}^\alpha\wedge {\mathcal{X}}^\beta,$$ $$\label{EL}
E_L={\frac{\partial L}{\partial y^\alpha}}y^\alpha-L.$$
Thus, a [<span style="font-variant:small-caps;">sode</span>]{} $\Gamma$ is a section of the form $$\Gamma=y^\alpha{\mathcal{X}}_\alpha+f^\alpha{\mathcal{V}}_\alpha.$$ The [<span style="font-variant:small-caps;">sode</span>]{} $\Gamma$ is a solution of the Euler-Lagrange equations if and only if the functions $f^\alpha$ satisfy the linear equations $$\label{free-forces} {\frac{\partial ^2L}{\partial y^\beta\partial
y^\alpha}}f^\beta+{\frac{\partial ^2L}{\partial x^i\partial y^\alpha}}\rho^i_\beta
y^\beta +{\frac{\partial L}{\partial y^\gamma}}C^\gamma_{\alpha\beta}y^\beta
-\rho^i_\alpha{\frac{\partial L}{\partial x^i}} =0, \mbox{ for all } \alpha.$$ The *Euler-Lagrange differential equations* are the differential equations for the integral curves of the vector field $\rho^1(\Gamma)$, where the section $\Gamma$ is the solution of the Euler-Lagrange equations. Thus, these equations may be written as $$\dot{x}^i=\rho_\alpha^iy^\alpha,\;\;\;\; \frac{d}{dt}(\frac{\partial
L}{\partial y^\alpha})-\rho_\alpha^i\frac{\partial L}{\partial x^i}
+ \frac{\partial L}{\partial y^\gamma}C_{\alpha\beta}^\gamma
y^\beta=0.$$
In other words, if $\delta L: {\operatorname{Adm}(E)}\to E^*$ is the *Euler-Lagrange operator*, which locally reads $$\delta L=(\frac{d}{dt}(\frac{\partial L}{\partial y^\alpha}) +
C_{\alpha\beta}^\gamma y^\beta \frac{\partial L}{\partial
y^\gamma}-\rho_\alpha^i\frac{\partial L}{\partial x^i})e^\alpha,$$ where $\{e^\alpha\}$ is the dual basis of $\{e_\alpha\}$, then the Euler-Lagrange differential equations read $$\delta L=0.$$ The function $L$ is said to be *regular Lagrangian* if $\omega_{L}$ is regular at every point as a bilinear map. In such a case, there exists a unique section $\Gamma_{L}$ of ${\mathcal{T}^E_{}E}$ which satisfies the equation $$i_{\Gamma_{L}}\omega_{L} - dE_{L} = 0.$$ Note that from (\[endverlo\]), (\[Lioulo\]), (\[omegaL\]) and (\[EL\]), it follows that $$\label{2.4'} i_{SX} \omega_{L} = -S^*(i_{X}\omega_{L}),
\makebox[.3cm]{} i_{\Delta}\omega_{L} = -S^*(dE_{L}),$$ for $X \in {\operatorname{Sec}_{}({{\@ifstar{\@proldf}{\@prolpf}}[E]{E}})}$. Thus, using (\[2.4’\]), we deduce that $$i_{S\Gamma_{L}}\omega_{L} = i_{\Delta}\omega_{L}$$ which implies that $\Gamma_{L}$ is a [<span style="font-variant:small-caps;">sode</span>]{} section. Therefore, for a regular Lagrangian function $L$ we will say that the dynamical equations (\[Euler-Lagrange\]) are just the Euler-Lagrange equations.
On the other hand, the vertical distribution is isotropic with respect to $\omega_L$, see [@LeMaMa]. This fact implies that the contraction of $\omega_L$ with a vertical vector is a semibasic form. This property allows us to define a symmetric 2-tensor ${G{^{\scriptscriptstyleL}}_{}}$ along $\tau$ by $$\label{GL}
{G{^{\scriptscriptstyleL}}_{a}}(b,c)=\omega_L(\tilde{b},c_a{{^{\scriptscriptstyleV}}}),$$ where $\tilde{b}$ is any element in ${\mathcal{T}^E_{a}E}$ which projects to $b$, i.e., ${\@ifstar{\@proldf}{\@prolpf}}{\tau}(\tilde{b})=b$, and $a \in E$. In coordinates ${G{^{\scriptscriptstyleL}}_{}}=W_{\alpha\beta}e^\alpha\otimes e^\beta$, where the matrix $W_{\alpha\beta}$ is given by $$\label{HessianL}
W_{\alpha\beta}={\frac{\partial ^2L}{\partial y^\alpha\partial y^\beta}}.$$ It is easy to see that the Lagrangian $L$ is regular if and only the matrix $W$ is regular at every point, that is, if the tensor ${G{^{\scriptscriptstyleL}}_{}}$ is regular at every point. By the kernel of ${G{^{\scriptscriptstyleL}}_{}}$ at a point $a$ we mean the vector space $$\operatorname{Ker}{G{^{\scriptscriptstyleL}}_{a}}={\{\,b\in E_{\tau(a)}\,|\,{G{^{\scriptscriptstyleL}}_{a}}(b,c)=0\text{ for all }c\in
E_{\tau(a)}\,\}}.$$ In the case of a regular Lagrangian, the Cartan 2-section $\omega_L$ is symplectic (non-degenerate and $d$-closed) and the vertical subbundle is Lagrangian. It follows that a 1-form is semi-basic if and only if it is the contraction of $\omega_L$ with a vertical element.
Finally, we mention that the *complete lift* $\sigma{{^{\scriptscriptstyleC}}}$ of a section $\sigma\in{\operatorname{Sec}_{}(E)}$ is the section of ${\mathcal{T}^E_{}E}$ characterized by the two following properties:
1. projects to $\sigma$, i.e., ${\@ifstar{\@proldf}{\@prolpf}}{\tau}\circ\sigma{{^{\scriptscriptstyleC}}}=\sigma\circ\tau$,
2. $d_{\sigma{{^{\scriptscriptstyleC}}}}\hat{\mu}=\widehat{d_\sigma\mu}$,
where by $\hat{\alpha}\in{C^\infty(E)}$ we denote the linear function associated to a 1-section $\alpha\in{\operatorname{Sec}_{}(E^*)}$. Note that $$\label{SODEcompl}
\Gamma \mbox{ {\sc sode} section, } \sigma \in {\operatorname{Sec}_{}(E)} \Rightarrow
S[\sigma^{c}, \Gamma ] = 0,$$
$$\label{Vertcompl}
S \gamma = 0,\;\;\; \sigma \in {\operatorname{Sec}_{}(E)} \Rightarrow S[\sigma^{c},
\gamma ] = 0.$$
Linearly constrained Lagrangian systems {#linear}
=======================================
Nonholonomic systems on Lie algebroids were introduced in [@CoMa]. This class of systems includes, as particular cases, standard nonholonomic systems defined on the tangent bundle of a manifold and systems obtained by the reduction of the action of a symmetry group. The situation is similar to the non-constrained case, where the general equation $\delta L=0$ comprises as particular cases the standard Lagrangian Mechanics, Lagrangian Mechanics with symmetry, Lagrangian systems with holonomic constraints, systems on semi-direct products and systems evolving on Lie algebras, see e.g., [@LMLA].
We start with a free Lagrangian system on a Lie algebroid $E$. As mentioned above, these two objects can describe a wide class of systems. Now, we plug in some nonholonomic linear constraints described by a subbundle $D$ of the bundle $E$ of admissible directions. If we impose to the solution curves $a(t)$ the condition to stay on the manifold $D$, we arrive at the equations $\delta
L_{a(t)}=\lambda(t)$ and $a(t)\in D$, where the constraint force $\lambda(t)\in E^*_{\tau(a(t))}$ is to be determined. In the tangent bundle geometry case ($E=TM$), the d’Alembert principle establishes that the mechanical work done by the constraint forces vanishes, which implies that $\lambda$ takes values in the annihilator of the constraint manifold $D$. Therefore, in the case of a general Lie algebroid, the natural equations one should pose are (see [@CoMa]) $$\delta L_{a(t)}\in{D_{\tau(a(t))}^\circ}\qquand a(t)\in D.$$ In more explicit terms, we look for curves $a(t)$ on $E$ such that
- they are admissible, $\rho(a(t))=\dot{m}(t)$, where $m=\tau\circ a$,
- they stay in $D$, $a(t)\in D_{m(t)}$,
- there exists $\lambda(t)\in {D_{m(t)}^\circ}$ such that $\delta
L_{{a}(t)}=\lambda(t)$.
If $a(t)$ is one of such curves, then $(a(t),\dot{a}(t))$ is a curve in ${\mathcal{T}^E_{}E}$. Moreover, since $a(t)$ is in $D$, we have $\dot{a}(t)$ is tangent to $D$, that is, $(a(t),\dot{a}(t))\in{\mathcal{T}^D_{}D}$. Under some regularity conditions (to be made precise later on), we may assume that the above curves are integral curves of a section $\Gamma$, which as a consequence will be a [<span style="font-variant:small-caps;">sode</span>]{} section taking values in ${\mathcal{T}^D_{}D}$. Based on these arguments, we may reformulate geometrically our problem as the search for a [<span style="font-variant:small-caps;">sode</span>]{} $\Gamma$ (defined at least on a neighborhood of $D$) satisfying $(i_\Gamma\omega_L-dE_L)_a\in{\widetilde{D_{\tau(a)}^\circ}}$ and $\Gamma(a)\in {\mathcal T}_{a}^{D}D$, at every point $a\in D$. In the above expression ${\widetilde{D_{}^\circ}}$ is the pullback of ${D_{}^\circ}$ to ${\mathcal{T}^E_{}E}$, that is, $\alpha\in{\widetilde{D_{\tau(a)}^\circ}}$ if and only if there exists $\lambda\in{D_{\tau(a)}^\circ}$ such that $\alpha=\lambda\circ{\@ifstar{\@proldf}{\@prolpf}}{\tau}[a]$.
A nonholonomically *constrained Lagrangian system* on a Lie algebroid $E$ is a pair $(L,D)$, where $L$ is a smooth function on $E$, *the Lagrangian*, and $i\colon D\hookrightarrow E$ is a smooth subbundle of $E$, known as the *constraint subbundle*. By a solution of the nonholonomically constrained Lagrangian system $(L,D)$ we mean a [<span style="font-variant:small-caps;">sode</span>]{} section $\Gamma\in{\mathcal{T}^E_{}E}$ which satisfies the *Lagrange-d’Alembert equations* $$\label{Lagrange-d'Alembert}
\begin{aligned}
&(i_\Gamma\omega_L-dE_L)|_D\in {\operatorname{Sec}_{}({\widetilde{D_{}^\circ}})},\\
&\Gamma|_D\in {\operatorname{Sec}_{}({\mathcal{T}^D_{}D})}.
\end{aligned}$$
With a slight abuse of language, we will interchangeably refer to a solution of the constrained Lagrangian system as a section or the collection of its corresponding integral curves. The restriction of the projection ${\tau\colonE\toM}$ to $D$ will be denoted by $\pi$, that is, ${\pi=\tau|_D\colonD\toM}$.
[We want to stress that a solution of the Lagrange-d’Alembert equations needs to be defined only over $D$, but for practical purposes we consider it extended to $E$ (or just to a neighborhood of $D$ in $E$). We will not make any notational distinction between a solution on $D$ and any of its extensions. Solutions which coincide on $D$ will be considered as equal. See [@GrMe; @LeMa] for a more in-depth discussion. In accordance with this convention, by a [<span style="font-variant:small-caps;">sode</span>]{} on $D$ we mean a section of ${\mathcal{T}^D_{}D}$ which is the restriction to $D$ of some [<span style="font-variant:small-caps;">sode</span>]{} defined in a neighborhood of $D$. Alternatively, a [<span style="font-variant:small-caps;">sode</span>]{} on $D$ is a section $\Gamma$ of ${\mathcal{T}^D_{}D}$ such that ${\@ifstar{\@proldf}{\@prolpf}}{\tau}(\Gamma(a))=a$ for every $a\in D$.]{}
A nonholonomically constrained Lagrangian system $(L, D)$ on a Lie algebroid $E$ is said to be *holonomic* if $D$ is a Lie subalgebroid of $E$. This means that $[X, Y] \in {\operatorname{Sec}_{}(D)}$, for $X,
Y \in {\operatorname{Sec}_{}(D)}$. Thus, the real function $L_{D} = L_{|D}: D \to \R$ defines an unconstrained (free) Lagrangian system on the Lie algebroid $D$. Moreover, it is easy to prove that $\mathcal{I}
\circ \Delta_{D} = \Delta \circ i$ and $\mathcal{I} \circ S_{D} =
S \circ \mathcal{I}$, where $\mathcal{I} = {{\@ifstar{\@proldf}{\@prolpf}}[i]{i}}: {\mathcal{T}^D_{}D}\to
{\mathcal{T}^E_{}E}$ is the prolongation of the Lie algebroid morphism $i\colon
D\hookrightarrow E$ and $\Delta_{D}$ (respectively, $S_{D}$) is the Liouville section (respectively, the vertical endomorphism) of the Lie algebroid ${\mathcal{T}^D_{}D}$. Therefore, since $L \circ i = L_D$, we deduce that $$\mathcal{I}^*(\theta_{L}) = \theta_{L_{D}}, \makebox[.4cm]{}
\mathcal{I}^*(\omega_{L}) = \omega_{L_{D}}, \makebox[.4cm]{}
\mathcal{I}^*(dE_{L}) = dE_{L_{D}}.$$ Consequently, if $\Gamma$ is a [<span style="font-variant:small-caps;">sode</span>]{} section of ${\mathcal{T}^E_{}E}$, $a, b \in
D$, $(b, X) \in {\@ifstar{\@proldf}{\@prolpf}}[E]{E}[a]$ and $(b, Y) \in {\@ifstar{\@proldf}{\@prolpf}}[E]{D}[a]$ then $$(i_{\Gamma}\omega_{L} - dE_{L})(a)(b, X) = (i_{\Gamma_{|D}}
\omega_{L_{D}} - dE_{L_{D}})(a)(b, Y) + (i_{\Gamma}\omega_{L} -
dE_{L})(a)(0, Z),$$ $(0, Z)$ being a vertical element of ${\@ifstar{\@proldf}{\@prolpf}}[E]{E}[a]$.
Now, using (\[2.4’\]), we have that $(i_{\Gamma}\omega_{L} -
dE_{L})(a)(0, Z) = 0$ which implies that $$(i_{\Gamma}\omega_{L} - dE_{L})(a)(b, X) = (i_{\Gamma_{|D}}
\omega_{L_{D}} - dE_{L_{D}})(a)(b, Y).$$ The above facts prove that a [<span style="font-variant:small-caps;">sode</span>]{} section $\Gamma$ of ${\mathcal{T}^E_{}E}$ is a solution of the holonomic Lagrangian system $(L, D)$ on $E$ if and only if $\Gamma_{|D}$ is a solution of the Euler-Lagrange equations for the (unconstrained) Lagrangian function $L_D$ on the Lie algebroid $D$. In other words, the holonomic Lagrangian system $(L, D)$ on $E$ may be considered as an unconstrained (free) Lagrangian system on the Lie algebroid $D$.
Next, suppose that $(L, D)$ is a nonholonomically constrained Lagrangian system on the Lie algebroid $E$. Then, the different spaces we will consider are shown in the following commutative diagram $$\xymatrix{ &&TM\ar@{=}[r]&TM&\\
&&D\ar[u]_{\rho_D}\ar[r]^i&E\ar[u]^\rho\\
&TD\ar[ruu]^{T\pi}\ar[rd]_{\tau_D}&
{\mathcal{T}^D_{}D}\ar[u]_{{\@ifstar{\@proldf}{\@prolpf}}{\pi}}\ar[r]^{\mathcal{I}}\ar[d]^{\pi^D_D}\ar[l]_{\rho^1}&
{\mathcal{T}^E_{}E}\ar[u]^{{\@ifstar{\@proldf}{\@prolpf}}{\tau}}\ar[d]_{\tau^E_E}\ar[r]^{\rho^1} &
TE\ar[luu]_{T\tau}\ar[dl]^{\tau_E}\\
&&D\ar[r]^i\ar[d]_{\pi}&E\ar[d]^{\tau}&\\
&&M\ar@{=}[r]&M }$$
As an intermediate space in our analysis of the regularity of the constrained systems, we will also consider ${\mathcal{T}^E_{}D}$, the $E$-tangent to $D$. The main difference between ${\mathcal{T}^E_{}D}$ and ${\mathcal{T}^D_{}D}$ is that the former has a natural Lie algebroid structure while the later does not.
The following two results are immediate consequences of the above form of the Lagrange-d’Alembert equations.
If $(L,D)$ is a constrained Lagrangian system and $\Gamma$ is a solution of the dynamics, then $d_\Gamma E_L=0$ (on $D$).
Indeed, for every $a\in D$, we have $\Gamma(a)\in{\mathcal{T}^D_{a}D}$, so that ${\@ifstar{\@proldf}{\@prolpf}}{\tau}(\Gamma(a))\in D$. Therefore $i_\Gamma{\widetilde{D_{}^\circ}}=0$ and contracting $0=i_\Gamma(i_\Gamma\omega_L-dE_L)=-d_\Gamma E_L$ at every point in $D$.
Let $(L,D)$ be a constrained Lagrangian system which admits a unique [<span style="font-variant:small-caps;">sode</span>]{} $\Gamma$ solution of the dynamics. If $\sigma$ is a section of $D$ such that there exists a function $f\in{C^\infty(M)}$ satisfying $$d_{\sigma{{^{\scriptscriptstyleC}}}}L=\dot{f},$$ then the function $F={\left\langle\theta_L,\sigma{{^{\scriptscriptstyleC}}}\right\rangle}-f$ is a constant of the motion, that is, $d_\Gamma F=0$ (on $D$).
Using that $\theta_L(\Gamma)=d_\Delta(L)$, we obtain $i_{\sigma{{^{\scriptscriptstyleC}}}}(i_\Gamma\omega_L-dE_L) = i_{\sigma{{^{\scriptscriptstyleC}}}}(-d_\Gamma
\theta_L + dL)= d_{\sigma{{^{\scriptscriptstyleC}}}}L-d_\Gamma{\left\langle\theta_L,\sigma{{^{\scriptscriptstyleC}}}\right\rangle} +
\theta_L [\Gamma,\sigma{{^{\scriptscriptstyleC}}}]$ and, since $[\Gamma,\sigma{{^{\scriptscriptstyleC}}}]$ is vertical, we deduce $$i_{\sigma^c}(i_\Gamma
\omega_L-dE_L)=d_{\sigma^c}L-d_\Gamma{\left\langle\theta_L,\sigma{{^{\scriptscriptstyleC}}}\right\rangle}.$$ Thus, taking into account that $i_{\sigma{{^{\scriptscriptstyleC}}}}{\widetilde{D_{}^\circ}}=0$, we get $0=d_\Gamma({\left\langle\theta_L,\sigma{{^{\scriptscriptstyleC}}}\right\rangle}-f)=-d_\Gamma F$.
Let ${\mathcal{G}}:E\times_M E\to \R$ be a bundle metric on $E$. The *Levi-Civita connection* $\nabla^{\mathcal{G}}$ is determined by the formula $$\begin{aligned}
2{\mathcal{G}}(\nabla^{{\mathcal{G}}}_\sigma\eta,\zeta)
&=\rho(\sigma)({\mathcal{G}}(\eta,\zeta)) +
\rho(\eta)({\mathcal{G}}(\sigma,\zeta))-\rho(\zeta)({\mathcal{G}}(\eta,\sigma))\\
& \qquad + {\mathcal{G}}(\sigma,[\zeta,\eta]) +
{\mathcal{G}}(\eta,[\zeta,\sigma])-{\mathcal{G}}(\zeta, [\eta,\sigma]) ,
\end{aligned}$$ for $\sigma,\eta,\zeta\in {\operatorname{Sec}_{}(E)}$. The coefficients of the connection $\nabla^{{\mathcal{G}}}$ are given by $$\Gamma_{\beta\gamma}^\alpha =
\frac{1}{2}{\mathcal{G}}^{\alpha\nu}([\nu,\beta;\gamma]+[\nu,\gamma; \beta] +
[\beta,\gamma; \nu]),$$ where ${\mathcal{G}}_{\alpha\nu}$ are the coefficients of the metric ${\mathcal{G}}$, $({\mathcal{G}}^{\alpha\nu})$ is the inverse matrix of $({\mathcal{G}}_{\alpha\nu})$ and $$[\alpha,\beta;\gamma]=\frac{\partial {\mathcal{G}}_{\alpha\beta}}{\partial
x^i}\rho_\gamma^i + C_{\alpha\beta}^\mu{\mathcal{G}}_{\mu\gamma}.$$
Using the covariant derivative induced by $\nabla^{{\mathcal{G}}}$, one may introduce the notion of a geodesic of $\nabla^{{\mathcal{G}}}$ as follows. An admissible curve $a:I\to E$ is said to be a *geodesic* if $\nabla_{a(t)}^{{\mathcal{G}}}a(t)=0$, for all $t\in I$. In local coordinates, the conditions for being a geodesic read $$\frac{da^\gamma}{dt} + \frac{1}{2}(\Gamma_{\alpha\beta}^\gamma +
\Gamma_{\beta\alpha}^\gamma)a^\alpha a^\beta=0,\;\;\; \mbox{for all
}\gamma.$$ The geodesics are the integral curves of a [<span style="font-variant:small-caps;">sode</span>]{} section $\Gamma_{\nabla^{{\mathcal{G}}}}$ of ${\@ifstar{\@proldf}{\@prolpf}}[E]{E}$, which is locally given by $$\Gamma_{\nabla^{{\mathcal{G}}}}=y^\gamma{\mathcal
X}_\gamma-\frac{1}{2}(\Gamma_{\alpha\beta}^\gamma +
\Gamma_{\beta\alpha}^\gamma)y^\alpha y^\beta{\mathcal{V}}_\gamma.$$ $\Gamma_{\nabla^{{\mathcal{G}}}}$ is called the *geodesic flow* (for more details, see [@CoMa]).
The class of systems that were considered in detail in [@CoMa] is that of *mechanical systems with nonholonomic constraints*[^2]. The Lagrangian function $L$ is of mechanical type, i.e., it is of the form $$L(a)=\frac{1}{2} {\mathcal{G}}(a,a) - V(\tau(a)),
\quad
a\in E,$$ with $V$ a function on $M$.
The Euler-Lagrange section for the unconstrained system can be written as $$\Gamma_L=\Gamma_{\nabla^{\mathcal{G}}} - (\operatorname{grad}_{\mathcal{G}}V) {{^{\scriptscriptstyleV}}}.$$ In this expression, by $\operatorname{grad}_{\mathcal{G}}{V}$ we mean the section of $E$ such that ${\left\langledV(m),a\right\rangle}={\mathcal{G}}(\operatorname{grad}_{\mathcal{G}}V(m),a)$, for all $m\in M$ and all $a\in E_m$, and where we remind that $d$ is the differential in the Lie algebroid. The Euler-Lagrange differential equations can be written as $$\label{Mechanical:eqs-motion}\begin{array}{ll}
\dot{x}^i & =\rho^i_\alpha y^\alpha , \\[5pt]
\dot{y}^\alpha &= -\displaystyle\frac{1}{2} \left(
\Gamma^\alpha_{\beta \gamma} +\Gamma^\alpha_{\gamma\beta}\right)
y^\beta y^\gamma - {\mathcal{G}}^{\alpha \beta} \rho^i_{\beta} {\frac{\partial V}{\partial x^i}}.
\end{array}$$ Alternatively, one can describe the dynamical behavior of the mechanical control system by means of an equation on $E$ via the covariant derivative. An admissible curve $a: t \mapsto a(t)$ with base curve $t\mapsto m(t)$ is a solution of the system if and only if $$\begin{aligned}
\label{Mechanical:eqs-motion-connection}
\nabla^{{\mathcal{G}}}_{a(t)}a(t) + \operatorname{grad}_{\mathcal{G}}V(m(t)) = 0.
\end{aligned}$$
Note that $${\mathcal{G}}(m(t))(\nabla_{a(t)}^{\mathcal{G}}a(t) + \operatorname{grad}_{\mathcal{G}}V(m(t)),b)=\delta
L(a(t))(b),\;\;\;\; \mbox{ for } b\in E_{m(t)}.$$ If this mechanical control system is subject to the constraints determined by a subbundle $D$ of $E$, we can do the following. Consider the orthogonal decomposition $E=D\oplus D^\perp$, and the associated orthogonal projectors ${P\colonE\toD}$, ${Q\colonE\toD^\perp}$. Using the fact that ${\mathcal{G}}(P\cdot,\cdot) =
{\mathcal{G}}(\cdot,P\cdot)$, one can write the Lagrange-d’Alembert equations in the form $$P(\nabla^{\mathcal{G}}_{a(t)}a(t)) + P(\operatorname{grad}_{\mathcal{G}}V(m(t))) = 0, \qquad Q(a)=0.$$ A specially well-suited form of these equations makes use of the *constrained connection* ${\check{\nabla}}$ defined by ${\check{\nabla}}_\sigma\eta
=P(\nabla^{\mathcal{G}}_\sigma\eta)+\nabla^{\mathcal{G}}_\sigma(Q\eta)$. In terms of ${\check{\nabla}}$, we can rewrite this equation as ${\check{\nabla}}_{a(t)}a(t) +
P(\operatorname{grad}_{\mathcal{G}}V(m(t))) = 0$, $Q(a)=0$, where we have used the fact that the connection ${\check{\nabla}}$ restricts to the subbundle $D$.
Moreover, following the ideas in [@Lewis], we proved in [@CoMa] that the subbundle $D$ is geodesically invariant for the connection ${\check{\nabla}}$, that is, any integral curve of the spray $\Gamma_{{\check{\nabla}}}$ associated with ${\check{\nabla}}$ starting from a point in $D$ is entirely contained in $D$. Since the terms coming from the potential $V$ also belongs to $D$, we have that the constrained equations of motion can be simply stated as $$\begin{aligned}
\label{eq:eqs-motion-connection-nh}
{\check{\nabla}}_{a(t)}a(t) + P(\operatorname{grad}_{\mathcal{G}}V(m(t))) = 0,
\qquad
a(0)\in D.
\end{aligned}$$ Note that one can write the constrained equations of the motion as follows $$\dot{a}(t)=\rho^1(\Gamma_{{\check{\nabla}}}(a(t))-{P}(\operatorname{grad}_{{\mathcal{G}}}V)^v(a(t)))$$ and that the restriction to $D$ of the vector field $\rho^1(\Gamma_{{\check{\nabla}}}-{P}(\operatorname{grad}_{{\mathcal{G}}}V)^v)$ is tangent to ${
D}.$
The coordinate expression of equations (\[eq:eqs-motion-connection-nh\]) is greatly simplified if we take a basis $\{e_\alpha\}=\{e_a,e_A\}$ of $E$ adapted to the orthogonal decomposition $E=D\oplus D^\perp$, i.e., $D
=\operatorname{span}\{e_a\}$, $D^\perp =
\operatorname{span}\{e_A\}$. Denoting by $(y^\alpha)=(y^a,y^A)$ the induced coordinates, the constraint equations $Q(a)=0$ just read $y^A=0$. The differential equations of the motion are then $$\begin{aligned}
\dot{x}^i & = \rho^i_a y^a , \\
\dot{y}^a & = - \frac{1}{2}
\left(\check{\Gamma}^a_{bc}+\check{\Gamma}^a_{cb}\right)
y^by^c-{\mathcal{G}}^{ab}\rho^i_b{\frac{\partial V}{\partial x^i}}, \\
y^A &= 0,
\end{aligned}$$ where $\check{\Gamma}^\alpha_{\beta \gamma}$ are the connection coefficients of the constrained connection $\check{\nabla}$.
In the above example the dynamics exists and is completely determined whatever the (linear) constraints are. As we will see in Section \[sec:regularity\], this property is lost in the general case.
Lagrange-d’Alembert equations in local coordinates
--------------------------------------------------
Let us analyze the form of the Lagrange-d’Alembert equations in local coordinates. Following the example above, let us choose a special coordinate system adapted to the structure of the problem as follows. We consider local coordinates $(x^i)$ on an open set $\mathcal{U}$ of $M$ and we take a basis $\{e_a\}$ of local sections of $D$ and complete it to a basis $\{e_a,e_A\}$ of local sections of $E$ (both defined on the open $\mathcal{U}$). In this way, we have coordinates $(x^i,y^a,y^A)$ on $E$. In this set of coordinates, the constraints imposed by the submanifold $D\subset E$ are simply $y^A=0$. If $\{e^a,e^A\}$ is the dual basis of $\{e_a,e_A\}$, then a basis for the annihilator ${D_{}^\circ}$ of $D$ is $\{e^A\}$ and a basis for ${\widetilde{D_{}^\circ}}$ is ${\mathcal{X}}^A$.
An element $z$ of ${\mathcal{T}^E_{}D}$ is of the form $z= u^\alpha{\mathcal{X}}_\alpha+z^a{\mathcal{V}}_a
= u^a{\mathcal{X}}_a+u^A{\mathcal{X}}_A+z^a{\mathcal{V}}_a$, that is, the component ${\mathcal{V}}_A$ vanishes since $\rho^1(z)$ is a vector tangent to the manifold $D$ with equations $y^A=0$. The projection of $z$ to $E$ is ${\@ifstar{\@proldf}{\@prolpf}}{\tau}(z)=u^a
e_a+u^A e_A$, so that the element $z$ is in ${\mathcal{T}^D_{}D}$ if and only if $u^A=0$. In other words, an element in ${\mathcal{T}^D_{}D}$ is of the form $z=u^a{\mathcal{X}}_a+z^a{\mathcal{V}}_a$.
Let us find the local expression of the Lagrange-d’Alembert equations in these coordinates. We consider a section $\Gamma$ such that $\Gamma_{|D} \in {\operatorname{Sec}_{}({\mathcal{T}^D_{}D})}$, which is therefore of the form $\Gamma=g^a{\mathcal{X}}_a+f^a{\mathcal{V}}_a$. From the local expression of the Cartan 2-form and the local expression of the energy function, we get $$0={\left\langlei_\Gamma\omega_L-dE_L,{\mathcal{V}}_\alpha\right\rangle}=
-y^B{\frac{\partial ^2L}{\partial y^\alpha\partial y^B}} -(y^b-g^b){\frac{\partial ^2L}{\partial y^\alpha\partial
y^b}}.$$ If we assume that the Lagrangian $L$ is regular, when we evaluate at $y^A=0$, we have that $g^a=y^a$ and thus $\Gamma$ is a [<span style="font-variant:small-caps;">sode</span>]{}. Moreover, contracting with ${\mathcal{X}}_a$, after a few calculations we get $$0={\left\langlei_\Gamma\omega_L-dE_L,{\mathcal{X}}_a\right\rangle} =-\left\{
d_\Gamma\left({\frac{\partial L}{\partial y^a}}\right)
+{\frac{\partial L}{\partial y^\gamma}}C^\gamma_{a\beta}y^\beta -\rho^i_a{\frac{\partial L}{\partial x^i}}
\right\},$$ so that (again after evaluation at $y^A=0$), the functions $f^a$ are solution of the linear equations $$\label{LD-explicit} {\frac{\partial ^2L}{\partial y^b\partial
y^a}}f^b+{\frac{\partial ^2L}{\partial x^i\partial y^a}}\rho^i_by^b
+{\frac{\partial L}{\partial y^\gamma}}C^\gamma_{ab}y^b -\rho^i_a{\frac{\partial L}{\partial x^i}} =0,$$ where all the partial derivatives of the Lagrangian are to be evaluated on $y^A=0$.
As a consequence, we get that there exists a unique solution of the Lagrange-d’Alembert equations if and only if the matrix $$\label{gld-local}
{\mathcal{C}}_{ab}(x^i,y^c)={\frac{\partial ^2L}{\partial y^a\partial y^b}}(x^i,y^c,0)$$ is regular. Notice that ${\mathcal{C}}_{ab}$ is a submatrix of $W_{\alpha\beta}$, evaluated at $y^A=0$ and that, as we know, if $L$ is of mechanical type then the Lagrange-d’Alembert equations have a unique solution. The differential equations for the integral curves of the vector field $\rho^1(\Gamma)$ are the Lagrange-d’Alembert differential equations, which read $$\label{LD-edo}
\begin{aligned}
&\dot{x}^i=\rho^i_ay^a,\\
&\frac{d}{dt}\left({\frac{\partial L}{\partial y^a}}\right) +
{\frac{\partial L}{\partial y^\gamma}}C^\gamma_{ab}y^b -\rho^i_a{\frac{\partial L}{\partial x^i}}=0,\\
&y^A=0.
\end{aligned}$$ Finally, notice that the contraction with ${\mathcal{X}}_A$ just gives the components $\lambda_A={\left\langlei_\Gamma\omega_L-dE_L,{\mathcal{X}}_A\right\rangle}|_{y^A=0}$ of the constraint forces $\lambda=\lambda_Ae^A$.
[In some occasions, it is useful to write the equations in the form $$\label{LD-edo2}
\begin{aligned}
&\dot{x}^i=\rho^i_ay^a,\\
&\frac{d}{dt}\left({\frac{\partial L}{\partial y^a}}\right) + {\frac{\partial L}{\partial y^c}}C^c_{ab}y^b
-\rho^i_a{\frac{\partial L}{\partial x^i}}=-{\frac{\partial L}{\partial y^A}}C^A_{ab}y^b,\\
&y^A=0,
\end{aligned}$$ where, on the left-hand side of the second equation, all the derivatives can be calculated from the value of the Lagrangian on the constraint submanifold $D$. In other words, we can substitute $L$ by the constrained Lagrangian $L_c$ defined by $L_c(x^i,y^a)=L(x^i,y^a,0)$.]{}
[A particular case of this construction is given by constrained systems defined in the standard Lie algebroid ${\tau_M\colonTM\toM}$. In this case, the equations are the Lagrange-d’Alembert equations written in quasicoordinates, where $C^{\alpha}_{\beta\gamma}$ are the so-called Hamel’s transpositional symbols, which obviously are nothing but the structure coefficients (in the Cartan’s sense) of the moving frame $\{e_\alpha\}$, see e.g., [@EhKoMoRi; @Ha].]{}
Solution of Lagrange-d’Alembert equations {#sec:regularity}
-----------------------------------------
In what follows, we will assume that the Lagrangian $L$ is regular at least in a neighborhood of $D$.
Let us now perform a precise global analysis of the existence and uniqueness of the solution of Lagrange-d’Alembert equations.
A constrained Lagrangian system $(L,D)$ is said to be *regular* if the Lagrange-d’Alembert equations have a unique solution.
In order to characterize geometrically those nonholonomic systems which are regular, we define the tensor ${G{^{\scriptscriptstyleL,D}}_{}}$ as the restriction of ${G{^{\scriptscriptstyleL}}_{}}$ to $D$, that is, ${G{^{\scriptscriptstyleL,D}}_{a}}(b,c)={G{^{\scriptscriptstyleL}}_{a}}(b,c)$ for every $a\in D$ and every $b,c\in D_{\tau(a)}$. In coordinates adapted to $D$, we have that the local expression of ${G{^{\scriptscriptstyleL,D}}_{}}$ is ${G{^{\scriptscriptstyleL,D}}_{}}={\mathcal{C}}_{ab}e^a\otimes e^b$ where the matrix ${\mathcal{C}}_{ab}$ is given by equation .
A second important geometric object is the subbundle $F\subset{\mathcal{T}^E_{}E}|_D\to D$ whose fiber at the point $a\in D$ is $F_a=\omega_L^{-1}({\widetilde{D_{\tau(a)}^\circ}})$. More explicitly, $$F_a = {\{\,z\in{\mathcal{T}^E_{a}E}\,|\,\text{exists $\zeta\in{D_{\tau(a)}^\circ}$ s.t.
$\omega_L(z,u)={\left\langle\zeta,{\@ifstar{\@proldf}{\@prolpf}}{\tau}(u)\right\rangle}$ for all
$u\in{\mathcal{T}^E_{a}E}$}\,\}}.$$ From the definition, it is clear that the rank of $F$ is ${\operatorname{rank}(F)}={\operatorname{rank}({D_{}^\circ})}={\operatorname{rank}(E)}-{\operatorname{rank}(D)}$.
Finally, we also consider the subbundle $({\mathcal{T}^D_{}D})^\perp\subset{\mathcal{T}^E_{}E}|_D\to
D$, the orthogonal to ${\mathcal{T}^D_{}D}$ with respect to the symplectic form $\omega_L$. The rank of $({\mathcal{T}^D_{}D})^\perp $ is ${\operatorname{rank}({\mathcal{T}^D_{}D})}^\perp={\operatorname{rank}({\mathcal{T}^E_{}E})} - {\operatorname{rank}({\mathcal{T}^D_{}D})} = 2({\operatorname{rank}(E)} - {\operatorname{rank}(D)})
= 2 {\operatorname{rank}({D_{}^\circ})}$.
The relation among these three objects is described by the following result.
\[F-TDD\] The following properties are satisfied:
1. The elements in $F$ are vertical. An element $\xi{{^{\scriptscriptstyleV}}}(a,b)\in
F_a$ if and only if ${G{^{\scriptscriptstyleL}}_{a}}(b,c)=0$ for all $c\in D_{\tau(a)}$.
2. $({\mathcal{T}^D_{}D})^\perp\cap{\operatorname{Ver}({\mathcal{T}^E_{}E})}=F$.
\(1) The elements in $F$ are vertical because the elements in ${\widetilde{D_{}^\circ}}$ are semi-basic. If $\xi{{^{\scriptscriptstyleV}}}(a,b)\in F_a$ then there exists $\zeta\in
{D_{\tau (a)}^\circ}$ such that $\omega_L(\xi{{^{\scriptscriptstyleV}}}(a,b),u)={\left\langle\zeta,{\@ifstar{\@proldf}{\@prolpf}}{\tau}(u)\right\rangle}$ for all $u\in{\mathcal{T}^E_{a}E}$. In terms of ${G{^{\scriptscriptstyleL}}_{}}$ and writing $c={\@ifstar{\@proldf}{\@prolpf}}{\tau}(u)$, the above equation reads $-{G{^{\scriptscriptstyleL}}_{a}}(b,c)={\left\langle\zeta,c\right\rangle}$. By taking $u\in{\@ifstar{\@proldf}{\@prolpf}}{\tau}^{-1}(D)$, then $c$ is in $D$ and therefore ${G{^{\scriptscriptstyleL}}_{a}}(b,c)=0$ for all $c\in D_{\tau(a)}$. Conversely, if ${G{^{\scriptscriptstyleL}}_{a}}(b,c)=0$ for all $c\in D_{\tau(a)}$, then the 1-form $\zeta=-{G{^{\scriptscriptstyleL}}_{a}}(b,\ )$ is in ${D_{\tau(a)}^\circ}$. Therefore $\omega_L(\xi{{^{\scriptscriptstyleV}}}(a,b),u) = - {G{^{\scriptscriptstyleL}}_{a}}(b,{\@ifstar{\@proldf}{\@prolpf}}{\tau}(u)) =
{\left\langle\zeta,{\@ifstar{\@proldf}{\@prolpf}}{\tau}(u)\right\rangle})$, which is the condition for $\xi{{^{\scriptscriptstyleV}}}(a,b)\in F_a$.
\(2) The condition for a vertical element $\xi{{^{\scriptscriptstyleV}}}(a,b)$ to be in $({\mathcal{T}^D_{}D})^\perp$ is $\omega_L(\xi{{^{\scriptscriptstyleV}}}(a,b),w)=0$ for all $w\in{\mathcal{T}^D_{a}D}$, or equivalently, ${G{^{\scriptscriptstyleL}}_{a}}(b,{\@ifstar{\@proldf}{\@prolpf}}{\tau}(w))=0$. The vector $c={\@ifstar{\@proldf}{\@prolpf}}{\tau}(w)$ is an arbitrary element of $D_{\tau(a)}$, so that the above condition reads ${G{^{\scriptscriptstyleL}}_{a}}(b,c)=0$, for all $c\in
D_{\tau(a)}$, which is precisely the condition for $\xi{{^{\scriptscriptstyleV}}}(a,b)$ to be in $F_a$.
\[regularity\] The following properties are equivalent:
1. The constrained Lagrangian system $(L,D)$ is regular,
2. $\operatorname{Ker}{G{^{\scriptscriptstyleL,D}}_{}}=\{0\}$,
3. ${\mathcal{T}^E_{}D}\cap F=\{0\}$,
4. ${\mathcal{T}^D_{}D}\cap({\mathcal{T}^D_{}D})^\perp=\{0\}$.
[\[(1)$\Leftrightarrow$(2)\]]{} The equivalence between the first two conditions is clear from the local form of the Lagrange-d’Alembert equations , since the coefficients of the unknowns $f^a$ are precisely the components of ${G{^{\scriptscriptstyleL,D}}_{}}$.
\[(2)$\Leftrightarrow$(3)\] ($\Rightarrow$) Let $a\in D$ and consider an element $z\in {\mathcal{T}^E_{a}D}\cap F_a$. Since the elements of $F$ are vertical, we have $z=\xi{{^{\scriptscriptstyleV}}}(a,b)$ for some $b\in E_{\tau(a)}$. Moreover, $z\in{\mathcal{T}^E_{a}D}$ implies that $b$ is an element in $D_{\tau(a)}$. On the other hand, if $z=\xi{{^{\scriptscriptstyleV}}}(a,b)$ is in $F_a$, then Lemma \[F-TDD\] implies that ${G{^{\scriptscriptstyleL}}_{a}}(b,c)=0$ for all $c\in
D_{\tau(a)}$. Thus ${G{^{\scriptscriptstyleL,D}}_{a}}(b,c)=0$ for all $c\in D_{\tau(a)}$, from where $b=0$, and hence $z=0$.
($\Leftarrow$) Conversely, if for some $a\in D$, there exists $b\in\operatorname{Ker}{G{^{\scriptscriptstyleL,D}}_{a}}$ with $b\neq0$ then, using Lemma \[F-TDD\], we deduce that $z=\xi{{^{\scriptscriptstyleV}}}(a,b)\in{\mathcal{T}^E_{a}D}\cap F_a$ and $z\neq0$.
\[(2)$\Leftrightarrow$(4)\] ($\Rightarrow$) Let $a\in D$ and consider an element $v\in {\mathcal{T}^D_{a}D}\cap({\mathcal{T}^D_{a}D})^\perp$, that is, $\omega_L(v,w)=0$ for all $w\in{\mathcal{T}^D_{a}D}$. If we take $w=\xi{{^{\scriptscriptstyleV}}}(a,b)$ for $b\in D_{\tau(a)}$ arbitrary, then we have $\omega_L(v,\xi{{^{\scriptscriptstyleV}}}(a,b)) = {G{^{\scriptscriptstyleL,D}}_{a}}({\@ifstar{\@proldf}{\@prolpf}}{\tau}(v),b)=0$ for all $b\in D_{\tau(a)}$, from where it follows that ${\@ifstar{\@proldf}{\@prolpf}}{\tau}(v)=0$. Thus $v$ is vertical, $v=\xi{{^{\scriptscriptstyleV}}}(a,c)$, for some $c\in D$ and then $\omega_L(\xi{{^{\scriptscriptstyleV}}}(a,c),w) = -{G{^{\scriptscriptstyleL,D}}_{a}}(c,{\@ifstar{\@proldf}{\@prolpf}}{\tau}(w))=0$ for all $w\in{\mathcal{T}^D_{a}D}$. Therefore $c=0$ and hence $v=0$.
($\Leftarrow$) Conversely, if for some $a\in D$, there exists $b\in\operatorname{Ker}{G{^{\scriptscriptstyleL,D}}_{a}}$ with $b\neq0$, then $0\neq\xi{{^{\scriptscriptstyleV}}}(a,b) \in {\mathcal{T}^D_{a}D}
\cap({\mathcal{T}^D_{a}D})^\perp$, because $\omega_L(\xi{{^{\scriptscriptstyleV}}}(a,b),w) =
{G{^{\scriptscriptstyleL,D}}_{}}(b,{\@ifstar{\@proldf}{\@prolpf}}{\tau}(w))=0$ for all $w\in{\mathcal{T}^D_{a}D}$.
In the case of a constrained mechanical system, the tensor ${G{^{\scriptscriptstyleL}}_{}}$ is given by ${G{^{\scriptscriptstyleL}}_{a}}(b,c) = {\mathcal{G}}_{\tau(a)}(b,c)$, so that it is positive definite at every point. Thus the restriction to any subbundle $D$ is also positive definite and hence regular. Thus, nonholonomic mechanical systems are always regular.
Conditions (3) and (4) in Theorem \[regularity\] are equivalent, respectively, to
- ${\mathcal{T}^E_{}E}|_D={\mathcal{T}^E_{}D}\oplus F$,
- ${\mathcal{T}^E_{}E}|_D={\mathcal{T}^D_{}D}\oplus({\mathcal{T}^D_{}D})^\perp$.
The equivalence between (4) and (4’) is obvious, since we are assuming that the free Lagrangian is regular, i.e., $\omega_L$ is symplectic. The equivalence of (3) and (3’) follows by computing the dimension of the corresponding spaces. The ranks of ${\mathcal{T}^E_{}E}$, ${\mathcal{T}^E_{}D}$ and $F$ are $$\begin{aligned}
&{\operatorname{rank}({\mathcal{T}^E_{}E})} = 2 {\operatorname{rank}(E)} , \\
&{\operatorname{rank}({\mathcal{T}^E_{}D})} = {\operatorname{rank}(E)} +{\operatorname{rank}(D)} , \\
&{\operatorname{rank}(F)} = {\operatorname{rank}({D_{}^\circ})} = {\operatorname{rank}(E)} - {\operatorname{rank}(D)}.
\end{aligned}$$ Thus ${\operatorname{rank}({\mathcal{T}^E_{}E})}={\operatorname{rank}({\mathcal{T}^E_{}D})} + {\operatorname{rank}(F)}$, and the result follows.
Projectors {#Projectors}
----------
We can express the constrained dynamical section in terms of the free dynamical section by projecting to the adequate space, either ${\mathcal{T}^E_{}D}$ or ${\mathcal{T}^D_{}D}$, according to each of the above decompositions of ${\mathcal{T}^E_{}E}|_D$. Of course, both procedures give the same result.
### Projection to ${\mathcal{T}^E_{}D}$ {#projection-to-mathcalte_d .unnumbered}
Assuming that the constrained system is regular, we have a direct sum decomposition $${\mathcal{T}^E_{a}E}={\mathcal{T}^E_{a}D}\oplus F_a,$$ for every $a\in D$, where we recall that the subbundle $F\subset{\mathcal{T}^E_{}D}$ is defined by $F=\omega_L^{-1}({\widetilde{D_{}^\circ}})$, or equivalently ${\widetilde{D_{}^\circ}}=\omega_L(F)$.
Let us denote by $P$ and $Q$ the complementary projectors defined by this decomposition, that is, $${P_a\colon{\mathcal{T}^E_{a}E}\to{\mathcal{T}^E_{a}D}} \quand {Q_a\colon{\mathcal{T}^E_{a}E}\toF_a}, \mbox{
for all }a\in D.$$ Then we have,
\[projection-external\] Let $(L,D)$ be a regular constrained Lagrangian system and let $\Gamma_L$ be the solution of the free dynamics, i.e., $i_{\Gamma_L}\omega_L=dE_L$. Then the solution of the constrained dynamics is the [<span style="font-variant:small-caps;">sode</span>]{} $\Gamma_{(L, D)}$ obtained by projection $\Gamma_{(L, D)}=P(\Gamma_L|_D)$.
Indeed, if we write $\Gamma_{(L, D)}(a) = \Gamma_L(a)-Q(\Gamma_L(a))$ for $a\in D$, then we have $$i_{\Gamma_{(L, D)}(a)}\omega_L-dE_L(a) =
i_{\Gamma_L(a)}\omega_L-i_{Q(\Gamma_L(a))}\omega_L-dE_L(a) =
-i_{Q(\Gamma_L(a))}\omega_L\in{\widetilde{D_{\tau(a)}^\circ}} ,$$ which is an element of ${\widetilde{D_{\tau(a)}^\circ}}$ because $Q(\Gamma_L(a))$ is in $F_a$. Moreover, since $\Gamma_L$ is a [<span style="font-variant:small-caps;">sode</span>]{} and $Q(\Gamma_L)$ is vertical (since it is in $F$), we have that $\Gamma_{(L, D)}$ is also a [<span style="font-variant:small-caps;">sode</span>]{}.
We consider adapted local coordinates $(x^i,y^a,y^A)$ corresponding to the choice of an adapted basis of sections $\{e_a,e_A\},$ where $\{e_a\}$ generate $D$. The annihilator ${D_{}^\circ}$ of $D$ is generated by $\{e^A\}$, and thus ${\widetilde{D_{}^\circ}}$ is generated by $\{{\mathcal{X}}^A\}$. A simple calculation shows that a basis $\{Z_A\}$ of local sections of $F$ is given by $$\label{zetas}
Z_A={\mathcal{V}}_A-Q^a_A{\mathcal{V}}_a,$$ where $Q^a_A = W_{Ab}{\mathcal{C}}^{ab}$ and ${\mathcal{C}}^{ab}$ are the components of the inverse of the matrix ${\mathcal{C}}_{ab}$ given by equation . The local expression of the projector over $F$ is then $$Q = Z_A\otimes {\mathcal{V}}^A.$$
If the expression of the free dynamical section $\Gamma_L$ in this local coordinates is $$\Gamma_L=y^\alpha{\mathcal{X}}_\alpha+f^\alpha{\mathcal{V}}_\alpha,$$ (where $f^\alpha$ are given by equation ), then the expression of the constrained dynamical section is $$\Gamma_{(L, D)} =y^a{\mathcal{X}}_a+(f^a+f^AQ^a_A){\mathcal{V}}_a,$$ where all the functions $f^\alpha$ are evaluated at $y^A=0$.
### Projection to ${\mathcal{T}^D_{}D}$ {#projection-to-mathcaltd_d .unnumbered}
We have seen that the regularity condition for the constrained system $(L,D)$ can be equivalently expressed by requiring that the subbundle ${\mathcal{T}^D_{}D}$ is a symplectic subbundle of $({\mathcal{T}^E_{}E},\omega_L)$. It follows that, for every $a\in D$, we have a direct sum decomposition $${\mathcal{T}^E_{a}E}={\mathcal{T}^D_{a}D}\oplus({\mathcal{T}^D_{a}D})^\perp.$$ Let us denote by $\bar{P}$ and $\bar{Q}$ the complementary projectors defined by this decomposition, that is, $${\bar{P}_a\colon{\mathcal{T}^E_{a}E}\to{\mathcal{T}^D_{a}D}} \qquand
{\bar{Q}_a\colon{\mathcal{T}^E_{a}E}\to({\mathcal{T}^D_{a}D})^\perp},\;\;\mbox{ for all }
a\in D.$$ Then, we have the following result:
\[projection-internal\] Let $(L,D)$ be a regular constrained Lagrangian system and let $\Gamma_L$ be the solution of the free dynamics, i.e., $i_{\Gamma_L}\omega_L=dE_L$. Then the solution of the constrained dynamics is the [<span style="font-variant:small-caps;">sode</span>]{} $\Gamma_{(L, D)}$ obtained by projection $\Gamma_{(L, D)}=\bar{P}(\Gamma_L|_D)$.
From Theorem \[projection-external\] we have that the solution $\Gamma_{(L, D)}$ of the constrained dynamics is related to the free dynamics by $\Gamma_L|_D=\Gamma+Q(\Gamma_L|_D)$. Let us prove that $Q(\Gamma_L|_D)$ takes values in $({\mathcal{T}^D_{}D})^\perp$. Indeed, $Q(\Gamma_L|_D)$ takes values in $F=({\mathcal{T}^D_{}D})^\perp\cap{\operatorname{Ver}({\mathcal{T}^E_{}E})}$, so that, in particular, it takes values in $({\mathcal{T}^D_{}D})^\perp$. Thus, since $\Gamma$ is a section of ${\mathcal{T}^D_{}D}$, it follows that $\Gamma_L|_D=\Gamma_{(L, D)}+Q(\Gamma_L|_D)$ is a decomposition of $\Gamma_L|_D$ according to ${\mathcal{T}^E_{}E}|_D={\mathcal{T}^D_{}D}\oplus({\mathcal{T}^D_{}D})^\perp$, which implies $\Gamma_{(L, D)}=\bar{P}(\Gamma_L|_D)$.
In adapted coordinates, a local basis of sections of $({\mathcal{T}^D_{}D})^\perp$ is $\{Y_A,Z_A\},$ where the sections $Z_A$ are given by and the sections $Y_A$ are $$Y_A = {\mathcal{X}}_A-Q^a_A{\mathcal{X}}_a+{\mathcal{C}}^{bc}(M_{Ab}-M_{ab}Q^a_A){\mathcal{V}}_c ,$$ with $M_{\alpha\beta}=\omega_L({\mathcal{X}}_\alpha,{\mathcal{X}}_\beta)$. Therefore the expression of the projector onto $({\mathcal{T}^D_{}D})^\perp$ is $$\bar{Q}=Z_A\otimes{\mathcal{V}}^A+Y_A\otimes{\mathcal{X}}^A.$$ Note that $S(Y_A) = Z_A$.
The distributional approach
---------------------------
The equations for the Lagrange-d’Alembert section $\Gamma$ can be entirely written in terms of objects in the manifold ${\mathcal{T}^D_{}D}$. Recall that ${\mathcal{T}^D_{}D}$ is not a Lie algebroid. In order to do this, define the 2-section $\omega{^{\scriptscriptstyleL,D}}$ as the restriction of $\omega_L$ to ${\mathcal{T}^D_{}D}$. If $(L,D)$ is regular, then ${\mathcal{T}^D_{}D}$ is a symplectic subbundle of $({\mathcal{T}^E_{}E},\omega_L)$. From this, it follows that $\omega{^{\scriptscriptstyleL,D}}$ is a symplectic section on that bundle. We also define $\varepsilon{^{\scriptscriptstyleL,D}}$ to be the restriction of $dE_L$ to ${\mathcal{T}^D_{}D}$. Then, taking the restriction of Lagrange-d’Alembert equations to ${\mathcal{T}^D_{}D}$, we get the following equation $$\label{LDBS}
i_\Gamma\omega{^{\scriptscriptstyleL,D}}=\varepsilon{^{\scriptscriptstyleL,D}},$$ which uniquely determines the section $\Gamma$. Indeed, the unique solution $\Gamma$ of the above equations is the solution of Lagrange-d’Alembert equations: if we denote by $\lambda$ the constraint force, we have for every $u\in{\mathcal{T}^D_{a}D}$ that $$\omega_L(\Gamma(a),u)-{\left\langledE_L(a),u\right\rangle} =
{\left\langle\lambda(a),{\@ifstar{\@proldf}{\@prolpf}}{\tau}(u)\right\rangle}=0 ,$$ where we have taken into account that ${\@ifstar{\@proldf}{\@prolpf}}{\tau}(u)\in D$ and $\lambda(a)\in{D_{}^\circ}$.
This approach, the so called *distributional approach*, was initiated by Bocharov and Vinogradov (see [@ViKu]) and further developed by Śniatycki and coworkers [@BaSn; @CuSn; @Sn98]. Similar equations, within the framework of Lie algebroids, are the base of the theory proposed in [@MeLa].
[One can also consider the restriction to ${\mathcal{T}^E_{}D}$, which is a Lie algebroid, but no further simplification is achieved by this. If $\bar{\omega}$ is the restriction of $\omega_L$ to ${\mathcal{T}^E_{}D}$ and $\bar{\varepsilon}$ is the restriction of $dE_L$ to ${\mathcal{T}^E_{}D}$, then the Lagrange-d’Alembert equations can be written in the form $i_\Gamma\bar{\omega}-\bar{\varepsilon}=\bar{\lambda}$, where $\bar{\lambda}$ is the restriction of the constraint force to ${\mathcal{T}^E_{}D}$, which, in general, does not vanish. Also notice that the 2-form $\bar{\omega}$ is closed but, in general, degenerated.]{}
The nonholonomic bracket {#NHBracket}
------------------------
Let $f, g$ be two smooth functions on $D$ and take arbitrary extensions to $E$ denoted by the same letters (if there is no possibility of confusion). Suppose that $X_f$ and $X_g$ are the Hamiltonian sections on ${\mathcal{T}^E_{}E}$ given respectively by $$i_{X_f} \, \omega_L = df
\qquand
i_{X_g} \, \omega_L = dg.$$ We define the *nonholonomic bracket* of $f$ and $g$ as follows: $$\label{nhb}
\{f, g\}_{nh} = \omega_L(\bar{P}(X_f), \bar{P}(X_g)).$$ Note that if $f'$ is another extension of $f$, then $(X_f-X_{f'})_{|D}$ is a section of $({\mathcal{T}^D_{}D})^{\perp}$ and, thus, we deduce that does not depend on the chosen extensions. The nonholonomic bracket is an almost-Poisson bracket, i.e., it is skew-symmetric, a derivation in each argument with respect to the usual product of functions and does not satisfy the Jacobi identity.
In addition, one can prove the following formula $$\label{evolution}
\dot{f} = \{f, E_L\}_{nh}.$$ Indeed, we have $$\begin{aligned}
\dot{f} & = d_{\Gamma_{(L, D)}} f = i_{\Gamma_{(L, D)}} df =
i_{\Gamma_{(L, D)}} i_{X_f} \omega_L \\
&= \omega_L(X_f, \Gamma_{(L, D)}) =
\omega_L(X_f, \bar{P}(\Gamma_L)) \\
&= \omega_L(\bar{P}(X_f), \bar{P}(\Gamma_L)) = \{f, E_L\}_{nh}.\end{aligned}$$ Equation implies once more the conservation of the energy (by the skew-symmetric character of the nonholonomic bracket).
Alternatively, since ${\mathcal{T}^D_{}D}$ is an anchored vector bundle, one can take the function $f \in {C^\infty(D)}$ and its differential $\bar{d}f\in{\operatorname{Sec}_{}(({\mathcal{T}^D_{}D})^*)}$. Since $\omega{^{\scriptscriptstyleL,D}}$ is regular, we have a unique section $\bar{X}_f \in {\operatorname{Sec}_{}({\mathcal{T}^D_{}D})}$ defined by $i_{\bar{X}_f}\omega{^{\scriptscriptstyleL,D}} = \bar{d}f$. Then the nonholonomic bracket of two functions $f$ and $g$ is $\{f,g\}_{nh} =
\omega{^{\scriptscriptstyleL,D}}(\bar{X}_f,\bar{X}_g)$. Note that if $\tilde{f}\in
C^\infty(E)$ (resp. $\tilde{g}\in C^\infty(E)$) is an extension to $E$ of $f$ (resp., $g$), then $\bar{X}_f = \bar{P}(X_{\tilde{f}})_{|D}$ (resp., $\bar{X}_g = \bar{P}(X_{\tilde{g}})_{|D}$).
Morphisms and reduction {#sec:reduction}
=======================
One important advantage of dealing with Lagrangian systems evolving on Lie algebroids is that the reduction procedure can be naturally handled by considering morphisms of Lie algebroids, as it was already observed by Weinstein [@Weinstein]. We study in this section the transformation laws of the different geometric objects in our theory and we apply these results to the study of the reduction theory.
\[transformation-S\] Let ${\Phi\colonE\toE'}$ be a morphism of Lie algebroids, and consider the $\Phi$-tangent prolongation of $\Phi$, i.e ${{\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}\colon{\mathcal{T}^E_{}E}\to{\mathcal T}^{E'}E'}$. Let $\xi{{^{\scriptscriptstyleV}}}$ and $\xi'{}{{^{\scriptscriptstyleV}}}$, $S$ and $S'$, and $\Delta$ and $\Delta'$, be the vertical liftings, the vertical endomorphisms, and the Liouville sections on $E$ and $E'$, respectively. Then,
1. ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(\xi{{^{\scriptscriptstyleV}}}(a,b)) =
\xi'{}{{^{\scriptscriptstyleV}}}(\Phi(a),\Phi(b))$, for all $(a,b)\in E\times_M E$,
2. ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}\circ\Delta = \Delta'\circ\Phi$,
3. ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}\circ S = S'\circ {\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}$.
For the first property, we notice that both terms are vertical, so that we just have to show that their action on functions coincide. For every function $f'\in{C^\infty(E')}$, we deduce that $$\begin{aligned}
\rho'^{1}({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(\xi{{^{\scriptscriptstyleV}}}(a,b)))f'
&=T\Phi(\rho^{1}(\xi{{^{\scriptscriptstyleV}}}(a,b)))f' =T\Phi(b{{^{\scriptscriptstyleV}}}_a)f'
=b{{^{\scriptscriptstyleV}}}_a(f'\circ\Phi)\\
&=\frac{d}{dt}f'(\Phi(a+tb)){\Big|_{t=0}}
=\frac{d}{dt}f'(\Phi(a)+t\Phi(b)){\Big|_{t=0}}\\
&=\Phi(b){{^{\scriptscriptstyleV}}}_{\Phi(a)}(f')
=\rho'^{1}(\xi'{}{{^{\scriptscriptstyleV}}}(\Phi(a),\Phi(b)))f'.
\end{aligned}$$ For the second property, we have $\Delta(a)=\xi{{^{\scriptscriptstyleV}}}(a,a)$ so that applying the first property it follows that $${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(\Delta(a))
={\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(\xi{{^{\scriptscriptstyleV}}}(a,a))
=\xi'{}{{^{\scriptscriptstyleV}}}(\Phi(a),\Phi(a))
=\Delta'(\Phi(a)).$$ Finally, for any $z=(a,b,V)\in{\mathcal{T}^E_{}E}$, we obtain that $$\begin{aligned}
{\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(S(z)) &={\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(\xi{{^{\scriptscriptstyleV}}}(a,b))
=\xi'{}{{^{\scriptscriptstyleV}}}(\Phi(a),\Phi(b))\\
&=S'(\Phi(a),\Phi(b),T\Phi(V)) =S'({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(z)),
\end{aligned}$$ which concludes the proof.
\[transformation-omegaL\] Let $L\in{C^\infty(E)}$ be a Lagrangian function, $\theta_L$ the Cartan form and $\omega_L=-d\theta_L$. Let ${\Phi\colonE\toE'}$ be a Lie algebroid morphism and suppose that $L=L'\circ \Phi$, with $L'\in
{C^\infty(E')}$ a Lagrangian function. Then, we have
1. $({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}){^\star}\theta_{L'}=\theta_L$,
2. $({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}){^\star}\omega_{L'}=\omega_L$,
3. $({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}){^\star}E_{L'}=E_L$,
4. $G{^{\scriptscriptstyleL'}}_{\Phi(a)}(\Phi(b),\Phi(c))={G{^{\scriptscriptstyleL}}_{a}}(b,c)$, for every $a\in E$ and every $b,c\in E_{\tau(a)}$.
Indeed, for every $Z\in{\mathcal{T}^E_{}E}$ we have $$\begin{aligned}
{\left\langle({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}){^\star}\theta_{L'},Z\right\rangle}
&={\left\langle\theta_{L'},{\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(Z)\right\rangle}
={\left\langledL',S'({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(Z))\right\rangle}
={\left\langledL',{\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(S(Z))\right\rangle}\\
&={\left\langle({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}){^\star}dL',S(Z)\right\rangle}
={\left\langled({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}){^\star}L',S(Z)\right\rangle}
={\left\langled(L'\circ\Phi),S(Z)\right\rangle}\\
&={\left\langledL,S(Z)\right\rangle} ={\left\langle\theta_L,Z\right\rangle},
\end{aligned}$$ where we have used the transformation rule for the vertical endomorphism. The second property follows from the fact that ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}$ is a morphism, so that $({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}){^\star}d=d({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}){^\star}$. The third one follows similarly and the fourth is a consequence of the second property and the definitions of the tensors ${G{^{\scriptscriptstyleL}}_{}}$ and $G{^{\scriptscriptstyleL'}}$.
Let $\Gamma$ be a [<span style="font-variant:small-caps;">sode</span>]{} and $L\in{C^\infty(E)}$ be a Lagrangian. For convenience, we define the 1-form $\delta_\Gamma L\in{\operatorname{Sec}_{}(({\mathcal{T}^E_{}E})^*)}$ by $${\left\langle\delta_\Gamma L,Z\right\rangle}={\left\langledE_L-i_\Gamma\omega_L,Z\right\rangle}
={\left\langledE_L,Z\right\rangle}-\omega_L(\Gamma,Z),$$ for every section $Z$ of ${\mathcal{T}^E_{}E}$. We notice that $\Gamma$ is the solution of the free dynamics if and only if $\delta_\Gamma L=0$. On the other hand, notice that the 1-form $\delta_\Gamma L$ is semibasic, because $\Gamma$ is a [<span style="font-variant:small-caps;">sode</span>]{}.
\[transformation-deltaL\] Let $\Gamma$ be a [<span style="font-variant:small-caps;">sode</span>]{} in $E$ and $\Gamma'$ a [<span style="font-variant:small-caps;">sode</span>]{} in $E'$. Let $L\in{C^\infty(E)}$ and $L'\in{C^\infty(E')}$ be Lagrangian functions defined on $E$ and $E'$, respectively, such that $L=L'\circ\Phi$. Then, $$\label{Gamma-relation}
{\left\langle\delta_\Gamma L-({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}){^\star}\delta_{\Gamma'} L',Z\right\rangle}
=\omega_{L'}(\Gamma'\circ\Phi-{\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi} \circ
\Gamma,{\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(Z)),$$ for every section $Z$ of ${\mathcal{T}^E_{}E}$.
Indeed, from $({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}){^\star}dE_{L'}=d E_L$, we have that $$\begin{aligned}
{\left\langle\delta_\Gamma L-({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}){^\star}\delta_{\Gamma'} L',Z\right\rangle}
&={\left\langle({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}){^\star}i_{\Gamma'}\omega_{L'}-i_\Gamma\omega_L,Z\right\rangle}\\
&={\left\langle({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}){^\star}i_{\Gamma'}\omega_{L'} -
i_\Gamma({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}){^\star}\omega_{L'},Z\right\rangle}\\
& = \omega_{L'}(\Gamma'\circ\Phi -
{\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}\circ\Gamma,{\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(Z)),
\end{aligned}$$ which concludes the proof.
Reduction of the free dynamics
------------------------------
Here, we build on Propositions \[transformation-omegaL\] and \[transformation-deltaL\] to identify conditions under which the dynamics can be reduced under a morphism of Lie algebroids. We first notice that, from Proposition \[transformation-omegaL\], if $\Phi$ is fiberwise surjective morphism and $L$ is a regular Lagrangian on $E$, then $L'$ is a regular Lagrangian on $E'$ (note that ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}:{\mathcal T}^EE\to {\mathcal T}^{E'}{E'}$ is a fiberwise surjective morphism). Thus, the dynamics of both systems is uniquely defined.
\[reduction-free-dynamics\] Suppose that the Lagrangian functions $L$ and $L'$ are $\Phi$-related, that is, $L = L' \circ \Phi$. If $\Phi$ is a fiberwise surjective morphism and $L$ is a regular Lagrangian then $L'$ is also a regular Lagrangian. Moreover, if $\Gamma_{L}$ and $\Gamma_{L'}$ are the solutions of the free dynamics defined by $L$ and $L'$ then $${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}\circ\Gamma_L=\Gamma_{L'}\circ\Phi.$$ Therefore, if $a(t)$ is a solution of the free dynamics defined by $L$, then $\Phi(a(t))$ is a solution of the free dynamics defined by $L'$.
If $\Gamma_L$ and $\Gamma_{L'}$ are the solutions of the dynamics, then $\delta_{\Gamma_L} L=0$ and $\delta_{\Gamma_{L'}}L'=0$ so that the left-hand side in equation vanishes. Thus $$\omega_{L'}(\Gamma_{L'}\circ\Phi-{\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi} \circ
\Gamma_L,{\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(Z)) = 0,$$ for every $Z\in{\operatorname{Sec}_{}({\mathcal{T}^E_{}E})}$. Therefore, using that $L'$ is regular and the fact that ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}$ is a fiberwise surjective morphism, we conclude the result.
We will say that the unconstrained dynamics $\Gamma_{L'}$ is the *reduction of the unconstrained dynamics* $\Gamma_L$ by the morphism $\Phi$.
Reduction of the constrained dynamics
-------------------------------------
The above results about reduction of unconstrained Lagrangian systems can be easily generalized to nonholonomic constrained Lagrangian systems whenever the constraints of one system are mapped by the morphism to the constraints of the second system. Let us elaborate on this.
Let $(L,D)$ be a constrained Lagrangian system on a Lie algebroid $E$ and let $(L',D')$ be another constrained Lagrangian system on a second Lie algebroid $E'$. Along this section, we assume that there is a fiberwise surjective morphism of Lie algebroids ${\Phi\colonE\toE'}$ such that $L=L'\circ\Phi$ and $\Phi(D)=D'$. The latter condition implies that the base map is also surjective, so that we will assume that $\Phi$ is an epimorphism (i.e., in addition to being fiberwise surjective, the base map $\varphi$ is a submersion).
As a first consequence, we have $G{^{\scriptscriptstyleL',D'}}_{\Phi(a)}
(\Phi(b),\Phi(c)) = {G{^{\scriptscriptstyleL,D}}_{a}}(b,c)$, for every $a\in D$ and every $b,c\in D_{\pi(a)}$, and therefore, if $(L,D)$ is regular, then so is $(L',D')$.
\[l5.5\] With respect to the decompositions ${\mathcal{T}^E_{}E}|_D={\mathcal{T}^E_{}D}\oplus F$ and ${\@ifstar{\@proldf}{\@prolpf}}[E']{E'}|_{D'}={\@ifstar{\@proldf}{\@prolpf}}[E']{D'}\oplus F'$, we have the following properties:
1. ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}({\mathcal{T}^E_{}D})={\@ifstar{\@proldf}{\@prolpf}}[E']{D'}$,
2. ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(F)=F'$,
3. If $P,Q$ and $P',Q'$ are the projectors associated with $(L,D)$ and $(L',D'),$ respectively, then $P'\circ{\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}={\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}\circ P$ and $Q'\circ{\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}={\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}\circ Q$.
With respect to the decompositions ${\mathcal{T}^E_{}E}|_D={\mathcal{T}^D_{}D}\oplus({\mathcal{T}^D_{}D})^\perp$ and ${\@ifstar{\@proldf}{\@prolpf}}[E']{E'}|_{D'}={\@ifstar{\@proldf}{\@prolpf}}[D']{D'}\oplus({\@ifstar{\@proldf}{\@prolpf}}[D']{D'})^\perp$ we have the following properties:
4. ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}({\mathcal{T}^D_{}D})={\@ifstar{\@proldf}{\@prolpf}}[D']{D'}$,
5. ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}\bigl(({\mathcal{T}^D_{}D})^\perp\bigr) =
({\@ifstar{\@proldf}{\@prolpf}}[D']{D'})^\perp$,
6. If $\bar{P},\bar{Q}$ and $\bar{P}',\bar{Q}'$ are the projectors associated with $(L,D)$ and $(L',D'),$ respectively, then $\bar{P}'\circ{\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}={\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}\circ\bar{P}$ and $\bar{Q}'\circ{\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}={\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}\circ\bar{Q}$.
From the definition of ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}$, it follows that $$({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi})({\mathcal T}^ED)\subseteq {\mathcal
T}^{E'}D',\;\;\;\;({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi})({\mathcal T}^DD)\subseteq
{\mathcal T}^{D'}D'.$$ Thus, one may consider the vector bundle morphisms $${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}:{\mathcal T}^ED\to {\mathcal T}^{E'}D',\;\;\;
{\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}:{\mathcal T}^DD\to {\mathcal T}^{D'}D'.$$ Moreover, using that $\Phi$ is fiberwise surjective and that $\varphi$ is a submersion, we deduce that the rank of the above morphisms is maximum. This proves (1) and (4).
The proof of (5) is as follows. For every $a'\in D'$, one can choose $a\in D$ such that $\Phi(a)=a'$, and one can write any element $w'\in{\@ifstar{\@proldf}{\@prolpf}}[D']{D'}[a']$ as $w'={\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(w)$ for some $w\in{\@ifstar{\@proldf}{\@prolpf}}[D]{D}[a]$. Thus, if $z\in({\mathcal{T}^D_{a}D})^\perp$, for every $w'\in{\@ifstar{\@proldf}{\@prolpf}}[D']{D'}[a']$ we have $$\omega_{L'}({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(z),w')
=\omega_{L'}({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(z),{\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(w))
=\omega_L(z,w)=0 ,$$ from where it follows that ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(z)\in({\@ifstar{\@proldf}{\@prolpf}}[D']{D'})^\perp$. In a similar way, using that ${\mathcal T}^\Phi\Phi:({\@ifstar{\@proldf}{\@prolpf}}[E]{E})_{|D}\to
({\@ifstar{\@proldf}{\@prolpf}}[E']{E'})_{|D'}$ is fiberwise surjective, (2) in Proposition \[transformation-omegaL\] and (4), we obtain that $({\@ifstar{\@proldf}{\@prolpf}}[D']{D'})^\perp\subseteq
({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi})(({\@ifstar{\@proldf}{\@prolpf}}[D]{D})^\perp)$.
For the proof of (2) we have that $${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(F) ={\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(({\mathcal{T}^D_{}D})^\perp\cap{\operatorname{Ver}({\mathcal{T}^E_{}E})})
\subseteq ({\@ifstar{\@proldf}{\@prolpf}}[D']{D'})^\perp\cap{\operatorname{Ver}({\@ifstar{\@proldf}{\@prolpf}}[E']{E'})} =F'.$$ Thus, using that ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}: ({\@ifstar{\@proldf}{\@prolpf}}[E]{E})_{|D}\to
({\@ifstar{\@proldf}{\@prolpf}}[E']{E'})_{|D'}$ is fiberwise surjective, the fact that $({\@ifstar{\@proldf}{\@prolpf}}[E]{E})_{|D}={\@ifstar{\@proldf}{\@prolpf}}[E]{D}\oplus F$ and (1), it follows that $$({\@ifstar{\@proldf}{\@prolpf}}[E']{E'})_{|D'}={\@ifstar{\@proldf}{\@prolpf}}[E']{D'} \oplus ({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi})(F).$$ Therefore, since $({\@ifstar{\@proldf}{\@prolpf}}[E']{E'})_{|D'}={\@ifstar{\@proldf}{\@prolpf}}[E']{D'}\oplus F'$, we conclude that (2) holds.
Finally, (3) is an immediate consequence of (1) and (2), and similarly, (6) is an immediate consequence of (4) and (5).
From the properties above, we get the following result.
\[t5.6\] Let $(L,D)$ be a regular constrained Lagrangian system on a Lie algebroid $E$ and let $(L',D')$ be a constrained Lagrangian system on a second Lie algebroid $E'$. Assume that a fiberwise surjective morphism of Lie algebroids ${\Phi\colonE\toE'}$ exists such that $L=L'\circ\Phi$ and $\Phi(D)=D'$. If $\Gamma_{(L, D)}$ is the constrained dynamics for $L$ and $\Gamma_{(L', D')}$ is the constrained dynamics for $L'$, then ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}\circ\Gamma_{(L, D)}=\Gamma_{(L', D')} \circ\Phi$. If $a(t)$ is a solution of Lagrange-d’Alembert differential equations for $L$, then $\Phi(a(t))$ is a solution of Lagrange-d’Alembert differential equations for $L'$.
For the free dynamics, we have that ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}\circ\Gamma_L=\Gamma_{L'}\circ\Phi$. Moreover, from property (3) in Lemma \[l5.5\], for every $a\in D$, we have that $$\begin{array}{l}
{\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(\Gamma_{(L, D)}(a)) ={\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(P(\Gamma_L(a)))
=P'({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(\Gamma_L(a))) \\[5pt] =P'(\Gamma_{L'}(\Phi(a)))
=\Gamma_{(L', D')}(\Phi(a)),
\end{array}$$ which concludes the proof.
We will say that the constrained dynamics $\Gamma_{(L', D')}$ is the *reduction of the constrained dynamics* $\Gamma_{(L, D)}$ by the morphism $\Phi$.
\[t5.7\] Under the same hypotheses as in Theorem \[t5.6\], we have that $$\{f'\circ \Phi,g'\circ \Phi\}_{nh}=\{f',g'\}'_{nh}\circ \Phi,$$ for $f',g'\in C^\infty(D'),$ where $\{\cdot,\cdot\}_{nh}$ (respectively, $\{\cdot,\cdot\}_{nh}')$ is the nonholonomic bracket for the constrained system $(L,D)$ (respectively, $(L',D')$). In other words, $\Phi:D\to D'$ is an almost-Poisson morphism.
Using (2) in Proposition \[transformation-omegaL\] and the fact that $\Phi$ is a Lie algebroid morphism, we deduce that $$(i_{X_{f'\circ
\Phi}}({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi})^*\omega_{L'})=i_{X_{f'}}\omega_{L'}\circ
\Phi.$$ Thus, since ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}$ is fiberwise surjective, we obtain that $${\@ifstar{\@proldf}{\@prolpf}}^\Phi\Phi\circ X_{f'\circ \Phi}=X_{f'}\circ \Phi.$$ Now, from (\[nhb\]) and Lemma \[l5.5\], we conclude that $$\{f'\circ \Phi,g'\circ \Phi\}_{nh}=\{f',g'\}_{nh}'\circ \Phi.$$
One of the most important cases in the theory of reduction is the case of reduction by a symmetry group. In this respect, we have the following result.
\[quotient-Lie-algebroid\] Let ${q^Q_G\colonQ\toM}$ be a principal $G$-bundle, let ${\tau\colonE\toQ}$ be a Lie algebroid, and assume that we have an action of $G$ on $E$ such that the quotient vector bundle $E/G$ is well-defined. If the set ${\operatorname{Sec}_{}(E)}^G$ of equivariant sections of $E$ is a Lie subalgebra of ${\operatorname{Sec}_{}(E)}$, then the quotient $E'=E/G$ has a canonical Lie algebroid structure over $M$ such that the canonical projection ${q^E_G\colonE\toE/G}$, given by $a\mapsto[a]_G$, is a (fiberwise bijective) Lie algebroid morphism over $q^Q_G$.
As a concrete example of application of the above theorem, we have the well-known case of the Atiyah or Gauge algebroid. In this case, the Lie algebroid $E$ is the standard Lie algebroid $TQ\to Q$, the action is by tangent maps $gv\equiv T\psi_g(v)$, the reduction is the Atiyah Lie algebroid $TQ/G\to Q/G$ and the quotient map ${q^{TQ}_G\colonTQ\toTQ/G}$ is a Lie algebroid epimorphism. It follows that if $L$ is a $G$-invariant regular Lagrangian on $TQ$ then the unconstrained dynamics for $L$ projects to the unconstrained dynamics for the reduced Lagrangian $L'$. Moreover, if the constraints $D$ are also $G$-invariant, then the constrained dynamics for $(L,D)$ reduces to the constrained dynamics for $(L',D/G)$.
On a final note, we mention that the pullback of the distributional equation $i_{\Gamma'}\omega{^{\scriptscriptstyleL',D'}}-\varepsilon{^{\scriptscriptstyleL',D'}}=0$ by ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}$ is precisely $(i_\Gamma\omega{^{\scriptscriptstyleL,D}}-\varepsilon{^{\scriptscriptstyleL,D}})\circ{\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}=0$.
Reduction by stages
-------------------
As a direct consequence of the results exposed above, one can obtain a theory of reduction by stages. In Poisson geometry, reduction by stages is a straightforward procedure. Given the fact that the Lagrangian counterpart of Poisson reduction is Lagrangian reduction, it is not strange that reduction by stages in our framework becomes also straightforward.
The Lagrangian theory of reduction by stages is a consequence of the following basic observation:
> Let ${\Phi_1\colonE_0\toE_1}$ and ${\Phi_2\colonE_1\toE_2}$ be a fiberwise surjective morphisms of Lie algebroids and let ${\Phi\colonE_0\toE_2}$ be the composition $\Phi=\Phi_2\circ\Phi_1$. The reduction of a Lagrangian system in $E_0$ by $\Phi$ can be obtained by first reducing by $\Phi_1$ and then reducing the resulting Lagrangian system by $\Phi_2$.
This result follows using that ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi} =
{\@ifstar{\@proldf}{\@prolpf}}[\Phi_2]{\Phi_2} \circ {\@ifstar{\@proldf}{\@prolpf}}[\Phi_1]{\Phi_1}$. Based on this fact, one can analyze one of the most interesting cases of reduction: the reduction by the action of a symmetry group. We consider a group $G$ acting on a manifold $Q$ and a closed normal subgroup $N$ of $G$. The process of reduction by stages is illustrated in the following diagram $$\xymatrix{ \ar@/_{15pt}/[dd]^{\ \cdot/G}_{\text{Total reduction }}
&{\tau_Q\colonE_0=TQ\toM_0=Q}\ar[d]^{\ \cdot/N}_{\text{First
reduction }}\\
&{\tau_1\colonE_1=TQ/N\toM_1=Q/N}\ar[d]^{\ \cdot/(G/N)}_{\text{Second
reduction }}\\
&{\tau_2\colonE_2=(TQ/N)/(G/N)\toM_2=(Q/N)/(G/N)} }$$
In order to prove our results about reduction by stages, we have to prove that $E_0, E_1$ and $E_2$ are Lie algebroids, that the quotient maps ${\Phi_1\colonE_0\toE_1}$, ${\Phi_2\colonE_1\toE_2}$ and ${\Phi\colonE_0\toE_2}$ are Lie algebroids morphisms and that the composition $\Phi_1\circ\Phi_2$ equals to $\Phi$. Our proof is based on the following well-known result (see [@CeMaRa]), which contains most of the ingredients in the theory of reduction by stages.
([@CeMaRa])\[principal-reduction\] Let ${q^Q_G\colonQ\toM}$ be a principal $G$-bundle and $N$ a closed normal subgroup of $G$. Then,
1. ${q^Q_N\colonQ\toQ/N}$ is a principal $N$-bundle,
2. $G/N$ acts on $Q/N$ by the rule $[g]_N[q]_N=[gq]_N$,
3. ${q^{Q/N}_{G/N}\colonQ/N\to(Q/N)/(G/N)}$ is a principal $(G/N)$-bundle.
4. The map ${i\colonQ/G\to(Q/N)/(G/N)}$ defined by $[q]_G\mapsto[[q]_N]_{G/N}$ is a diffeomorphism.
Building on the previous results, one can deduce the following theorem, which states that the reduction of a Lie algebroid can be done by stages.
Let ${q^Q_G\colonQ\toM}$ be a principal $G$-bundle and $N$ be a closed normal subgroup of $G$. Then,
1. ${\tau_{TQ/G}\colonTQ/G\toQ/G}$ is a Lie algebroid and ${q^{TQ}_G\colonTQ\toTQ/G}$ is a Lie algebroid epimorphism,
2. ${\tau_{TQ/N}\colonTQ/N\toQ/N}$ is a Lie algebroid and ${q^{TQ}_N\colonTQ\toTQ/N}$ is a Lie algebroid epimorphism,
3. $G/N$ acts on $TQ/N$ by the rule $[g]_N[v]_N=[gv]_N$,
4. ${\tau_{(TQ/N)/(G/N)}\colon(TQ/N)/(G/N)\to(Q/N)/(G/N)}$ is a Lie algebroid and ${q^{TQ/N}_{G/N}\colonTQ/N\to(TQ/N)/(G/N)}$ is a Lie algebroid epimorphism,
5. The map ${I\colonTQ/G\to(TQ/N)/(G/N)}$ defined by $[v]_G\mapsto[[v]_N]_{G/N}$ is an isomorphism of Lie algebroids over the map $i$.
The vector bundle $\tau_{TQ/G}:TQ/G\to Q/G$ (respectively, $\tau_{TQ/N}:TQ/N\to Q/N)$ is the Atiyah algebroid for the principal $G$-bundle $q_G^Q:Q\to Q/G$ (respectively, $q_N^Q:Q\to Q/N$), so that (1) and (2) are obvious. Condition (3) is just condition (2) of Theorem \[principal-reduction\] applied to the principal $N$-bundle $TQ\to TQ/N$. To prove condition (4), we notice that the action of $G/N$ on the Lie algebroid $TQ/N$ is free and satisfies the conditions of Theorem \[quotient-Lie-algebroid\]. Finally, the Lie algebroid morphism ${j\colonTQ\toTQ/N}$ is equivariant with respect to the $G$-action on $TQ$ and the $(G/N)$-action on $TQ/N$. Thus, it induces a morphism of Lie algebroids in the quotient. It is an isomorphism since it is a diffeomorphism by Theorem \[principal-reduction\].
The following diagram illustrates the above situation: $$\xymatrix{ TQ\ar@/^{20pt}/[rr]^{\Phi} \ar[d]_{\tau_Q}\ar[r]_{\kern-10pt\Phi_1}
&
TQ/N\ar[d]^{\tau_1}\ar[r]_{\kern-20pt \Phi_2} & (TQ/N)/(G/N)\ar[d]^{\tau_2}\\
Q\ar@/_{20pt}/[rr]_{q^Q_G}\ar[r]^{q_N^Q} & Q/N\ar[r]^{\kern-15pt
q_{G/N}^{Q/N}}&(Q/N)/(G/N) }$$
In particular, for the unconstrained case one has the following result.
Let ${q^Q_G\colonQ\toQ/G}$ be a principal $G$-bundle, and $N$ a closed normal subgroup of $G$. Let $L$ be a Lagrangian function on $Q$ which is $G$-invariant. Then the reduction by the symmetry group $G$ can be performed in two stages:
- reduce by the normal subgroup $N$,
- reduce the resulting dynamics from 1. by the residual symmetry group $G/N$.
Since the dynamics of a constrained system is obtained by projection of the free dynamics, we also the following result.
Let ${q^Q_G\colonQ\toQ/G}$ be a principal $G$-bundle and $N$ a closed normal subgroup of $G$. Let $(L,D)$ be a $G$-invariant constrained Lagrangian system. Then the reduction by the symmetry group $G$ can be performed in two stages:
- reduce by the normal subgroup $N$,
- reduce the resulting dynamics from 1. by the residual symmetry group $G/N$.
The momentum equation {#momentum-equation}
=====================
In this section, we introduce the momentum map for a constrained system on a Lie algebroid, and examine its evolution along the dynamics. This gives rise to the so-called momentum equation.
Unconstrained case
------------------
Let us start by discussing the unconstrained case. Let $\tau_E:E\to
M$ be a Lie algebroid over a manifold $M$ and $L:E\to \R$ be a regular Lagrangian function. Suppose that $\tau_K:K\to M$ is a vector bundle over $M$ and that $\Psi:K\to E$ is a vector bundle morphism (over the identity of $M$) between $K$ and $E$. Then, we can define *the unconstrained momentum map* $J_{(L,\Psi)}:E\to K^*$ *associated with $L$ and $\Psi$* as follows $$J_{(L,\Psi)}(a)\in K_x^*, \mbox{ for } a\in E_x,$$ and $$(J_{(L,\Psi)}(a))(k)=\frac{d}{dt}_{|t=0}L(a+t\Psi(k))=\Psi(k)_a{{^{\scriptscriptstyleV}}}(L),\mbox{
for }k\in K_x.$$ If $\sigma:M\to K$ is a section of $\tau_K:K\to M$ then, using the momentum map $J_{(L,\Psi)}$, we may introduce the real function $J_{(L,\Psi)}^\sigma:E\to \R$ given by $$\label{Momentum1}
J_{(L,\Psi)}^\sigma(a) =
J_{(L,\Psi)}(a)(\sigma(x))=\Psi(\sigma(x))_a{{^{\scriptscriptstyleV}}}(L),\mbox{
for }a\in E_x.$$
\[theorem6.1\] Let $\Gamma_L$ be the Euler-Lagrange section associated with the regular Lagrangian function $L:E\to \R$. If $\sigma:M\to K$ is a section of $\tau_K:K\to M$ and $(\Psi\circ \sigma)^c\in
Sec({\@ifstar{\@proldf}{\@prolpf}}[E]{E})$ is the complete lift of $(\Psi\circ \sigma)\in
Sec(E)$, we have that $$\label{EqMomen1}
<d^{{\@ifstar{\@proldf}{\@prolpf}}[E]{E}}J_{(L,\Psi)}^\sigma,\Gamma_L>=<d^{{\@ifstar{\@proldf}{\@prolpf}}[E]{E}}L,
(\Psi\circ \sigma)^c>,$$ where $d^{{\mathcal T}^EE}$ is the differential of Lie algebroid ${{\@ifstar{\@proldf}{\@prolpf}}[E]{E}}\to E.$ In particular, if $<d^{{\@ifstar{\@proldf}{\@prolpf}}[E]{E}}L,(\Psi\circ \sigma)^c>=0$, then the real function $J_{(L,\Psi)}^\sigma$ is a constant of the motion for the Lagrangian dynamics associated with the Lagrangian function $L.$
Let $S:{{\@ifstar{\@proldf}{\@prolpf}}[E]{E}}\to {{\@ifstar{\@proldf}{\@prolpf}}[E]{E}}$ be the vertical endomorphism. If $(\Psi\circ \sigma)^v\in Sec({\@ifstar{\@proldf}{\@prolpf}}[E]{E})$ is the vertical lift of $(\Psi\circ \sigma)\in Sec(E)$ then, using (\[Momentum1\]) and the fact that $S(\Psi\circ \sigma)^c=(\Psi\circ \sigma)^v,$ it follows that $$\label{Momen2}
J_{(L,\Psi)}^\sigma=\theta_L((\Psi\circ \sigma)^c),$$ where $\theta_L$ is the Cartan $1$-form associated with $L$.
Thus, from (\[Momen2\]), we deduce that $$d^{{\@ifstar{\@proldf}{\@prolpf}}[E]{E}}J_{(L,\Psi)}^\sigma={\mathcal L}_{(\Psi\circ
\sigma)^c}^{{\@ifstar{\@proldf}{\@prolpf}}[E]{E}}\theta_L + i(\Psi\circ
\sigma)^c(\omega_L),$$ $\omega_L$ being the Cartan $2$-form associated with $L$.
Therefore, if $E_L:E\to \R$ is the Lagrangian energy, we obtain that $$\label{Formula}
\begin{array}{rcl}
<d^{{\@ifstar{\@proldf}{\@prolpf}}[E]{E}}J^\sigma_{(L,\Psi)},\Gamma_L> & = &
<d^{{\@ifstar{\@proldf}{\@prolpf}}[E]{E}}(\theta_L(\Gamma_L)),
(\Psi\circ \sigma)^c>-<d^{{\@ifstar{\@proldf}{\@prolpf}}[E]{E}}E_L,(\Psi\circ \sigma)^c>\\[8pt]
&&-<\theta_L,[(\Psi\circ
\sigma)^c,\Gamma_L]>.
\end{array}$$ Now, from (\[SODEcompl\]) and since $\Gamma_L$ is a [<span style="font-variant:small-caps;">sode</span>]{} section, it follows that $$\Theta_L(\Gamma_L)=<d^{{\@ifstar{\@proldf}{\@prolpf}}[E]{E}}L,\Delta>,\;\;\;\;
<\theta_L, [(\Psi\circ \sigma)^c,\Gamma_L]> =0,$$ where $\Delta\in Sec({{\@ifstar{\@proldf}{\@prolpf}}[E]{E}})$ is the Liouville section. Consequently, using (\[Formula\]) we deduce that (\[EqMomen1\]) holds.
\[remark6.2\]
Let $L:TM\to \R$ be an standard regular Lagrangian function on $TM.$ Suppose that $G$ is a Lie group with Lie algebra ${\frak g}$ and that $\psi:G\times M\to M$ is a (left) action of $G$ on $M.$ Then, we may consider the trivial vector bundle over $M$ $$K=M\times {\frak g}\to M$$ and the vector bundle morphism $\Psi:K\to TM$ (over the identity of $M$) defined by $$\label{Psi}
\Psi(x,\xi)=\xi_M(x),$$ where $\xi_M\in {\frak X}(M)$ is the infinitesimal generator of the action $\psi$ associated with $\xi\in {\frak g}$.
A direct computation proves that the (unconstrained) momentum map $J_{L,\Psi)}:E=TM\to K^*=M\times {\frak g}^*$ associated with $L$ and $\Psi$ is given by $$J_{(L,\Psi)}(v_x)=(x,J(v_x)),\mbox{ for }
v_x\in T_xM,$$ where $J:TM\to {\frak g}^*$ is the standard momentum map associated with $L$ and the action $\psi$ defined by $$J(v_x)(\xi)=\frac{d}{dt}_{|t=0}L(v_x+t\xi_M(x)), \mbox{ for }
v_x\in T_xM \mbox{ and } \xi\in {\frak g}$$ (see, for instance, [@AM]).
Now, each $\xi\in {\frak g}$ defines a (constant) section $\sigma$ of the vector bundle $K=M\times {\frak g}\to M$ and the real function $J_{(L,\Psi)}^\sigma$ is just the momentum $J_\xi:TM\to
\R$ in the direction of $\xi$.
On the other hand, if $\eta\in {\frak g}$, then the infinitesimal generator $\eta_{TM}$ of the tangent action $T\psi:G\times TM\to
TM$ associated with $\eta$ is the (standard) complete lift $\eta_M^c\in {\frak X}(TM)$ of $\eta_M.$ Therefore, using Theorem \[theorem6.1\], we deduce a well-known result [@AM]: “If the Lagrangian function $L:TM\to \R$ is invariant under the tangent action $T\psi$ of $G$ on $TM$ then, for every $\xi\in
{\frak g}$, the momentum $J_\xi:TM\to \R$ in the direction of $\xi$ is a constant of the motion of the Lagrangian dynamics.”
Constrained case
----------------
Next, let us discuss the constrained case. Suppose that $L:E\to \R$ is a regular Lagrangian function on a Lie algebroid $\tau_E:E\to M$, that $\tau_K:K\to M$ is a vector bundle over $M$ and that $\Psi:K\to
E$ is a vector bundle morphism (over the identity of $M$) between $K$ and $E$.
In addition, let $\tau_D:D\to M$ be a vector subbundle of $\tau_E:E\to
M$ such that the nonholonomic Lagrangian system $(L,D)$ is regular.
If $x$ is point of $M$ we consider the vector subspace $K_x^D$ of $K_x$ given by $$K_x^D=\{k\in K_x/\Psi(k)\in D_x\}.$$ We will denote by $i_x:K_x^D\to K_x$ the canonical inclusion, by $i_x^*:K_x^*\to (K_x^D)^*$ the canonical projection and by $K^D$ and $(K^D)^*$ the sets $$K^D=\bigcup_{x\in M}K_x^D,\;\;\;\; (K^D)^*=\bigcup_{x\in
M}(K_x^D)^*.$$ Then, we define the *nonholonomic momentum map $J_{(L,D,\Psi)}:E\to (K^D)^*$ associated with the system $(L,D)$ and the morphism $\Psi$* as follows $$(J_{(L,D,\Psi)})_{|E_x}=i_x^*\circ (J_{(L,\Psi)})_{|E_x},\;\;\;
\mbox{ for } x\in M.$$ Now, if $\sigma:M\to K$ is a section of $\tau_{K}:K\to M$ such that $\sigma(x)\in K_x^D$, for all $x\in M,$ we may introduce the real function $J^\sigma_{(L,D,\Psi)}:E\to \R$ given by $$J^\sigma_{(L,D,\Psi)}(a)=J_{(L,D,\Psi)}(a)(\sigma(x)),\mbox{ for }
a\in E_x,$$ that is, $J_{(L,D,\Psi)}^\sigma=J_{(L,\Psi)}^\sigma.$
\[theorem6.3\] Let $\Gamma_{(L,D)}$ be the solution of the constrained dynamics for the nonholonomic Lagrangian system $(L,D)$. If $\sigma:M\to K$ is a section of $\tau_K:K\to M$ such that $\sigma(x)\in K_x^D$, for all $x\in M$, and $(\Psi\circ \sigma)^c\in Sec({{\@ifstar{\@proldf}{\@prolpf}}[E]{E}})$ is the complete lift of $(\Psi\circ \sigma)\in Sec(E)$ then we have that $$\label{EqMomen2}
<d^{{\@ifstar{\@proldf}{\@prolpf}}[E]{D}}((J_{(L,D,\Psi)})_{|D}),\Gamma_{(L,D)}> =
<d^{{\@ifstar{\@proldf}{\@prolpf}}[E]{E}}L,(\Psi\circ \sigma)^c>_{|D},$$ where $d^{{\@ifstar{\@proldf}{\@prolpf}}[E]{D}}$ (respectively, $d^{{\@ifstar{\@proldf}{\@prolpf}}[E]{E}}$) is the differential of Lie algebroid ${{\@ifstar{\@proldf}{\@prolpf}}[E]{D}}\to D$ (respectively, ${{\@ifstar{\@proldf}{\@prolpf}}[E]{E}}\to E$). In particular, if $<d^{{\@ifstar{\@proldf}{\@prolpf}}[E]{E}}L,(\Psi\circ \sigma)^c>_{|D}=0$, then the real function $J_{(L,D,\Psi)}^\sigma$ is a constant of the motion for the constrained dynamics associated with the nonholonomic Lagrangian system $(L,D)$.
Denote by $j:D\to E$ and by ${\mathcal J}:{{\@ifstar{\@proldf}{\@prolpf}}[E]{D}}\to {{\@ifstar{\@proldf}{\@prolpf}}[E]{E}}$ the canonical inclusions and by $Q:{{\mathcal T}_D^EE}\to F$ the corresponding projector, where $F=\omega_L^{-1}(\widetilde{D}^0)$ (see Section \[Projectors\]). Then, as we know, $$\Gamma_{(L,D)}=(\Gamma_L-Q\Gamma_L)_{|D}.$$ Moreover, the pair $({\mathcal J},j)$ is a Lie algebroid monomorphism which implies that $$d^{{\@ifstar{\@proldf}{\@prolpf}}[E]{D}}((J_{(L,D,\Psi)}^\sigma)_{|D})=({\mathcal
J},j)^*(d^{{\@ifstar{\@proldf}{\@prolpf}}[E]{E}}J^\sigma_{(L,D,\Psi)}).$$ Thus, using that $J_{(L,D,\Psi)}^\sigma=J^\sigma_{(L,\Psi)}$ and proceedings as in the proof of Theorem \[theorem6.1\], we deduce that $$\label{step1}
\begin{array}{rcl}
<d^{{\@ifstar{\@proldf}{\@prolpf}}[E]{D}}((J_{(L,D,\Psi)}^\sigma)_{|D}),\Gamma_{(L,D)}> &
= &<d^{{\@ifstar{\@proldf}{\@prolpf}}[E]{E}}L,(\Psi\circ
\sigma)^c>_{|D}\\&&\kern-80pt-\{({\mathcal L}_{(\Psi\circ
\sigma)^c}^{{\@ifstar{\@proldf}{\@prolpf}}[E]{E}}\theta_L)(Q\Gamma_L) + (i(\Psi\circ
\sigma)^c(\omega_L))(Q\Gamma_L)\}_{|D}.
\end{array}$$ Now, since $S(Q\Gamma_L)=0$, then $S[(\Psi\circ
\sigma)^c,Q\Gamma_L]=0$ (see (\[Vertcompl\])) and it follows that $$\theta_L(Q\Gamma_L)=0,\;\;\;\; \theta_L[(\Psi\circ
\sigma)^c,Q\Gamma_L]=0.$$ Therefore, $$\label{Lider}
({\mathcal L}_{(\Psi\circ
\sigma)^c}^{{\@ifstar{\@proldf}{\@prolpf}}[E]{E}}\theta_L)(Q\Gamma_L)=0.$$ On the other hand, we have that $$(i(Q\Gamma_L)\omega_L)_{|D}=S^*(\alpha_{(L,D)}),\mbox{ with }
\alpha_{(L,D)}\in Sec(({{\@ifstar{\@proldf}{\@prolpf}}[E]{D}})^0).$$ Consequently, $$\{(i(\psi\circ
\sigma)^c\omega_L)(Q\Gamma_L)\}_{|D}=-\alpha_{(L,D)}((\Psi\circ
\sigma)^v_{|D}).$$
But, since $\Psi\circ \sigma$ is a section of $\tau_D:D\to M$, it follows that $(\Psi\circ \sigma)^v_{|D}$ is a section of ${{\@ifstar{\@proldf}{\@prolpf}}[E]{D}}\to D.$ This implies that $$\label{QGamma}
\{(i(\Psi\circ \sigma)^c\omega_L)(Q\Gamma_L)\}_{|D}=0.$$ Finally, using (\[step1\]), (\[Lider\]) and (\[QGamma\]), we conclude that (\[EqMomen2\]) holds.
\[Remark6.4\]
Suppose that $L:TM\to \R$ is an standard regular Lagrangian function on $E=TM$ and that $\psi:G\times M\to M$ is a (left) action of a Lie group $G$ on $M$. Then, we consider the trivial vector bundle $\tau_K:K=M\times {\frak g}\to M$ and the vector bundle morphism $\Psi:K\to TM$ (over the identity of $M$) defined by (\[Psi\]).
Now, let $D$ be a vector subbundle (over $M$) of the vector bundle $\tau_{M}:TM\to M$, that is, $D$ is a distribution on $M$, and assume that the nonholonomic Lagrangian system $(L,D)$ is regular. If $x$ is a point of $M,$ we have that $K_x^D=\{x\}\times {\frak
g}^x,$ where ${\frak g}^x$ is the vector subspace of ${\frak g}$ given by $${\frak g}^x=\{\xi\in {\frak g}/\xi_M(x)\in D_x\}.$$ We also remark that the sets $K^D$ and $(K^D)^*$ may be identified with the sets $${\frak g}^D=\bigcup_{x\in M}{\frak g}^x,\;\;\;\; ({\frak
g}^D)^*=\bigcup_{x\in M}({\frak g}^x)^*.$$ Under this identification, the nonholonomic momentum map $J_{(L,D,\Psi)}:E\to (K^D)^*$ associated with the system $(L,D)$ and the morphism $\Psi$ is just the standard nonholonomic momentum map $J^{nh}:TM\to ({\frak g}^D)^*$ associated with the system $(L,D)$ and the action $\psi$ (see [@BlKrMaMu; @CaLeMaMa; @CaLeMaMa2]).
Now, if $\widetilde{\xi}:M\to {\frak g}$ is an smooth map the $\widetilde{\xi}$ defines, in a natural way, a section $\sigma_{\widetilde{\xi}}:M\to K=M\times {\frak g}$ of the vector bundle $\tau_K:K=M\times {\frak g}\to M.$ We denote by $J^{nh}_{\widetilde{\xi}}:TM\to \R$ the real function $J_{(L,D,\Psi)}^{\sigma_{\widetilde{\xi}}}:E\to \R$ and by $\Xi_{\widetilde{\xi}}$ the vector field $\Psi\circ
\sigma_{\widetilde{\xi}}$ on $M.$ Then, using Theorem \[theorem6.3\], we deduce a well-known result (see [@BlKrMaMu; @CaLeMaMa; @CaLeMaMa2]): “If $\Gamma_{(L,D)}$ is the solution of the constrained dynamics for the nonholonomic system $(L,D)$, we have that $$\Gamma_{L,D}((J_{\widetilde{\xi}}^{nh})_{|D})=(\Xi_{\widetilde{\xi}})^c_{|D}(L)."$$ The above equality is an intrinsic expression of the *standard nonholonomic momentum equation*. In addition, using again Theorem \[theorem6.3\] we also deduce another well-known result (see [@BlKrMaMu; @CaLeMaMa; @CaLeMaMa2]): “If the Lagrangian function $L:TM\to \R$ is invariant under the tangent action $T\psi$ of $G$ on $TM$ and $\xi\in {\frak g}$ is *a horizontal symmetry* (that is, $\xi\in {\frak g}^x$, for all $x\in M$) then the real function $(J^{nh}_{\widetilde{\xi}})_{|D}$ is a constant of the motion for the constrained Lagrangian dynamics, where $\widetilde{\xi}:M\to
{\frak g}$ is the constant map $$\widetilde{\xi}(x)=\xi, \mbox{ for all } x\in M." \eqoprocend$$
Examples
========
As in the unconstrained case, constrained Lagrangian systems on Lie algebroids appear frequently. We show some examples next.
Nonholonomic Lagrangian systems on Lie algebras
-----------------------------------------------
Let ${\frak g}$ be a real algebra of finite dimension. Then, it is clear that ${\frak g}$ is a Lie algebroid over a single point. Now, suppose that $(l,{\frak d})$ is a nonholonomic Lagrangian system on ${\frak g}$, that is, $l:{\frak g}\to \R$ is a Lagrangian function and ${\frak d}$ is a vector subspace of ${\frak g}$. If $w:I\to {\frak
g}$ is a curve on ${\frak g}$ then $$dl(\omega(t))\in T_{\omega(t)}^*{\frak g}\cong {\frak
g}^*,\;\;\;\; \forall t\in I,$$ and thus, the map $dl\circ \omega$ may be considered as a curve on ${\frak g}^*$ $$dl\circ \omega:I\to {\frak g}^*.$$ Therefore, $$(dl\circ \omega)'(t)\in T_{dl(\omega(t))}{\frak g}^*\cong {\frak
g}^*,\;\;\;\forall t\in I.$$ Moreover, from (\[LD-edo\]), it follows that $\omega$ is a solution of the Lagrange-d’Alembert equations for the system $(l,{\frak d})$ if and only if $$\label{EPSeq}
(dl\circ \omega)'(t)-ad^*_{\omega(t)}(dl(\omega(t)))\in {\frak
d}^\circ,\;\;\; \omega(t)\in {\frak d}, \; \; \; \; \forall t$$ where $ad^*:{\frak g}\times {\frak g}^*\to {\frak g}^*$ is the infinitesimal coadjoint action.
The above equations are just the so-called *Euler-Poincar[é]{} -Suslov equations* for the system $(l,{\frak d})$ (see [@FeZe]). We remark that in the particular case when the system is unconstrained, that is, ${\frak d}={\frak g}$, then one recovers the *the standard Euler-Poincar[é]{} equations* for the Lagrangian function $l:{\frak g}\to \R.$
If $G$ is a Lie group with Lie algebra ${\frak g}$ then nonholonomic Lagrangian systems on ${\frak g}$ may be obtained (by reduction) from nonholonomic LL mechanical systems with configuration space the Lie group $G.$
In fact, let $e$ be the identity element of $G$ and $\mathbb{I}:{\frak
g} \to {\frak g}^*$ be a symmetric positive definite inertia operator. Denote by $g_e:{\frak g}\times {\frak g}\to \R$ the corresponding scalar product on ${\frak g}$ given by $$g_e(\omega,\omega')=<\mathbb{I}(\omega),\omega'>, \mbox{ for
}\omega,\omega'\in {\frak g}\cong T_eG.$$ $g_e$ induces a left-invariant Riemannian metric $g$ on $G$. Thus, we way consider the Lagrangian function $L:TG\to \R$ defined by $$L(v_h)=\frac{1}{2}g_h(v_h,v_h),\;\;\; \mbox{ for }v_h\in T_hG.$$ In other words, $L$ is the kinetic energy associated with the Riemannian metric $g$.
Now, let $D$ be a left-invariant distribution on $G$. Then, since $L$ is a left-invariant function, the pair $(L,D)$ is an standard nonholonomic LL system in the terminology of [@FeZe].
On the other hand, the Lagrangian momentum map $\Phi:TG\to {\frak g}$ given by $$\Phi(v_h)=(T_h l_{h^{-1}})(v_h),\mbox{ for } v_h\in T_hG$$ is a fiberwise bijective morphism of Lie algebroids. Moreover, if $l=L_{|{\frak g}}$ and ${\frak d}=D_e$ then the pair $(l,{\frak d} )$ is a nonholonomic Lagrangian system on ${\frak g}$ and $$l\circ \Phi=L \mbox{ and } \Phi(D)={\frak d}.$$ Thus, the system $(l,{\frak d})$ is regular. In addition, if $v:I\to TG$ is a solution of the Lagrange-d’Alembert equations for the system $(L,D)$ then, using Theorem \[t5.6\], we deduce that the curve $\Phi\circ v:I\to {\frak g}$ is a solution of the Lagrange-d’Alembert equations for the system $(l,{\frak d}).$
We remark that $$l(\omega)=\frac{1}{2}g_e(\omega,\omega) =
\frac{1}{2}<\mathbb{I}(\omega),\omega>,\mbox{ for }\omega\in {\frak
g}.$$ Therefore, if $\omega:I\to {\frak g}$ is a curve on ${\frak g},$ we have that $$(dl\circ \omega)(t)=\mathbb{I}(\omega(t)),\mbox{ for all } t$$ and, using (\[EPSeq\]), it follows that $\omega$ is a solution of the Lagrange-d’Alembert equations for the system $(l,{\frak d})$ if and only if $$\dot{\omega}-\mathbb{I}^{-1}(ad_{\omega(t)}^*\mathbb{I}(\omega(t)))\in
{\frak d}^{\perp},\;\;\; \omega(t)\in {\frak d}, \mbox{ for all }
t,$$ where ${\frak d}^\perp$ is the orthogonal complement of the subspace ${\frak d}$, that is, $${\frak d}^\perp=\{\omega'\in {\frak
g}/<\mathbb{I}(\omega'),\omega>=0,\forall \omega\in {\frak d}\}.$$ Two simple examples of the above general situation are the following ones.
The Suslov system {#the-suslov-system .unnumbered}
-----------------
The most natural example of LL system is the *nonholonomic Suslov problem*, which describes the motion of a rigid body about a fixed point under the action of the following nonholonomic constraint: the body angular velocity vector is orthogonal to a some fixed direction in the body frame.
The configuration space of the problem is the group $G=SO(3)$. Thus, in this case, the Lie algebra ${\frak g}$ may be identified with $\R^3$ and, under this identification, the Lie bracket on ${\frak g}$ is just the cross product $\times$ on $\R^3.$
Moreover, if $\mathbb{I}:\R^3\to (\R^3)^*\cong \R^3$ is the inertia tensor of the body then a curve $\omega:I\to \R^3$ on $\R^3$ is a solution of the Euler-Poincar[é]{}-Suslov equations for the system if and only if $$\label{Suseq}
\dot{\omega}=\mathbb{I}^{-1}((\mathbb{I}\omega) \times \omega)) +
\lambda \mathbb{I}^{-1}(\Gamma),\;\;\;\; <\omega,\Gamma>=0,$$ where $\lambda$ is the Lagrange multiplier, $\Gamma$ is a fixed unit vector in $\R^3$ and $<\cdot,\cdot>$ is the standard scalar product in $\R^3$. Since the nonholonomic system is regular, the Lagrange multiplier $\lambda$ is uniquely determined. In fact, differentiating the equation $<\omega,\Gamma>=0$, we find $$\lambda=-\frac{<\mathbb{I}\omega\times \omega,
\mathbb{I}^{-1}\Gamma>}{<\Gamma, \mathbb{I}^{-1}\Gamma>}$$ and, consequently, Eqs. (\[Suseq\]) are equivalent to $$\dot{\omega}=\mathbb{I}^{-1}(<\mathbb{I}\omega,\Gamma>\omega\times
\mathbb{I}^{-1}\Gamma),\;\;\;\; <\omega,\Gamma>=0.$$ Multidimensional generalizations of the Suslov problem have been discussed by several authors (see [@FeKo; @Jo; @ZeBl1]).
The Chaplygin sleigh {#the-chaplygin-sleigh .unnumbered}
--------------------
The Chaplygin sleigh is a rigid body sliding on a horizontal plane. The body is supported at three points, two of which slide freely without friction while the third is a knife edge, a constraint that allows no motion orthogonal to this edge. This mechanical system was introduced and studied in 1911 by Chaplygin [@Ch] (see also [@NF]).
The configuration space of this system is the group $SE(2)$ of Euclidean motions of the two-dimensional plane $\R^2.$ As we know, we may choose local coordinates $(\theta,x,y)$ on $SE(2)$. $\theta$ and $(x,y)$ are the angular orientation of the blade and position of the contact point of the blade on the plane, respectively.
Now, we introduce a coordinate system called the body frame by placing the origin at the contact point and choosing the first coordinate axis in the direction of the knife edge. Denote the angular velocity of the body by $\omega=\dot{\theta}$ and the components of the linear velocity of the contact point relative to the body frame by $v_1,v_2$. The set $(\omega,v_1,v_2)$ is regarded as an element of the Lie algebra ${\frak{se}}(2)$. Note that $$v_1=\dot{x}\cos\theta + \dot{y} \sin \theta,\;\;\;\;
v_2=\dot{y}\cos\theta - \dot{x}\sin\theta.$$ The position of the center of mass is specified by the coordinates $(a,b)$ relative to the body frame. Let $m$ and $J$ denote the mass and moment of inertia of the sleigh relative to the contact point. Then, the corresponding symmetric positive definite inertia operator $\mathbb{I}:{\frak se}(2)\to {\frak se}(2)^*$ and the reduced nonholonomic Lagrangian system $(l,{\frak d})$ on ${\frak se}(2)$ are given by $$\begin{array}{rcl}
\mathbb{I} (\omega,v_1,v_2)&=&\left(
\begin{array}{ccc}
J + m(a^2+b^2)&-bm&am\\
-bm&m&0\\
am&0&m \end{array}
\right)\left(\begin{array}{c}\omega\\v_1\\v_2\end{array}\right),
\\[20pt]
l(\omega,v_1,v_2)&=&\frac{1}{2}[(J+m(a^2+b^2))\omega^2 + m(v_1^2 +
v_2^2)\\[5pt]&&-2mb\omega v_1 + 2am\omega v_2],\\[8pt]
{\frak d}&=&\{(\omega, v_1,v_2)\in {\frak
se}(2)/v_2=0\},\end{array}$$ (see [@FeZe]). Thus, the Lagrange-d’Alembert equations for the system $(l,{\frak d})$ are $$\begin{array}{rcl}
\dot{\omega}&=&\displaystyle\frac{am\omega}{J+ma^2}(b\omega-v_1),\\[8pt]
\dot{v_1} & = &
\displaystyle\frac{a\omega}{J+ma^2}((J+m(a^2+b^2))\omega-mbv_1),\\[8pt]
v_2&=&0.
\end{array}$$ Multidimensional generalizations of the Chaplygin sleigh were discussed in [@FeZe] (see also [@NF] and [@ZeBl2]).
Nonholonomic LR systems and right action Lie algebroids
-------------------------------------------------------
Here, we show how the reduction of a nonholonomic LR system produces a nonholonomic Lagrangian system on a right action Lie algebroid.
Let us start by recalling the definition of a right action Lie algebroid (see [@HiMa]). Let $(F, [\cdot, \cdot]_{F}, \rho_{F})$ be a Lie algebroid over a manifold $N$ and $\pi: M \to N$ be a smooth map. *A right action of $F$ on $\pi: M \to N$* is a $\R$-linear map $$\Psi: Sec (F) \to {\frak X}(M), \; \; \; X\in Sec (F) \to
\Psi(X) \in {\frak X}(M)$$ such that $$\begin{array}{l}
\Psi (f X) = (f \circ \pi)\Psi (X), \; \; \Psi([X, Y]_{F}) =
[\Psi(X), \Psi(Y)], \\[5pt] (T_{m}\pi)(\Psi(X)(m)) =
\rho_{F}(X(\pi(m))),
\end{array}$$ for $f \in C^{\infty}(N)$, $X, Y \in Sec (F)$ and $m \in M$. If $\Psi: Sec (E) \to {\frak X}(M)$ is a right action of $F$ on $\pi: M \to N$ and $\tau_{F}: F \to N$ is the vector bundle projection then the pullback vector bundle of $F$ over $\pi$, $$E = \pi^{*}F = \{(m, f) \in M \times F / \tau_{F}(f) = \pi (m)\}$$ is a Lie algebroid over $M$ with Lie algebroid structure $([\cdot, \cdot]_{E}, \rho_{E})$ which is characterized by $$[X, Y]_{E} = [X, Y]_{F} \circ \pi, \; \; \; \rho_{E}(X)(m) =
\Psi(X)(m),$$ for $X, Y \in Sec(E)$ and $m \in M$. The triple $(E, [\cdot,
\cdot]_{E}, \rho_{E})$ is called the *right action Lie algebroid of $F$ over $\pi$* and it is denoted by $\pi_{\Psi}F$ (see [@HiMa]).
Note that if the Lie algebroid $F$ is a real Lie algebra ${\frak g}$ of finite dimension and $\pi: M \to \{\mbox{a point}\}$ is the constant map then a right action of ${\frak g}$ on $\pi$ is just a right infinitesimal action $\Psi: {\frak g} \to {\frak X}(M)$ of ${\frak g}$ on the manifold $M$. In such a case, the corresponding right action Lie algebroid is the trivial vector bundle $pr_{1}: M
\times {\frak g} \to M$.
Next we recall the definition of a nonholonomic LR system following [@FeJo; @Jo2]. Let $G$ be a compact connected Lie group with Lie algebra ${\frak g}$ and $<\cdot,\cdot>:{\frak g}\times {\frak
g}\to \R$ be an $Ad_G$-invariant scalar product on ${\frak g}$. Now, suppose that ${\mathcal I}:{\frak g} \to {\frak g}$ is a inertia operator which is symmetric and definite positive with respect to the scalar product $<\cdot,\cdot>$. Denote by $g$ the left-invariant Riemannian metric given by $$\label{metrica}
g_h(v_h,v_{h'})=<{\mathcal
I}(T_hl_{h^{-1}}(v_h)),(T_hl_{h^{-1}})(v_h')>$$ for $h\in G$ and $v_h,v_h'\in T_hG.$
Then, the Lagrangian function $L:TG\to \R$ of the system is $$\label{L}
L(v_h)=\frac{1}{2}g_h(v_h,v_h)-V(h), \mbox{ for }v_h\in T_hG,$$ $V:G\to \R$ being the potential energy. The constraint distribution $D$ is a right-invariant distribution on $G$. Thus, if $e$ is the identity element of $G$ and ${\frak d} =D_e$, we have that $$\label{eq:D}
D_h=(T_er_h)({\frak d} )=(T_el_h)(Ad_{h^{-1}}({\frak d})),\mbox{
for }h\in G$$ where $Ad:G\times {\frak g}\to {\frak g}$ is the adjoint action.
The nonholonomic Lagrangian system $(L,D)$ on $TG$ is called a *nonholonomic LR system* in the terminology of [@FeJo; @Jo2]. Note that, since $L$ is a Lagrangian function of mechanical type, the system $(L,D)$ is regular. Now, assume that $${\frak s}={\frak d}^{\perp}=\{\omega' \in {\frak g} / <
\omega,\omega'>=0,\forall \omega\in {\frak d}\}$$ is a Lie subalgebra of ${\frak g}$, that $S$ is a closed Lie subgroup of $G$ with Lie algebra ${\frak s}$ and that the potential energy $V$ is $S$-invariant.
Next, let us show that the nonholonomic LR system $(L,D)$ may be reduced to a nonholonomic Lagrangian system on a right action Lie algebroid. In fact, consider the Riemannian homogeneous space $M=S\setminus G$ and the standard transitive right action $\psi$ of $G$ on $M=S\setminus G$. Denote by $\Psi:{\frak g}\to {\frak
X}(S\setminus G)$ the corresponding right infinitesimal action of ${\frak g}$ on $S\setminus G$. Then, $\Psi$ induces a Lie algebroid structure on the trivial vector bundle $pr_1:S\setminus G\times {\frak
g}\to S\setminus G$.
On the other hand, using that the potential energy $V$ is $S$ invariant, we deduce that $V$ induces a real function $\widetilde{V}:S\setminus G\to \R$ on $S\setminus G$ such that $$\label{vtil}
\widetilde{V}\circ \pi=V,$$ where $\pi:G\to S\setminus G$ is the canonical projection. Thus, we can introduce the Lagrangian function $\tilde{L}: S\setminus G\times
{\frak g}\to \R$ on the action Lie algebroid $pr_1:S\setminus G\times
{\frak g}\to S\setminus G$ defined by $$\label{Ltil}
\widetilde{L}(\widetilde{h},\omega)=\frac{1}{2}<{\mathcal
I}(\omega),\omega>-\widetilde{V}(\widetilde{h}),\;\;\; \mbox{ for
}\widetilde{h}\in S\setminus G \mbox{ and }\omega\in {\frak g}.$$
Now, for every $h\in G$, we consider the subspace ${\frak d} (h)$ of ${\frak g}$ given by $$\label{Deltah}
{\frak d}(h)=Ad_{h^{-1}}({\frak d}).$$ The dimension of ${\frak d}(h)$ is equal to the dimension of ${\frak d}$. Moreover, since $<\cdot,\cdot>$ is $Ad_G$-invariant, it follows that $${\frak d}(h)=(Ad_{h^{-1}}({\frak s}))^\perp=\{\omega'\in {\frak
g}/<\omega',Ad_{h^{-1}}(\omega)>=0,\;\;\forall \omega\in {\frak
s}\}.$$ In particular, we have that $${\frak d}(s)={\frak s}^\perp={\frak d}, \;\;\forall s\in S$$ which implies that ${\frak d} (sh)={\frak d} (h)$, for all $h\in
G.$
Therefore, we can define a vector subbundle $D$ of the Lie algebroid $pr_1:S\setminus G\times {\frak g}\to S\setminus G$ as follows $$\label{Dtil}
\widetilde{D}_{\widetilde{h}}=\{\widetilde{h}\}\times {\frak d}
(h),\mbox{ for } \widetilde{h}\in S\setminus G$$ with $h\in G$ and $\pi(h)=\widetilde{h}.$ Consequently, the pair $(\widetilde{L},\widetilde{D})$ is a nonholonomic Lagrangian system on the action Lie algebroid $pr_1:S\setminus G\times {\frak g}\to
S\setminus G.$
In addition, we may prove the following result
\[propositon7.1\]
1. If $\widetilde{\Phi}:TG\to S\setminus G\times {\frak g}$ is the map given by $$\label{Fitil}
\widetilde{\Phi}(v_h)=(\pi(h),(T_hl_{h^{-1}}(v_h)),\mbox{ for all
} v_h\in T_hG$$ then $\widetilde{\Phi}$ is a fiberwise bijective Lie algebroid morphism over $\pi$.
2. The nonholonomic Lagrangian systems $(L,D)$ and $(\tilde{L},\widetilde{D})$ on $TG$ and $S/G\times
{\frak g}$ are $\widetilde{\Phi}$-related, that is, $$\tilde{L} \circ \widetilde{\Phi}=L,\;\;\;
\widetilde{\Phi}(D)=\widetilde{D}.$$
3. The system $(\widetilde{L},\widetilde{D})$ is regular and if $\gamma:I\to TG$ is a solution of the Lagrange-d’Alembert equations for the system $(L,D)$ then $\widetilde{\Phi}\circ
\gamma: I \to S\setminus G\times {\frak g}$ is a solution of the Lagrange-d’Alembert equations for the system $(\widetilde{L},\widetilde{D}).$
$(1)$ Consider the standard (right) action $r$ of $G$ on itself $$r:G\times G\to G, \; \; \; (h,h')\in G\times G\to r_{h'}(h)=hh'\in
G.$$ As we know, the infinitesimal generator of $r$ associated with an element $\omega$ of ${\frak g}$ is $$\omega_G={\overleftarrow{\omega}},$$ where ${\overleftarrow{\omega}}$ is the left-invariant vector field on $G$ such that ${\overleftarrow{\omega}}(e)=\omega$.
On the other hand, it is clear that the projection $\pi:G\to
S\setminus G$ is equivariant with respect to the actions $r$ and $\psi$. Thus, $$(T_h\pi)({\overleftarrow{\omega}}(h))=\Psi(\omega)(\pi(h)), \mbox{ for } h\in
G.$$ Therefore, if $\rho:S/G\times {\frak g}\to T(S\setminus G)$ is the anchor map of the Lie algebroid $pr_{1}: S\setminus G\times {\frak
g} \to S\setminus G$, it follows that $$\rho(\widetilde{\Phi}({\overleftarrow{\omega}}(h))) =
\rho(\pi(h),\omega)=(T_h\pi)({\overrightarrow{\omega}}(h)), \; \; \; \mbox{ for
} h\in G.$$ Furthermore, since $$[{\overleftarrow{\omega}},{\overleftarrow{\omega}}']={\overleftarrow{[\omega,\omega']_{\frak g}}},
\; \; \; \; \mbox{ for } \omega,\omega'\in {\frak g},$$ we conclude that $\widetilde{\Phi}$ is a Lie algebroid morphism over $\pi.$
In addition, it is obvious that if $h\in G$ then $$\widetilde{\Phi}_{|T_hG}:T_hG\to \{\pi(h)\}\times {\frak g}$$ is a linear isomorphism.
$(2)$ From (\[metrica\]), (\[L\]), (\[vtil\]), (\[Ltil\]) and $(\ref{Fitil})$, we deduce that $$\widetilde{L}\circ \widetilde{\Phi}=L.$$ Moreover, using (\[eq:D\]), (\[Deltah\]), (\[Dtil\]) and (\[Fitil\]), we obtain that $$\widetilde{\Phi}(D)=\widetilde{D}.$$ $(3)$ It follows from $(1),$ $(2)$ and using the results of Section \[sec:reduction\] (see Theorem \[t5.6\]).
Next, we obtain the necessary and sufficient conditions for a curve $(\widetilde{h},\omega):I\to S\setminus G\times {\frak g}$ to be a solution of the Lagrange-d’Alembert equations for the system $(\widetilde{L},\widetilde{D})$. Let $\flat_{<\cdot,\cdot>}:{\frak
g}\to {\frak g}^*$ be the linear isomorphism induced by the scalar product $<\cdot,\cdot>:{\frak g}\times {\frak g}\to \R$ and $\mathbb{I}:{\frak g}\to {\frak g}^*$ be the inertia operator given by $$\label{Inertia2}
\mathbb{I}(w_1)(w_2)=<{\mathcal I}(\omega_1),\omega_2>, \mbox{ for
}\omega_1,\omega_2\in {\frak g}.$$ On the other hand, if $\widetilde{h}'\in S\setminus G$ we will denote by $\Psi_{\tilde{h}'}:{\frak g}\to T_{\widetilde{h}'}(S\setminus G)$ the linear epimorphism defined by $$\Psi_{\widetilde{h}'}(\omega')=\Psi({\omega'})(\widetilde{h}'),
\mbox{ for } \omega'\in {\frak g}.$$
In addition, if $\pi(h')=\widetilde{h}'$, we identify the vector space $\widetilde{D}_{\widetilde{h}'}$ with the vector subspace ${\frak
d}(h')$ of ${\frak g}$. Then, using (\[LD-edo\]), (\[Ltil\]) and (\[Inertia2\]), we deduce that the curve $(\widetilde{h},\omega)$ is a solution of the Lagrange-d’Alembert equations for the system $(\widetilde{L},\widetilde{D})$ if and only if $$\begin{array}{l}
\dot{\widetilde{h}}(t)=\Psi_{\widetilde{h}(t)}(\omega(t))\\
\{\dot{\omega}(t)-\mathbb{I}^{-1}(ad_{\omega(t)}^{*}\mathbb{I}
(\omega(t)))-\mathbb{I}^{-1}(\Psi^{*}_{\widetilde{h}(t)}
(d\widetilde{V}(\widetilde{h}(t))))\}\in
\widetilde{D}_{\widetilde{h}(t)}^\perp,\\
\omega(t)\in \widetilde{D}_{\widetilde{h}(t)},
\end{array}$$ for all $t$, where $\widetilde{D}_{\widetilde{h}(t)}^\perp$ is the orthogonal complement of the vector subspace $\widetilde{D}_{\widetilde{h}(t)}\subseteq {\frak g}$ with respect to the scalar product $<\cdot,\cdot>$. These equations will be called *the reduced Poincaré-Chetaev equations*.
We treat next a simple example of the above general situation.
The Veselova system {#the-veselova-system .unnumbered}
-------------------
The most descriptive illustration of an LR system is the Veselova problem on the motion of a rigid body about a fixed point under the action of the nonholonomic constraint $$<\omega,\gamma>=0.$$ Here, $\omega$ is the vector of the angular velocity in the body frame, $\gamma$ is a unit vector which is fixed in an space frame and $<\cdot,\cdot>$ denotes the standard scalar product in $\R^3$ (see [@VeVe]).
The Veselova system is an LR system on the Lie group $G=SO(3)$ which is the configuration space of the rigid body motion. Thus, in this case, the Lie algebra ${\frak g}$ may be identified with $\R^3$ and, under this identification, the Lie bracket $[\cdot,\cdot]_{\frak g}$ is the cross product $\times$ on $\R^3$. Moreover, the adjoint action of $G=SO(3)$ on ${\frak g}\cong \R^3$ is the standard action of $SO(3)$ on $\R^3$. This implies that $<\cdot,\cdot>$ is an $Ad_{SO(3)}$-invariant scalar product on ${\frak g}\cong \R^3.$
The vector subspace ${\frak d}$ of $\R^3$ is just the orthogonal complement (with respect to $<\cdot,\cdot>$) of a vector subspace $<\gamma_0>$ of dimension $1$, with $\gamma_0$ a unit vector in $\R^3$, that is, $${\frak d}=\{\omega\in \R^3/<\omega,\gamma_0>=0\}.$$ Therefore, $${\frak s}={\frak d}^\perp = <\gamma_0>$$ is a Lie subalgebra of ${\frak g}\cong \R^3$. Furthermore, the isotropy group $S$ of $\gamma_0$ with respect to the adjoint action of $G=SO(3),$ $$S=\{s\in SO(3)/s\gamma_0^T=\gamma_0^T\},$$ is a closed Lie subgroup with Lie algebra ${\frak s}$. We remark that $S$ is isomorphic to the circle $S^1.$
Consequently, the corresponding homogeneous space $M=S\setminus SO(3)$ is the orbit of the adjoint action of $SO(3)$ on $\R^3$ over the point $\gamma_0$ and, it is well-known that, such an orbit may be identified with the unit sphere $S^2$. In fact, the map $$S\setminus SO(3)\to S^2, \;\;\; [h]\to \gamma_0
h=(h^{-1}\gamma_0^T)^T$$ is a diffeomorphism (see, for instance, [@MaRa]).
Under the above identification the (right) action of SO(3) on $M=S\setminus SO(3)$ is just the standard (right) action of $SO(3)$ on $S^2$. Thus, our action Lie algebroid is the trivial vector bundle $pr_1:S^2\times \R^3\to S^2$ and the Lie algebroid structure on it is induced by the standard infinitesimal (right) action $\Psi:\R^3\to
{\frak X}(S^2)$ of the Lie algebra $(\R^3,\times)$ on $S^2$ defined by $$\Psi(\omega)(\gamma)=\gamma\times \omega,\mbox{ for } \omega\in \R^3
\mbox{ and } \gamma\in S^2.$$ In the presence of a potential $\tilde{V}: \gamma\to
\widetilde{V}(\gamma)$ the nonholonomic Lagrangian system $(\widetilde{L},\widetilde{D})$ on the Lie algebroid $pr_1:S^2\times \R^3\to \R^3$ is given by $$\widetilde{L}(\gamma,\omega)=\frac{1}{2}
\mathbb{I}(\omega)(\omega)-\widetilde{V}(\gamma), \; \; \;
\widetilde{D}(\gamma)=\{\gamma\}\times
\{\omega\in\R^3/<\omega,\gamma>=0\},$$ $\mathbb{I}:\R^3\to\R^3$ being the inertia tensor of the rigid body.
The Lagrange-d’Alembert equations for $(\widetilde{L},\widetilde{D})$ are $$\label{Vese1}
\dot{\gamma}=\gamma\times \omega,\;\;\;
\dot{\omega}=\mathbb{I}^{-1}\{(\mathbb{I}\omega\times\omega) +
\omega\times \frac{\partial \widetilde{V}}{\partial \gamma} +
\lambda\gamma\},\;\;\; <\omega,\gamma>=0$$ where $\lambda$ is the Lagrange multiplier. Since the system $(\widetilde{L}, \widetilde{D})$ is regular, $\lambda$ is uniquely determined. In fact, $$\label{Vese2}
\lambda=-\frac{<\mathbb{I}\omega\times \omega + \gamma\times
\frac{\partial \widetilde{V}}{\partial \gamma},
\mathbb{I}^{-1}\gamma>}{<\mathbb{I}^{-1}\gamma,\gamma>}.$$ Eqs (\[Vese1\]) and (\[Vese2\]) are just the classical dynamical equations for the Veselova system (see [@VeVe]; see also [@FeJo; @Jo2]).
Semidirect product symmetry and left action Lie algebroids
----------------------------------------------------------
Here, we show how the reduction of some nonholonomic mechanical systems with semidirect product symmetry produces nonholonomic Lagrangian systems on left action Lie algebroids.
Let us start by recalling the definition of a left action Lie algebroid (see [@HiMa]). Let $(F, [\cdot, \cdot]_{F}, \rho_{F})$ be a Lie algebroid over a manifold $N$ and $\pi: M \to N$ be a smooth map. *A left action of $F$ on $\pi: M \to N$* is a $\R$-linear map $$\Psi: Sec (F) \to {\frak X}(M), \; \; \; X\in Sec (F) \to
\Psi(X) \in {\frak X}(M)$$ such that $$\begin{array}{l}
\Psi (f X) = (f \circ \pi)\Psi (X), \; \; \Psi([X, Y]_{F}) =
-[\Psi(X), \Psi(Y)], \\[5pt] (T_{m}\pi)(\Psi(X)(m)) =
-\rho_{F}(X(\pi(m))),
\end{array}$$ for $f \in C^{\infty}(N)$, $X, Y \in Sec (F)$ and $m \in M$. If $\Psi: Sec (E) \to {\frak X}(M)$ is a left action of $F$ on $\pi: M \to N$ and $\tau_{F}: F \to N$ is the vector bundle projection then the pullback vector bundle of $F$ over $\pi$, $$E = F^{*}\pi = \{(f, m) \in F \times M / \tau_{F}(f) = \pi (m)\}$$ is a Lie algebroid over $M$ with Lie algebroid structure $([\cdot, \cdot]_{E}, \rho_{E})$ which is characterized by $$[X, Y]_{E} = [X, Y]_{F} \circ \pi, \; \; \; \rho_{E}(X)(m) =
-\Psi(X)(m),$$ for $X, Y \in Sec(E)$ and $m \in M$. The triple $(E, [\cdot,
\cdot]_{E}, \rho_{E})$ is called *the left action Lie algebroid of $F$ over $\pi$* and it is denoted by $F_{\Psi}\pi$ (see [@HiMa]).
Next, we consider a particular class of nonholonomic Lagrangian systems on left action Lie algebroids. Let $V$ be a real vector space of finite dimension and $\cdot: G \times V \to V$ be a left representation of a Lie group $G$ on $V$. We also denote by $\cdot:
{\frak g} \times V \to V$ the left infinitesimal representation of the Lie algebra ${\frak g}$ of $G$ on $V$. Then, we can consider the semidirect Lie group $S= G \circledS V$ with the multiplication $$(g, v) (g', v') = (gg', v + g \cdot v').$$ The Lie algebra ${\frak s}$ of $S$ is the semidirect product ${\frak s} = {\frak g} \circledS V$ with the Lie bracket $[\cdot,
\cdot]_{\frak s}: {\frak s} \times {\frak s} \to {\frak s}$ given by $$[(\omega, \dot{v}), (\omega', \dot{v}')]_{\frak s} = ([\omega,
\omega']_{\frak g}, \omega \cdot \dot{v}' - \omega' \cdot \dot{v})$$ for $\omega, \omega' \in {\frak g}$ and $\dot{v}, \dot{v}' \in V$. Here, $[\cdot, \cdot]_{\frak g}$ is the Lie bracket on ${\frak
g}$.
Moreover, we use the following notation. If $v \in V$ then $\rho_{v}:
{\frak g} \to V$ is the linear map defined by $$\rho_{v}(\omega) = \omega \cdot v, \; \; \; \mbox{ for } \omega
\in {\frak g},$$ and $\rho_{v}^{*}: V^{*} \to {\frak g}^{*}$ is the dual map of $\rho_{v}: {\frak g} \to V$.
Now, let $N$ be a smooth manifold. Then, it is clear that the product manifold $F = {\frak s} \times TN$ is the total space of a vector bundle over $N$. Moreover, if $(\omega, \dot{v}) \in {\frak
s}$ and $X$ is a vector field on $N$ then the pair $((\omega,
\dot{v}), X)$ defines a section of the vector bundle $\tau_{F}: F
= {\frak s} \times TN \to N$. In fact, if $\{\omega_{i}\}$ is a basis of ${\frak g}$, $\{\dot{v}_{j}\}$ is a basis of $V$ and $\{X_{k}\}$ is a local basis of ${\frak X}(N)$ then $\{((\omega_{i}, 0), 0), ((0, \dot{v}_{j}), 0), ((0, 0), X_{k})\}$ is a local basis of $Sec (F)$.
The vector bundle $\tau_{F}: F \to N$ admits a Lie algebroid structure $([\cdot, \cdot]_{F}, \rho_{F})$, which is characterized by the following relations $$\label{LiealgF}
\begin{array}{rcl}
[((\omega, \dot{v}), X), ((\omega', \dot{v}'), X')]_{F} & = &
([(\omega, \dot{v}), (\omega', \dot{v}')]_{\frak s}, [X, X']) \\
& = &([\omega, \omega']_{\frak g}, \omega \cdot \dot{v}' - \omega'
\cdot \dot{v},
[X, X']), \\[5pt]
\rho_{F}((\omega, \dot{v}), X) & = & X,
\end{array}$$ for $((\omega, \dot{v}), X), ((\omega', \dot{v}'), X') \in {\frak
s} \times {\frak X}(N)$.
Next, suppose that $v_{0}$ is a point of $V$ and that ${\mathcal
O}_{v_{0}}$ is the orbit of the action of $G$ on $V$ by $v_{0}$, that is, $${\mathcal O}_{v_{0}} = \{g \cdot v_{0} \in V / g \in G \}.$$ Denote by $\pi: M = N \times {\mathcal O}_{v_{0}} \to N$ the canonical projection on the first factor and by $\Psi: Sec (F) \to
{\frak X}(M)$ the left action of $F$ on $\pi$, which is characterized by the following relation $$\Psi ((\omega, \dot{u}), X)(n, v) = (-X(n), \omega \cdot v)$$ for $((\omega, \dot{u}), X) \in {\frak s} \times {\frak X}(N)$ and $(n, v) \in N \times {\mathcal O}_{v_{0}} = M$.
Then, we have the corresponding left action Lie algebroid $\tau_{E}: E = ({\frak s} \times TN)_{\Psi} \pi \to M = N \times
{\mathcal O}_{v_{0}}$. Note that $E = ({\frak s} \times
TN)_{\Psi}\pi = ({\frak s} \times TN) \times {\mathcal O}_{v_{0}}$ and that the anchor map $\rho_{E}: E = ({\frak s} \times TN)
\times {\mathcal O}_{v_{0}} \to TM = TN \times T{\mathcal
O}_{v_{0}}$ of $\tau_{E}: E \to M$ is given by $$\label{AnclaE}
\rho_{E}((\omega, \dot{u}), X_{n}, v) = (X_{n}, -\omega \cdot v)$$ for $((\omega, \dot{u}), X_{n}, v) \in {\frak s} \times T_{n}N
\times {\mathcal O}_{v_{0}}$.
Now, let $L: ({\frak s} \times TN) \times {\mathcal O}_{v_{0}} \to
\R$ be a Lagrangian function and $D$ be the vector subbundle of $\tau_{E}: E \to M$ whose fiber $D_{(n, v)}$ over the point $(n,
v) \in N \times {\mathcal O}_{v_{0}} = M$ is defined by $$\label{D}
\begin{array}{l}
D_{(n, v)} = \{(((\omega, \omega \cdot v), X_{n}), v) / \omega \in
{\frak g}, X_{n} \in T_{n}N \} \\ [5pt] \; \; \; \; \subseteq
E_{(n, v)} = ({\frak s} \times T_{n}N) \times \{v\}.
\end{array}$$ Next, we obtain the Lagrange-d’Alembert equations for the system $(L,
D)$. For this purpose, we choose a basis $\{\omega_{\alpha}\}$ of ${\frak g}$, a basis $\{u_{A}\}$ of $V$, a system of local fibred coordinates $(x^{i}, \dot{x}^{i})$ on $TN$ and a system of local coordinates $(v^{i})$ on ${\mathcal O}_{v_{0}}$. Denote by $\omega^{\alpha}$ (respectively, $u^{A}$) the global coordinates on ${\frak g}$ (respectively, $V$) induced by the basis $\{\omega_{\alpha}\}$ (respectively, $\{u_{A}\}$).
Suppose that $$[\omega_{\alpha}, \omega_{\beta}]_{\frak g} = c_{\alpha
\beta}^{\gamma} \omega_{\gamma}, \; \; \; \omega_{\alpha} \cdot
u_{A} = a_{\alpha A}^{B} u_{B}.$$ Then, we have that $$c_{\alpha \beta}^{\gamma} a_{\gamma A}^{B} = a_{\beta
A}^{C}a_{\alpha C}^{B} - a_{\alpha A}^{C}a_{\beta C}^{B}.$$ Next, we consider the local basis of sections $\{e_{i},
e_{\alpha}, e_{A}\}$ of $E$ given by $$\begin{array}{rcl}
e_{i}(n, v) & = & ((0, 0, \frac{\partial}{\partial x^{i}}_{|n}),
v), \; \; \; e_{\alpha}(n, v) \; = \; ((\omega_{\alpha},
\omega_{\alpha} \cdot v, 0_{n}), v) \\ [5pt] e_{A}(n, v) & = &
((0, u_{A}, 0_{n}), v)
\end{array}$$ for $(n, v) \in N \times {\mathcal O}_{v_{0}} = M$. Note that $\{e_{i}, e_{\alpha}\}$ is a local basis of sections of the constraint subbundle $D$. In addition, if $(x^{i}, v^{j}; y^{i}, y^{\alpha},
y^{A})$ are the local coordinates on $E$ induced by the basis $\{e_{i}, e_{\alpha}, e_{A}\}$, it follows that $$\label{Changecoor}
y^{i} = \dot{x}^{i}, \; \; \; y^{\alpha} = \omega^{\alpha}, \; \;
\; y^{A} = u^{A} - a_{\alpha B}^{A}u^{B}_{0} \omega^{\alpha},$$ where $u^{B}_{0}$ is the local function on $M = N \times {\mathcal
O}_{v_{0}}$ defined by $u^{B}_{0} = (u^{B})_{|{\mathcal
O}_{v_{0}}}.$ Moreover, $$\begin{array}{l}
\rho_{E}(e_{i}) = -\frac{\partial}{\partial
x^{i}}, \; \; \; \rho_{E}(e_{\alpha}) =
\rho^{i}_{\alpha}\frac{\partial}{\partial v^{i}}, \; \; \;
\rho_{E}(e_{A}) = 0, \\[5pt]
[e_{\alpha}, e_{\beta}]_{E} = c_{\alpha \beta}^{\gamma}
(e_{\gamma} + a_{\gamma A}^{B}u^{A}_{0} e_{B}), \; \; \;
[e_{\alpha}, e_{A}]_{E} = -[e_{A}, e_{\alpha}]_{E} = a_{\alpha
A}^{B} e_{B},
\end{array}$$ and the rest of the fundamental Lie brackets are zero. Thus, a curve $$t \to (x^{i}(t), v^{j}(t); y^{i}(t), y^{\alpha}(t), y^{A}(t))$$ is a solution of the Lagrange-d’Alembert equations for the system $(L, D)$ if and only if $$\begin{array}{l}
\dot{x}^{i} = y^{i}, \; \; \; \dot{v}^{j} =
\rho_{\alpha}^{j}y^{\alpha}, \mbox{ for all } i \mbox{ and } j,
\\[5pt]
\displaystyle \frac{d}{dt}(\displaystyle \frac{\partial
L}{\partial y^{i}}) - \displaystyle \frac{\partial L}{\partial
x^{i}} = 0, \mbox{ for all } i,
\\[8pt]
\displaystyle \frac{d}{dt}(\displaystyle \frac{\partial
L}{\partial y^{\alpha}}) + \displaystyle (\frac{\partial
L}{\partial y^{\gamma}} + \displaystyle \frac{\partial L}{\partial
y^{B}} a_{\gamma A}^{B}u^{A}_{0})c_{\alpha
\beta}^{\gamma}y^{\beta} - \rho_{\alpha}^{i}\displaystyle
\frac{\partial L}{\partial v^{i}} = 0, \mbox{ for all } \alpha,
\\[8pt]
y^{A} = 0, \mbox{ for all } A.
\end{array}$$ If we consider the local expression of the curve in the coordinates $(x^{i}, v^{j}; \dot{x}^{i}, \omega^{\alpha}, u^{A})$ then, from (\[Changecoor\]), we deduce that the above equations are equivalent to $$\begin{array}{l}
\dot{x}^{i} = y^{i}, \; \; \; \dot{v}^{j} =
\rho_{\alpha}^{j}\omega^{\alpha}, \mbox{ for all } i \mbox{ and }
j,
\\[5pt]
\displaystyle \frac{d}{dt}(\displaystyle \frac{\partial
L}{\partial \dot{x}^{i}}) - \displaystyle \frac{\partial
L}{\partial x^{i}} = 0, \mbox{ for all } i,
\\[8pt]
\displaystyle \frac{d}{dt}(\displaystyle \frac{\partial
L}{\partial \omega^{\alpha}}) + \displaystyle \frac{\partial
L}{\partial \omega^{\gamma}} c_{\alpha \beta}^{\gamma}
\omega^{\beta} + \displaystyle \frac{d}{dt}(a_{\alpha
B}^{A}u^{B}_{0} \displaystyle \frac{\partial L}{\partial u^{A}})
\\[8pt]
\hspace{1cm} + 2a_{\gamma B}^{A}u^{B}_{0} \displaystyle
\frac{\partial L}{\partial u^{A}} c_{\alpha
\beta}^{\gamma}\omega^{\beta} - \rho_{\alpha}^{i} \displaystyle
\frac{\partial L}{\partial v^{i}} = 0, \mbox{ for all } \alpha,
\\[5pt]
u^{A} = a_{\alpha B}^{A}u^{B}_{0} \omega^{\alpha}, \mbox{ for all
} A,
\end{array}$$ or, in vector notation, $$\begin{array}{l}
\dot{v} = - \omega \cdot v,
\\[5pt]
\displaystyle \frac{d}{dt}(\displaystyle \frac{\partial
L}{\partial \dot{x}}) - \displaystyle \frac{\partial L}{\partial
x} = 0,
\\[8pt]
\displaystyle \frac{d}{dt}(\displaystyle \frac{\partial
L}{\partial \omega}) + (ad^{*}_{\omega} \displaystyle
\frac{\partial L}{\partial \omega}) = -\displaystyle \frac{d}{dt}
(\rho^{*}_{v} \displaystyle \frac{\partial L}{\partial u}) -2
ad^{*}_{\omega}(\rho^{*}_{v} \displaystyle \frac{\partial
L}{\partial u}) - \rho^{*}_{v} \displaystyle \frac{\partial
L}{\partial v},
\\[5pt]
u = \rho_{v}\omega.
\end{array}$$ Nonholonomic Lagrangian systems, of the above type, on the left action Lie algebroid $\tau_{E}: E = ({\frak s} \times TN) \times
{\mathcal O}_{v_{0}} \to M = N \times {\mathcal O}_{v_{0}}$ may be obtained (by reduction) from an standard nonholonomic Lagrangian system with semidirect product symmetry.
In fact, let $Q$ be the product manifold $S \times N$ and suppose that we have a Lagrangian function $\tilde{L}: TQ \to \R$ and a distribution $\tilde{D}$ on $Q$ whose characteristic space $\tilde{D}_{((g, v), n)} \subseteq T_{g}G \times T_{v}V \times
T_{n}N \simeq T_{g}G \times V \times T_{n}N$ at the point $((g,
v), n) \in S \times N$ is $$\label{Dtilv0}
\tilde{D}_{((g, v), n)} = \{((\dot{g}, \dot{v}), \dot{n}) \in
T_{g}G \times V \times T_{n}N / \dot{v} =
(T_{g}r_{g^{-1}})(\dot{g}) \cdot v_{0} \},$$ where $v_{0}$ is a fixed point of $V$.
We can consider the natural left action of the Lie group $S$ on $Q$ and, thus, the left action $A$ of the Lie subgroup $H_{v_{0}}
= G_{v_{0}} \circledS V$ of $S$ on $Q$, where $G_{v_{0}}$ is the isotropy group of $v_{0}$ with respect to the action of $G$ on $V$. The tangent lift $TA$ of $A$ is given by $$\label{TA}
TA ((\tilde{g}, \tilde{u}), (v_{g}, (v, \dot{v}), X_{n})) =
((T_{g}l_{\tilde{g}})(v_{g}), (\tilde{u} + \tilde{g} \cdot v,
\tilde{g} \cdot \dot{v}), X_{n})$$ for $(\tilde{g}, \tilde{u}) \in H_{v_{0}}$ and $(v_{g}, (v,
\dot{v}), X_{n}) \in T_{((g, v), n)}Q \simeq T_{g}G \times V
\times T_{n}N$.
Using (\[TA\]), it follows that the distribution $\tilde{D}$ is invariant under the action $TA$ of $H_{v_{0}}$ on $TQ$. Moreover, we will assume that the Lagrangian function is also $H_{v_{0}}$-invariant. Therefore, we have a nonholonomic Lagrangian system $(\tilde{L}, \tilde{D})$ on the standard Lie algebroid $TQ \to Q$ which is $H_{v_{0}}$-invariant. This type of systems were considered in [@MT:04].
Since the function $\tilde{L}$ is $H_{v_{0}}$-invariant, we deduce that there exists a real function $L: ({\frak s} \times TN) \times
{\mathcal O}_{v_{0}} \to \R$ on the left action Lie algebroid $\tau_{E}: E = ({\frak s} \times TN) \times {\mathcal O}_{v_{0}}
\to M = N \times {\mathcal O}_{v_{0}}$ which is defined by $$\label{Ele}
L(((\omega, \dot{v}), X_{n}), u) = \tilde{L}((T_{e}l_{g})(\omega),
(v, g \cdot \dot{v}), X_{n}),$$ for $(((\omega, \dot{v}), X_{n}), u) \in {\frak s} \times T_{n}N
\times {\mathcal O}_{v_{0}}$, with $g \in G$, $u = g^{-1}v_{0}$ and $v
\in V$.
Moreover, we may prove the following result.
1. If $\Phi: TQ \simeq TG \times (V \times V) \times TN \to E =
({\frak s} \times TN) \times {\mathcal O}_{v_{0}}$ and $\varphi: G
\times V \times N \to N \times {\mathcal O}_{v_{0}}$ are the maps defined by $$\label{Defmor}
\begin{array}{l}
\Phi(u_{g}, (v, \dot{v}), X_{n}) = ((((T_{g}l_{g^{-1}})(u_{g}),
g^{-1} \cdot \dot{v}), X_{n}), g^{-1} \cdot v_{0}), \\[5pt]
\varphi(g, v, n) = (n,
g^{-1} \cdot v_{0}),
\end{array}$$ then $\Phi$ is a fiberwise bijective Lie algebroid morphism over $\varphi$.
2. The nonholonomic Lagrangian systems $(\tilde{L}, \tilde{D})$ and $(L, D)$ on $TQ$ and $E = ({\frak s} \times TN)\times
{\mathcal O}_{v_{0}}$ are $\Phi$-related, that is, $$L \circ \Phi = \tilde{L}, \; \; \; \Phi(\tilde{D}) = D.$$ Here, $D$ is the vector subbundle of the vector bundle $E$ whose fiber at the point $(n, v) \in N \times {\mathcal O}_{v_{0}}$ is given by (\[D\]).
3. If the system $(\tilde{L}, \tilde{D})$ is regular then the system $(L, D)$ is also regular. In addition, if $\gamma: I \to
TQ$ is a solution of the Lagrange-d’Alembert equations for $(\tilde{L}, \tilde{D})$ then $\Phi \circ \gamma: I \to ({\frak s}
\times TN) \times {\mathcal O}_{v_{0}}$ is a solution of the Lagrange-d’Alembert equations for $(L, D)$.
$(1)$ Suppose that $\omega_{1}$ and $\omega_{2}$ are elements of ${\frak g}$, that $\dot{v}_{1}$ and $\dot{v}_{2}$ are vectors of $V$ and that $X_{1}$ and $X_{2}$ are vector fields on $N$. Then, we consider the vector fields $Z_{1}$ and $Z_{2}$ on $Q$ defined by $$\begin{array}{l}
Z_{1}(g, v, n) = ({\overleftarrow{\omega}}_{1}(g), g \cdot \dot{v}_{1},
X_{1}(n))
\in T_{g}G \times V \times T_{n}N, \\[5pt]
Z_{2}(g, v, n) = ({\overleftarrow{\omega}}_{2}(g), g \cdot \dot{v}_{2},
X_{2}(n)) \in T_{g}G \times V \times T_{n}N,
\end{array}$$ for $(g, v, n) \in G \times V \times N = Q$, where ${\overleftarrow{\omega}}_{1}$ (respectively, ${\overleftarrow{\omega}}_{2}$) is the left-invariant vector field on $G$ such that ${\overleftarrow{\omega}}_{1}(e)
= \omega_{1}$ (respectively, ${\overleftarrow{\omega}}_{2}(e) = \omega_{2}$), $e$ being the identity element of $G$.
A direct computation proves that $$[Z_{1}, Z_{2}] (g, v, n) = ({\overleftarrow{[\omega_{1},
\omega_{2}]}}_{{\frak g}}(g), g(\omega_{1} \cdot \dot{v}_{2} -
\omega_{2} \cdot \dot{v}_{1}), [X_{1}, X_{2}](n)).$$ Moreover, if $((\omega_{1}, \dot{v}_{1}), X_{1})$ (respectively, $((\omega_{2}, \dot{v}_{2}), X_{2})$) is the section of the vector bundle $\tau_{E}: E \to M$ induced by $\omega_{1}$, $\dot{v}_{1}$ and $X_{1}$ (respectively, $\omega_{2}$, $\dot{v}_{2}$ and $X_{2}$) then it is clear that $$\Phi \circ Z_{1} = ((\omega_{1}, \dot{v}_{1}), X_{1}) \circ
\varphi, \; \; \; \Phi \circ Z_{2} = ((\omega_{2}, \dot{v}_{2}),
X_{2}) \circ \varphi .$$ Thus, using (\[LiealgF\]), it follows that $$\label{Bracket}
\Phi \circ [Z_{1}, Z_{2}] = [((\omega_{1}, \dot{v}_{1}), X_{1}),
((\omega_{2}, \dot{v}_{2}), X_{2})]_{E} \circ \varphi.$$ On the other hand, we have that $$\begin{array}{l}
(T_{(g, v, n)}\varphi)(u_{g}, \dot{v}, X_{n}) = (X_{n},
-(T_{g}l_{g^{-1}})(u_{g}) \cdot (g^{-1} \cdot v_{0})) \\[5pt]
\in T_{n}N \times T_{g^{-1} \cdot v_{0}}{\mathcal O}_{v_{0}}
\subseteq T_{n}N \times V,
\end{array}$$ for $(g, v, n) \in Q$ and $(u_{g}, \dot{v}, X_{n}) \in T_{g}G \times
V \times T_{n}N \simeq T_{(g, v, n)}Q$.
Therefore, from (\[AnclaE\]) and (\[Defmor\]), we deduce that $$\label{Anchor}
T\varphi = \rho_{E} \circ \Phi.$$ Consequently, using (\[Bracket\]) and (\[Anchor\]), we conclude that the pair $(\Phi, \varphi)$ is a Lie algebroid morphism. Note that one may choose a local basis $\{Z_{i}\}$ of vector fields on $Q$ such that $$Z_{i}(g, v, n) = ({\overleftarrow{\omega}}_{i}(g), g \cdot \dot{v}_{i},
X_{i}(n)), \; \; \mbox{ for } (g, v, n) \in Q$$ with $\omega_{i} \in {\frak g}$, $\dot{v}_{i} \in V$ and $X_{i}
\in {\frak X}(N)$.
Finally, if $(g, v, n) \in Q$, it is clear that $$\Phi_{|T_{(g,v,n)}Q}: T_{(g,v,n)}Q \simeq T_{g}G \times V \times
T_{n}N \to E_{(n, g^{-1} \cdot v_{0})} \simeq {\frak g} \times V
\times T_{n}N$$ is a linear isomorphism.
$(2)$ It follows from (\[D\]), (\[Dtilv0\]), (\[Ele\]) and (\[Defmor\]).
$(3)$ It follows using $(1)$, $(2)$ and the results of Section \[sec:reduction\] (see Theorem \[t5.6\]).
The above theory may be applied to a particular example of a mechanical system: *the Chaplygin Gyro* (see [@Ma; @MT:04]). This system consists of a Chaplygin sphere (that is, a ball with nonhomogeneous mass distribution) with a gyro-like mechanism, consisting of a gimbal and a pendulous mass, installed in it. The gimbal is a circle-like structure such that its center coincides with the geometric center of the Chaplygin sphere. It is free to rotate about the axis connecting the north and south poles of the Chaplygin sphere. The pendulous mass can move along the smooth track of the gimbal. For this particular example, the vector space $V$ is $\R^{3}$, the Lie group $G$ is $SO(3)$ and the manifold $N$ is $\R^{2}$. The action of $SO(3)$ on $\R^{3}$ is the standard one and $v_{0} = (0, 0,
1)$ is the advected parameter, see [@MT:04] for more details.
Chaplygin-type systems
----------------------
A frequent situation is the following. Consider a constrained Lagrangian system $(L,D)$ on a Lie algebroid ${\tau\colonE\toM}$ such that the restriction of the anchor to the constraint distribution, ${\rho|D\colonD\toTM}$, is an isomorphism of vector bundles. Let ${h\colonTM\toD\subset E}$ be the right-inverse of $\rho|_D$, so that $\rho\circ h=\operatorname{id}_{TM}$. It follows that $E$ is a transitive Lie algebroid and $h$ is a splitting of the exact sequence $$\xymatrix{0\ar[r]&\operatorname{Ker}(\rho)\ar[r]&E\ar[r]^\rho&TM\ar[r]&0\,.}$$
Let us define the function $\bar{L}\in{C^\infty(TM)}$ by $\bar{L}=L\circ
h$. The dynamics defined by $L$ does not reduce to the dynamics defined by $\bar{L}$ because, while the map $\Phi=\rho$ is a morphism of Lie algebroids and $\Phi(D)=TM$, we have $\bar{L}\circ\Phi=L\circ
h\circ\rho\neq L$. Nevertheless, we can use $h$ to express the dynamics on $TM$, by finding relations between the dynamics defined by $L$ and $\bar{L}$.
We need some auxiliary properties of the splitting $h$ and its prolongation. We first notice that $h$ is an admissible map over the identity in $M$, because $\rho_E\circ h=\operatorname{id}_{TM}$ and $T\operatorname{id}_M\circ\rho_{TM}=id_{TM}$, but in general $h$ is not a morphism. We can define the tensor $K$, a $\ker(\rho)$-valued differential 2-form on $M$, by means of $$K(X,Y)=[h\circ X,h\circ Y]-h\circ[X,Y]$$ for every $X,Y\in\mathfrak{X}(M)$. It is easy to see that $h$ is a morphism if and only if $K=0$. In coordinates $(x^i)$ in $M$, $(x^i,v^i)$ in $TM$, and linear coordinates $(x^i,y^i,y^A)$ on $E$ corresponding to a local basis $\{e^i,e^A\}$ of sections of $E$ adapted to the splitting $h$, we have that $$K=\frac{1}{2}\Omega_{ij}^A\,dx^i\wedge dx^j\otimes e_A,$$ where $\Omega_{ij}^A$ are defined by $[e_i,e_j]=\Omega_{ij}^Ae_A$.
Since $h$ is admissible, its prolongation ${\@ifstar{\@proldf}{\@prolpf}}[h]{h}$ is a well-defined map from $T(TM)$ to ${\mathcal{T}^E_{}E}$. Moreover, it is an admissible map, which is a morphism if and only if $h$ is a morphism. In what respect to the energy and the Cartan 1-form, we have that $({\@ifstar{\@proldf}{\@prolpf}}[h]{h}){^\star}E_L=E_{\bar{L}}$ and $({\@ifstar{\@proldf}{\@prolpf}}[h]{h}){^\star}\theta_L=\theta_{\bar{L}}$. Indeed, notice that by definition, $({\@ifstar{\@proldf}{\@prolpf}}[h]{h}){^\star}E_L= E_L\circ h$ and $$\begin{aligned}
E_L(h(v))
&=\frac{d}{dt}L(h(v)+t(h(v))|_{t=0}-L(h(v))
=\frac{d}{dt}L(h(v+tv))|_{t=0}-L(h(v))\\
&=\frac{d}{dt}\bar{L}(v+tv)|_{t=0}-\bar{L}(v) =E_{\bar{L}}(v).\end{aligned}$$ On the other hand, for every $V_v\equiv (v,w,V)\in T(TM)\equiv
{\@ifstar{\@proldf}{\@prolpf}}[TM]{(TM)}$ where $w=T\tau(V)$, we have $$\begin{aligned}
{\left\langle({\@ifstar{\@proldf}{\@prolpf}}[h]{h}){^\star}\theta_L,V\right\rangle}
&={\left\langle\theta_L,{\@ifstar{\@proldf}{\@prolpf}}[h]{h}(v,w,V)\right\rangle}
={\left\langle\theta_L,(h(v),h(w),Th(V))\right\rangle}\\
&=\frac{d}{dt}L(h(v)+t(h(w))|_{t=0}
=\frac{d}{dt}L(h(v+tw))|_{t=0}\\
&=\frac{d}{dt}\bar{L}(v+tw)|_{t=0} ={\left\langle\theta_{\bar{L}},V\right\rangle}.\end{aligned}$$
Nevertheless, since $h$ is not a morphism, and hence $({\@ifstar{\@proldf}{\@prolpf}}[h]{h}){^\star}\circ d\neq d\circ ({\@ifstar{\@proldf}{\@prolpf}}[h]{h}){^\star}$, we have that $({\@ifstar{\@proldf}{\@prolpf}}[h]{h}){^\star}\omega_L\neq\omega_{\bar{L}}$. Let ${J\!K}$ be the 2-form on $TM$ defined by $${J\!K}_v(V,W)={\left\langleJ_{h(v)},K_{h(v)}(T\tau_M(V),T\tau_M(W))\right\rangle}$$ where $J$ is the momentum map defined by $L$ and $\operatorname{Ker}{\rho}$ and $V,W\in T_{h(v)}(TM)$. The notation resembles the contraction of the momentum map $J$ with the curvature tensor $K$. Instead of being symplectic, the map ${\@ifstar{\@proldf}{\@prolpf}}[h]{h}$ satisfies $$({\@ifstar{\@proldf}{\@prolpf}}[h]{h}){^\star}\omega_L=\omega_{\bar{L}}+{J\!K}.$$ Indeed, we have that $$({\@ifstar{\@proldf}{\@prolpf}}[h]{h}){^\star}\omega_L-\omega_{\bar{L}}=
[d\circ({\@ifstar{\@proldf}{\@prolpf}}[h]{h}){^\star}-({\@ifstar{\@proldf}{\@prolpf}}[h]{h}){^\star}\circ d]\,\theta_L$$ and on a pair of projectable vector fields $U,V$ projecting onto $X,Y$ respectively, one can easily prove that $$[d\circ({\@ifstar{\@proldf}{\@prolpf}}[h]{h}){^\star}-({\@ifstar{\@proldf}{\@prolpf}}[h]{h}){^\star}\circ d]\,\theta_L(U,V)
={\left\langle\theta_L,[{\@ifstar{\@proldf}{\@prolpf}}[h]{h}(U),{\@ifstar{\@proldf}{\@prolpf}}[h]{h}(V)]-{\@ifstar{\@proldf}{\@prolpf}}[h]{h}([U,V])\right\rangle}$$ from where the result follows by noticing that ${\@ifstar{\@proldf}{\@prolpf}}[h]{h}\circ
U$ is a projectable section and projects to $h\circ X$, and similarly ${\@ifstar{\@proldf}{\@prolpf}}[h]{h}\circ V$ projects to $h\circ Y$. Hence $[{\@ifstar{\@proldf}{\@prolpf}}[h]{h}(U),{\@ifstar{\@proldf}{\@prolpf}}[h]{h}(V)]-{\@ifstar{\@proldf}{\@prolpf}}[h]{h}([U,V]$ is projectable and projects to $K(X,Y)$.
Let now $\Gamma$ be the solution of the nonholonomic dynamics for $(L,
D)$, so that $\Gamma$ satisfies the equation $i_\Gamma\omega_L-dE_L\in{\widetilde{D_{}^\circ}}$ and the tangency condition $\Gamma\big|_D\in{\mathcal{T}^D_{}D}$. From this second condition we deduce the existence of a vector field $\bar{\Gamma}\in\mathfrak{X}(TM)$ such that ${\@ifstar{\@proldf}{\@prolpf}}[h]{h}\circ\bar{\Gamma}=\Gamma\circ h$. Explicitly, the vector field $\bar{\Gamma}$ is defined by $\bar{\Gamma}={\@ifstar{\@proldf}{\@prolpf}}[\rho]{\rho}\circ\Gamma\circ h$, from where it immediately follows that $\bar{\Gamma}$ is a [<span style="font-variant:small-caps;">sode</span>]{} vector field on $M$.
Taking the pullback by ${\@ifstar{\@proldf}{\@prolpf}}[h]{h}$ of the first equation we get $({\@ifstar{\@proldf}{\@prolpf}}[h]{h}){^\star}\bigl( i_\Gamma\omega_L-dE_L\bigr)=0$ since $({\@ifstar{\@proldf}{\@prolpf}}[h]{h}){^\star}{\widetilde{D_{}^\circ}}=0$. Therefore $$\begin{aligned}
0
&=({\@ifstar{\@proldf}{\@prolpf}}[h]{h}){^\star}i_\Gamma\omega_L-({\@ifstar{\@proldf}{\@prolpf}}[h]{h}){^\star}dE_L\\
&=i_{\bar{\Gamma}}({\@ifstar{\@proldf}{\@prolpf}}[h]{h}){^\star}\omega_L-d ({\@ifstar{\@proldf}{\@prolpf}}[h]{h}){^\star}E_L\\
&=i_{\bar{\Gamma}}\bigl(\omega_{\bar{L}}+{J\!K})-dE_{\bar{L}}\\
&=i_{\bar{\Gamma}}\omega_{\bar{L}}-dE_{\bar{L}}+i_{\bar{\Gamma}}{J\!K}.\end{aligned}$$ Therefore, the vector field $\bar{\Gamma}$ is determined by the equations $$i_{\bar{\Gamma}}\omega_{\bar{L}}-dE_{\bar{L}}=-{\left\langleJ,K(\mathbb{T},\,
\cdot\, )\right\rangle},$$ where $\mathbb{T}$ is the identity in $TM$ considered as a vector field along the tangent bundle projection $\tau_M$ (also known as the total time derivative operator). Equivalently we can write these equations in the form $$d_{\bar{\Gamma}}\theta_{\bar{L}}-d\bar{L}={\left\langleJ,K(\mathbb{T},\,
\cdot\, )\right\rangle}.$$
Note that if $\bar{a}: I \to TM$ is an integral curve of $\bar{\Gamma}$ then $a = h \circ \bar{a}: I \to D$ is a solution of the constrained dynamics for the nonholonomic Lagrangian system $(L,
D)$ on $E$. Conversely, if $a: I \to D$ is a solution of the constrained dynamics then $\rho \circ a: I \to TM$ is an integral curve of the vector field $\bar{\Gamma}$.
Finally we mention that extension of the above decomposition for non transitive Lie algebroids is under development.
Chaplygin systems and Atiyah algebroids {#chaplygin-systems-and-atiyah-algebroids .unnumbered}
---------------------------------------
A particular case of the above theory is that of ordinary Chaplygin systems(see [@BlKrMaMu; @CaCoLeMa; @cortes; @Ko] and references there in). In such case we have a principal $G$-bundle ${\pi\colonQ\toM=Q/G}$. Then, we may consider the quotient vector bundle $E = TQ/G \to M=Q/G$ and, it is well-known that, the space of sections of this vector bundle may be identified with the set of $G$-invariant vector fields on $Q$. Thus, using that the Lie bracket of two $G$-invariant vector fields is also $G$-invariant and the fact that a $G$-invariant vector field is $\pi$-projectable, we may define a Lie algebroid structure $([\cdot , \cdot], \rho)$ on the vector bundle $E= TQ/G \to M = Q/G$. The resultant Lie algebroid is called the [**Atiyah (gauge) algebroid**]{} associated with the principal bundle $\pi: Q \to M = Q/G$ (see [@Mackenzie]). Note that the canonical projection $\Phi: TQ \to E= TQ/G$ is a fiberwise bijective Lie algebroid morphism. Now, suppose that $(L_{Q}, D_{Q})$ is an standard nonholonomic Lagrangian system on $TQ$ such that $L_{Q}$ is $G$-invariant and $D_{Q}$ is the horizontal distribution of a principal connection on $\pi: Q \to M= Q/G$. Then, we have a reduced nonholonomic Lagrangian system $(L, D)$ on $E$. In fact, $L_{Q} = L
\circ \Phi$ and $\Phi((D_{Q})_{q}) = D_{\pi(q)}$, for all $q \in Q$. Moreover, $\rho_{|D}: D \to TM = T(Q/G)$ is an isomorphism (over the identity of M) between the vector bundles $D \to M$ and $TM \to M$. Therefore, we may apply the above general theory.
Next, we describe the nonholonomic Lagrangian system on the Atiyah algebroid associated with a particular example of a Chaplygin system: a two-wheeled planar mobile robot (see [@cortes] and the references there in). Consider the motion of two-wheeled planar mobile robot which is able to move in the direction in which it points and, in addition, can spin about a vertical axis. Let $P$ be the intersection point of the horizontal symmetry axis of the robot and the horizontal line connecting the centers of the two wheels. The position and orientation of the robot is determined, with respect to a fixed Cartesian reference frame by $(x, y, \theta) \in SE(2)$, where $\theta \in S^1$ is the heading angle, the coordinates $(x, y) \in
\R^{2}$ locate the point $P$ and $SE(2)$ is the group of Euclidean motions of the two-dimensional plane $\R^{2}$. Let $\psi_{1}, \psi_{2}
\in S^1$ denote the rotation angles of the wheels which are assumed to be controlled independently and roll without slipping on the floor. The configuration space of the system is $Q = \mathbb{T}^{2} \times
SE(2)$, where $\mathbb{T}^{2}$ is the real torus of dimension $2$.
The Lagrangian function $L_{Q}$ is the kinetic energy corresponding to the metric $g_{Q}$ $$\begin{array}{rcl}
g_{Q} &=& m dx \otimes dx + m dy \otimes dy + m_{0}l \cos \theta
(dy \otimes d\theta + d\theta \otimes dy)\\&& - m_{0}l \sin \theta
(dx \otimes d\theta + d\theta \otimes dx) + J d\theta \otimes
d\theta + J_{2} d\psi_{1} \otimes d\psi_{1} + J_{2} d\psi_{2}
\otimes d\psi_{2},
\end{array}$$ where $m = m_{0} + 2m_{1}$, $m_{0}$ is the mass of the robot without the wheels, $J$ its momenta of inertia with respect to the vertical axis, $m_{1}$ the mass of each wheel, $J_{2}$ the axial moments of inertia of the wheels, and $l$ the distance between the center of mass $C$ of the robot and the point $P$. Thus, $$\begin{array}{rcl}
L_{Q} &=& \displaystyle \frac{1}{2} (m \dot{x}^{2} + m \dot{y}^{2}
+ 2 m_{0}l \dot{y}\dot{\theta} \cos \theta - 2 m_{0}l
\dot{x}\dot{\theta}\sin \theta \\ && + J\dot{\theta}^{2} +
J_{2}\dot{\psi}_{1}^{2} + J_{2}\dot{\psi}_{2}^{2}).
\end{array}$$
The constraints, induced by the conditions that there is no lateral sliding of the robot and that the motion of the wheels also consists of a rolling without sliding, are $$\begin{array}{rcl}
\dot{x}\sin \theta - \dot{y} \cos \theta &=& 0,\\
\dot{x}\cos \theta + \dot{y}\sin \theta + c\dot{\theta} +
R\dot{\psi}_{1} &= & 0,
\\
\dot{x}\cos \theta + \dot{y}\sin \theta - c\dot{\theta} +
R\dot{\psi}_{2} &=& 0,
\end{array}$$ where $R$ is the radius of the wheels and $2c$ the lateral length of the robot. The constraint distribution $D$ is then spanned by $$\begin{array}{lr}
\{H_{1} = \displaystyle \frac{\partial}{\partial \psi_{1}} -
\frac{R}{2}(\cos \theta \frac{\partial}{\partial x} +\sin \theta
\frac{\partial}{\partial y} + \frac{1}{c} \frac{\partial}{\partial
\theta}),& \\
& \kern-38pt H_{2} = \displaystyle \frac{\partial}{\partial \psi_{2}} -
\frac{R}{2}(\cos \theta \frac{\partial}{\partial x} +\sin \theta
\frac{\partial}{\partial y} - \frac{1}{c} \frac{\partial}{\partial
\theta})\},
\end{array}$$
Note that if $\{\xi_{1}, \xi_{2}, \xi_{3}\}$ is the canonical basis of ${\frak se}(2)$, $$[\xi_{1}, \xi_{2}] = 0, \; \; [\xi_{1}, \xi_{3}] = -\xi_{2}, \; \;
[\xi_{2}, \xi_{3}] = \xi_{1},$$ then $$H_{1} = \displaystyle \frac{\partial}{\partial \psi_{1}} -
\frac{R}{2} {\overleftarrow{\xi_{1}}} - \frac{R}{2c} {\overleftarrow{\xi_{3}}}, \; \;
H_{2} = \displaystyle \frac{\partial}{\partial \psi_{2}} -
\frac{R}{2} {\overleftarrow{\xi_{1}}} + \frac{R}{2c} {\overleftarrow{\xi_{3}}},$$ where ${\overleftarrow{\xi_{i}}}$ ($i = 1, 2, 3$) is the left-invariant vector field of $SE(2)$ such that ${\overleftarrow{\xi_{i}}}(e) = \xi_{i}$, $e$ being the identity element of $SE(2)$.
On the other hand, it is clear that $Q = \mathbb{T}^{2} \times SE(2)$ is the total space of a trivial principal $SE(2)$-bundle over $M =
\mathbb{T}^{2}$. Moreover, the metric $g_{Q}$ is $SE(2)$-invariant and $D_{Q}$ is the horizontal distribution of a principal connection on $Q
= \mathbb{T}^{2} \times SE(2) \to \mathbb{T}^{2}$.
Now, we consider the corresponding Atiyah algebroid $$E = TQ/SE(2) \simeq (T\mathbb{T}^{2} \times TSE(2))/SE(2) \to M =
\mathbb{T}^{2}.$$ Using the left-translations on $SE(2)$, we have that the tangent bundle to $SE(2)$ may be identified with the product manifold $SE(2)
\times {\frak se}(2)$ and, under this identification, the Atiyah algebroid is isomorphic to the trivial vector bundle $$\tilde{\tau}_{\mathbb{T}^{2}} = \tau_{\mathbb{T}^{2}} \circ
pr_{1}: T\mathbb{T}^{2} \times {\frak se}(2) \to \mathbb{T}^{2},$$ where $\tau_{\mathbb{T}^{2}}: T\mathbb{T}^{2} \to \mathbb{T}^{2}$ is the canonical projection. In addition, if $([ \cdot , \cdot ],
\rho)$ is the Lie algebroid structure on $\tilde{\tau}_{\mathbb{T}^{2}}: T\mathbb{T}^{2} \times {\frak
se}(2) \to \mathbb{T}^{2}$ and $\{\displaystyle
\frac{\partial}{\partial \psi_{1}}, \frac{\partial}{\partial
\psi_{2}}, \xi_{1}, \xi_{2}, \xi_{3} \}$ is the canonical basis of sections of $\tilde{\tau}_{\mathbb{T}^{2}}: T\mathbb{T}^{2} \times
{\frak se}(2) \to \mathbb{T}^{2}$ then $$\begin{array}{rclrclrclrcl}
\rho(\displaystyle \frac{\partial}{\partial \psi_{1}}) &=&
\displaystyle \frac{\partial}{\partial \psi_{1}}, \; \; &
\rho(\displaystyle \frac{\partial}{\partial \psi_{2}}) & = &
\displaystyle \frac{\partial}{\partial
\psi_{2}}, \; \; &\rho(\xi_{i}) & = &0, \; \;& i = 1, 2, 3 \\[8pt]
[\xi_{1}, \xi_{3}] & = & -\xi_{2}, \; \; & [\xi_{2}, \xi_{3}] & =
& \xi_{1}, && &&
\end{array}$$ and the rest of the fundamental Lie brackets are zero.
Denote by $(\psi_{1}, \psi_{2}, \dot{\psi}_{1}, \dot{\psi}_{2},
\omega^{1}, \omega^{2}, \omega^{3})$ the (local) coordinates on $T\mathbb{T}^{2} \times {\frak se}(2)$ induced by the basis $\{\displaystyle \frac{\partial}{\partial \psi_{1}},
\frac{\partial}{\partial \psi_{2}}, \xi_{1}, \xi_{2}, \xi_{3} \}$. Then, the reduced Lagrangian $L: T\mathbb{T}^{2} \times {\frak se}(2)
\to \R$ is given by $$L = \displaystyle \frac{1}{2} (m (\omega^{1})^{2} + m
(\omega^{2})^{2} + 2 m_{0}l \omega^{2}\omega^{3} +
J(\omega^{3})^{2} + J_{2}\dot{\psi}_{1}^{2} +
J_{2}\dot{\psi}_{2}^{2})$$ and the constraint vector subbundle $D$ is generated by the sections $$e_{1} = \displaystyle \frac{\partial}{\partial \psi_{1}} -
\frac{R}{2}\xi_{1} - \frac{R}{2c} \xi_{3}, \; \; e_{2} =
\displaystyle \frac{\partial}{\partial \psi_{2}} - \frac{R}{2}
\xi_{1} + \frac{R}{2c} \xi_{3}.$$ Since the system $(L_{Q}, D_{Q})$ is regular on the standard Lie algebroid $\tau_{Q}: TQ \to Q$, we deduce that the nonholonomic Lagrangian system $(L, D)$ on the Atiyah algebroid $\tilde{\tau}_{\mathbb{T}^{2}}: T\mathbb{T}^{2} \times {\frak se}(2)
\to \mathbb{T}^{2}$ is also regular.
Now, as in Section \[linear\], we consider a basis of sections of $\tilde{\tau}_{\mathbb{T}^{2}}: T\mathbb{T}^{2} \times {\frak se}(2)
\to \mathbb{T}^{2}$ which is adapted to the constraint subbundle $D$. This basis is $$\{e_{1}, e_{2}, \xi_{1}, \xi_{2}, \xi_{3}\}.$$ The corresponding (local) coordinates on $T\mathbb{T}^{2} \times
{\frak se}(2)$ are $(\psi_{1}, \psi_{2}, \dot{\psi}_{1},
\dot{\psi}_{2}, \tilde{\omega}^{1}, \tilde{\omega}^{2},
\tilde{\omega}^{3})$, where $$\omega^{1} = \tilde{\omega}^{1} - \displaystyle \frac{R}{2}
\dot{\psi}_{1} - \frac{R}{2}\dot{\psi}_{2}, \; \; \omega^{2} =
\tilde{\omega}^{2}, \; \; \omega^{3} = \tilde{\omega}^{3} -
\displaystyle \frac{R}{2c} \dot{\psi}_{1} +
\frac{R}{2c}\dot{\psi}_{2}.$$ Therefore, using (\[LD-edo\]), we deduce that the Lagrange-d’Alembert equations for the system $(L, D)$ are $$\begin{array}{rclrcl}
\ddot{\psi}_{1} & = & \displaystyle \frac{U(\dot{\psi}_{2} -
\dot{\psi}_{1})}{P^{2} - S^{2}} (P\dot{\psi}_{2} +
S\dot{\psi}_{1}), \; \; & \ddot{\psi}_{2} & = & \displaystyle -
\frac{U(\dot{\psi}_{2} - \dot{\psi}_{1})}{P^{2} - S^{2}}
(P\dot{\psi}_{1} + S\dot{\psi}_{2}),\\ [8pt]
\tilde{\omega}^{1}
&=& \tilde{\omega}^{2} = \tilde{\omega}^{3} = 0,&&&
\end{array}$$ where $P$, $S$ and $U$ are the real numbers $$P = \displaystyle \frac{R^{2}}{4} (m + \frac{J}{c^{2}}) + J_{2},
\; \; S = \displaystyle \frac{R^{2}}{4} (m - \frac{J}{c^{2}}), \;
\; U = \displaystyle \frac{R^{3}}{4c^{2}}m_{0}l.$$ On the other hand, the Lagrangian function $\bar{L}:
T\mathbb{T}^{2} \to \R$ on $T\mathbb{T}^{2}$ is given by $$\bar{L}(\psi_{1}, \psi_{2}, \dot{\psi}_{1}, \dot{\psi}_{2}) =
\displaystyle \frac{1}{2} (P \dot{\psi}_{1}^{2} + P
\dot{\psi}_{2}^{2} + 2S \dot{\psi}_{1} \dot{\psi}_{2})$$ and the $1$-form ${\left\langleJ,K(\mathbb{T},\, \cdot\, )\right\rangle}$ on $T\mathbb{T}^{2}$ is $${\left\langleJ,K(\mathbb{T},\, \cdot\, )\right\rangle} = -U (\dot{\psi}_{2} -
\dot{\psi}_{1}) (\dot{\psi}_{1} d\psi_{2} - \dot{\psi}_{2}
d\psi_{1}).$$
Nonlinearly constrained Lagrangian systems {#nonlinear}
==========================================
We show in this section how the main results for linearly constrained Lagrangian systems can be extended to the case of Lagrangian systems with nonlinear nonholonomic constraints. This is true under the assumption that a suitable version of the classical Chetaev’s principle in nonholonomic mechanics is valid (see e.g., [@LeMa2] for the study of standard nonholonomic Lagrangian systems subject to nonlinear constraints).
Let $\tau: E \to M$ be a Lie algebroid and ${\mathcal{M}}$ be a submanifold of $E$ such that ${\pi=\tau|_{\mathcal{M}}\colon{\mathcal{M}}\toM}$ is a fibration. ${\mathcal{M}}$ is the constraint submanifold. Since $\pi$ is a fibration, the prolongation ${\@ifstar{\@proldf}{\@prolpf}}[E]{{\mathcal{M}}}$ is well-defined. We will denote by $r$ the dimension of the fibers of ${\pi\colon{\mathcal{M}}\toM}$, that is, $r=\operatorname{dim}{\mathcal{M}}-\operatorname{dim} M$.
We define the bundle ${\mathcal{V}_{}}\to {\mathcal{M}}$ of *virtual displacements* as the subbundle of $\tau^*E$ of rank $r$ whose fiber at a point $a\in{\mathcal{M}}$ is $${\mathcal{V}_{a}}={\{\,b\in E_{\tau(a)}\,|\,b_a{{^{\scriptscriptstyleV}}}\in T_a{\mathcal{M}}\,\}}.$$ In other words, the elements of ${\mathcal{V}_{}}$ are pairs of elements $(a,b)\in E\oplus E$ such that $$\frac{d}{dt}\phi(a+tb){\Big|_{t=0}}=0,$$ for every local constraint function $\phi$.
We also define the bundle of *constraint forces* ${\boldsymbol{\Psi}}$ by ${\boldsymbol{\Psi}}=S^*(({\@ifstar{\@proldf}{\@prolpf}}[E]{{\mathcal{M}}})^\circ)$, in terms of which we set the Lagrange-d’Alembert equations for a regular Lagrangian function $L \in
C^{\infty}(E)$ as follows: $$\label{8.1}
\begin{array}{ll}
&(i_\Gamma\omega_L-dE_L)|_{\mathcal{M}}\in{\operatorname{Sec}_{}({\boldsymbol{\Psi}})}, \\[5pt]
&\Gamma|_{\mathcal{M}}\in{\operatorname{Sec}_{}({\@ifstar{\@proldf}{\@prolpf}}[E]{{\mathcal{M}}})},
\end{array}$$ the unknown being the section $\Gamma$. The above equations reproduce the corresponding ones for standard nonlinear constrained systems.
From (\[2.4’\]) and (\[8.1\]), it follows that $$(i_{S\Gamma}\omega_{L} - i_{\Delta}\omega_{L})|_{\mathcal{M}}= 0,$$ which implies that a solution $\Gamma$ of equations (\[8.1\]) is a [<span style="font-variant:small-caps;">sode</span>]{} section along ${\mathcal{M}}$, that is, $(S\Gamma - \Delta)| {\mathcal{M}}= 0$.
Note that the rank of the vector bundle $({\@ifstar{\@proldf}{\@prolpf}}[E]{{\mathcal{M}}})^\circ \to
{\mathcal{M}}$ is $s = {\operatorname{rank}(E)}-r$ and, since $\pi$ is a fibration, the transformation $S^*: ({\@ifstar{\@proldf}{\@prolpf}}[E]{{\mathcal{M}}})^{\circ} \to \Psi $ defines an isomorphism between the vector bundles $({\@ifstar{\@proldf}{\@prolpf}}[E]{{\mathcal{M}}})^{\circ}
\to {\mathcal{M}}$ and $\Psi \to {\mathcal{M}}$. Therefore, the rank of $\Psi$ is also $s$. Moreover, if $a \in {\mathcal{M}}$ we have $$\label{psia}
\Psi_a=S^*(({\@ifstar{\@proldf}{\@prolpf}}[E]{{\mathcal{M}}}[a])^\circ)={\{\,\zeta\circ
{\@ifstar{\@proldf}{\@prolpf}}{\tau}\,|\,\zeta\in {\mathcal V}_a^\circ\,\}}.$$ In fact, if $\alpha_a\in ({\@ifstar{\@proldf}{\@prolpf}}[E]{{\mathcal{M}}}[a])^\circ$, we may define $\zeta\in E_{\tau(a)}^*$ by $$\zeta(b)=\alpha_a(\xi{{^{\scriptscriptstyleV}}}(a,b)),\mbox{
for } b\in E_{\tau(a)}.$$ Then, a direct computation proves that $\zeta\in {\mathcal V}_a^\circ$ and $S^*(\alpha_a)=\zeta\circ {\@ifstar{\@proldf}{\@prolpf}}{\tau}$. Thus, we obtain $$\Psi_a\subseteq {\{\,\zeta\circ {\@ifstar{\@proldf}{\@prolpf}}{\tau}\,|\,\zeta\in {\mathcal
V}_a^\circ\,\}}$$ and, using that the dimension of both spaces is $s$, we deduce (\[psia\]) holds. Note that, in the particular case when the constraints are linear, we have ${\mathcal V}=\tau^*(D)$ and $\Psi=\widetilde{D^\circ}.$
Next, we consider the vector bundles $F$ and ${\@ifstar{\@proldf}{\@prolpf}}[{\mathcal{V}_{}}]{{\mathcal{M}}}$ over ${{\mathcal{M}}}$ whose fibers at the point $a\in {{\mathcal{M}}}$ are $$F_{a} = \omega_{L}^{-1}(\Psi_{a}), \makebox[.4cm]{}
{\@ifstar{\@proldf}{\@prolpf}}[{\mathcal{V}_{}}]{{\mathcal{M}}}[a]={\{\,(b,v)\in{\mathcal{V}_{a}}\times
T_a{\mathcal{M}}\,|\,T\pi(v)=\rho(b)\,\}}.$$ It follows that $$F_{a} = {\{\, z \in {\@ifstar{\@proldf}{\@prolpf}}[E]{E}[a]\,|\,\mbox{ exists } \zeta \in {\mathcal{V}}_{a}^0
\mbox{ and } i_{z}\omega_{L}(a) = \zeta \circ {\@ifstar{\@proldf}{\@prolpf}}{\tau}\,\}}$$ and $$\label{8.1'}
{\@ifstar{\@proldf}{\@prolpf}}[{\mathcal{V}_{}}]{{\mathcal{M}}}[a]={\{\,z\in{\@ifstar{\@proldf}{\@prolpf}}[E]{{\mathcal{M}}}[a]\,|\,{\@ifstar{\@proldf}{\@prolpf}}{\pi}(z)\in{\mathcal{V}_{a}}\,\}}={\{\,z\in
{\@ifstar{\@proldf}{\@prolpf}}[E]{{\mathcal{M}}}[a]\,|\,S(z)\in {\@ifstar{\@proldf}{\@prolpf}}[E]{{\mathcal{M}}}[a]\,\}}.$$ Note that the dimension of ${\@ifstar{\@proldf}{\@prolpf}}[{\mathcal{V}_{}}]{{\mathcal{M}}}[a]$ is $2r$ and, when the constraints are linear, i.e., ${\mathcal M}$ is a vector subbundle $D$ of $E$, we obtain $${\@ifstar{\@proldf}{\@prolpf}}[{\mathcal{V}_{}}]{{\mathcal{M}}}[a]={\@ifstar{\@proldf}{\@prolpf}}[D]{D}[a], \mbox{ for all } a\in {\mathcal{M}}=
D.$$ Moreover, from (\[8.1’\]), we deduce that the vertical lift of an element of ${\mathcal{V}_{}}$ is an element of ${\@ifstar{\@proldf}{\@prolpf}}[{\mathcal{V}_{}}]{{\mathcal{M}}}$. Thus we can define for $b,c\in{\mathcal{V}_{a}}$ $${G_{a}{^{\scriptscriptstyleL,{\mathcal{V}_{}}}}}(b,c)=\omega_L(a)(\tilde{b},\xi{{^{\scriptscriptstyleV}}}(a,c)),$$ where $\tilde{b} \in {\@ifstar{\@proldf}{\@prolpf}}[E]{E}[a]$ and ${\@ifstar{\@proldf}{\@prolpf}}{\tau}(\tilde{b}) =
b$.
Dynamics in local coordinates
-----------------------------
Here we analyze the local nature of equations (\[8.1\]). We consider local coordinates $(x^i)$ on an open subset $U$ of $M$ and take a basis $\{e_{\alpha}\}$ of local sections of $E$. In this way, we have local coordinates $(x^i, y^{\alpha})$ on $E$. Suppose that the local equations defining ${\mathcal{M}}$ as a submanifold of $E$ are $$\phi^{A}= 0, \makebox[.4cm]{} A = 1, \dots , s ,$$ where $\phi^{A}$ are independent local constraint functions. Since $\pi: {\mathcal{M}}\to M$ is a fibration, it follows that the matrix $\displaystyle (\frac{\partial \phi^{A}}{\partial y^{\alpha}})$ is of rank $s$. Thus, if $d$ is the differential of the Lie algebroid ${\@ifstar{\@proldf}{\@prolpf}}[E]{E} \to E$, we deduce that $\{d\phi^{A}|_{\mathcal{M}}\}_{A=1, \dots ,
s}$ is a local basis of sections of the vector bundle $({\@ifstar{\@proldf}{\@prolpf}}[E]{{\mathcal{M}}})^{0} \to {\mathcal{M}}$. Note that $$d\phi^{A} = \displaystyle \rho^j_{\alpha}\frac{\partial
\phi^{A}}{\partial x^j}{\mathcal{X}}^{\alpha} + \frac{\partial
\phi^{A}}{\partial y^{\alpha}}{\mathcal{V}}^{\alpha}.$$ Moreover, $\{S^*(d\phi^{A})|_{\mathcal{M}}= \displaystyle \frac{\partial
\phi^{A}}{\partial y^{\alpha}} {\mathcal{X}}^{\alpha}|_{\mathcal{M}}\}_{A=1, \dots ,
s}$ is a local basis of sections of the vector bundle $\Psi \to
{\mathcal{M}}$.
Next, we introduce the local sections $\{Z_{A}\}_{A=1, \dots , s}$ of ${\@ifstar{\@proldf}{\@prolpf}}[E]{E} \to E$ defined by $$i_{Z_{A}} \omega_{L} = S^*(d\phi^{A}) = \displaystyle
\frac{\partial \phi^{A}}{\partial y^{\alpha}} {\mathcal{X}}^{\alpha}.$$ A direct computation, using (\[omegaL\]), proves that $$\label{Zeta}
Z_{A} = \displaystyle -\frac{\partial \phi^{A}}{\partial
y^{\alpha}} W^{\alpha \beta}{\mathcal{V}}_{\beta}, \makebox[.4cm]{} \mbox{
for all } A,$$ where $(W^{\alpha \beta})$ is the inverse matrix of $(W_{\alpha
\beta} = \displaystyle \frac{\partial^{2}L}{\partial
y^{\alpha}y^{\beta}})$. Furthermore, it is clear that $\{Z_{A}|_{\mathcal{M}}\}$ is a local basis of sections of the vector bundle $F \to {\mathcal{M}}$.
On the other hand, if $\Gamma_L$ is the Euler-Lagrange section associated with the regular Lagrangian $L$, then a section $\Gamma$ of ${\@ifstar{\@proldf}{\@prolpf}}[E]{{\mathcal{M}}}\to {\mathcal{M}}$ is a solution of equations (\[8.1\]) if and only if $$\Gamma=(\Gamma_L + \lambda^A Z_A)|_{\mathcal{M}}$$ with $\lambda^A$ local real functions on $E$ satisfying $$(\lambda^Ad\phi^B(Z_A) + d\phi^B(\Gamma_L))|_{\mathcal{M}}\ =0, \mbox{ for
all } B=1,\dots ,s.$$ Therefore, using (\[Zeta\]), we conclude that there exists a unique solution of the Lagrange-d’Alembert equations (\[8.1\]) if and only if the matrix $$\begin{aligned}
\label{eq:matrix}
\big({\mathcal C}^{AB} = \frac{\partial \phi^A}{\partial
y^\alpha}W^{\alpha\beta}\frac{\partial \phi^B}{\partial
y^\beta}\big)_{A,B=1,\dots ,s}\end{aligned}$$ is regular. We are now ready to prove the following result.
\[regular-nonl\] The following properties are equivalent:
1. The constrained Lagrangian system $(L,{\mathcal{M}})$ is regular, that is, there exists a unique solution of the Lagrange-d’Alembert equations,
2. $\operatorname{Ker}{G_{}{^{\scriptscriptstyleL,{\mathcal{V}_{}}}}}=\{0\}$,
3. ${\@ifstar{\@proldf}{\@prolpf}}[E]{{\mathcal{M}}} \cap F=\{0\}$,
4. ${\@ifstar{\@proldf}{\@prolpf}}[{\mathcal V}]{{\mathcal{M}}} \cap({\@ifstar{\@proldf}{\@prolpf}}[{\mathcal
V}]{{\mathcal{M}}})^\perp=\{0\}$.
It is clear that the matrix $({\mathcal C}^{AB})$ in is regular if and only if ${\@ifstar{\@proldf}{\@prolpf}}[E]{{\mathcal{M}}}\cap
F=\{0\}$. Thus, the properties (1) and (3) are equivalent. Moreover, proceeding as in the proof of Theorem \[regularity\], we deduce that the properties (2) and (3) (respectively, (2) and (4)) also are equivalent.
\[mechty\] [If $L$ is a Lagrangian function of mechanical type, then, using Theorem \[regular-nonl\], we deduce (as in the case of linear constraints) that the constrained system $(L, {\mathcal{M}})$ is always regular.]{}
Lagrange-d’Alembert solutions and nonholonomic bracket
------------------------------------------------------
Assume that the constrained Lagrangian system $(L, {\mathcal{M}})$ is regular. Then (3) in Theorem \[regular-nonl\] is equivalent to $({\@ifstar{\@proldf}{\@prolpf}}[E]{E})|_{\mathcal{M}}\ = {\@ifstar{\@proldf}{\@prolpf}}[E]{{\mathcal{M}}} \oplus F$. Denote by $P$ and $Q$ the complementary projectors defined by this decomposition $$P_{a}: {\@ifstar{\@proldf}{\@prolpf}}[E]{E}[a] \to {\@ifstar{\@proldf}{\@prolpf}}[E]{{\mathcal{M}}}[a], \makebox[.3cm]{} Q_{a}:
{\@ifstar{\@proldf}{\@prolpf}}[E]{E}[a] \to F_a, \; \; \mbox{ for all } a \in {\mathcal{M}}.$$ As in the case of linear constraints, we may prove the following.
\[dym-nonl\] Let $(L, {\mathcal{M}})$ be a regular constrained Lagrangian system and let $\Gamma_{L}$ be the solution of the free dynamics, i.e., $i_{{\Gamma}_{L}}\omega_{L} = dE_{L}$. Then, the solution of the constrained dynamics is the [<span style="font-variant:small-caps;">sode</span>]{} $\Gamma_{(L, {\mathcal{M}})}$ obtained as follows $$\Gamma_{(L, {\mathcal{M}})} = P(\Gamma_{L}|_{\mathcal{M}}).$$
On the other hand, $(4)$ in Theorem \[regular-nonl\] is equivalent to $({\@ifstar{\@proldf}{\@prolpf}}[E]{E})|_{\mathcal{M}}\ ={\@ifstar{\@proldf}{\@prolpf}}[{\mathcal{V}_{}}]{{\mathcal{M}}}\oplus
({\@ifstar{\@proldf}{\@prolpf}}[{\mathcal{V}_{}}]{{\mathcal{M}}})^\perp$ and we will denote by $\bar{P}$ and $\bar{Q}$ the corresponding projectors induced by this decomposition, that is, $$\bar{P}_a:{\@ifstar{\@proldf}{\@prolpf}}[E]{E}[a]\to {\@ifstar{\@proldf}{\@prolpf}}[{\mathcal{V}_{}}]{{\mathcal{M}}}[a],\;\;\;
\bar{Q}_a:{\@ifstar{\@proldf}{\@prolpf}}[E]{E}[a]\to ({\@ifstar{\@proldf}{\@prolpf}}[{\mathcal{V}_{}}]{{\mathcal{M}}}[a])^\perp, \mbox{ for
all } a\in {\mathcal{M}}.$$
\[t8.3\] Let $(L,{\mathcal{M}})$ be a regular constrained Lagrangian system, $\Gamma_L$ (respectively, $\Gamma_{(L, {\mathcal{M}})}$) be the solution of the free (respectively, constrained) dynamics and $\Delta$ be the Liouville section of ${\@ifstar{\@proldf}{\@prolpf}}[E]{E}\to E$. Then, $\Gamma_{(L,
{\mathcal{M}})}=\bar{P}(\Gamma_L|_{{\mathcal{M}}})$ if and only if the restriction to ${{\mathcal{M}}}$ of the vector field $\rho^1(\Delta)$ on $E$ is tangent to ${\mathcal{M}}$.
Proceeding as in the proof of Lemma \[F-TDD\], we obtain that $$({\@ifstar{\@proldf}{\@prolpf}}[{\mathcal{V}_{}}]{{\mathcal{M}}}[a])^\perp \cap {\operatorname{Ver}({\@ifstar{\@proldf}{\@prolpf}}[E]{E}[a])}=F_a,\mbox{
for all }a\in {\mathcal{M}}.$$ Thus, it is clear that $$Q(\Gamma_L(a))\in F_a\subseteq ({\@ifstar{\@proldf}{\@prolpf}}[{\mathcal{V}_{}}]{{\mathcal{M}}}[a])^\perp, \mbox{
for all } a\in {\mathcal{M}}.$$ Moreover, from (\[8.1’\]) and using the fact that the solution of the constrained dynamics is a [<span style="font-variant:small-caps;">sode</span>]{} along ${\mathcal{M}}$, we deduce $$\Gamma_{(L, {\mathcal{M}})}(a) = P(\Gamma_{L}(a)) \in {\@ifstar{\@proldf}{\@prolpf}}[{\mathcal{V}_{}}]{{\mathcal{M}}}[a], \;
\; \mbox{ for all } a \in {\mathcal{M}},$$ if and only if the restriction to ${\mathcal{M}}$ of the vector field $\rho^1(\Delta)$ on $E$ is tangent to ${\mathcal{M}}$. This proves the result.
[Note that if ${\mathcal{M}}$ is a vector subbundle $D$ of $E$, then the vector field $\rho^1(\Delta)$ is always tangent to ${\mathcal{M}}= D$.]{}
As in the case of linear constraints, one may develop the distributional approach in order to obtain the solution of the constrained dynamics. In fact, if $(L, {\mathcal{M}})$ is regular, then ${\@ifstar{\@proldf}{\@prolpf}}[{\mathcal{V}_{}}]{{\mathcal{M}}} \to {\mathcal{M}}$ is a symplectic subbundle of $({\@ifstar{\@proldf}{\@prolpf}}[E]{E}, \omega_{L})$ and, thus, the restriction $\omega^{L,
{\mathcal{M}}}$ of $\omega_{L}$ to ${\@ifstar{\@proldf}{\@prolpf}}[{\mathcal{V}_{}}]{{\mathcal{M}}}$ is a symplectic section on that bundle. We may also define $\varepsilon^{L, {\mathcal{M}}}$ as the restriction of $dE_{L}$ to ${\@ifstar{\@proldf}{\@prolpf}}[{\mathcal{V}_{}}]{{\mathcal{M}}}$. Then, taking the restriction of Lagrange-d’Alembert equations to ${\@ifstar{\@proldf}{\@prolpf}}[{\mathcal{V}_{}}]{{\mathcal{M}}}$, we get the following equation $$\label{disapp}
i_{\bar{\Gamma}}\omega^{L, {\mathcal{M}}} = \varepsilon ^{L,
{\mathcal{M}}} ,$$ which uniquely determines a section $\bar{\Gamma}$ of ${\@ifstar{\@proldf}{\@prolpf}}[{\mathcal{V}_{}}]{{\mathcal{M}}}
\to {\mathcal{M}}$. It is not difficult to prove that $\bar{\Gamma} =
\bar{P}(\Gamma_L|_{{\mathcal{M}}})$. Thus, the unique solution of equation (\[disapp\]) is the solution of the constrained dynamics if and only if the vector field $\rho^{1}(\Delta)$ is tangent to ${\mathcal{M}}$.
Let $(L,{\mathcal{M}})$ a regular constrained Lagrangian system. Since $S^*:
({\@ifstar{\@proldf}{\@prolpf}}[E]{{\mathcal{M}}})^0 \to \Psi$ is a vector bundle isomorphism, it follows that there exists a unique section $\alpha_{(L, {\mathcal{M}})}$ of $({\@ifstar{\@proldf}{\@prolpf}}[E]{{\mathcal{M}}})^0 \to {\mathcal{M}}$ such that $$i_{Q(\Gamma_{L}|_{{\mathcal{M}}})}\omega_{L} = S^*(\alpha_{(L, {\mathcal{M}})}).$$ Moreover, we have the following result.
\[conser-ener\] If $(L, {\mathcal{M}})$ is a regular constrained Lagrangian system and $\Gamma_{(L, {\mathcal{M}})}$ is the solution of the dynamics, then $d_{\Gamma_{(L, {\mathcal{M}})}}(E_{L}|_{{\mathcal{M}}}) = 0$ if and only if $\alpha_{(L, {\mathcal{M}})}(\Delta|_{{\mathcal{M}}}) = 0$. In particular, if the vector field $\rho^{1}(\Delta)$ is tangent to ${\mathcal{M}}$, then $d_{\Gamma_{(L,
{\mathcal{M}})}}(E_{L}|_{{\mathcal{M}}}) = 0$.
From Theorem \[dym-nonl\], we deduce $$(i_{\Gamma_{(L, {\mathcal{M}})}}\omega_{L} - dE_{L})|_{{\mathcal{M}}} = -S^*(\alpha_{(L, {\mathcal{M}})}).$$ Therefore, using that $\Gamma_{(L, {\mathcal{M}})}$ is a [<span style="font-variant:small-caps;">sode</span>]{} along ${\mathcal{M}}$, we obtain $$d_{\Gamma_{(L, {\mathcal{M}})}}(E_{L}|_{{\mathcal{M}}}) = \alpha_{(L, {\mathcal{M}})}(\Delta|_{{\mathcal{M}}}).$$
Now, let $(L, {\mathcal{M}})$ be a regular constrained Lagrangian system. In addition, suppose that $f$ and $g$ are two smooth functions on ${\mathcal{M}}$ and take arbitrary extensions to $E$ denoted by the same letters. Then, as in Section \[NHBracket\], we may define *the nonholonomic bracket* of $f$ and $g$ as follows $$\{f,g\}_{nh}=\omega_L(\bar{P}(X_f),\bar{P}(X_g))|_{{\mathcal{M}}},$$ where $X_f$ and $X_g$ are the Hamiltonian sections on ${\@ifstar{\@proldf}{\@prolpf}}[E]{E}$ associated with $f$ and $g$, respectively.
Moreover, proceeding as in the case of linear constraints, one can prove that $$\dot{f}=\rho^1(R_L)(f)+\{f,E_L\}_{nh}, \;\;\; f\in C^\infty({\mathcal{M}}),$$ where $R_L$ is the section of ${\@ifstar{\@proldf}{\@prolpf}}[E]{{\mathcal{M}}}\to {\mathcal{M}}$ defined by $R_L=P(\Gamma_L|_{\mathcal{M}})-\bar{P}(\Gamma_L|_{{\mathcal{M}}}).$ Thus, in the particular case when the restriction to ${\mathcal{M}}$ of the vector field $\rho^1(\Delta)$ on $E$ is tangent to ${{\mathcal{M}}}$, it follows that $$\dot{f}=\{f,E_L\}_{nh},\;\;\; \mbox{ for } f\in C^\infty({\mathcal{M}}).$$
Alternatively, since ${\@ifstar{\@proldf}{\@prolpf}}[{\mathcal{V}_{}}]{{\mathcal{M}}}$ is an anchored vector bundle, we may consider the differential $\bar{d}f\in
{\operatorname{Sec}_{}(({\@ifstar{\@proldf}{\@prolpf}}[{\mathcal{V}_{}}]{{\mathcal{M}}})^*)}$ for a function $f\in C^\infty({\mathcal{M}})$. Thus, since the restriction $\omega^{L,{\mathcal{M}}}$ of $\omega_L$ to ${\@ifstar{\@proldf}{\@prolpf}}[{\mathcal{V}_{}}]{{\mathcal{M}}}$ is regular, we have a unique section $\bar{X}_f \in
{\operatorname{Sec}_{}({\@ifstar{\@proldf}{\@prolpf}}[{\mathcal{V}_{}}]{{\mathcal{M}}})}$ given by $i_{\bar{X}_f}\omega^{L,{\mathcal{M}}}=\bar{d}f$ and it follows that $$\{f,g\}_{nh}=\omega^{L,{\mathcal{M}}}(\bar{X}_f,\bar{X}_g).$$
Morphisms and reduction {#morphisms-and-reduction}
-----------------------
Let $(L,{\mathcal{M}})$ be a regular constrained Lagrangian system on a Lie algebroid $\tau:E\to M$ and let $(L',{\mathcal{M}}')$ be another constrained Lagrangian system on a second Lie algebroid $\tau':E'\to M'$. Suppose also that we have a fiberwise surjective morphism of Lie algebroids $\Phi:E\to E'$ over a surjective submersion $\phi:M\to
M'$ such that:
- $L=L'\circ \Phi$,
- $\Phi|_{\mathcal{M}}:{\mathcal{M}}\to {\mathcal{M}}'$ is a surjective submersion,
- $\Phi({\mathcal{V}_{a}})={\mathcal{V}_{\Phi(a)}}'$, for all $a\in {\mathcal{M}}$.
Note that condition (ii) implies that $\Phi({\mathcal{V}_{a}})\subseteq
{\mathcal{V}_{\Phi(a)}}'$, for all $a\in {\mathcal{M}}$. Moreover, if $V(\Phi)$ is the vertical bundle of $\Phi$ and $$V_a(\Phi)\subset T_a{\mathcal{M}}, \mbox{ for all } a\in {\mathcal{M}},$$ then condition (ii) also implies that ${\mathcal{V}_{\Phi(a)}}'\subseteq
\Phi({\mathcal{V}_{a}})$, for all $a\in {\mathcal{M}}$.
On the other hand, using condition (iii) and Proposition \[transformation-omegaL\], it follows that $\ker
G^{L',{{\mathcal{V}_{}}'}}=\{0\}$ and, thus, the constrained Lagrangian system $(L',{\mathcal{M}}')$ is regular. Moreover, proceeding as in the proof of Lemma \[l5.5\] and Theorem \[t5.6\], we deduce the following results.
\[l8.4\] With respect to the decompositions $$({\@ifstar{\@proldf}{\@prolpf}}[E]{E})|_{\mathcal{M}}={\@ifstar{\@proldf}{\@prolpf}}[E]{\mathcal{M}}\oplus F \quad \text{and} \quad
({\@ifstar{\@proldf}{\@prolpf}}[E']E')|_{{\mathcal{M}}'}={\@ifstar{\@proldf}{\@prolpf}}[E']{\mathcal{M}}'\oplus F'$$ we have the following properties
1. ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}({\@ifstar{\@proldf}{\@prolpf}}[E]{\mathcal{M}})={\@ifstar{\@proldf}{\@prolpf}}[E']{{\mathcal{M}}'},$
2. ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}(F)=F'$,
3. If $P$, $Q$ and $P',Q'$ are the projectors associated with $(L,{\mathcal{M}})$ and $(L',{\mathcal{M}}')$, respectively, then $P'\circ
{\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}={\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}\circ P$ and $Q'\circ
{\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}={\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}\circ Q$.
With respect to the decompositions $$({\@ifstar{\@proldf}{\@prolpf}}[E]{E})|_{{\mathcal{M}}}={\@ifstar{\@proldf}{\@prolpf}}[\mathcal V]{\mathcal{M}}\oplus ({\@ifstar{\@proldf}{\@prolpf}}[\mathcal V
]{\mathcal{M}})^\perp \mbox{ and } ({\@ifstar{\@proldf}{\@prolpf}}[E']{E'})|_{{\mathcal{M}}'} = {\@ifstar{\@proldf}{\@prolpf}}[\mathcal
V']{{\mathcal{M}}'}\oplus ({\@ifstar{\@proldf}{\@prolpf}}[\mathcal V']{{\mathcal{M}}'})^\perp$$ we have the following properties
1. $({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi})({\@ifstar{\@proldf}{\@prolpf}}[\mathcal
V]{{\mathcal{M}}})={\@ifstar{\@proldf}{\@prolpf}}[\mathcal V']{{\mathcal{M}}'}$,
2. $({\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi})(({\@ifstar{\@proldf}{\@prolpf}}[\mathcal
V]{{\mathcal{M}}})^\perp)=({\@ifstar{\@proldf}{\@prolpf}}[\mathcal V']{{\mathcal{M}}'})^\perp$,
3. If $\bar{P},\bar{Q}$ and $\bar{P}'$ and $\bar{Q'}$ are the projectors associated with $(L,{{\mathcal{M}}})$ and $(L',{{\mathcal{M}}'})$, respectively, then $\bar{P}'\circ {\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi} =
{\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}\circ \bar{P}$ and $\bar{Q}'\circ
{\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi} = {\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}\circ \bar{Q}$.
\[t8.5\] Let $(L, {\mathcal{M}})$ be a regular constrained Lagrangian system on a Lie algebroid $E$ and let $(L',{\mathcal{M}}')$ be a constrained Lagrangian system on a second Lie algebroid $E'$. Assume that we have a fiberwise surjective morphism of Lie algebroids $\Phi:E\to E'$ over $\phi:M\to
M'$ such that conditions (i)-(iii) hold. If $\Gamma_{(L, {\mathcal{M}})}$ is the constrained dynamics for $L$ and $\Gamma_{(L', {\mathcal{M}}')}$ is the constrained dynamics for $L'$, respectively, then ${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}\circ
\Gamma_{(L, {\mathcal{M}})}=\Gamma_{(L', {\mathcal{M}}')}\circ \Phi$. If $a(t)$ is a solution of Lagrange-d’Alembert differential equations for $L$, then $\Phi(a(t))$ is a solution of Lagrange-d’Alembert differential equations for $L'$.
We will say that the constrained dynamics $\Gamma_{(L', {\mathcal{M}}')}$ is *the reduction of the constrained dynamics* $\Gamma_{(L, {\mathcal{M}})}$ by the morphism $\Phi$. As in the case of linear constraints (see Theorem \[t5.7\]), we also may prove the following result
\[t8.6’\] Under the same hypotheses as in Theorem \[t8.5\], we have that $$\{f'\circ \Phi,g'\circ \Phi\}_{nh}=\{f',g'\}_{nh}'\circ \Phi ,$$ for $f',g'\in C^\infty({\mathcal{M}}')$, where $\{\cdot,\cdot\}_{nh}$ (respectively, $\{\cdot,\cdot\}_{nh}'$) is the nonholonomic bracket for the constrained system $(L,{\mathcal{M}})$ (respectively, $(L',{\mathcal{M}}')$). In other words, $\Phi:{\mathcal{M}}\to {\mathcal{M}}'$ is an almost-Poisson morphism.
Now, let $\phi:Q\to M$ be a principal $G$-bundle and $\tau:E\to Q$ be a Lie algebroid over $Q$. In addition, assume that we have an action of $G$ on $E$ such that the quotient vector bundle $E/G$ is defined and the set ${\operatorname{Sec}_{}(E)}^G$ of equivariant sections of $E$ is a Lie subalgebra of ${\operatorname{Sec}_{}(E)}$. Then, $E'=E/G$ has a canonical Lie algebroid structure over $M$ such that the canonical projection $\Phi:E\to E'$ is a fiberwise bijective Lie algebroid morphism over $\phi$ (see Theorem \[quotient-Lie-algebroid\]).
Next, suppose that $(L,{\mathcal{M}})$ is a $G$-invariant regular constrained Lagrangian system, that is, the Lagrangian function $L$ and the constraint submanifold ${\mathcal{M}}$ are $G$-invariant. Then, one may define a Lagrangian function $L':E'\to \R$ on $E'$ such that $$L=L'\circ \Phi.$$ Moreover, $G$ acts on ${\mathcal{M}}$ and if the set of orbits ${\mathcal{M}}'={\mathcal{M}}/G$ of this action is a quotient manifold, that is, ${\mathcal{M}}'$ is a smooth manifold and the canonical projection $\Phi_{|{\mathcal{M}}}:{\mathcal{M}}\to {\mathcal{M}}'={{\mathcal{M}}}/{G}$ is a submersion, then one may consider the constrained Lagrangian system $(L',{{\mathcal{M}}}')$ on $E'$.
\[r8.5’\] [If ${\mathcal{M}}$ is a closed submanifold of $E$, then, using a well-known result (see [@AM Theorem 4.1.20]), it follows that the set of orbits ${\mathcal{M}}'={\mathcal{M}}/G$ is a quotient manifold.]{}
Since the orbits of the action of $G$ on $E$ are the fibers of $\Phi$ and ${\mathcal{M}}$ is $G$-invariant, we deduce that $$V_a(\Phi)\subseteq T_a{\mathcal{M}}, \mbox{ for all } a\in {\mathcal{M}},$$ which implies that $\Phi_{|{\mathcal{V}_{a}}}:{{\mathcal{V}_{a}}}\to {\mathcal{V}_{\Phi(a)}}'$ is a linear isomorphism, for all $a\in {\mathcal{M}}.$
Thus, from Theorem \[t8.5\], we conclude that the constrained Lagrangian system $(L',{\mathcal{M}}')$ is regular and that $${\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}\circ \Gamma_{(L, {\mathcal{M}})}=\Gamma_{(L', {\mathcal{M}}')}\circ
\Phi,$$ where $\Gamma_{(L, {\mathcal{M}})}$ (resp., $\Gamma_{(L', {\mathcal{M}}')}$) is the constrained dynamics for $L$ (resp., $L'$). In addition, using Theorem \[t8.6’\], we obtain that $\Phi: {\mathcal{M}}\to {\mathcal{M}}'$ is an almost-Poisson morphism when on ${\mathcal{M}}$ and ${\mathcal{M}}'$ we consider the almost-Poisson structures induced by the corresponding nonholonomic brackets.
We illustrate the results above in a particular example in the following subsection.
Example: a ball rolling on a rotating table
-------------------------------------------
The following example is taken from [@BlKrMaMu; @CLMM; @NF]. A (homogeneous) sphere of radius $r>0$, unit mass $m=1$ and inertia about any axis $k^2,$ rolls without sliding on a horizontal table which rotates with constant angular velocity $\Omega$ about a vertical axis through one of its points. Apart from the constant gravitational force, no other external forces are assumed to act on the sphere.
Choose a Cartesian reference frame with origin at the center of rotation of the table and $z$-axis along the rotation axis. Let $(x,y)$ denote the position of the point of contact of the sphere with the table. The configuration space for the sphere on the table is $Q=\R^2\times SO(3)$, where $SO(3)$ may be parameterized by the Eulerian angles $\theta,\varphi$ and $\psi$. The kinetic energy of the sphere is then given by $$T=\frac{1}{2}(\dot{x}^2 + \dot{y}^2 + k^2(\dot\theta^2 + \dot\psi^2
+ 2 \dot\varphi\dot\psi \cos \theta)) ,$$ and with the potential energy being constant, we may put $V=0.$ The constraint equations are $$\begin{aligned}
\dot{x}-r\dot{\theta}\sin \psi + r \dot\varphi\sin \theta \cos
\psi&=-\Omega y,\\
\dot{y} + r\dot\theta \cos\psi + r \dot{\varphi}\sin \theta \sin
\psi&=\Omega x.
\end{aligned}$$ Since the Lagrangian function is of mechanical type, the constrained system is regular. Note that the constraints are affine, and hence not linear, and that the restriction to the constraint submanifold ${\mathcal{M}}$ of the Liouville vector field on $TQ$ is not tangent to ${\mathcal{M}}$. Indeed, the constraints are linear if and only if $\Omega = 0$.
Now, we can proceed from here to construct to equations of motion of the sphere, following the general theory. However, the use of the Eulerian angles as part of the coordinates leads to very complicated expressions. Instead, one may choose to exploit the symmetry of the problem, and one way to do this is by the use of appropriate *quasi-coordinates* (see [@NF]). First of all, observe that the kinetic energy may be expressed as $$T=\frac{1}{2}(\dot{x}^2+\dot{y}^2 + k^2(\omega_x^2 + \omega_y^2 +
\omega_z^2)),$$ where $$\begin{aligned}
\omega_x&= \dot{\theta}\cos\psi + \dot{\varphi}\sin\theta\sin\psi
, \\
\omega_y&= \dot{\theta}\sin\psi - \dot{\varphi}\sin\theta\cos\psi
, \\
\omega_z&= \dot{\varphi}\cos\theta + \dot\psi ,
\end{aligned}$$ are the components of the angular velocity of the sphere. The constraint equations expressing the rolling conditions can be rewritten as $$\begin{aligned}
\dot{x}-rw_y & = -\Omega y,\\
\dot{y} + r\omega_x& = \Omega x.
\end{aligned}$$ Next, following [@CLMM], we consider local coordinates $(\bar{x},\bar{y},\bar{\theta}, \bar{\varphi}, \bar{\psi};
\pi_i)_{i=1,\dots ,5}$ on $TQ=T\R^2 \times T(SO(3))$, where $$\bar{x}=x,\;\;\; \bar{y}=y,\;\;\; \bar{\theta}=\theta,\;\;\;
\bar{\varphi}=\varphi,\;\;\; \bar{\psi}=\psi,$$ $$\pi_1=r\dot{x}+k^2\dot{q}_2,\;\;\;
\pi_2=r\dot{y}-k^2\dot{q}_1,\;\;\; \pi_3=k^2\dot{q}_3,$$ $$\pi_4=\frac{k^2}{(k^2+ r^2)}(\dot{x}-r\dot{q}_2 + \Omega y),\;\;\;
\pi_5=\frac{k^2}{(k^2+ r^2)}(\dot{y}+r\dot{q}_1 - \Omega x),$$ and $(\dot{q}_1,\dot{q}_2,\dot{q}_3)$ are the quasi-coordinates defined by $$\dot{q}_1=\omega_x,\;\;\; \dot{q}_2=\omega_y,\;\;\;
\dot{q}_3=\omega_z.$$ As is well-known, the coordinates $q_i$ only have a symbolic meaning. In fact, $ \displaystyle \{\frac{\partial }{\partial
q_1}, \frac{\partial }{\partial q_2}, \frac{\partial }{\partial
q_3}\}$ is the basis of left-invariant vector fields on $SO(3)$ given by $$\begin{aligned}
\frac{\partial }{\partial q_1} &= (\cos\psi) \frac{\partial
}{\partial \theta} + \frac{\sin \psi }{\sin \theta}(\frac{\partial
}{\partial \varphi}-\cos\theta \frac{\partial }{\partial
\psi}),\\
\frac{\partial }{\partial q_2} &= (\sin \psi) \frac{\partial
}{\partial \theta}-\frac{\cos\psi }{\sin \theta}(\frac{\partial
}{\partial \varphi}-\cos\theta \frac{\partial }{\partial
\psi}),\\
\frac{\partial }{\partial q_3} &= \frac{\partial }{\partial \psi} ,
\end{aligned}$$ and we have that $$[\frac{\partial }{\partial q_2}, \frac{\partial }{\partial
q_1}]=\frac{\partial }{\partial q_3} , \quad [\frac{\partial
}{\partial q_1}, \frac{\partial }{\partial q_3}]=\frac{\partial
}{\partial q_2} , \quad [\frac{\partial }{\partial q_3},
\frac{\partial }{\partial q_2}]=\frac{\partial }{\partial q_1}.$$
Note that in the new coordinates the local equations defining the constraint submanifold ${\mathcal{M}}$ are $\pi_4=0,$ $\pi_5=0$. On the other hand, if $P:({\@ifstar{\@proldf}{\@prolpf}}[TQ]{TQ})|_{\mathcal{M}}=T_{\mathcal{M}}(TQ)\to {\@ifstar{\@proldf}{\@prolpf}}[TQ]{{\mathcal{M}}}=T{\mathcal{M}}$ and $Q:T_{\mathcal{M}}(TQ)\to F$ are the projectors associated with the decomposition $T_{\mathcal{M}}(TQ)=T{\mathcal{M}}\oplus F$, then we have that (see [@CLMM]) $$\begin{aligned}
Q&= \frac{\partial }{\partial \pi_4}\otimes d\pi_4 +
\frac{\partial }{\partial \pi_5}\otimes
d\pi_5,\\
P&= \text{Id} - \frac{\partial }{\partial \pi_4}\otimes
d\pi_4-\frac{\partial }{\partial \pi_5}\otimes d\pi_5.
\end{aligned}$$ Moreover, using that the unconstrained dynamics $\Gamma_L$ is given by $$\begin{aligned}
\Gamma_L &=\dot{x}\displaystyle\frac{\partial }{\partial \bar{x}} +
\dot{y}\displaystyle\frac{\partial }{\partial \bar{y}}+
\dot{\theta}\displaystyle\frac{\partial }{\partial \bar{\theta}} +
\dot{\varphi}\displaystyle\frac{\partial }{\partial \bar{\varphi}}+
\dot{\psi}\displaystyle\frac{\partial }{\partial \bar{\psi}} +
\displaystyle\frac{k^2\Omega}{(k^2 +
r^2)}\dot{y}\displaystyle\frac{\partial }{\partial \pi_4} +
\displaystyle\frac{k^2\Omega}{(k^2 +
r^2)}\dot{x}\displaystyle\frac{\partial }{\partial \pi_5}\\
&= \dot{x}\displaystyle\frac{\partial }{\partial \bar{x}} +
\dot{y}\displaystyle\frac{\partial }{\partial \bar{y}}+
\dot{q_1}\displaystyle\frac{\partial }{\partial q_1} +
\dot{q_2}\displaystyle\frac{\partial }{\partial q_2}+
\dot{q_3}\displaystyle\frac{\partial }{\partial q_3} +
\displaystyle\frac{k^2\Omega}{(k^2 +
r^2)}\dot{y}\displaystyle\frac{\partial }{\partial \pi_4} +
\displaystyle\frac{k^2\Omega}{(k^2 +
r^2)}\dot{x}\displaystyle\frac{\partial }{\partial \pi_5},
\end{aligned}$$ we deduce that the constrained dynamics is the [<span style="font-variant:small-caps;">sode</span>]{} $\Gamma_{(L, {\mathcal{M}})}$ along ${\mathcal{M}}$ defined by $$\begin{aligned}
\label{Gamma}
\Gamma_{(L, {\mathcal{M}})} = (P\Gamma_L|_{{\mathcal{M}}}) &=(\dot{x}\displaystyle\frac{\partial
}{\partial \bar{x}} + \dot{y}\displaystyle\frac{\partial }{\partial
\bar{y}}+ \dot{\theta}\displaystyle\frac{\partial }{\partial
\bar{\theta}} + \dot{\varphi}\displaystyle\frac{\partial
}{\partial \bar{\varphi}}+ \dot{\psi}\displaystyle\frac{\partial
}{\partial \bar{\psi}})|_{{\mathcal{M}}} \nonumber
\\&=(\dot{x}\displaystyle\frac{\partial }{\partial \bar{x}} +
\dot{y}\displaystyle\frac{\partial }{\partial \bar{y}}+
\dot{q_1}\displaystyle\frac{\partial }{\partial q_1} +
\dot{q_2}\displaystyle\frac{\partial }{\partial q_2}+
\dot{q_3}\displaystyle\frac{\partial }{\partial q_3})|_{{\mathcal{M}}}.
\end{aligned}$$ This implies that $$d_{\Gamma_{(L, {\mathcal{M}})}}(E_{L}|_{{\mathcal{M}}}) = d_{\Gamma_{(L, {\mathcal{M}})}}(L|_{{\mathcal{M}}}) = \displaystyle
\frac{\Omega^{2}k^{2}}{(k^{2} + r^{2})}(x \dot{x} + y
\dot{y})|_{{\mathcal{M}}}.$$ Consequently, the Lagrangian energy is a constant of the motion if and only if $\Omega = 0$.
When constructing the nonholonomic bracket on ${\mathcal{M}}$, we find that the only non-zero fundamental brackets are $$\label{Cornonh}
\begin{array}{ll}
\{x,\pi_1\}_{nh}=r,&\kern-50pt\{y,\pi_2\}_{nh}=r,\\
\{q_1,\pi_2\}_{nh}=-1,&\kern-50pt\{q_2,\pi_1\}_{nh}=1,\;\;\;\;\;\;\;\;
\{q_3,\pi_3\}_{nh}=1,\\[5pt]
\{\pi_1,\pi_2\}_{nh}=\pi_3,&\kern-50pt\{\pi_2,\pi_3\}_{nh} =
\displaystyle\frac{k^2}{(k^2+r^2)}\pi_1
+ \displaystyle\frac{rk^2\Omega}{(k^2+
r^2)}y,\\\{\pi_3,\pi_1\}_{nh}=\displaystyle\frac{k^2}{(k^2+r^2)}\pi_2
- \displaystyle\frac{rk^2\Omega}{(k^2+ r^2)}x,
\end{array}$$ in which the “appropriate operational” meaning has to be attached to the quasi-coordinates $q_i$. As a result, we have $$\dot{f}=R_L(f) + \{f,L\}_{nh}, \mbox{ for $f\in C^\infty({\mathcal{M}})$}$$ where $R_L$ is the vector field on ${\mathcal{M}}$ given by $$\begin{aligned}
R_L&=\displaystyle(\frac{k^2\Omega}{(k^2+
r^2)}(x\displaystyle\frac{\partial }{\partial
y}-y\displaystyle\frac{\partial }{\partial x}) +
\displaystyle\frac{r\Omega}{(k^2+
r^2)}(x\displaystyle\frac{\partial }{\partial q_1} + y
\frac{\partial
}{\partial q_2} \\[5pt]
& + x(\pi_3-k^2\Omega)\displaystyle\frac{\partial }{\partial \pi_1}
+ y (\pi_3-k^2\Omega)\displaystyle\frac{\partial }{\partial
\pi_2}-k^2(\pi_1x+\pi_2y)\displaystyle\frac{\partial }{\partial
\pi_3}))|_{{\mathcal{M}}}.
\end{aligned}$$ Note that $R_{L} = 0$ if and only if $\Omega = 0$.
Now, it is clear that $Q=\R^2\times SO(3)$ is the total space of a trivial principal $SO(3)$-bundle over $\R^2$ and the bundle projection $\phi:Q\to M=\R^2$ is just the canonical projection on the first factor. Therefore, we may consider the corresponding Atiyah algebroid $E'=TQ/SO(3)$ over $M=\R^2$. Next, we describe this Lie algebroid.
Using the left-translations in $SO(3)$, one may define a diffeomorphism $\lambda$ between the tangent bundle to $SO(3)$ and the product manifold $SO(3)\times \R^3$ (see [@AM]). In fact, in terms of the Euler angles, the diffeomorphism $\lambda$ is given by $$\label{lambda}
\lambda(\theta,\varphi,\psi;\dot\theta,\dot\varphi,\dot\psi) =
(\theta,\varphi,\psi;
\omega_x,\omega_y,\omega_z).$$ Under this identification between $T(SO(3))$ and $SO(3)\times \R^3$, the tangent action of $SO(3)$ on $T(SO(3))\cong SO(3)\times \R^3$ is the trivial action $$\label{Action}
SO(3)\times (SO(3)\times \R^3)\to SO(3)\times \R^3,\;\;\;
(g,(h,\omega))\mapsto (gh,\omega).$$ Thus, the Atiyah algebroid $TQ/SO(3)$ is isomorphic to the product manifold $T\R^2\times \R^3$, and the vector bundle projection is $\tau_{\R^2}\circ pr_1$, where $pr_1:T\R^2\times \R^3\to T\R^2$ and $\tau_{\R^2}:T\R^2\to \R^2$ are the canonical projections.
A section of $E'=TQ/SO(3)\cong T\R^2\times \R^3\to \R^2$ is a pair $(X,u)$, where $X$ is a vector field on $\R^2$ and $u:\R^2\to \R^3$ is a smooth map. Therefore, a global basis of sections of $T\R^2\times \R^3\to \R^2$ is $$\begin{array}{rrr}
e_1'=(\displaystyle\frac{\partial}{\partial x},0),&
e_2'=(\displaystyle\frac{\partial}{\partial y},0),&\\[5pt]
e_3'=(0,u_1),& e_4'=(0,u_2),& e_5'=(0,u_3),
\end{array}$$ where $u_1,u_2,u_3:\R^2\to \R^3$ are the constant maps $$u_1(x,y)=(1,0,0),\;\;\; u_2(x,y)=(0,1,0),\;\;\; u_3(x,y)=(0,0,1).$$ In other words, there exists a one-to-one correspondence between the space ${\operatorname{Sec}_{}(E'=TQ/SO(3))}$ and the $G$-invariant vector fields on $Q$. Under this bijection, the sections $e_1'$ and $e_2'$ correspond with the vector fields $\displaystyle\frac{\partial }{\partial x}$ and $\displaystyle\frac{\partial }{\partial y}$ and the sections $e_3'$, $e_4'$ and $e_5'$ correspond with the vertical vector fields $\displaystyle\frac{\partial }{\partial q_1}$, $\displaystyle\frac{\partial }{\partial q_2}$ and $\displaystyle\frac{\partial }{\partial q_3}$, respectively.
The anchor map $\rho':E'=TQ/SO(3)\cong T\R^2\times \R^3\to T\R^2$ is the projection over the first factor and, if $\lcf\cdot, \cdot \rcf'$ is the Lie bracket on the space ${\operatorname{Sec}_{}(E'=TQ/SO(3))}$, then the only non-zero fundamental Lie brackets are $$\lcf e_4',e_3'\rcf'=e_5',\;\;\;\lcf e_5',e_4'\rcf'=e_3',\;\;\; \lcf
e_3',e_5'\rcf'=e_4'.$$ From (\[lambda\]) and (\[Action\]), it follows that the Lagrangian function $L=T$ and the constraint submanifold ${\mathcal{M}}$ are $SO(3)$-invariant. Consequently, $L$ induces a Lagrangian function $L'$ on $E'=TQ/SO(3)$ and, since ${\mathcal{M}}$ is closed on $TQ$, the set of orbits ${\mathcal{M}}'={\mathcal{M}}/SO(3)$ is a submanifold of $E'=TQ/SO(3)$ in such a way that the canonical projection $\Phi|_{\mathcal{M}}:{\mathcal{M}}\to {\mathcal{M}}'={\mathcal{M}}/SO(3)$ is a surjective submersion.
Under the identification between $E'=TQ/SO(3)$ and $T\R^2\times
\R^3$, $L'$ is given by $$L'(x,y,\dot{x},\dot{y};\omega_1,\omega_2,\omega_3) =
\frac{1}{2}(\dot{x}^2 + \dot{y}^2) + \frac{k^2}{2} (\omega_1^2 +
\omega_2^2 + \omega_3^2) ,$$ where $(x,y,\dot{x},\dot{y})$ and $(\omega_1,\omega_2,\omega_3)$ are the standard coordinates on $T\R^2$ and $\R^3$, respectively. Moreover, the equations defining ${\mathcal{M}}'$ as a submanifold of $T\R^2\times \R^3$ are $$\dot{x}-r\omega_2 + \Omega y=0,\;\;\; \dot{y} + r\omega_1-\Omega x=0.$$ So, we have the constrained Lagrangian system $(L',{\mathcal{M}}')$ on the Atiyah algebroid $E'=TQ/SO(3)\cong T\R^2\times \R^3$. Note that the constraints are not linear, and that, if $\Delta'$ is the Liouville section of the prolongation ${\@ifstar{\@proldf}{\@prolpf}}[E']{E'}$, then the restriction to ${\mathcal{M}}'$ of the vector field $(\rho')^1(\Delta')$ is not tangent to ${\mathcal{M}}'$.
Now, it is clear that the tangent bundle $TQ=T\R^2\times
T(SO(3))\cong T\R^2\times (SO(3)\times \R^3)$ is the total space of a trivial principal $SO(3)$-bundle over $E'=TQ/SO(3)\cong T\R^2\times
\R^3$ and, in addition (see [@LeMaMa Theorem 9.1]), the prolongation ${\@ifstar{\@proldf}{\@prolpf}}[E']{E'}$ is isomorphic to the Atiyah algebroid associated with this principal $SO(3)$-bundle. Therefore, the sections of the prolongation ${\@ifstar{\@proldf}{\@prolpf}}[E']{E'}\to E'$ may be identified with the $SO(3)$-invariant vector fields on $TQ\cong T\R^2\times
(SO(3)\times \R^3)$. Under this identification, the constrained dynamics $\Gamma_{(L', {\mathcal{M}}')}$ for the system $(L',{\mathcal{M}}')$ is just the $SO(3)$-invariant vector field $\Gamma_{(L, {\mathcal{M}})}=P(\Gamma_L|_{{\mathcal{M}}})$. We recall that if $\Phi:TQ\to TQ/SO(3)$ is the canonical projection, then $$\label{Reduc}
{\@ifstar{\@proldf}{\@prolpf}}[\Phi]{\Phi}\circ \Gamma_{(L, {\mathcal{M}})}=\Gamma_{(L', {\mathcal{M}}')}\circ \Phi.$$
Next, we give a local description of the vector field $(\rho')^1(\Gamma_{(L', {\mathcal{M}}')})$ on $E'=TQ/SO(3)\cong T\R^2\times
\R^3$ and the nonholonomic bracket $\{\cdot,\cdot\}_{nh}'$ for the constrained system $(L',{\mathcal{M}}')$. For this purpose, we consider a suitable system of local coordinates on $TQ/SO(3)\cong T\R^2\times
\R^3.$ If we set $$\begin{array}{lll}
x'=x,&y'=y,&\\
\pi_1'=r\dot{x} +
k^2\omega_2,&\pi_2'=r\dot{y}-k^2\omega_1,&\pi_3'=k^2\omega_3,\\
\pi_4=\frac{k^2}{(k^2 + r^2)}(\dot{x}-r\omega_2 + \Omega y),
&\pi_5'=\frac{k^2}{(k^2+r^2)}(\dot{y} + r \omega_1 -\Omega x),
\end{array}$$ then $(x',y',\pi_1',\pi_2',\pi_3',\pi_4',\pi_5')$ is a system of local coordinates on $TQ/SO(3)\cong T\R^2\times \R^3$. In these coordinates the equations defining the submanifold ${\mathcal{M}}'$ are $\pi_4'=0$ and $\pi_5'=0$, and the canonical projection $\Phi:TQ\to
TQ/SO(3)$ is given by $$\label{Phi}
\Phi(\bar{x},\bar{y},\bar{\theta},\bar{\varphi},\bar{\psi};
\pi_1,\pi_2,\pi_3,\pi_4,\pi_5)=(\bar{x},\bar{y};{\pi}_1, {\pi}_2,
{\pi}_3, {\pi}_4, {\pi}_5).$$
Thus, from (\[Gamma\]) and (\[Reduc\]), it follows that $$(\rho')^1(\Gamma_{(L', {\mathcal{M}}')})=(\dot{x}'\frac{\partial }{\partial x'} +
\dot{y}'\frac{\partial }{\partial y'})|_{{\mathcal{M}}'} ,$$ or, in the standard coordinates $(x,y,\dot{x},\dot{y};
\omega_1,\omega_2,\omega_3)$ on $T\R^2\times \R^3,$ $$\begin{aligned}
(\rho')^1(\Gamma_{(L', {\mathcal{M}}')})&=\{\dot{x}(\displaystyle\frac{\partial
}{\partial x} + \displaystyle\frac{\Omega k^2}{(k^2 +
r^2)}\displaystyle\frac{\partial }{\partial \dot{y}} +
\displaystyle\frac{\Omega r}{(k^2 +
r^2)}\displaystyle\frac{\partial }{\partial \omega_1})\\
& \quad + \dot{y}(\displaystyle\frac{\partial }{\partial y} -
\displaystyle\frac{\Omega k^2}{(k^2 +
r^2)}\displaystyle\frac{\partial }{\partial \dot{x}} +
\displaystyle\frac{\Omega r}{(k^2 +
r^2)}\displaystyle\frac{\partial }{\partial \omega_2})\}|_{{\mathcal{M}}'}.
\end{aligned}$$ On the other hand, from (\[Cornonh\]), (\[Phi\]) and Theorem \[t8.6’\], we deduce that the only non-zero fundamental nonholonomic brackets for the system $(L',{\mathcal{M}}')$ are $$\begin{array}{lll}
\{x',\pi_1'\}'_{nh}=r,& \{y',\pi_2'\}'_{nh}=r,&\\
\{\pi_1',\pi_2'\}_{nh}'=\pi_3',&\{\pi_2',\pi_3'\}_{nh}' =
\displaystyle\frac{k^2}{(k^2+
r^2)}\pi_1' + \displaystyle\frac{rk^2\Omega}{(k^2+ r^2)}y',&\\
\{\pi_3',\pi_1'\}_{nh}'=\displaystyle\frac{k^2}{(k^2 + r^2)}
\pi_2'- \displaystyle\frac{rk^2\Omega}{(k^2+ r^2)}x'.&&
\end{array}$$ Therefore, we have that $$\dot{f}'=(\rho')^1(R_{L'})(f') + \{f',L'\}_{nh}',\mbox{ for } f'\in
C^\infty({\mathcal{M}}'),$$ where $(\rho')^1(R_{L'})$ is the vector field on ${\mathcal{M}}'$ given by $$\begin{aligned}
(\rho')^1(R_{L'}) &= \{ \displaystyle\frac{k^2\Omega}{k^2+
r^2}(x'\displaystyle\frac{\partial }{\partial
y'}-y'\displaystyle\frac{\partial }{\partial x'}) +
\displaystyle\frac{r\Omega}{(k^2 +
r^2)}(x'(\pi_3'-k^2\Omega)\displaystyle\frac{\partial }{\partial
\pi_1'} \\[5pt]
& \quad + y' (\pi_3'-k^2\Omega)\displaystyle\frac{\partial
}{\partial \pi_2'}-k^2(\pi_1' x' + \pi_2'
y')\displaystyle\frac{\partial }{\partial \pi_3'})\}|_{{\mathcal{M}}'}.
\end{aligned}$$
Conclusions and outlook {#conclusions}
=======================
We have developed a geometrical description of nonholonomic mechanical systems in the context of Lie algebroids. This formalism is the natural extension of the standard treatment on the tangent bundle of the configuration space. The proposed approach also allows to deal with nonholonomic mechanical systems with symmetry, and perform the reduction procedure in a unified way. The main results obtained in the paper are summarized as follows:
- we have identified the notion of regularity of a nonholonomic mechanical system with linear constraints on a Lie algebroid, and we have characterized it in geometrical terms;
- we have obtained the constrained dynamics by projecting the unconstrained one using two different decompositions of the prolongation of the Lie algebroid along the constraint subbundle;
- we have developed a reduction procedure by stages and applied it to nonholonomic mechanical systems with symmetry. These results have allowed us to get new insights in the technique of quasicoordinates;
- we have defined the operation of nonholonomic bracket to measure the evolution of observables along the solutions of the system;
- we have examined the setup of nonlinearly constrained systems;
- we have illustrated the main results of the paper in several examples.
Current and future directions of research include the in-depth study of the reduction procedure following the steps of [@BlKrMaMu; @CaLeMaMa] for the standard case; the synthesis of so-called nonholonomic integrators [@cortes; @CoSo; @LeMaSa] for systems evolving on Lie algebroids, and the development of a comprehensive treatment of classical field theories within the Lie algebroid formalism following the ideas by E. Mart[í]{}nez [@CFTLAMF].
[99]{}\
,
[<span style="font-variant:small-caps;">Abraham R and Marsden JE</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Bates L and Śniatycki J</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Bloch AM</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Bloch AM, Krishnaprasad PS, Marsden JE and Murray RM</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Bojowald M, Kotov A and Strobl T</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Cannas da Silva A and Weinstein A</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Cantrijn F, Cortés J., de León M and Martín de Diego D</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Cantrijn F, de León M, Marrero JC and Martín de Diego D</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Cantrijn F, de León M, Marrero JC and Martín de Diego D</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Cantrijn F, de León M and Martín de Diego D</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Cantrijn F, de León M, Marrero, JC and Martín de Diego D</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Cariñena JF and Rañada M</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Cariñena JF, Nunes da Costa J.M. and Santos P</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Cariñena JF, Nunes da Costa J.M. and Santos P</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Cendra H, Marsden JE and Ratiu TS</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Chaplygin SA</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Cortés J</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Cortés J and de León M</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Cortés J, de León M, Marrero JC, Mart[í]{}n de Diego D and Mart[í]{}nez E</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Cortés J and Martínez E</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Cortés J and Martínez S</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Crampin M</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Cushman R, Kemppainen D, Śniatycki J and Bates L</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Cushman R and Śniatycki J</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Ehlers K, Koiller J, Montgomery R and Rios PM</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Fedorov YN and Jovanović B</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Fedorov YN and Kozlov VV</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Fedorov YN and Zenkov D</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Fernandes RL</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Grifone J and Mehdi M</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Hamel G</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Higgins PJ and Mackenzie K</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Ibort A, de León M, Marrero JC and Martín de Diego D</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Jovanović B</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Jovanović B</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Klein J</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Koiller J</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Koon WS and Marsden JE</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Koon WS and Marsden JE</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">de León M and Martín de Diego D</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">de León M, Marrero JC and Martín de Diego D</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">de León M, Martín de Diego D and Santamar[í]{}a-Merino A</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">de León M, Marrero JC and Martínez E</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">de León M and Rodrigues, PR</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Lewis AD</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Libermann P</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Marle ChM</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Mackenzie K</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Markeev AP</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Marsden JE and Ratiu TS</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Martínez E</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Martínez E, Mestdag T and Sarlet W</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Martínez E</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Mart[í]{}nez E</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Mestdag T</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Mestdag T and Langerock B</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Neimark J and Fufaev N</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Nijenhuis A</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Popescu M and Popescu P</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Saunders D</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Saunders DJ, Sarlet W and Cantrijn F</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Śniatycki J</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Suslov GK</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Tai M</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Van der Schaft AJ and Maschke BM</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Veselov AP and Veselova LE</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Vinogradov A. M., Kupershmidt B.A.</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Weinstein A</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Zenkov D, Bloch AM</span>\
]{}
[.]{}
[<span style="font-variant:small-caps;">Zenkov D, Bloch AM</span>\
]{}
[.]{}
[^1]: This work has been partially supported by Spanish Ministry of Education and Culture grants MTM2004-7832, BFM2003-01319, MTM2006-03322 and BFM2003-02532. J. Cortés was partially supported by faculty research funds granted by the University of California, Santa Cruz.
[^2]: In fact, in [@CoMa], we considered controlled mechanical systems with nonholonomic constraints, that is, mechanical systems evolving on Lie algebroids and subject to some external control forces.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
$^a$, Jens Zuther $^a$, Sebastian Fischer $^a$, Gerold Busch $^a$, Monica Valencia-S.$^a$, Andreas Eckart$^{a,b}$, Melanie Krips $^b$, Julia Scharwächter $^c$\
1. Physikalisches Institut, University of Cologne, Germany\
Max-Planck-Institut für Radioastronomie, Germany\
Institut de Radioastronomie Millimétrique (IRAM), Grenoble, France\
Observatoire de Paris, Paris, France\
E-mail: , , , , , , ,
title: 'A low-redshift low luminosity QSO sample: Comparison with NUGA galaxies and PG QSOs and first interferometric images of three sample members'
---
Introduction
============
The strong correlation between the black hole mass and the bulge velocity dispersion of galaxies suggests a coeval growth of supermassive black holes (SMBH) and the surrounding stellar bulges. The growth is believed to be regulated by an interplay of nuclear fueling, i.e. inflow of gas which is then consumed by star formation and accretion onto the SMBH, and feedback from these regions by winds, outflows and radiation. The diverse mechanisms that might be involved in these processes (e.g. galactic interactions or secular processes) are topics of current research. While for high-luminosity active galactic nuclei (AGN) the gas inflow is likely to be triggered by large-scale bars or major galaxy interactions (see e.g. \[1\], \[2\]) , secular evolution seems to be the dominating fueling mechanism for low-luminosity AGN (LLAGN, see e.g. \[3\], \[1\], \[4\]). The goal of our low luminosity QSO (LLQSO) sample is to investigate the nuclear fueling in AGN more powerful than the local LLAGN in order to study their relation to the local population and the more active galaxies at higher redshifts.
The sample (\[1\]) has been selected from the Hamburg/ESO QSO survey (flux limit of $B_J\leq 17.3$, \[6\]) with the selection criterion of $z\leq 0.06$ which yielded 99 sources. Up to this redshift limit the CO(2–0) band head from stellar atmospheres in the near–infrared (NIR) $K$–band is observable so that stellar and AGN properties can be analyzed. Furthermore, the sub-kpc scale can still be sufficiently resolved by the current telescopes and arrays to disentangle the starburst from the AGN component and to investigate the interstellar medium (ISM) in the circumnuclear environment.
In fact, the Atacama Large Millimeter Array (ALMA) enables us to observe the LLQSOs with spatial scales (few 10 pc) comparable to those achieved for the NUGA sources (NUclei of GAlaxies, see e.g. \[7\]) with the Plateau de Bure Interferometer (PdBI) and with a signal-to-noise ratio (SNR) for the high density tracers of the order of the SNR obtained for NUGA submillimeter observations of CO. In terms of redshift and activity range our sample adds to the NUGA survey and enlarges the volume from the local to the nearby universe as well as the part of the AGN activity sequence that can be studied in detail to higher activity levels, that are typical for e.g. the Palomar Green (PG) QSO sample (\[8\], \[9\], \[10\]). The PG QSOs are already at a redshift at which the ALMA could only resolve them on few 100 pc scales.
40% of the LLQSO sample has already been observed in the CO(1–0) and (2–1) transition with the IRAM 30 m telescope and with a detection rate of 70% (\[5\]). This paper compares some redshift dependent global properties of our CO-detected HE sources with the NUGA and the PG QSO sample. In addition, we present first results from an interferometric observation of CO line emission in three galaxies from our sample with the Submillimeter Array (SMA).
$
\begin{array}{cc}
% trim=l b r t - trim = 12mm 12mm 12mm 20mm, clip,
\includegraphics[trim = 0mm 20mm 20mm 30mm, clip, width=0.46\textwidth]{z-MH2.eps}&
\includegraphics[trim = 0mm 20mm 20mm 30mm, clip, width=0.46\textwidth]{z-MHI.eps}\\
\includegraphics[trim = 0mm 20mm 20mm 30mm, clip, width=0.46\textwidth]{z-LIR.eps}&
\includegraphics[trim = 0mm 20mm 20mm 30mm, clip, width=0.46\textwidth]{z-LFIR.eps}\\
%\includegraphics[width=0.3\textwidth]{images/z-SFR.eps}& 6
\includegraphics[trim = 0mm 20mm 20mm 30mm, clip, width=0.46\textwidth]{z-LX.eps}&
\includegraphics[trim = 0mm 20mm 20mm 30mm, clip, width=0.46\textwidth]{z-LR.eps}\\
%\includegraphics[width=0.3\textwidth]{images/z-f_12_25.eps}&
%\includegraphics[width=0.3\textwidth]{images/z-f_25_60.eps}\\
%&\includegraphics[width=0.3\textwidth]{images/z-f_60_100.eps}& \\
\end{array} $
Comparison to the NUGA and PG QSO sample
========================================
Figure \[vs-z\] visualizes that our CO-detected LLQSOs cover the apparent gap between the distributions of NUGA sources and PG QSOs not only in terms of redshift but also in terms of gas masses and luminosities. The LLQSOs clearly follow the trend of increasing AGN and star forming activity with increasing redshift. There are two possible explanations for this behavior. First, the probability to detect a more powerful AGN and with this most likely a more massive supermassive black hole and central spheroid (Schechter-function) increases with the volume size. The samples cover volumes of different size, NUGA the smallest, PG QSOs the largest. Therefore, this trend can be explained as due to a volume/luminosity effect. Second, the trend might demonstrate the “cosmic downsizing” in the luminosity and mass distribution (\[27\], \[28\], \[29\], \[30\], \[31\], \[32\]). In this scenario massive black holes and spheroids grow first, i.e. at high redshift, and low mass systems later, see e.g. local/nearby universe. Consequently, our sample would not only contribute to the study of the evolutionary connection between local and more distant LLAGN, but also to the test of the anti-hierarchical black hole growth hypothesis for the population of LLAGN in the nearby universe. In order to shed light on these questions a statistically relevant sample of the order of 100 objects is essential.
Interferometric data and ongoing work
=====================================
Three galaxies, belonging to the brightest detections in CO emission, have been observed with the SMA as a follow-up to our IRAM 30m single dish study. Currently, the CO data of these three galaxies is analyzed, soon to be published in Moser et al. 2013 (in prep., see \[HE-images\]). We find the molecular gas with $M_{\mathrm{H}_2+\mathrm{He}} \sim 10^8 - 10^9 M_\odot$ to be concentrated within regions $\leq$ 1.6 kpc. The emission region size as well as the consistency between the gas and dust mass when using a gas mass conversion factor and gas to dust ratio typical for ULIRGs indicate that these LLQSOs could be downsized versions of ULIRGs or high luminosity QSO. In terms of morphology the comparison between the SMA data and public optical/NIR images imply that HE 0433–1028 and HE 1108–2813 might contain a nested bar whereas HE1029–1831 seems to display a lopsidedness in the CO emission (for details, see Moser et al. 2013 (in prep.)).
$
\begin{array}{ccc}
% trim=l b r t - trim = 12mm 12mm 12mm 20mm, clip,
\includegraphics[%trim = 0mm 20mm 20mm 30mm, clip,
height=0.23\textheight]{HE0433-contour-mom0-cripple.eps}&
\includegraphics[%trim = 0mm 20mm 20mm 30mm, clip,
height=0.23\textheight]{HE1029-contour-mom0.eps}&
\includegraphics[trim = 0mm 15mm 0mm 20mm, clip, height=0.23\textheight]{HE1108-contour-mom0.eps}\\
\end{array} $
Apart from this submillimeter work, the sample has been investigated with regard to many other aspects: First results of a near-infrared study of a sub-sample of our LLQSOs are presented in Busch et al. 2013 (these proceeding), Zuther et al. 2013 (these proceeding) investigated the NLS1 galaxies in our LLQSO sample and Tremou et al. 2013 (in prep.) discuss the optical spectra of the LLQSOs.
[99]{} Hopkins, P. F. & Quataert, E., 2010, *How do massive black holes get their gas?*, MNRAS, 407:1529 Hilz, M. et al. 2013, *How do minor mergers promote inside-out growth of ellipticals, transforming the size, density profile and dark matter fraction?*, MNRAS, 429:2924 Hopkins, P. F. et al. 2008, *A Cosmological Framework for the Co-Evolution of Quasars, Supermassive Black Holes, and Elliptical Galaxies. I. Galaxy Mergers and Quasar Activity*, ApJS, 175:356 Kormendy, J. et al. 2011, *Supermassive black holes do not correlate with galaxy disks or pseudobulges*, Nature, 469:374 Bertram, T. et al. 2007, *Molecular gas in nearby low-luminosity QSO host galaxies*, A&A, 470:571 Wisotzki et al. 2000, *The Hamburg/ESO survey for bright QSOs. III. A large flux-limited sample of QSOs*, A&A, 358:77 García-Burillo, S. et al. 2003, *Molecular Gas in NUclei of GAlaxies (NUGA). I. The counter-rotating LINER NGC 4826*, A&A, 407:485 Evans, A. S. et al. 2001, *Molecular Gas in Infrared-Excess, Optically Selected and the Quasars Connection with Infrared-Luminous Galaxies*, AJ, 121:1893 Evans, A. S. et al. 2006, *Dense Molecular Gas and the Role of Star Formation in the Host Galaxies of Quasi-stellar Objects*, AJ, 132:2398 Scoville, N. Z. et al. 2003, *The Host Galaxies of Optically Bright Quasi-stellar Objects: Molecular Gas in z$\leq$0.1 Palomar-Green Quasi-stellar Objects*, ApJ, 585:L105 Casasola, V. et al. 2011, *Molecular Gas in NUclei of GAlaxies (NUGA). XIV. The barred LINER/Seyfert 2 galaxy NGC 3627*, A&A, 527:A92 Combes, F. et al. 2004, *Molecular gas in NUclei of GAlaxies (NUGA). II. The ringed LINER NGC 7217*, A&A, 414:857 Krips, M. et al. 2005, *Molecular gas in NUclei of GAlaxies (NUGA). III. The warped LINER NGC 3718*, A&A, 442:479 Boone, F. et al. 2007, *Molecular gas in NUclei of GAlaxies (NUGA). VII. NGC 4569, a large scale bar funnelling gas into the nuclear region*, A&A, 471:113 Krips, M. et al. 2007, *Molecular gas in NUclei of GAlaxies (NUGA): VI. Detection of a molecular gas disk/torus via HCN in the Seyfert 2 galaxy NGC 6951?*, A&A, 468:L63 García-Burillo, S. et al. 2009, *Molecular gas in NUclei of GAlaxies (NUGA). XI. A complete gravity torque map of NGC 4579: new clues to bar evolution*, A&A, 496:85 Casasola, V. et al. 2010, *Molecular gas in NUclei of GAlaxies (NUGA) XIII. The interacting Seyfert 2/LINER galaxy NGC 5953*, A&A, 510:A52 Lindt-Krieg, E. et al. 2008, *Molecular gas in NUclei of GAlaxies (NUGA). VIII. The Seyfert 2 NGC 6574*, A&A, 479:377 Hunt, L. K. et al. 2008, *Molecular Gas in NUclei of GAlaxies (NUGA). IX. The decoupled bars and gas inflow in NGC 2782*, A&A, 482:133 Casasola, V. et al. 2008, *Molecular gas in NUclei of GAlaxies (NUGA). X. The Seyfert 2 galaxy NGC 3147*, A&A, 490:61 Combes, F. et al. 2009, *Molecular gas in NUclei of GAlaxies (NUGA). XII. The head-on collision in NGC 1961*, A&A, 503:73 Krips, M. et al. 2007, *Barred CO emission in HE 1029-1831*, A&A, 464:187 Haan, S. et al. 2008, *Atomic Hydrogen Properties of Active Galactic Nuclei Host Galaxies: H I in 16 Nuclei of Galaxies (nuga) Sources*, AJ, 135:232 König, S. et al. 2009, *H I in nearby low-luminosity QSO host galaxies*, A&A, 507:757 Sanders, D. B. & Mirabel, I. F., 1996, *Luminous Infrared Galaxies*, ARA&A, 34:749 Helou, G. et al. 1988, *IRAS observations of galaxies in the Virgo cluster area*, ApJS, 68:151 Ueda, Y. et al. 2003, *Cosmological Evolution of the Hard X-Ray Active Galactic Nucleus Luminosity Function and the Origin of the Hard X-Ray Background*, ApJ, 598:886 Merloni, A., 2004, *The anti-hierarchical growth of supermassive black holes*, MNRAS, 353:1035 Marconi, A. et al. 2004, *Local supermassive black holes, relics of active galactic nuclei and the X-ray background*, MNRAS, 351:169 Hunt, M. P. et al. 2004, *The Faint End of the QSO Luminosity Function at z=3*, ApJ, 605:625 Hasinger, G. et al. 2005, *Luminosity-dependent evolution of soft X-ray selected AGN. New Chandra and XMM-Newton surveys*, A&A, 441:417 Di Matteo, T. et al. 2008, *Direct Cosmological Simulations of the Growth of Black Holes and Galaxies*, ApJ, 676:33
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Based on density functional theory calculations, we systematically investigate the behaviors of a H atom in Ag-doped ZnO, involving the preference sites, diffusion behaviors, the electronic structures and vibrational properties. We find that a H atom can migrate to the doped Ag to form a Ag-H complex by overcoming energy barriers of 0.3 - 1.0 eV. The lowest-energy site for H location is the bond center of a Ag-O in the basal plane. Moreover, H can migrate between this site and its equivalent sites with energy cost of less than 0.5 eV. In contrast, dissociation of such a Ag-H complex needs energy of about 1.1 - 1.3 eV. This implies that the Ag-H complexes can commonly exist in the Ag-doped ZnO, which have a negative effect on the desirable p-type carrier concentrations of Ag-doped ZnO. In addition, based on the frozen phonon calculation, the vibrational properties of ZnO with a Ag-H complex are predicted. Some new vibrational modes associated with the Ag-H complex present in the vibrational spectrum of the system.'
author:
- 'H.Y. He, J. Hu and B.C. Pan'
title: ' Hydrogen in Ag-doped ZnO: theoretical calculations'
---
Introduction
============
ZnO is a promising material for short wavelength optoelectronic device, due to its various attractive properties, such as optical [@Look1; @Lee1; @Kang], photoelectric [@Wenas] and piezoelectric [@Mitsuyu] properties. However, the difficulty in fabricating p-type ZnO restricts the application of ZnO in future [@Zhang]. Therefore, many efforts have been made to achieve p-type ZnO, with using many techniques and dopants. Among various dopants, either group I elements (Li, Na, and K) or group IB elements (Cu, Ag, and Au) [@Zunger; @Zhang2; @Wei] are good candidates for substitution of Zn. However, with smaller ionic radii, group I elements prefer to occupy the interstitial sites rather than substitutional sites, acting as donors [@Look]. In contrast, group IB elements have suitable size for substitution of Zn. So far, the Ag doped p-type ZnO has been successfully achieved in the experiment [@Kang2].
On the other hand, the role of H in ZnO was paid much attention, because H is usually to be unintentionally-doped in as-grown ZnO bulk, and it always acts as a “hole-killer" in ZnO [@Walle1; @Cox; @Hofmann]. Previous studies revealed that the activation energy for H diffusion in pure ZnO was about 0.17 - 0.50 eV [@Ip; @Nickel; @Wardle], which indicated that a H atom could migrate in ZnO easily at low temperature. In ZnO with p-type dopants such as N, Cu and Li, H can be easily trapped by the dopants to form the impurity complexes. For instance, in N doped ZnO, H preferably located at the antibond site of N-Zn bond to form a N-H complex [@Zhang1], and the N-H vibrational modes were observed by Raman spectroscopy in experiment [@Kumar]. While in the case of Li doped ZnO, the preferred site for H was the bond center of a Li-O bond parallel to the $c$ axis. Such a Li-H complex was studied both experimentally and theoretically [@Lavrov3; @Halliburton; @Wardle2; @Shi; @Martin]. Except for the Li-H and N-H complexes, the complexes of H with transition metals such as Cu, Fe and Ni, were also studied [@Wardle1; @Lavrov2], in which presence of H significantly reduced the effect of the spin polarization. Up to date, behavior of H in Ag doped p-type ZnO has not been reported yet.
In this paper, based on density functional theory calculations, we find that a H atom can be trapped by the doped Ag atom in ZnO to form a Ag-H complex, and the lowest-energy site (ground state site) is the bond center of Ag-O in the basal plane. In addition, the vibrational features associated with the Ag-H complex are also discussed in this paper.
Computational details
=====================
Our calculations are based on the density functional theory implemented in the SIESTA program [@Siesta]. Double- $\zeta$ basis plus polarization (DZP) sets[@Soler] are used for all the concerned atoms, and spin polarization is taken into account in our calculations. A 72-atom supercell consisting of $3\times3\times2$ primitive unit cells is employed, in which a Zn atom is substituted by a Ag atom, and a H atom is located at some typical sites around the doped Ag respectively. The lattice constants for each case are optimized and periodic boundary condition is applied. The Brillouin zone is sampled with a set of $k$-point grids ($3\times3\times2$) according to the Monkhorst-Pack scheme [@Monkhorst]. The configurations are fully relaxed with using the conjugate gradient method until the Hellman-Feynman force on each atom is less than 0.02 eV/[Å]{}.
With these settings, we optimize the geometry of the perfect wurtzite ZnO ($w$-ZnO) both with the local density approximation (LDA) and with the Perdew-Burke-Ernzerhos generalized gradient approximation (GGA-PBE). The obtained optimal crystallographic parameters of $w$-ZnO from LDA calculations are $a$= 3.278 [Å]{}, $c$=5.237 [Å]{}, and $u$=0.382, which are slightly larger than those ( $a$ =3.25 [Å]{}, $c$=5.207 [Å]{}, and $u$=0.382) measured from the experiment [@Albertsson] at room temperature. It is noted that the previous experiment demonstrated that the lattice constants of ZnO decreased slightly when the temperature of the system increased [@Yan]. Therefore, our calculated parameters are consistent with the experimental values at low temperature. In contrast, the lattice constants evaluated at the level of GGA are larger than those at the level of LDA. On the other hand, the formation enthalpy (-3.43 eV/ZnO) of $w$-ZnO obtained from the LDA calculation does also match that (-3.61 eV/ZnO) derived from the experiments. Therefore, we perform our calculations at the level of LDA. In addition, we test the convergence of the super cell size by calculating the defect formation energy of the $Ag_Zn $ in the supercells of $2\times2\times2$, $3\times2\times2$ and $3\times3\times3$ respectively, in which the lattice constants are optimized in each case. Here the defect formation energy is defined as the difference of the total energies between the defect case and the corresponding perfect case. We find that the difference of the defect formation energies between $2\times2\times2$ and $3\times2\times2$ supercells is as large as 0.5 eV, whereas for the supercells of $3\times3\times2$ and $3\times3\times3$, the difference of the defect formation energies is about 0.02 eV only. This shows that the interaction of the doped Ag with its images is very weak in supercell $3\times3\times2$, and thus the influence of the images of Ag can be neglected. In addition, the obtained bond lengths as well as the electronic structures in $3\times3\times2$ are almost the same as that in $3\times3\times2$ supercell. So the results achieved from the $3\times3\times2$ supercell are reliable.
The frozen phonon approximation [@Yin] is employed to explore the vibrational properties of the typical systems with Ag-H complexes. In our calculations, the atoms in each of the systems are displaced one by one from their equilibrium positions along three Cartesian directions and the reverse directions, with an amplitude of 0.04 Bohr. By the numerical derivatives for the displacements of each atom, the force constants are obtained, which are used to build up the dynamical matrix of this system. By solving the dynamical equation
$$\omega^{2}M_{i}u_{i,\alpha} = \sum _{j,\beta} C_{i,\alpha;j,\beta}u_{j,\beta},$$
the vibrational frequencies $\omega$ and the corresponding eigenmodes $u_{i,\alpha}$ are yielded, where the $M_{i}$ is the mass of $ith$ atom, $C_{i,\alpha;j,\beta}$ is the force constant, and $\alpha$ ($\beta$) means the direction of x, y or z.
Diffusion of a H atom in Ag-doped ZnO
=====================================
Firstly, we replace a Zn atom in the perfect ZnO supercell by a Ag atom, then optimize the system with the considerations above. Then a H atom is located around some typical O atoms, which are notated as $O_I$, $O_{II}$, $O_{III}$, $O_{IV}$ and $O_{V}$ respectively, as shown in Fig.1. These O atoms are near the doped Ag within the fifth neighbour. For the structure of ZnO bulk, there are two kinds of inequivalent $Zn-O$ bonds which is either parallel to $c$ axis or within the basal plane respectively. Therefore, for the Ag-doped ZnO bulk, there are four inequivalent sites available for the location of H around each concerned O atom: bond center (BC) sites and antibond (AB) sites of $Zn(Ag)-O$ bonds parallel to the $c$ axis and those of $Zn(Ag)-O$ bonds in the basal plane. From the calculations, we find that some of these sites are unstable for H, where the located H atom spontaneously moves to other places during full relaxation. For example, when a H atom is located at either AB site of $Zn-O_{II}$ bond parallel to the $c$ axis or BC site of $Zn-O_{II}$ bond in the basal plane, it spontaneously migrates to a BC site or AB site of $Ag-O_{II}$ bond during geometry optimization. For convenience, some stable or metastable sites for H are marked with characters of $A_{n}$ (n=1 to 3) for AB sites and $B_{n}$ (n=1 to 7) for BC sites in Fig.1. The coordinates of these sites and the doped Ag are listed in Table I.
A1 A2 A3 B1 B2 B3 B4 B5 B6 B7 Ag
--- ------ ------ ------- ------ ------ ------- ------ ------- ------ ------ ------
x 1.71 1.64 -1.13 1.65 0.63 -0.58 1.07 -0.03 2.68 1.03 1.64
y 2.87 2.84 4.65 2.86 3.43 4.13 3.14 5.55 3.42 5.96 2.85
z 8.80 3.37 4.18 6.62 4.91 3.86 1.92 6.62 4.92 4.57 5.32
: The optimized coordinates of H at the stable (metastable) sites near the doped Ag. The coordinates of the doped Ag is also listed. The positions for H refer to Fig.1. The lattice vectors of the supercell are $a$=(9.85, 0.0, 0.0), $b$=(-4.93, 8.53, 0.0), and $c$=(0.0, 0.0, 10.65). The unit of the length is in angstrom.
Among these sites, site $B2$ in Fig.1 is the ground state site for H location, which is similar to that of H in Cu doped ZnO [@Hu]. Meanwhile, site $A3$ is another preference site for H, where the energy of the system is only 0.03 eV higher than that at site $B2$. Nevertheless, the obtained energies for the other metastable sites are higher than that at site $B2$ by less than 1.0 eV.
![(Color online) Schematic showing of the concerned diffusion paths of H around the doped Ag in ZnO, indicated by path-1 to path-11. As labeled, green, purple, red and gray balls represent O, Zn, Ag and H atoms respectively. Some O atoms near the Ag atom are marked with Romanic numbers. The stable and metastable sites for H location in the diffusion paths are marked with symbols of $A_{n}$ (n=1 to 3) and $B_{n}$ (n=1 to 7).[]{data-label="Fig.1"}](Fig1.eps){width="45.00000%"}
Since the ground state and metastable sites for H are obtained, we now study the diffusion behaviors of H near the doped Ag. Eleven diffusion paths are considered, which are marked in Fig.1. By using the climbing image nudged elastic band (CI-NEB) scheme [@Jonsson1; @Jonsson2], the energies for the images in each diffusion path are calculated. Taking the lowest energy of the system (with H at site $B2$) as a reference, the energy profiles for each concerned diffusion path are plotted in Fig.2. As shown in Fig.2 (a), once a H atom arrives at the site of either $A1$ or $B1$ near $O_I$, it can migrate to the ground state site ($B2$), by overcoming energy barriers of less than 1.0 eV.
In contrast, when a H atom reaches site $B4$ near $O_{III}$, it can diffuse to site $B2$ via site $B3$ (path-4, path-3), with energy cost of less than 0.5 eV. For a H atom at site $A2$, it prefers to migrate towards site $B2$ directly through path-5, rather than towards site $A3$ via site $B3$ (path-6, path-7), because the energy barrier in the former case is about 0.4 eV, while that in the latter case is about 0.7 eV, which can be seen in Fig.2 (b). Meanwhile, if a H atom reaches near $O_{V}$ (at site $B5$) or $O_{IV}$ (site $B7$), it can reach site $A3$ through path-8 or path-11, with overcoming a low barrier of about 0.4 or 0.3 eV, as shown in Fig.2 (b,c). Interestingly, the energy barriers for H diffusing between the equivalent sites of $B2$ and $B6$ are less than 0.5 eV (Fig.2 (c), path-9), which indicates that a H atom can move “freely” between $B2$ and its equivalent sites near the doped Ag at room temperature. Such local motion of H near the doped Ag was also found in the Cu-doped ZnO [@Hu]. Moreover, H can diffuse between the low energy sites of $B2$ and $A3$ through two pathways. One is from site $B2$ to site $A3$ directly via path-10 with energy barriers of less than 0.9 eV, and the other is to site $B3$ through path-3 firstly, then to site $A3$ through path-7 by overcoming the energy barriers of no more than 0.7 eV. On the other hand, one can find from the Fig.2 that once a H atom locates at either site $B2$ or $A3$, the dissociation of such a Ag-H complex needs the energy cost as high as about 1.1 - 1.3 eV.
![Energy profiles calculated with CI-NEB for the diffusion paths of H in Ag-doped ZnO: (a) from path-1 to path-4, (b) from path-5 to path-8, and (c) from path-9 to path-11. The calculated relative energies for the images in each path are indicated with stars. The diffusion paths and reaction coordinates shown in this figure refer to Fig.1.[]{data-label="Fig.2"}](Fig2.eps){width="40.00000%"}
From above, we can conclude that a H atom can diffuse to the doped Ag to form a Ag-H complex in ZnO through many paths, and the lowest diffusion barrier is 0.3 eV only. In contrast, the dissociation of a Ag-H complex requires energy costs of more than 1.1 eV. This indicates that the Ag-H complex may easily exist in the Ag-doped ZnO. We stress that such stable complexes of Ag-H in ZnO play a negative role in p-type ZnO with doped Ag.
To go further, we pay our attention to the influence of a Ag-H complex on electronic structures of the system. The calculated total density of states (TDOS) for the systems with a H atom at site $B2$ and $A3$ are shown in Fig.3 a and b respectively. For comparison, the TDOS for the Ag monodoped ZnO and the perfect ZnO with the same size are shown in Fig.3 (c). Obviously shown in Fig.3 (a and b), some occupied defect states appear near Fermi level for the cases of H at site $B2$ and site $A3$. By analysis of the local density of states (LDOS), we find that these defect states mainly come from the contribution of the doped Ag atom and its neighbouring O atoms, and little from H and its bonded O atom, as displayed in Fig.3 (d and e). Furthermore, the projected density of states (PDOS) analysis reveals that these states are mainly contributed from 3d orbitals of the Ag atom and 2p orbitals of the O atoms.
![(Color online) The total density of states (TDOS) of the system for (a) the Ag-H complex with H at site $B2$, (b) the Ag-H complex with H at site $A3$, (c) ZnO with and without Ag doping, and the corresponding local density of states (LDOS) (d), (e) and (f). The dash line in (c) stands for the TDOS of the perfect ZnO, and the solid lines in (c) stand for Ag-doped ZnO with majority spin and minority spin. O1 refers to the O atom being far away from the the doped Ag; O2 and O3 stand for the O atoms in Ag-O parallel and perpendicular to the $c$ axis respectively, and O4 for the O atom in H-O bond. The dot-dash lines indicate the Fermi levels, which are shifted to be zero.[]{data-label="Fig.3"}](Fig3.eps){width="35.00000%"}
vibrational properties
======================
Vibration is a fundamental property of a system, which is closely coupled with the structure of the system. For such a Ag-H complex in ZnO, some new vibrational features should appear in the vibrational spectrum. With the frozen phonon approximation, we explore the vibrational properties of the ZnO containing a Ag-H complex. As a reference, we firstly calculate the total vibrational density of states (TVDOS) for the perfect ZnO, which is plotted in Fig.4 (b). The high frequency range for the perfect ZnO is below 570 $cm^{-1}$, and a silent region exists between 270 and 400 $cm^{-1}$. This is consistent with the previous report [@Hewot].
![(Color online) (a) The local vibrational density of states (LVDOS) and (b) total vibrational density of states (TVDOS) for the Ag-H complex with H at site $B2$, where frequency range is below 800 cm$^{-1}$. The stretch modes of H-O are not included. 1,2,3,4 and 5 in (a) represent the averaged LVDOS for the O atoms far from Ag, the O atoms around Ag, the O atom bonding with H, Ag and H respectively. In (b), the red dash line stands for the TVDOS of the perfect ZnO, and the solid line for that of the complex system with Ag-H. The states marked with the characters indicate the new states induced by the Ag-H complex.[]{data-label="Fig.4"}](Fig4.eps){width="35.00000%"}
For the system containing a Ag-H complex with H at site $B2$, a typical stretching mode of H-O bond is found to be at 3275 $cm^{-1}$, and no state between 782 and 3275 $cm^{-1}$. The calculated TVDOS in the region below 782 $cm^{-1}$ is shown in Fig.4 (b). From Fig.4 (b), we find that some new vibrational states, which are marked with $A$, $B$, $C$ and $D$ respectively, appear in the silent frequency region of the perfect ZnO.
The local vibrational density of states (LVDOS) analysis reveals that the states at lower frequency region of TVDOS (below 270 $cm^{-1}$) for the perfect ZnO comes from the contribution of the Zn atoms, and those at higher frequency region (above 400 $cm^{-1}$) from the O atoms. This remains in the TVDOS for ZnO containing a Ag-H complex. Furthermore, we find that the vibrational states marked with characters of $A$ (below 400 $cm^{-1}$) and $B$ (above 570 $cm^{-1}$) in Fig.4 (b) are mainly from the contribution of four nearest neighbouring O atoms around Ag, and little from the O atoms far from Ag, which can be seen obviously in Fig.4 (a). The presence of these modes are ascribed to the local structural distortion induced by the Ag-H complex. By examining the configuration of the system, we find that the Ag-O bonds both parallel and perpendicular to $c$ axis are all enlarged by about 0.2 [Å]{}. Consequently, the Zn-O bonds in the vicinity are shortened or enlarged within 0.04 [Å]{}, and the bond lengths of the next neighbouring Zn-O are altered slightly. Such structural changes result in presence of these new states near $A$ and $B$ in Fig.4 (b). Meanwhile, the vibrational states marked with $C$ (at about 646 $cm^{-1}$) and $D$ (782 $cm^{-1}$) are attributed to the bending modes of the H-O bond. In addition, the states contributed from the doped Ag mainly emerge at the low frequency region, as shown in Fig.4 (a).
For the case of a H atom at site $A3$, we also calculate the TVDOS of the system, and its main features are similar to those of H at site $B2$. The stretching mode of the H-O bond locates at the frequency of 3287 $cm^{-1}$, and the bending modes of the H-O bond are at 768 and 879 $cm^{-1}$ respectively. Similarly, some new states associated with the neighbouring O atoms around the doped Ag emerge near the frequency regions of below 400 and above 570 $cm^{-1}$ respectively. Especially, a mode corresponding to displacement of H and its bonded O atom in the same direction is found at the frequency of 273 $cm^{-1}$. These typical vibrational modes may drip a hint to detect the presence of such a Ag-H complex in experiment.
Summary
=======
In summary, the site preferences and diffusion behaviors of a H atom in Ag-doped ZnO are investigated based on density functional theory calculations. We find that the lowest-energy site for H in the Ag-doped ZnO is bond-centered sites of Ag-O in the basal plane. H can migrate to these sites through some typical diffusion paths with energy barriers of 0.3 - 1.0 eV, most of which are less than 0.7 eV. H can diffuse between its ground state sites with the energy cost of less than 0.5 eV. In contrast, releasing of this H atom from the doped Ag requires energy of about 1.1 - 1.3 eV, which indicates that the Ag-H complex may commonly exist in Ag-doped ZnO. By calculating the vibrational properties of ZnO with a Ag-H complex, we find some new vibrational modes in the silent frequency region of perfect ZnO, which are attributed to the distortion of the O atoms nearby induced by the Ag-H complex.
Acknowledgments
===============
This work is supported by the University of Science and Technology of China, the Chinese academy of Science, National Science Foundation of China (Grant Nos. NSFC10574115 and NSFC50721091). B. C. Pan thanks the support of National Basic Research Program of China (2006CB922000). This indicates that the Ag-H complexes can commonly exist in ZnO. The HP-LHPC of USTC is acknowledged for computational support.
[99]{} D. C. Look, D. C. Reynolds, J. R. Sizelove, R. L. Jones, C. W. Litton, G. Cantwell, and W. C. Harsch, Solid State Commun. **105**, 399 (1998). S. Y. Lee, Y. Li, J.S. Lee, J.K. Lee, M. Nastasi, S.A. Crooker, Q.X. Jia, H.S. Lee, and J.S. Kang, Appl. Phys. Lett. **85**, 218 (2003). H. S. Kang, J. S. Kang, J. W. Kim, and S. Y. Lee, J. Appl. Phys. **95**, 1246 (2004). W. W. Wenas, A. Yamada, and K. Takahashi, J. Appl. Phys. **70**, 7119 (1991). T. Mitsuyu, S. Ono, and K. Wasa, J. Appl. Phys. **51**, 2646 (1980). S. B. Zhang, S. H. Wei, and A. Zunger, J. Appl. Phys. **83**, 3192 (1998), and references therein. A. Zunger, Appl. Phys. Lett. **83**, 57 (2003). S.B. Zhang, S.-H. Wei, and A. Zunger, Phys. Rev. B **63**, 075205 (2001). S.-H. Wei, Comput. Mater. Sci. **30**, 337 (2004). D.C. Look, R.L. Jones, J.R. Sizelove, N.Y. Garces, N.C. Giles, and L. E. Halliburton, Phys. Status Solid A **195**, 171 (2003). H.S. Kang, B.D. Ahn, J.H. Kim, G.H. Kim, S.H. Lim, H.W. Chang, and S.Y. Lee Appl. Phys. Lett. **88**, 202108 (2006). C. G. Van de Walle, Phys. Rev. Lett. **85**, 1012 (2000). S. F. J. Cox, E. A. Davis, S. P. Cottrell, P. J. C. King, J. S. Lord, J. M. Gil, H. V. Alberto, R. C. Vilao, J. Piroto Duarte, N. Ayres de Campos, A. Weidinger, R. L. Lichti, and S. J. C. Irvine, Phys. Rev. Lett. **86**, 2601 (2001). D. M. Hofmann, A. Hofstaetter, F. Leiter, H. Zhou, F. Henecker, B. K. Meyer, S. B. Orlinskii, J. Schmidt and P. G. Baranov, Phys. Rev. Lett. **88**, 045504 (2002). K. Ip, M. E. Overberg, Y. M. Heo, D. P. Norton, S. J. Pearton, C. E. Stutz, B. Luo, F. Ren, D.C. Look, and J.M. Zavada, Appl. Phys. Lett. **82**, 385 (2003). N. H. Nickel, Phys. Rev. B **73**, 195204 (2006). M.G. Wardle, J.P. Goss, and P.R. Briddon, Phys. Rev. Lett. 96, 205504 (2006). X. N. Li, B. Keyes, S. Asher, S. B. Zhang, S. H. Wei, T. J. Coutts, S. Limpijumnong, C. G. Van de Walle, Appl. Phys. Lett. **86**, 122107 (2005). O. S. Kumar, E. Watanabe, R. Nakai, N. Nishimoto, and Y. Fujita, J. Cryst. Growth **298**, 491 (2007). L. E. Hallibuton, L. J. Wang, L. H. Bai, N. Y. Garces, N. C. Giles, M. J. Callahan, and B. G. Wang, J. Appl. Phys. **96**, 7168 (2004). M. G. Wardle, J. P. Goss, and P. R. Briddon, Phys. Rev. B **71**, 155205 (2005). G. A. Shi, M. Stavola, and W. B. Fowler, Phys. Rev. B **73**, 081201(R) (2006). K. R. Martin, P. Baney, G. Shi, M. Stavola, and W. B. Fowler, Phys. Rev. B **73**, 235209 (2006). E. V. Lavrov, F. Borrnert, and J. Weber, Phys. Rev. B **71**, 035205 (2005). M. G. Wardle, J. P. Goss, and P. R. Briddon, Phys. Rev. B **72**, 155108 (2005). F. Borrnert, E. V. Lavrov, and J. Weber, Phys. Rev. B **75**, 205202 (2007). D. Sanchez-Portal, P. Ordejon, E. Artacho and J. M. Soler, Int. J. Quantum Che **65**, 453 (1997); N.Troullier and J. L. Martins, Phys. Rev. B **43**, 1993 (1991). J. M. Soler, E. Artacho, J. D Gale, A. Garcia, J. Junquera, P. Ordejon, and D. Sanchez-Portal, Phys. Condens. Matter **14**, 2745 (2002), and references therein. H. J. Monkhorst and J. D. Pack, Phys. Rev. B **13**, 5188 (1976). J. Albertsson, S.C. Abrahams, and A. Kvick, Acta Crystallogr., sect. B: Struct. Sci. 45, 34 (1989); Handbook of Chemistry and Physics, 85th ed., edited by D.R. Lide (CRC Boca Raton, FL, 2004). L. Yan, C.K. Ong, X.S. Rao, J. Appl. Phys. 96, 508 (2004). M.T. Yin, M.L. Cohen, Phys. Rev. B **26**, 3259 (1982). J. Hu, and B.C. Pan, J. Phys. Chem. C 112, 19142 (2008). G. Henkelman, B. P. Uberuaga, and H. Jonsson, J. Chem. Phys. **113**, 9901 (2000). G. Henkelman, and H. Jonsson, J. Chem. Phys. **113**, 9978 (2000). A. W. Hewot, Solid State Commun. **8**, 187 (1970).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Matthias Steinmetz
- Tomaž Zwitter
- Gal Matijevic
- Alessandro Siviero
- Ulisse Munari
title: A comparison between RAVE DR5 and Gaia DR2 Radial Velocities
---
The Radial Velocity Experiment ([@2006ApJ...132..1645]) is a massive spectroscopic campaign of some 460000 stars in the southern hemisphere. RAVE spectra were taken with the 6dF multi-object spectrograph on the 1.2m UK Schmidt telescope of the Australian Astronomical Observatory (formerly Anglo-Australian Observatory). RAVE spectra are taken at a resolution of $R\approx 7500$ in the IR Calcium triplet region (8410 - 8795Å). Observations for RAVE were taken between April 2003 and April 2013.
On 25 April 2018, the 2nd data release of the ESA mission Gaia was published (), featuring radial velocities for some 7 million targets. The Radial Velocity Spectrometer (RVS) of Gaia also operates in the Ca triplet region, though at a somewhat higher resolution of $R=11000$. This research note presents a comparison of radial velocities between the most recent data release of RAVE ([@2017AJ...153..75K]) with Gaia DR2. RAVE DR5 and Gaia DR2 have 450587 stars in common. In particular we would like to shed light on a small subset of joint targets (707 stars) that exhibit a constant velocity offset of $+105$ [kms$^{-1}$]{}(see also [@arXiv:1807.11716]) or $-76$ [kms$^{-1}$]{}, respectively. Since RAVE provided the largest subset of targets for the Gaia pipeline verification, the two data sets are not fully independent.
1. Figure 1 compares the radial velocities published in Gaia DR2 with those published in RAVE DR5. Overall this comparison confirms the excellent agreement between those two data sets. The velocity differences can well be matched with two Gaussians with FWHM of $1.2$ [kms$^{-1}$]{}and $3.6$ [kms$^{-1}$]{}, respectively. There is a systematic offset of about $-0.32$ [kms$^{-1}$]{}, consistent with the findings published in [@arXiv:1807.11716], who find an offset of $-0.29$ [kms$^{-1}$]{}. The offset is also comparable to the offset found between Gaia DR2 and other ground-based spectroscopic surveys in a similar magnitude range, such as APOGEE (), indicative that the source for this offset may at least partially be related to the radial velocity zeropoint of Gaia DR2. The difference is also within the error estimates published in the RAVE data release papers based in internal template matching errors, errors with external samples, and errors derived from a subset of stars with repeat observations.
2. A further analysis exhibits no systematic tendency with RAVE derived effective temperatures for stars with 4000K$<T_{\rm eff}<$7000K. Stars cooler than 4000K exhibit a somewhat smaller shift of $-0.1$ [kms$^{-1}$]{}. For stars hotter than 7000K (a small subset of the RAVE sample), the accuracy of the radial velocity deteriorates resulting in a larger systematic shift and a considerably increased FWHM, owing to the increasing dominance of broad Paschen lines at the expense of a less prominent Calcium triplet. With increasing SNR, the prominence of the $1.2$ [kms$^{-1}$]{}Gaussian increases, and that of the $3.6$ [kms$^{-1}$]{}Gaussian decreases. Dwarfs stars (log $g >3.5$) have a lower fraction of stars in the $1.2$ [kms$^{-1}$]{}Gaussian than giant stars. There is a very mild tendency that the velocity shift between RAVE DR5 and Gaia DR2 changes with metallicity. This effect amounts to about $0.5$ [kms$^{-1}$]{}between $[Fe/H]<-1$ and $[Fe/H]>0$.
3. Figure 1 (left) plots the radial velocities published in RAVE DR5 against those in Gaia DR2. This figure shows a small group of 707 stars that exhibit a velocity offset of $\sim 105$ [kms$^{-1}$]{}or $\sim -76$ [kms$^{-1}$]{}. The first group has independently been identified by [@arXiv:1807.11716]. Thanks to the availability of additional data including spectra, observing data and setup, and fiber allocation, we are in a position to identify the cause for this velocity offset. While this group of stars exhibit on average a somewhat lower signal-to-noise than the RAVE DR5 catalog in general, there is no clear correlation of observing date or epoch, stellar parameters, signal-to-noise or any other parameter published in the DR5 catalog, technically or astronomically. However a deeper inspection exhibits that these objects can be found almost exclusively near the edge of the robotically positioned field plate, i.e. at very low or very high fiber number. The camera used in the 6dF spectrograph has a very fast focal ratio resulting in a large field curvature at the edges of the focal plane. The wavelength calibration, i.e. the mapping between the column number of a CCD pixel to the corresponding wavelength, thus rapidly changes between adjacent fibres near the edges of the focal plane. In a small number of fields our reduction pipeline encountered a problem in propagating the wavelength calibration between neighboring fibres near the edges of the field plate, in particular for observations done just before fibre repairs, i.e. when live fibres were separated by large gaps of broken ones. Such instances are clearly indicated in the corresponding log files which have now been recovered. In total we identified [**707**]{} stars with questionable wavelength calibration. In 82% of these cases the offset between RAVE DR5 and Gaia DR2 radial velocity is in the $+105 \pm 15$ [kms$^{-1}$]{}range. These spectra belong to the first 10 fibres of the field plate. Another 3% of the spectra have velocity offset in the $-76 \pm 4$ [kms$^{-1}$]{}range and were observed with the fibres 148-150. Even though the majority of these do not exhibit any particularity in the radial velocity, we recommend excluding them from future studies.
We have published a CSV-file with the RAVE\_OBS\_IDs of the suspicious targets on the RAVE DR5 survey website <http://www.rave-survey.org>. The corresponding author is also happy to email this file upon request. A more detailed analysis will be published in an upcoming paper featuring the final data release of the RAVE survey (Steinmetz et al., in preparation).
![Left: Radial Velocity derived from Gaia DR2 vs those from RAVE DR5. Right: Distribution of radial velocity differences between Gaia DR2 and RAVE DR5. The black line compares this distribution function with a fit using two Gaussian with a FWHM of x (red) and y(blue) km/s, respectively. A list with the RAVE\_OBS\_IDs of the outliers is attached. \[fig:1\]](rvcomp_dr5_clean.pdf)
Gaia Collaboration, Brown, A. G. A., et al. 2018, , 616, A1 (Gaia DR2)
Kunder, A. et al., C. 2017, , 153, 75 (RAVE DR5)
Reedy, D. & Reedy, B., 2018, arXiv:1807.11716
Sartoretti, P., et al. 2018, , 616, A6
Steinmetz, M. et al., C. 2006, , 132, 1645
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Seed black holes formed in the collapse of population III stars have been invoked to explain the presence of supermassive black holes at high redshift. It has been suggested that a seed black hole can grow up to $10^{5\sim 6}\sunm$ through highly super-Eddington accretion for a period of $\sim 10^{6\sim 7}$ yr between redshift $z=20\sim 24$. We studied the feedback of radiation pressure, Compton heating and outflow during the seed black hole growth. It is found that its surrounding medium fueled to the seed hole is greatly heated by Compton heating. For a super-critical accretion onto a $10^3\sunm$ seed hole, a Compton sphere (with a temperature $\sim 10^6$K) forms in a timescale of $1.6\times 10^3$yr so that the hole is only supplied by a rate of $10^{-3}$ Eddington limit from the Compton sphere. Beyond the Compton sphere, the kinetic feedback of the strong outflow heats the medium at large distance, this leads to a dramatical decrease of the outer Bondi accretion onto the black hole and avoid the accumulation of the matter. The highly super-critical accretion will be rapidly halted by the strong feedback. The seed black holes hardly grow up at the very early universe unless the strong feedback can be avoided.'
author:
- 'Jian-Min Wang, Yan-Mei Chen and Chen Hu'
title: Feedback Limits Rapid Growth of Seed Black Holes at High Redshift
---
Introduction
============
Black holes are regarded as an extremely important population in modern cosmological physics. The reionization of the Universe may get started from $z\sim 17$ deduced from [*Wilkinson Microwave Anisotropic Probe*]{} ([*WMAP*]{}) which gives Thomson scattering depth $\tau=0.17\pm 0.04$ (Spergel et al. 2003). Such an early reionization epoch needs a large population of seed black holes collapsed from population III stars (Madau et a. 2004). Second, the discovery of the currently known highest redshift quasar, SDSS 1148+3251 at $z=6.4$ (roughly 1 Gyr) from [*Sloan Digital Sky Survey*]{} (SDSS; Fan et al. 2001) indicates that there are already supermassive black holes with ${\ifmmode M_{\rm BH} \else $M_{\rm BH}$\ \fi}>10^9\sunm$ (Netzer 2003, Barth et al. 2003, Willott et al. 2003). What is the relation between the seed and supermassive black holes? How to form supermassive black holes?
Rees’ diagram shows several possible ways to form supermassive black holes (Rees 1984). A direct collapse of primordial clouds could form supermassive black holes after cosmic background radiation photons remove enough angular momentum through Compton drag (Loeb 1993, Loeb & Rassio 1994). This scenario is favored by Ly$\alpha$ fuzz of the extended emission in quasar where the supermassive black hole has been formed and the galaxy is assembling (Weidinger, Moller & Fynbo 2004), especially the recent discovery of an isolated black hole of the quasar HE 0450-2958 without a massive host galaxy (Magain et al. 2005). Second, a rapid growth of a seed black hole with highly super-Eddington accretion rates is used to explain the existence of the black hole $>10^9\sunm$ at high redshift. Third a compact cluster of main sequence stars, or neutron stars/black holes will inevitably evolve into a supermassive black hole (Duncan & Shapiro 1983, Quinlan & Shapiro 1990), and this gets supports from the quasar’s metallicity properties (Wang 2001). The different ways to form a supermassive black hole may apply to different redshifts or environments.
A rapid growth of seed black holes is quite a promising model to issue the formation of supermassive black holes at redshift $z\sim 6$ (Volonteri & Rees 2005). The Bondi accretion rate is $\dot{m}=\dotmb/\dotmedd\sim 40$, where $\dot{M}_{\rm Edd}=L_{\rm Edd}/c^2$ and $L_{\rm Edd}$ is the Eddington limit, for a $10^3\sunm$ black hole surrounded by medium cooled by the hydrogen atomic lines. The seed black hole is able to grow up exponentially ${\ifmmode M_{\rm BH} \else $M_{\rm BH}$\ \fi}=M_0\exp(\dot{m} t/t_{\rm Salp})$, where the Salpeter timescale $t_{\rm Salp}=0.45$Gyr. However, such a high accretion rate inevitably gives rise to strong interactions of radiation and outflows with the Bondi accretion flow. Consequently, the strong feedback seriously constrains the matter supply to the black hole, even stops the accretion.
In this Letter we discuss how the feedback impacts on the growth of the seed black holes. We find they can not grow up between redshift $z=20\sim 24$ through accretion. The implications of the present results are discussed.
Growth: Feedback Limit
======================
Angular momentum and accretion onto seed black holes
----------------------------------------------------
The very first population III stars rapidly evolve into intermediate mass black holes (IMBHs), $20<{\ifmmode M_{\rm BH} \else $M_{\rm BH}$\ \fi}/\sunm<70$ and $130<{\ifmmode M_{\rm BH} \else $M_{\rm BH}$\ \fi}/\sunm<600$ (Fryer et al. 2001, Omukai & Palla 2003). These IMBHs, as shown in Volonteri & Rees (2005), can rapidly grow up to supermassive black holes via Bondi accretion with a super-critical rate.
The strong tidal torque due to nearby clouds efficiently offer angular momentum $J$ to each other (Peebles 1969, Barnes & Efstathiou 1987). For a rigid virialized sphere with mass $M$, total energy $E$ and radius $R$, its angular momentum is characterized by $J_0=GM^{5/2}/|E|^{1/2}$, where $G$ is the gravitational constant (Peebles 1969). It is useful to define a dimensionless parameter, $\lambda\equiv J/J_0=J|E|^{1/2}G^{-1}M^{-5/2}$, for a cloud with mass $M$, angular momentum $J$ and the total energy $E$. Following Oh & Haiman (2002), a cold fat disk is formed with an isothermal exponential radial distribution $$n(R,h)=n_0\exp{\left(-\frac{2R}{R_d}\right)}
\sec^2\left(\frac{h}{\sqrt{2}H_0}\right),$$ where $n_0$ and $H_0$ are the central density and vertical scale height of the disk at radius $R$, $R_d$ is the radial scale length. The central
\[fig1\] 0.5cm
density is given by $$n_0\approx 6\times 10^4f_{d,0.5}^2~\lambda_{0.05}^{-4}\left(T_3/8\right)^{-1}
R_{\rm vir,6}^{-4}M_{\rm H,9}^2~~~{\rm cm^{-3}},$$ where $f_{d,0.5}=f_d/0.5$ is the gas fraction settled into the isothermal disk, $\lambda_{0.05}=\lambda/0.05$, $T_3=T/10^3{\rm K}$, $R_{\rm vir,6}=R_{\rm vir}/6{\rm kpc}$ and $M_{\rm H,9}=M_{\rm H}/10^9\sunm$ are the virialized radius and mass of the halo, respectively. The disk’s radius scale is $R_d\sim \lambda R_{\rm vir}/\sqrt{2}$.
The Bondi accretion radius of the seed black hole is $$R_{\rm Bondi}^{\rm I}=1.4\times 10^9R_{\rm G} ~m_3\left(T_3/8\right)^{-1},$$ where $m_3={\ifmmode M_{\rm BH} \else $M_{\rm BH}$\ \fi}/10^3\sunm$, $R_{\rm G}=G{\ifmmode M_{\rm BH} \else $M_{\rm BH}$\ \fi}/c^2$ and $c$ is the light speed. $R_{\rm Bondi}^{\rm I}$ is much smaller than the vertical scale of the cold fat disk. This means that the density within Bondi accretion radius is almost constant with a quasi-spherical geometry. As argued by Volonteri & Rees (2005), the transverse velocity at the Bondi radius ($R_{\rm Bondi}^{\rm I}$) is of the order of $\left(R_{\rm Bondi}^{\rm I}/R_d\right)v_{\rm D}$, where $v_{\rm D}$ is the rotation velocity of the fat gas disk at $R_d$. Due to the specific angular momentum of the quasi-Bondi flow, a tiny disk will form within $R_{\rm Bondi}^{\rm I}$. Figure 1 shows a cartoon of the present case with feedback. Assuming conservation of the specific angular momentum within $R_{\rm Bondi}^{\rm I}$, we have the outer radius of the tiny accretion disk $$R_{\rm disk}=6\times 10^2R_{\rm G}~ v_{\rm D,10}^2m_3^2\lambda_{0.05}^{-2}R_{\rm vir,6}^{-2},$$ where $v_{10}=v/10{\ifmmode {\rm km\ s}^{-1} \else km s$^{-1}$\ \fi}$, which is much smaller than the Bondi accretion radius. The quasi-sphere flow will switch on the tiny disk with a Bondi accretion rate, given by $\dot{M}_{\rm Bondi}=\gamma 4\pi nm_p\left(G{\ifmmode M_{\rm BH} \else $M_{\rm BH}$\ \fi}\right)^2/c_s^3$ (Bondi 1952) $$\frac{\dot{M}_{\rm Bondi}^{\rm I}}{\dot{M}_{\rm Edd}}=188~\gamma m_3(n_4/4)(T_3/8)^{-3/2},$$ where $c_s$ is the sound speed, $m_p$ is the proton mass, $\gamma$ is a dimensionless constant of order unity and $n_4=n/10^4{\rm cm^{-3}}$. The tiny accretion disk with such a high rate is characterized by photon-trapping self-similar structure within trapping radius, where the photon escaping timescale from the disk is equal to that of the radial moving to the black hole (Begelman & Meier 1982, Wang & Zhou 1999, Ohsuga et al. 2003). We note that the disk radius ($R_{\rm disk}$) is comparable to the photon trapping radius $R_{\rm tr}\approx 3.2\times 10^2 R_{\rm G}\left(\dot{m}/180\right)$ (Wang & Zhou 1999), indicating that the entire tiny disk may be advection-dominated via photon trapping process. According to the self-similar solution of the optically thick advection-dominated accretion flow (Wang & Zhou 1999), the radiated luminosity is only a small fraction of the released gravitational energy via viscosity dissipation $$L_{\rm Bol}\approx 4.0\times 10^{40}m_3\left[1+\frac{1}{4}\ln \left(\frac{\dot{m}}{180}\right)\right]~~~{\ifmmode {\rm erg\ s}^{-1} \else erg s$^{-1}$\ \fi},$$ where we have used the photon trapping radius $R_{\rm tr}$ implied above. This is a saturated luminosity weakly dependent on the accretion rate. The total released luminosity is given by $L_{\rm grav}=1.4\times 10^{42}(\eta/0.1)\left(\dot{m}/10^2\right)m_3$ [\^[-1]{} erg s$^{-1}$ ]{}, most of them are advected into the black hole.
The emergent spectrum from the slim disk is characterized by a universe shape of $F_{\nu}\propto \nu^{-1}$ (Wang & Zhou 1999, Wang et al. 1999, see also Shimura & Manmoto 2003, Kawaguchi 2003), the hard X-ray emission is highly dependent on the factor $f$, which indicates how many energy is released in the hot corona (Wang & Netzer 2003, Chen & Wang 2004, hereafter, Socrates & Davis 2005). Even $f=1\%$, the released energy from hot corona is close to Eddington limit. The physics related with the factor $f$ remains open. The Compton temperature reads $T_{\rm C}=\pe/4k$, where $k$ is the Boltzmann constant and $\pe$ is the mean energy of photons from the disk which depend on $F_{\nu}$. The mean energy of the photons are in a range of $0.4\sim 10$keV for $f=0.01\sim 0.025$ (Wang & Netzer 2003). We take the lower value $\pe=0.4$keV, so the medium will be heated up to $T_C\approx 10^6$K at least. It should be pointed out that the higher the Compton temperature is, the stronger the feedback from the tiny accretion disk has influence on its surroundings.
Feedback: radiation pressure and Compton heating
------------------------------------------------
The radiation pressure acting on the medium inside the Bondi accretion sphere is of order $P_{\rm rad}\sim \tau_{\rm es}L/4\pi R^2c$, where the Thomson scattering depth is $\tau_{\rm es}=Rn\sigma$ and $\sigma$ is Thomson cross section. We find the ratio of $P_{\rm rad}/P_{\rm gas}=L\sigma/4\pi R ckT$, beyond the radius $R_R=8.5\times 10^{7}R_{\rm G}~L_{40}T_4^{-1}m_3^{-1}$, the radiation pressure is negligible, where $T_4=T/10^4{\rm K}$. Since the photon trapping effects, the radiated luminosity from the highly super-critical accretion disk is lower than Eddington limit by a factor of 3 (see equation 6). So radiation pressure hardly prevents the infalling matter from accretion onto the black hole. The feedback from the radiation pressure is negligible.
The Bondi flow will be heated by the radiation from the tiny accretion disk. In turn, the heated medium determines a new Bondi accretion rate switched on the tiny disk. The thermal status of the photoionized medium can be conveniently described by the ionization parameter defined as $\Xi=L/4\pi R^2c nkT$ (Krolik, McKee & Tarter 1981). We focus on a region with $\Xi\ge \Xi_{\rm C}=1.1\times 10^{3}T_{C,6}^{-3/2}$ (for pure hydrogen gas), here $T_{C,6}=T_C/10^6{\rm K}$, where the entire medium will be heated up to Compton temperature $T_C$. It follows from $$R_{\rm Comp}^{\Xi}=\left(\frac{L}{4\pi c\Xi_c nkT}\right)^{1/2}=2.8\times 10^{8}R_{\rm G}~L_{40}^{1/2}
n_4^{-1/2}T_4^{-1/2}m_3^{-1},$$ which is called the Compton radius (and sphere). It needs a Compton timescale $t_C=6\pi m_ec^2R_{\rm Comp}^{{\Xi}^2}/\sigma L\sim 1.3\times 10^{5}{\rm yr}$ to form such a Compton sphere. However the Compton sphere is also constrained by the balance of the heating and infalling timescale of the Bondi flow. It follows from $t_C=t_{\rm infall}$, $$R_{\rm Comp}^{\rm infall}=\frac{\sigma L}{6\pi m_ec^2\beta_s c_s}
=3.1\times 10^7 R_{\rm G}~\beta_{s,0.1}^{-1}L_{40}T_4^{-1/2}m_3^{-1}$$ where $\beta_s=v_R/c_s$ is the infalling velocity of the Bondi flow normalized by the sound speed and $\beta_{s,0.1}=\beta_s/0.1$. The effects of the infalling on the Compton sphere can be neglected if the infalling velocity is less than $\beta_s \le 10^{-2}$. For $\beta_s=0.1$, we find $R_{\rm Comp}^{\Xi}>R_{\rm Comp}^{\rm infall}$. This means that [*not* ]{} entire region with $\Xi>\Xi_c$ can be heated up to the Compton temperature ($T_C$). So the Compton radius is given by $R_{\rm Comp}=\min\left(R_{\rm Comp}^{\Xi},R_{\rm Comp}^{\rm infall}\right)=R_{\rm Comp}^{\rm infall}$, corresponding to a timescale of $1.6\times 10^3$yr. We note that $R_{\rm Comp}< R_{\rm Bondi}^{\rm I}\ll R_d\approx 0.2\lambda_{0.05}R_{\rm vir,6}$ kpc.
Within the Compton sphere, the seed black hole has a new Bondi accretion radius $$R_{\rm Bondi}^{\rm II}=\frac{G{\ifmmode M_{\rm BH} \else $M_{\rm BH}$\ \fi}}{c_s^2}=1.1\times 10^7R_{\rm G}~T_6^{-1},$$ and a Bondi accretion rate, $$\frac{\dot{M}_{\rm Bondi}^{\rm II}}{\dot{M}_{\rm Edd}}=0.13~m_3 (n_4/4)T_6^{-3/2}.$$ We note that this radius is still much larger than the tiny disk $R_{\rm disk}$ and this accretion rate is much smaller than that in the pre-feedbacked. The timescale of the Compton heating at $R_{\rm Bondi}^{\rm II}$ is $t_C=2.0\times 10^2$yr. That is to say the beginning Bondi accretion will be halted suddenly. After the termination of the super-critical accretion, the hot plasma in Bondi sphere II will be cooled at a timescale of $1.8\times 10^4n_{C,2}^{-1}T_{C,6}^{1/2}$yr via free-free emission, where $n_{C,2}=n_C/10^2{\rm cm^{-3}}$ and $T_{C,6}=T_C/10^6{\rm K}$. Therefore it is expected that there is a steady accretion onto the seed black hole, but with a sub-Eddington rate, which is self-regulated.
When the Compton sphere forms, the pressure balance with its surroundings should hold, we have $n_CkT_C=n_0kT_0$, where the subscript $"C"$ and $"0"$ represent Compton sphere and its pre-feedbacked surroundings, respectively. So we can get the Bondi accretion rate $\dot{M}_{\rm Bondi}^{\rm II}$ inside the Compton sphere in the form of $n_0$ and $T_0$, $$\frac{\dot{M}_{\rm Bondi}^{\rm II}}{\dot{M}_{\rm Edd}}=1.3\times 10^{-3}m_3(n_{0,4}/4)T_{0,4}^{-3/2},$$ where $n_{0,4}=n_0/10^4{\rm cm^{-3}}$ and $T_{0,4}=T_0/10^4{\rm K}$. With such a low accretion rate, the seed black hole can not grow significantly.
Since $\dot{M}_{\rm Bondi}^{\rm I}\gg \dot{M}_{\rm Bondi}^{\rm II}$, the accretion onto the black hole from the Bondi sphere I is continuing, then the matter will pile up between the region of $R_{\rm Bondi}^{\rm II}\le R\le R_{\rm Bondi}^{\rm I}$. Such a dense shell will collapse when pressures are not powerful enough to balance it. Thus the accretion onto the seed black holes could be pulsive with a period. However, the oscillating accretion will be avoided if the outflow is supplied.
Feedback: outflow
-----------------
Wind/outflow is a generic property of the accretion disk. There are many pieces of evidence for the presence of strong outflows from accretion disks. In the quasar PG 1211+143, the outflow is indicated by the blueshifted X-ray absorption lines with a velocity of $0.1c$ and has an outflow mass rate $\dot{M}_{\rm out}\sim 1.6\sunm/{\rm yr}$ (Pounds et al. 2003). Similar strong outflows are found in ultraluminous X-ray sources (Mukai et al. 2003, Fabbiano et al. 2003) and SS 433 which clearly has strong outflows ($>5\times 10^{-7}\sunm{\rm yr^{-1}}$) from the highly super-critical accretion disk (Kotani et al. 1996). A strong outflow is expected to develop from the tiny accretion disk, leading to efficient kinetic feedback to its surroundings. Here we neglect the detailed micro-physics of the interaction between outflow and medium. The kinetic luminosity is given by $L_{\rm kin}\approx (\dot{m}f_m\beta^2)L_{\rm Edd}\approx L_{\rm Edd}$, where $f_m$ is the fraction of the outflow to the accretion rate and we take $\dot{m}f_m\beta^2=1$. The simple energy budget can provide a rough estimation of the outflow kinetic energy. When the outflow is damped by its surroundings, its most kinetic energy will dissipate at the sonic radius $$R_{\rm Sonic}\approx 1.0 {\rm pc}~L_{40}^{1/2}(n_4/6)^{-1/2}T_{0.1}^{-1/4},$$ where $T_{0.1}=T/0.1{\rm keV}$ is the temperature of the surroundings (Begelman 2004). This radius is much smaller than the characterized radius of the fat disk ($R_d\sim 0.2{\rm kpc}$), but it is comparable to the inner quasi-spheric part of the cold fat disk. The kinetic luminosity ($L_{\rm kin}$) can compensate the free-free radiation loss of the hot plasma within the radius $R_{\rm Sonic}$. For a typical outflow, the timescale of reaching the sonic point is $R_{\rm Sonic}/\beta c\approx 30(\beta/0.1)$yr. It would be interesting to note that this feedback timescale is much shorter than the Compton heating in the present case. We could draw a conclusion that the kinetic feedback from the outflow does play an important role in the growth of the seed hole.
On the other hand, the mass loss through the strong outflow definitely prevents the seed holes from growth. However the mass rate of the outflow is uncertain, we do not know exactly how the growth of the seed holes is affected via mass loss.
Finally, we would like to point out that we use the self-similar solutions of the super-Eddington accretion (Wang & Zhou 1999) based on the slim disk (Abramowicz et al 1988). The super-critical accretion remains a matter of debate since there are several instabilities, especially the photon bubble instability (Gammie 1998). The instability results in inhomogeneity of the disk, allowing the trapped photons to escape from the disk and enhance the disk radiation. If we employ the radiation leakage model at a level of $L_{\rm Bol}\sim 300 L_{\rm Edd}$ (Begelman 2002), the feedback will be much stronger than we have discussed above. Thus the results from the present paper is a lower limit of feedback.
Discussions
===========
For simplicity, we assume the radiation is isotropic from the tiny accretion disk. The anisotropic radiation from the slim disk could suppress the Compton heating in some degrees. The self-similar solution gives the ratio of the height to the radius $H/R=(5+\alpha^2/2)^{-1/2}\approx 0.45$, where $\alpha$ is the viscosity parameter (Wang & Zhou 1999). The half-opening angle is about $\theta\approx 66^{\circ}$, the radiation will be confined within $1-\cos\theta\approx 60\%$ of the $4\pi$ solid angle. This assumption is reasonable. However the anisotropic radiation from the tiny disk may lower the feedback of the Compton heating. Future work will consider how the anisotropy effects the feedback for a time-dependent situation, which is beyond the scope of the present paper.
There are other possible ways to form supermassive black holes. The direct collapse of primordial clouds in which the angular momentum is removed by the cosmic background photons (Loeb 1993, Loeb & Rassio 1994) can form a $10^6\sunm$ black hole at $z\sim 15$. If the seed black hole accrets dark matter, its growth is not suffering from the strong feedback (Hu et al. 2005). The inevitable fate of a compact cluster is the formation of a supermassive black hole (Quinlan & Shapiro 1990). These possible ways might apply to different environments in the universe, it depends on future observations to test them.
Conclusions
===========
We show that the feedback of the radiation pressure, Compton heating and outflows from the super-critical accretion disks will result in strong influence on the accretion itself. We show that the Compton heating almost quenches the super-critical accretion and the outflow from the tiny disk heats up the outer region so that accumulation of matter is avoided. A rapid growth of seed black holes is greatly suppressed and they are hardly able to grow up unless the strong feedback can be avoided, for example considering the anisotropy of the radiation from the tiny accretion disk.
Abramowicz, M. A., Czerny, B., Lasota, J. P. & Szuszkiewicz, E., 1988, ApJ, 332, 646 Barnes, J. & Efstathiou, G., 1987, ApJ, 319, 575 Barth, A. J., Martini, P., Nelson, C. H. & Ho, L. C., 2003, ApJ, 594, L95 Begelman, M. C., 2002, ApJ, 568, L97 Begelman, M. C., in Coevolution of Black Holes and Galaxies, from the Carnegie Observatories Centennial Symposia. Cambridge University Press, Edited by L. C. Ho, 2004, p. 375. Begelman, M. C. & Meier, D. L., 1982, ApJ, 253, 873 Bondi, H., 1952, MNRAS, 112, 195 Chen, L.-H. & Wang, J-.M., 2004, ApJ, 614, 101 Duncan, M. J. & Shapiro, S. L., 1983, ApJ, 268, 565 Fabbiano, G., et al., 2003, ApJ, 591, 843 Fan, X. et al., 2001, AJ, 122, 2833 Fryer, C. L., Woosley, S. E. & Heger, A., 2001, ApJ, 550, 372 Gammie, C. F., 1998, MNRAS, 297, 929 Hu, J., Shen, Y., Luo, Y. Q. & Zhang, S. N., 2005, MNRAS, in press (astro-ph/0510222) Kawaguchi, T., 2003, ApJ, 593, 69 Kotani, T., et al., 1996, PASJ, 48, 619 Krolik, J. H., McKee, J. F. & Tarter, C. B., 1981, ApJ, 249, 422 Loeb, A., 1993, ApJ, 403, 542 Loeb, A. & Rassio, F. A., 1994, ApJ, 432, 52 Madau, P., et al. 2004, ApJ, 604, 484 Magain, P. et al., 2005, Nature, 437, 381 McLure, R. J. & Dunlop, J. S., 2004, MNRAS, 352, 1390 Mukai, K., et al., 2003, ApJ, 582, 184 Netzer, H., 2003, ApJ, 583, L5 Oh, S. P. & Haiman, Z., 2002, ApJ, 569, 558 Ohsuga, K., Mineshige, S. & Watarai, K.-Y., 2003, ApJ, 596, 429 Omukai, K. & Palla, F., 2003, ApJ, 589, 677 Pounds, K., et al., 2003, MNRAS, 345, 750 Peebles, P. J. E., 1969, ApJ, 155, 393 Quinlan, G. & Shapiro, S. L., 1990, ApJ, 356, 483 Rees, M. J., 1984, ARA&A, 22, 471 Shimura, T. & Manmoto, T., 2003, MNRAS, 338, 1013 Socrates, A. & Davis, S. W., astro-ph/0511549 Spergel, D. N., et al., 2003, ApJS, 148, 175 Volonteri, M. & Rees, M. J., 2005, ApJ, 633, 624 Wang, J.-M., 2001, A&A, 376, L39 Wang, J.-M. & Netzer, H., 2003, A&A, 398, 927 Wang, J.-M., Szuszkiewicz, E., Lu, F.-J. & Zhou, Y.-Y., 1999, ApJ, 522, 839 Wang, J.-M. & Zhou, Y.-Y., 1999, ApJ, 516, 420 Weidinger, M, Moller, P. & Fynbo, J. P. U., 2004, Nature, 430, 999 Willott, C. J., McLure, R. J. & Jarvis, M. J., 2003, ApJ, 614, L15
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Brownian motion in ${\bf R}_+^2$ with covariance matrix $\Sigma$ and drift $\mu$ in the interior and reflection matrix $R$ from the axes is considered. The asymptotic expansion of the stationary distribution density along all paths in ${\bf R}_+^2$ is found and its main term is identified depending on parameters $(\Sigma, \mu, R)$. For this purpose the analytic approach of Fayolle, Iasnogorodski and Malyshev in [@fayolle_random_1999] and [@malyshev_asymptotic_1973], restricted essentially up to now to discrete random walks in ${\bf Z}_+^2$ with jumps to the nearest-neighbors in the interior is developed in this article for diffusion processes on ${\bf R}_+^2$ with reflections on the axes.'
address:
- 'Laboratoire de Probabilités et Modèles Aléatoires, Université Pierre et Marie Curie, 4 Place Jussieu, 75252 Paris Cedex 05, France & Laboratoire de Mathématiques et Physique Théorique, Université de Tours, Parc de Grandmont, 37200 Tours, France'
- 'Laboratoire de Probabilités et Modèles Aléatoires, Université Pierre et Marie Curie, 4 Place Jussieu, 75252 Paris Cedex 05, France'
author:
- 'S. Franceschi'
- 'I. Kurkova'
bibliography:
- 'biblio2.bib'
title: |
Asymptotic expansion of stationary distribution for reflected Brownian motion in the quarter plane\
via analytic approach
---
[^1]
Introduction and main results
=============================
Context
-------
Two-dimensional semimartingale reflecting Brownian motion (SRBM) in the quarter plane received a lot of attention from the mathematical community. Problems such as SRBM existence [@reiman_boundary_1988; @taylor_existence_1993], stationary distribution conditions [@harrison_diffusion_1978; @harrison_brownian_1987], explicit forms of stationary distribution in special cases [@dai_stationary_2013; @dieker_reflected_2009; @harrison_diffusion_1978; @harrison_multidimensional_1987; @latouche_product-form_2013], large deviations [@avram_explicit_2001; @dai_stationary_2013; @majewski_large_1996; @majewski_large_1998] construction of Lyapunov functions [@dupuis_lyapunov_1994], and queueing networks approximations [@harrison_diffusion_1978; @harrison_brownian_1993; @lieshout_tandem_2007; @lieshout_asymptotic_2008; @williams_approximation_1996] have been intensively studied in the literature. References cited above are non-exhaustive, see also [@williams_semimartingale_1995] for a survey of some of these topics. Many results on two-dimensional SRBM have been fully or partially generalized to higher dimensions.
In this article we consider stationary SRBMs in the quarter plane and focus on the asymptotics of their stationary distribution along any path in ${\bf R}_+^2$. Let $Z(\infty)=(Z_1(\infty), Z_2(\infty))$ be a random vector that has the stationary distribution of the SRBM. In [@dai_reflecting_2011], Dai et Myazawa obtain the following asymptotic result: for a given directional vector $c \in {\bf R}_+^2$ they find the function $f_c(x)$ such that $$\lim_{ x \to \infty} \frac{\mathbf{P}( \langle c \mid Z(\infty) \rangle {\geqslant}x)}{f_c(x)}=1$$ where $\langle \cdot \mid \cdot \rangle$ is the inner product. In [@dai_stationary_2013] they compute the exact asymptotics of two boundary stationary measures on the axes associated with $Z(\infty)$. In this article we solve a harder problem arisen in [@dai_reflecting_2011 §8 p.196], the one to compute the asymptotics of $$\mathbf{P}(Z(\infty) \in x c+B), \text{ as } x\to \infty,$$ where $c \in {\bf R}_+^2$ is any directional vector and $B \subset {\bf R}_+^2$ is a compact subset. Furthermore, our objective is to find [*the full asymptotic expansion of the density $\pi(x_1,x_2)$ of $Z(\infty)$ as $x_1,x_2 \to \infty$ and $x_2/x_1 \to {\rm \tan } (\alpha)$ for any given angle $\alpha \in ]0,\pi/2[$.*]{}
Our main tool is the analytic method developed by V. Malyshev in [@malyshev_asymptotic_1973] to compute the asymptotics of stationary probabilities for discrete random walks in ${\bf Z}_+^2$ with jumps to the nearest-neighbors in the interior and reflections on the axes. This method proved to be fruitful for the analysis of Green functions and Martin boundary [@kourkova_random_2011; @kurkova_martin_1998], and also useful for studying some joining the shortest queue models [@kurkova_malyshevs_2003]. The article [@malyshev_asymptotic_1973] has been a part of Malyshev’s global analytic approach to study discrete-time random walks in ${\bf Z}_+^2$ with four domains of spatial homogeneity (the interior of ${\bf Z}_+^2$, the axes and the origin). Namely, in the book [@malyshev_sluchainye_1970] he made explicit their stationary probability generating functions as solutions of boundary problems on the universal covering of the associated Riemann surface and studied the nature of these functions depending on parameters. G. Fayolle and R. Iasnogorodski [@fayolle_two_1979] determined these generating functions as solutions of boundary problems of Riemann-Hilbert-Carleman type on the complex plane. Fayolle, Iansogorodski and Malyshev merged together and deepened their methods in the book [@fayolle_random_1999]. The latter is entirely devoted to the explicit form of stationary probabilities generating functions for discrete random walks in ${\bf Z}_+^2$ with nearest-neighbor jumps in the interior. The analytic approach of this book has been further applied to the analysis of random walks absorbed on the axes in [@kourkova_random_2011]. It has been also especially efficient in combinatorics, where it allowed to study all models of walks in ${\bf Z}_+^2$ with small steps by making explicit the generating functions of the numbers of paths and clarifying their nature, see [@raschel_counting_2012] and [@kourkova_functions_2012].
However, the methods of [@fayolle_random_1999] and [@malyshev_asymptotic_1973] seem to be essentially restricted to discrete-time models of walks in the quarter plane with jumps in the interior only to the nearest-neighbors. They can hardly be extended to discrete models with bigger jumps, even at distance 2, nevertheless some attempts in this direction have been done in [@fayolle_about_2015]. In fact, while for jumps at distance 1 the Riemann surface associated with the random walk is the torus, bigger jumps lead to Riemann surfaces of higher genus, where the analytic procedures of [@fayolle_random_1999] seem much more difficult to carry out. Up to now, as far as we know, neither the analytic approach of [@fayolle_random_1999], nor the asymptotic results [@malyshev_asymptotic_1973] have been translated to the continuous analogs of random walks in ${\bf Z}_+^2$, such as SRBMs in ${\bf R}_+^2$, except for some special cases in [@baccelli_analysis_1987] and in [@foddy_analysis_1984]. This article is the first one in this direction. Namely, the asymptotic expansion of the stationary distribution density for SRBMs is obtained by methods strongly inspired by [@malyshev_asymptotic_1973]. The aim of this work goes beyond the solution of this particular problem. It provides the basis for the development of the analytic approach of [@fayolle_random_1999] for diffusion processes in cones of ${\bf R}_+^2$ which is continued in the next articles [@franceschi_tuttes_2016] and [@franceschi_explicit_2017]. In [@franceschi_explicit_2017] the first author and K. Raschel make explicit Laplace transform of the invariant measure for SRBMs in the quarter plane with general parameters of the drift, covariance and reflection matrices. Following [@fayolle_random_1999], they express it in an integral form as a solution of a boundary value problem and then discuss possible simplifications of this integral formula for some particular sets of parameters. The special case of orthogonal reflections from the axes is the subject of [@franceschi_tuttes_2016]. Let us note that the analytic approach for SRBMs in ${\bf R}_+^2$ which is developed in the present paper and continued by the next ones [@franceschi_tuttes_2016] and [@franceschi_explicit_2017], looks more transparent than the one for discrete models and deprived of many second order details. Last but not the least, contrary to random walks in ${\bf Z}_+^2$ with jumps at distance 1, it can be easily extended to diffusions in any cones of ${\bf R}^2$ via linear transformations, as we observe in the concluding remarks, see Section \[subsec:concludingremarks\].
Reflected Brownian motion in the quarter plane
----------------------------------------------
We now define properly the two-dimensional SRBM and present our results. Let $$\begin{aligned}
\begin{cases}
\Sigma = \left( \begin{array}{cc} \sigma_{11} & \sigma_{12} \\ \sigma_{12} & \sigma_{22} \end{array} \right) \in {\bf R}^{2 \times 2}
\text{ be a non-singular covariance matrix,}
\\
\mu= \left( \begin{array}{c} \mu_1 \\ \mu_2 \end{array} \right) \in {\bf R}^2
\text{ be a drift,}
\\
R= (R^1,R^2)=\left( \begin{array}{cc} r_{11} & r_{12} \\ r_{21} & r_{22} \end{array} \right) \in {\bf R}^{2 \times 2}
\text{ be a reflection matrix.}
\end{cases}\end{aligned}$$
The stochastic process $Z(t)=(Z^1(t), Z^2(t))$ is said to be a reflected Brownian motion with drift in the quarter plane ${\bf R}_+^2$ associated with data $(\Sigma, \mu, R)$ if $$Z(t)=Z_0 + W(t) + \mu t + R L(t) \ \in {\bf R}_+^2,$$ where
- $(W(t))_{t \in \mathbb{R^+}} $ is an unconstrained planar Brownian motion with covariance matrix $\Sigma$, starting from $0$;
- $L(t)=(L^1(t), L^2(t))$; for $i=1,2$, $L^i(t)$ is a continuous and non-decreasing process that increases only at time $t$ such as $Z^i(t)=0$, namely $\int_{0}^t 1_{\{Z^i(s) \ne 0 \}} \mathrm{d} L^i(s)=0$ $\forall t{\geqslant}0$;
- $Z(t)\in {\bf R}_+^2$ $\forall t{\geqslant}0$.
Process $Z(t)$ exists if and only if $r_{11} > 0$, $r_{22} > 0$ and either $r_{12},r_{21}>0$ or $r_{11} r_{22} - r_{12} r_{21} > 0$ (see [@taylor_existence_1993] and [@reiman_boundary_1988] which obtain an existence criterion in any dimension). In this case the process is unique in distribution for each given initial distribution of $Z_0$.
Columns $R^1$ and $R^2$ represent the directions where the Brownian motion is pushed when it reaches the axes, see Figure \[rebond\].
![Drift $\mu$ and reflection vectors $R^1$ and $R^2$[]{data-label="rebond"}](rebondpasorth.pdf)
The reflected Brownian motion $Z(t)$ associated with $(\Sigma, \mu, R)$ is well defined, and its stationary distribution $\Pi$ exists and is unique if and only if the data satisfy the following conditions: $$r_{11} > 0, \ \ r_{22} > 0, \ \ r_{11} r_{22} - r_{12} r_{21} > 0, \label{u}$$ $$r_{22} \mu_1 - r_{12} \mu_2 < 0, \ \ r_{11} \mu_2 - r_{21} \mu_1 < 0 . \label{v}$$
The proof and some more detailed statements can be found in [@hobson_recurrence_1993; @williams_recurrence_1985; @harrison_reflected_2009]. From now on we assume that conditions (\[u\]) and (\[v\]) are satisfied. The stationary distribution $\Pi$ is absolutely continuous with respect to Lebesgue measure as it is shown in [@harrison_brownian_1987] and [@dai_steady-state_1990]. We denote its density by $\pi(x_1,x_2)$.
Functional equation for the stationary distribution
---------------------------------------------------
Let $A$ be the generator of $(W_t +\mu t)_{t{\geqslant}0}$. For each $f \in \mathcal{C}^2_b ({\bf R}_+^2)$ (the set of twice continuously differentiable functions $f$ on ${\bf R}_+^2$ such that $f$ and its first and second order derivatives are bounded) one has $$Af(z)= \frac{1}{2} \sum_{i,j=1}^{2} \sigma_{i,j} \frac{\partial^2 f}{\partial z_i \partial z_j}(z) + \sum_{i=1}^2 \mu_i \frac{\partial f}{\partial z_i} (z).$$ Let us define for $i=1,2$, $$D_i f(x)= \langle R^i | \nabla f \rangle$$ that may be interpreted as generators on the axes. We define now $\nu_1$ and $\nu_2$ two finite boundary measures with their support on the axes: for any Borel set $B\subset {\bf R}_+^2$, $$\nu_i (B) = \mathbb{E}_{\Pi} [ \int_0^1 1_{\{Z(u) \in B\}} \mathrm{d}L^i (u)].$$ By definition of stationary distribution, for all $t\geqslant 0$, $\mathbb{E}_{\Pi} [f(Z(t))]= \int_{{\bf R}_+^2} f(z) \Pi (\mathrm{d} z)$. A similar formula holds true for $\nu_i$: $\mathbb{E}_{\Pi} [\int_0^t f(Z(u)) \mathrm{d}L^i(u) ]= t \int_{{\bf R}_+^2} f(x) \nu_i (\mathrm{d} x)$. Therefore $\nu_1$ and $\nu_2$ may be viewed as a kind of boundary invariant measures. The basic adjoint relationship takes the following form: for each $f \in \mathcal{C}^2_b ({\bf R}_+^2)$, $$\label{da}
\int_{{\bf R}_+^2} Af(z) \Pi (\mathrm{d} z) + \sum_{i=1,2} \int_{{\bf R}_+^2} D_i f(z) \nu_i (\mathrm{d} z) =0 .$$ The proof can be found in [@harrison_brownian_1987] in some particular cases and then has been extended to a general case, for example in [@dai_reflected_1992]. We now define $\varphi(\theta)$ the two-dimensional Laplace transform of $\Pi$ also called its moment generating function. Let $$\varphi (\theta) = \mathbb{E}_{\Pi} [\exp (\langle \theta | Z\rangle)] = \iint_{{\bf R}_+^2} \exp (\langle \theta | z \rangle) \Pi (\mathrm{d} z)$$ for all $\theta=(\theta_1,\theta_2) \in \mathbb{C}^2$ such that the integral converges. It does of course for any $\theta$ with ${\Re}\,\theta_1{\leqslant}0, {\Re}\,\theta_2{\leqslant}0$. We have set $\langle \theta | Z \rangle=\theta_1 Z^1 + \theta_2 Z^2$. Likewise we define the moment generating functions for $\nu_1(\theta_2)$ and $\nu_2(\theta_1)$ on ${\bf C}$: $$\varphi_2 (\theta_1) =
\mathbb{E}_{\Pi} [ \int_0^1 e^{\theta_1 Z_t^1} \mathrm{d} L^2(t)]
=\int_{{\bf R}_+^2} e^{\theta_1 z} \nu_2 (\mathrm{d} z),
\
\varphi_1 (\theta_2) =
\mathbb{E}_{\Pi} [ \int_0^1 e^{\theta_2 Z_t^2} \mathrm{d} L^1(t)]
=\int_{{\bf R}_+^2} e^{\theta_2 z} \nu_1 (\mathrm{d} z).$$ Function $\phi_2(\theta_1)$ exists a priori for any $\theta_1$ with ${\Re}\,\theta_1{\leqslant}0$. It is proved in [@dai_reflecting_2011] that it also does for $\theta_1$ with ${\Re}\, \theta_1 \in [0,\epsilon_1]$, up to its first singularity $\epsilon_1>0$, the same is true for $\phi_1(\theta_2)$. The following key functional equation (proven in [@dai_reflecting_2011]) results from the basic adjoint relationship (\[da\]).
For any $\theta\in{\bf R}_+^2$ such $\varphi (\theta)<\infty$, $\varphi_2 (\theta_1)<\infty$ and $\varphi_1 (\theta_2)<\infty$ we have the following fundamental functional equation: $$\label{maineq}
\gamma (\theta) \varphi (\theta) =\gamma_1 (\theta) \varphi_1 (\theta_2) + \gamma_2 (\theta) \varphi_2 (\theta_1),$$ where $$\begin{aligned}
\begin{cases}
\gamma (\theta)=- \frac{1}{2} \langle \theta | \Sigma \theta \rangle- \langle \theta | \mu \rangle =-\frac{1}{2}(\sigma_{11} \theta_1^2 + \sigma_{22} \theta_2^2 +2\sigma_{12}\theta_1\theta_2)
-
(\mu_1\theta_1+\mu_2\theta_2), \\
\gamma_1 (\theta)= \langle R^1 | \theta \rangle=r_{11} \theta_1 + r_{21} \theta_2, \\
\gamma_2 (\theta)=\langle R^2 | \theta \rangle=r_{12} \theta_1 + r_{22} \theta_2.
\end{cases}\end{aligned}$$
This equation holds true a priori for any $\theta=(\theta_1, \theta_2)$ with ${\Re}\,\theta_1{\leqslant}0, {\Re}\, \theta_2{\leqslant}0$. It plays a crucial role in the analysis of the stationary distribution.
Results
-------
Our aim is to obtain the asymptotic expansion of the stationary distribution density $\pi(x)=\pi(x_1,x_2)$ as $x_1,x_2 \to \infty$ and $x_2/x_1 \to {\rm tan}\,(\alpha_0)$ for any given angle $\alpha_0 \in [0, \pi/2]$.
[**Notation.**]{} We write the asymptotic expansion $f(x) \sim \sum_{k=0}^n g_k (x) $ as $x \to x_0$ if $g_k(x)=o(g_{k-1}(x))$ as $x \to x_0$ for all $k=1,\ldots, n$ and $f(x)-\sum_{k=0}^n g_k(x)=o(g_n(x))$ as $x \to x_0$.
It will be more convenient to expand $\pi(r \cos \alpha, r \sin \alpha)$ as $r \to \infty$ and $\alpha \to \alpha_0$. We give our final results in Section \[sec:asymptexpansion\], Theorems \[thmresults1\]–\[thmresults4\]: we find the expansion of $\pi(r \cos \alpha, r \sin \alpha)$ as $r \to \infty$ and prove it uniform for $\alpha$ fixed in a small neighborhood $\mathcal {O}(\alpha_0) \subset ]0, \pi/2[$ of $\alpha_0 \in ]0, \pi/2[$.
In this section, Theorem \[thmmain\] below announces the main term of the expansion depending on parameters $(\mu, \Sigma, R)$ and a given direction $\alpha_0$. Next, in Section \[subsec:sketch\] we sketch our analytic approach following the main lines of this paper in order to get the full asymptotic expansion of $\pi$. We present at the same time the organization of the article.
Now we need to introduce some notations. The quadratic form $\gamma(\theta)$ is defined in (\[maineq\]) via the covariance matrix $\Sigma$ and the drift $\mu$ of the process in the interior of ${\bf R}_+^2$. Let us restrict ourselves on $\theta \in {\bf R}^2$. The equation $\gamma(\theta)=0$ determines an ellipse $\mathcal{ E}$ on ${\bf R}^2$ passing through the origin, its tangent in it is orthogonal to vector $\mu$, see Figure \[ellipse\]. Stability conditions (\[u\]) and (\[v\]) imply the negativity of at least one of coordinates of $\mu$, see [@dai_reflecting_2011 Lemma 2.1]. In this article, in order to shorten the number of pictures and cases of parameters to consider, we restrict ourselves to the case $$\label{mumu}
\mu_1<0 \text{ and } \mu_2<0,$$ although our methods can be applied without any difficulty to other cases, we briefly sketch some different details at the end of Section 2.4.
![Ellipse $\mathcal{E}$, straight lines $\gamma_1(\theta)=0$, $\gamma_2(\theta)= 0$, points $\theta^{*}, \theta^{**}, \eta \theta^*, \zeta \theta ^{**}$[]{data-label="ellipse"}](ellipsedroites.pdf)
Let us call $s_1^+=(\theta_1(s_1^+), \theta_2(s_1^+))\in {{\mathcal}E}$ the point of the ellipse with the maximal first coordinate: $\theta_1(s_1^+)=\sup\{\theta_1 : \gamma(\theta_1,\theta_2)=0\}$. Let us call $s_2^+$ the point of the ellipse with the maximal second coordinate. Let ${{\mathcal}A}$ be the arc of the ellipse with endpoints $s_1^+$, $s_2^+$ [*not*]{} passing through the origin, see Figure \[pointpolgeom\]. For a given angle $\alpha \in [0,\pi/2]$ let us define the point $\theta(\alpha)$ on the arc ${{\mathcal}A}$ as $$\label{thetaalpha}
\theta(\alpha)={\rm argmax}_{\theta \in {{\mathcal}A}}\langle \theta \mid e_\alpha \rangle \ \ \hbox{where }e_\alpha=(\cos \alpha, \sin \alpha).$$ Note that $\theta(0)=s_1^+$, $\theta(\pi/2)=s_2^+$, and $\theta(\alpha)$ is an isomorphism between $[0, \pi/2]$ and ${{\mathcal}A}$. Coordinates of $\theta(\alpha)$ are given explicitly in . One can also construct $\theta(\alpha)$ geometrically: first draw a ray $r(\alpha)$ on ${\bf R}_+^2$ that forms the angle $\alpha$ with $\theta_1$-axis, and then the straight line $l(\alpha)$ orthogonal to this ray and tangent to the ellipse. Then $\theta(\alpha)$ is the point where $l(\alpha)$ is tangent to the ellipse, see Figure \[pointpolgeom\].
![Arc ${\mathcal}{A}$ and point $\theta(\alpha)$ on ${\mathcal}{E}$[]{data-label="pointpolgeom"}](pointcol1.pdf)
Secondly, consider the straight lines $\gamma_1(\theta)=0$, $\gamma_2(\theta)=0$ defined in (\[maineq\]) via the reflection matrix $R$. They cross the ellipse ${{\mathcal}E}$ in the origin. Furthermore, due to stability conditions (\[u\]) and (\[v\]) the line $\gamma_1(\theta)=0$ \[resp. $\gamma_2(\theta)=0$\] intersects the ellipse at the second point $\theta^{*}=(\theta_1^*, \theta_2^*)$ (resp. $\theta^{**}=(\theta_1^{**}, \theta_2^{**})$) where $\theta_2^*>0$ (resp. $\theta_1^{**}>0$). Stability conditions also imply that the ray $\gamma_1(\theta)=0$ is always “above” the ray $\gamma_2(\theta)=0$, see [@dai_reflecting_2011 Lemma 2.2]. To present our results, we need to define the images of these points via the so-called Galois automorphisms $\zeta$ and $ \eta$ of ${{\mathcal}E}$. Namely, for point $\theta^*=(\theta_1^*,\theta_2^*) \in {{\mathcal}E}$ there exists a unique point $\eta \theta^*=(\eta \theta_1^*, \theta_2^*) \in {{\mathcal}E}$ that has the same second coordinate. Clearly, $\theta_1^*$ and $ \eta \theta_1^*$ are two roots of the second degree equation $\gamma(\cdot, \theta_2^{*})=0.$ In the same way for point $\theta^{**}=(\theta_1^{**},\theta_2^{**}) \in {{\mathcal}E}$ there exists a unique point $\zeta \theta^{**}=( \theta_1^{**}, \zeta \theta_2^{**}) \in {{\mathcal}E}$ with the same first coordinate. Points $\theta_2^{**}$ and $ \zeta \theta_2^{**}$ are two roots of the second degree equation $\gamma(\theta_1^{**}, \cdot)=0.$ Points $\theta^*$, $\theta^{**}$, $\eta \theta^*$ and $\zeta \theta^{**}$ are pictured on Figure \[ellipse\]. Their coordinates are made explicit in (\[zerostheta\]) and (\[zerosthetaauto\]).
Finally let $s'_0=(0,-2\frac{\mu_{22}}{\sigma_{22}}
)$ be the point of intersection of the ellipse $\mathcal{E}$ with $\theta_2$-axis and let $s''_0=(-2\frac{\mu_{11}}{\sigma_{11}},0)$ be the point of intersection of the ellipse with $\theta_1$-axis, see Figure \[pointpolgeom\]. The following theorem provides the main asymptotic term of $\pi(r\cos \alpha, r \sin \alpha)$.
\[thmmain\] Let $\alpha_0 \in ]0, \pi/2[$. Let $\theta(\alpha)$ be defined in (\[thetaalpha\]). Let $\red \{ \theta(\alpha_0), s'_0 \}$ [(]{}resp. $\blue \{s''_0, \theta(\alpha_0)\}$[)]{} be the arc of the ellipse ${{\mathcal}E}$ with end points $s'_0$ and $\theta(\alpha_0)$ [(]{}resp. $s''_0$ and $\theta(\alpha_0)$[)]{} passing through the origin. We have the following results.
- If $\zeta\theta^{**}\notin \{ \theta(\alpha_0), s'_0 \}$ and $\eta\theta^{*}\notin \{s''_0, \theta(\alpha_0) \}$, then there exists a constant $c(\alpha_0)$ such that $$\label{aspointcol}
\pi(r \cos \alpha , r \sin \alpha)\sim \frac{c(\alpha_0)}{\sqrt{r} }
\exp\Big(- r \langle e_\alpha \mid \theta(\alpha) \rangle \Big) \ \ \ r\to \infty, \alpha \to \alpha_0.$$ The function $c(\alpha)$ varies continuously on $[0, \pi/2]$, $\lim_{\alpha \to 0} c(\alpha) =\lim_{\alpha \to \pi/2} c(\alpha)=0$. \[1\]
- \[2\] If $\zeta\theta^{**}\in \}\theta(\alpha_0), s'_0 \}$ and $\eta\theta^{*}\notin \{ s''_0,\theta(\alpha_0) \{$, then with some constant $c_1>0$ $$\label{aspol1}
\pi(r \cos \alpha , r \sin \alpha )\sim c_1
\exp\Big(-r \langle e_\alpha \mid \zeta \theta^{**} \rangle \Big) \ \ r\to \infty, \alpha \to \alpha_0.$$
- \[3\] If $\zeta\theta^{**}\notin \} \theta(\alpha_0), s'_0\}$ and $\eta\theta^{*}\in \{ s''_0, \theta(\alpha_0) \{ $, then with some constant $c_2>0$ $$\label{aspol2}
\pi(r \cos \alpha , r \sin \alpha)\sim c_2
\exp\Big(- r \langle e_\alpha
\mid \eta \theta^{*} \rangle \Big) \ \ r\to \infty, \alpha \to \alpha_0.$$
- \[4\] Let $\zeta\theta^{**}\in \} \theta(\alpha_0), s'_0 \}$ and $\eta\theta^{*}\in \{ s''_0,\theta(\alpha_0) \{$. If $\langle \zeta \theta^{**} \mid e_{\alpha_0} \rangle < \langle \eta \theta^{*} \mid e_{\alpha_0} \rangle$, then the asymptotics (\[aspol1\]) is valid with some constant $c_1>0$. If $\langle \zeta \theta^{**} \mid e_{\alpha_0} \rangle > \langle \eta \theta^{*} \mid e_{\alpha_0} \rangle$, then the asymptotics (\[aspol2\]) is valid with some constant $c_2>0$. If $\langle \zeta \theta^{**} \mid e_{\alpha_0} \rangle = \langle \eta \theta^{*} \mid e_{\alpha_0} \rangle$, then then with some constants $c_1>0$ and $c_2>0$ $$\label{aspol12}
\pi(r \cos \alpha , r \sin \alpha)\sim c_1
\exp\Big(-r \langle e_\alpha \mid \zeta \theta^{**} \rangle \Big) +
c_2 \exp\Big(-r \langle e_\alpha \mid \eta \theta^{*} \rangle \Big) \ \ \ r\to \infty, \alpha \to \alpha_0.$$
See Figure \[figthmprin\] for the different cases. (The arcs $\}a, b\}$ or $\{a, b\{$ of $\mathcal{E}$ are those not passing through the origin where the left or the right end respectively is excluded).
![Cases (1),(2),(3),(4)[]{data-label="figthmprin"}](ellipsethmA1.pdf "fig:") ![Cases (1),(2),(3),(4)[]{data-label="figthmprin"}](ellipsethmBa1.pdf "fig:") ![Cases (1),(2),(3),(4)[]{data-label="figthmprin"}](ellipsethmBb1.pdf "fig:") ![Cases (1),(2),(3),(4)[]{data-label="figthmprin"}](ellipsethmC1.pdf "fig:")
Let us note that the exponents in Theorem \[thmmain\] are the same as in the large deviation rate function found in [@dai_stationary_2013 Thm 3.2]. The same phenomenon is observed for discrete random walks, cf. [@malyshev_asymptotic_1973] and [@ignatyuk_influence_1994].
Sketch of the analytic approach. Organization of the paper {#subsec:sketch}
----------------------------------------------------------
The starting point of our approach is the main functional equation (\[maineq\]) valid for any $\theta= (\theta_1, \theta_2) \in {\bf C}^2$ with ${\Re}\, \theta_1{\leqslant}0$, ${\Re}\, \theta_2{\leqslant}0$. The function $\gamma(\theta_1, \theta_2)$ in the left-hand side is a polynomial of the second order of $\theta_1$ and $\theta_2$. The algebraic function $\Theta_1(\theta_2)$ defined by $\gamma(\Theta_1(\theta_2), \theta_2)\equiv 0$ is $2$-valued and its Riemann surface ${\bf S}_{\theta_2}$ is of genus $0$. The same is true about the $2$-valued algebraic function $\Theta_2(\theta_1)$ defined by $\gamma(\theta_1, \Theta_2(\theta_1))=0$ and its Riemann surface ${\bf S}_{\theta_1}$. The surfaces ${\bf S}_{\theta_1}$ and ${\bf S}_{\theta_2}$ being equivalent, we will consider just one surface ${\bf S}$ defined by the equation $\gamma(\theta_1, \theta_2)=0$ with two different coverings. Each point $s \in {\mathbf{S}}$ has two “coordinates” $(\theta_1(s), \theta_2(s))$, both of them are complex or infinite and satisfy $\gamma(\theta_1(s), \theta_2(s))=0$. For any point $s=(\theta_1,\theta_2) \in {\mathbf{S}}$, there exits a unique point $s'=(\theta_1, \theta'_2)\in {\mathbf{S}}$ with the same first coordinate and there exists a unique point $s''=(\theta''_1, \theta_2)\in {\mathbf{S}}$ with the same second coordinate. We say that $s'=\zeta s$, i.e. $s'$ and $s$ are related by Galois automorphism $\zeta$ of ${\bf S}$ that leaves untouched the first coordinate, and that $s''=\eta s$, i.e. $s''$ and $s$ are related by Galois automorphism $\eta$ of ${\bf S}$ that leaves untouched the second coordinate. Clearly $\zeta^2=Id$, $\eta^2=Id$ and the branch points of $\Theta_1(\theta_2)$ and of $\Theta_2(\theta_1)$ are fixed points of $\zeta$ and $\eta$ respectively. The ellipse $\mathcal{ E}$ is the set of points of ${\bf S}$ where both “coordinates” are real. The construction of ${\bf S}$ and definition of Galois automorphisms are carried out in Section \[riemannsurfaceS\].
Next, unknown functions $\varphi_1(\theta_2)$ and $\varphi_2(\theta_1)$ are lifted in the domains of ${\mathbf{S}}$ where $\{s \in {\mathbf{S}}: {\Re}\, \theta_2(s){\leqslant}0\}$ and $\{s \in {\mathbf{S}}: {\Re}\, \theta_1(s){\leqslant}0\}$ respectively. The intersection of these domains on ${\bf S}$ is non-empty, both $\varphi_2$ and $\varphi_1$ are well defined in it. Since for any $s=(\theta_1(s),\theta_2(s))\in {\mathbf{S}}$ we have $\gamma(\theta_1(s),\theta_2(s))=0$, the main functional equation (\[maineq\]) implies: $$\gamma_1(\theta_1(s), \theta_2(s) ) \phi_1(\theta_2(s))+\gamma_2(\theta_1(s), \theta_2(s) )\phi_2(\theta_1(s))=0 \ \ \forall
s \in {\mathbf{S}},{\Re}\,\theta_1(s){\leqslant}0, {\Re}\,\theta_2(s){\leqslant}0.$$ Using this relation, Galois automorphisms and the facts that $\varphi_1$ and $\varphi_2$ depend just on one ”coordinate” ($\phi_1$ depends on $\theta_2$ and $\phi_2$ on $\theta_1$ only), we continue $\varphi_1$ and $\varphi_2$ explicitly as meromorphic on the whole of ${\mathbf{S}}$. This meromorphic continuation procedure is the crucial step of our approach, it is the subject of Section \[meromcontinuation\]. It allows to recover $\varphi_1$ and $\varphi_2$ on the complex plane as multivalued functions and determines all poles of all its branches. Namely, it shows that poles of $\phi_1$ and $\phi_2$ may be only at images of zeros of $\gamma_1$ and $\gamma_2$ by automorphisms $\eta$ and $\zeta$ applied several times. We are in particular interested in the poles of their first (main) branch, and more precisely in the most ”important” pole (from the asymptotic point of view, to be explained below), that turns out to be at one of points $\zeta \theta^{**}$ or $\eta \theta^*$ defined above. The detailed analysis of the “main” poles of $\phi_1$ and $\phi_2$ is furnished in Section \[mainpole\].
Let us now turn to the asymptotic expansion of the density $\pi(x_1,x_2)$. Its Laplace transform comes from the right-hand side of the main equation (\[maineq\]) divided by the kernel $\gamma(\theta_1,\theta_2)$. By inversion formula the density $\pi(x_1,x_2)$ is then represented as a double integral on $\{\theta : {\Re}\, \theta_1={\Re}\, \theta_2=0\}$. In Section \[invlaplacetransf\], using the residues of the function $\frac{1}{\gamma(\theta_1, \cdot)}$ or $\frac{1}{\gamma(\cdot, \theta_2)}$ we transform this double integral into a sum of two single integrals along two cycles on ${\mathbf{S}}$, those where ${\Re}\, \theta_1(s)=0$ or ${\Re}\, \theta_2(s)=0$. Putting $(x_1,x_2)=r e_\alpha$ we get the representation of the density as a sum of two single integrals along some contours on ${\bf S}$: $$\label{121}
\pi(r e_\alpha)=
\frac{1}{2\pi \sqrt{\det \Sigma}}
\big(
\int_{\mathcal{I}_{\theta_1}^+}
\frac{\varphi_2(s) \gamma_2(\theta(s))}{s} e^{- r \langle \theta(s) \mid e_\alpha \rangle }
{\mathrm{d}s} +
\int_{\mathcal{I}_{\theta_2}^+}
\frac{\varphi_1(s) \gamma_1(\theta(s))}{s} e^{- r \langle \theta(s) \mid e_\alpha \rangle }
\mathrm{d}s \big) .$$ We would like to compute their asymptotic expansion as $r \to \infty$ and prove it to be uniform for $\alpha$ fixed in a small neighborhood $\mathcal{O}(\alpha_0)$, $\alpha_0\in ]0, \pi/2[$.
These two integrals are typical to apply the saddle-point method, see [@fedoryuk_asymptotic_1989; @pemantle_analytic_2013]. The point $\theta(\alpha) \in \mathcal{E}$ defined above is the saddle-point for both of them, this is the subject of Section \[subsec:saddle\]. The integration contours on ${\bf S}$ are then shifted to new ones $\Gamma_{\theta_1, \alpha}$ and $\Gamma_{\theta_2,\alpha}$ which are constructed in such a way that they pass through the saddle-point $\theta(\alpha)$, follow the steepest-descent curve in its neighborhood $\mathcal{O}(\theta(\alpha))$ and are “higher” than the saddle-point w.r.t. the level curves of the function $\langle \theta(s) \mid e_\alpha \rangle$ outside $\mathcal{O}(\theta(\alpha))$, see Section \[subsec:shiftingcontours\]. Applying Cauchy Theorem, the density is finally represented as a sum of integrals along these new contours and the sum of residues at poles of the integrands we encounter deforming the initial ones: $$\begin{gathered}
\label{122}
\pi(r e_\alpha )= \sum_{p \in {{\mathcal}P'}_\alpha} {\rm res}_p \phi_2(\theta_1(s))
\frac{\gamma_2(p)}{\sqrt{d(\theta_1(p))}} e^{-r \langle \theta(p) \mid e_\alpha \rangle } +
\sum_{p \in {{\mathcal}P''}_\alpha} {\rm res}_p \phi_1(\theta_2(s))
\frac{\gamma_1(p)}{\sqrt{\tilde d(\theta_2(p))}} e^{-r \langle \theta(p) \mid e_\alpha \rangle }
\\
\frac{1}{2\pi \sqrt{\det \Sigma}}
\left(
\int_{\Gamma_{\theta_1, \alpha} }
\frac{\varphi_2(s) \gamma_2(\theta(s))}{s} e^{- r \langle \theta(s) \mid e_\alpha \rangle }
{\mathrm{d}s} +
\int_{ \Gamma_{\theta_2, \alpha} }
\frac{\varphi_1(s) \gamma_1(\theta(s))}{s} e^{- r \langle \theta(s) \mid e_\alpha \rangle }
\mathrm{d}s \right).\end{gathered}$$ Here ${{\mathcal}P}'_\alpha$ (resp. ${{\mathcal}P''}_\alpha$) is the set of poles of the first order of $\phi_1$ (resp. $\phi_2$) that are found when shifting the initial contour $\mathcal{I}_{\theta_1}^+$ to the new one $\Gamma_{\theta_1,\alpha}$ (resp. $\mathcal{I}_{\theta_2}^+ $ to $
\Gamma_{\theta_2, \alpha} $), all of them are on the arc $\{s'_0, \theta(\alpha)\}$ (resp. $\{\theta(\alpha), s''_0\}$) of ellipse $\mathcal{E}$.
The asymptotic expansion of integrals along $\Gamma_{\theta_1,\alpha }$ and $\Gamma_{\theta_2,\alpha}$ is made explicit by the standard saddle-point method in Section \[subsec:asymptintalongshift\]. The set of poles ${{\mathcal}P}'_\alpha \cup {{\mathcal}P''}_\alpha$ is analyzed in Section \[subsec:contributionpole\] . In Case (1) of Theorem \[thmmain\] this set is empty, thus the asymptotic expansion of the density is determined by the saddle-point, its first term is given in Theorem \[thmmain\]. In Cases (2), (3) and (4) this set is not empty. The residues at poles over ${{\mathcal}P}'_\alpha \cup {{\mathcal}P''}_\alpha $ in (\[122\]) bring all more important contribution to the asymptotic expansion of $\pi(r e_\alpha)$ than the integrals along $\Gamma_{\theta_1,\alpha}$ and $\Gamma_{\theta_2,\alpha}$. Taking into account the monotonicity of function $\langle \theta \mid e_\alpha \rangle$ on the arcs $\{s''_0, \theta(\alpha)\}$ and on $\{\theta(\alpha), s'_0\}$, they can be ranked in order of their importance: clearly, the term associated with a pole $p'$ is more important than the one with $p''$ if $\langle p' \mid e_\alpha \rangle < \langle p'' \mid e_\alpha \rangle$. In Case (2) (resp. (3)) the most important pole is $\zeta \theta^{**} $ (resp. $\eta \theta^* $), as announced in Theorem \[thmmain\]. In Case (4) the most important of them is among $\zeta \theta^{**} $ and $ \eta \theta^{*} $, as stated in Theorem \[thmmain\] as well. The expansion of integrals in (\[122\]) along $\Gamma_{\theta_1,\alpha}$ and $\Gamma_{\theta_2,\alpha} $ via the saddle-point method provides all smaller asymptotic terms than those coming from the poles. Section \[sec:asymptexpansion\] is devoted to the results: they are stated from two points of view in Sections \[subsec:givenangle\] and \[subsec:givenparameters\] respectively. First, given an angle $\alpha_0$, we find the uniform asymptotic expansion of the density $\pi(r \cos (\alpha), r \sin (\alpha))$ as $r \to \infty$ and $\alpha \in \mathcal{O}(\alpha_0)$ depending on parameters $(\Sigma, \mu, R)$: Theorems \[thmresults1\] –\[thmresults4\] of Section \[subsec:givenangle\] state it in all cases of parameters (1)–(4). Second, in Section \[subsec:givenparameters\], given a set parameters $(\Sigma, \mu, R)$, we compute the asymptotics of the density for all angles $\alpha_0 \in ]0, \pi/2[$, see Theorems \[thmresultsnew1\]-\[thmresultsnew3\].
[**Remark.**]{} The constants mentioned in Theorem \[thmmain\] and all those in asymptotic expansions of Theorems \[thmresults1\]–\[thmresultsnew3\] are specified in terms of functions $\varphi_1$ and $\varphi_2$. In the present paper we leave unknown these functions in their initial domains of definition although we carry out explicitly their meromorphic continuation procedure and find all their poles. In [@franceschi_explicit_2017] the first author and K. Raschel make explicit these functions solving some boundary value problems. This determines the constants in asymptotic expansions in Theorems \[thmmain\], \[thmresults1\]–\[thmresultsnew3\] .
[**Future works.**]{} The case of parameters such that $\zeta\theta^{**}=\theta(\alpha)$ and $\eta \theta^{*} \not \in \{s'_0, \theta(\alpha)\{$ or the case such that $\eta \theta^*=\theta(\alpha)$ and $\zeta \theta^{**} \not\in \{s''_0, \theta(\alpha)\{$ are not treated in Theorem \[thmmain\]. Theorem \[thmresults4\] gives a partial result but not at all as satisfactory as in all other cases. In fact, in these cases the saddle-point $\theta(\alpha)$ coincides with the “main” pole of $\phi_1$ or $\phi_2$. The analysis is then reduced to a technical problem of computing the asymptotics of an integral whenever the saddle-point coincides with a pole of the integrand or approaches to it. We leave it for the future work.
In the cases $\alpha=0$ and $\alpha=\pi/2$, the tail asymptotics of the boundary measures $\nu_1$ and $\nu_2$ has been found in [@dai_stationary_2013] and the constants have been specified in [@franceschi_explicit_2017]. It would be also possible to find the asymptotics of $\pi(r \cos \alpha, r \sin \alpha)$ where $r \to \infty$ and $\alpha \to 0$ or $\alpha \to \pi/2$. This problem is reduced to obtaining the asymptotics of an integral when the saddle-point $\theta(0)$ or $\theta(\pi/2)$ coincides with a branch point of the integrand $\phi_1$ or $\phi_2$. It can be solved by the same methods as in [@kourkova_random_2011] for discrete random walks.
Riemann surface S {#riemannsurfaceS}
=================
Kernel gamma
------------
The kernel of the main functional equation $$\gamma(\theta_1,\theta_2)=
\frac{1}{2}(\sigma_{11} \theta_1^2 + \sigma_{22} \theta_2^2 +2\sigma_{12}\theta_1\theta_2)
+
\mu_1\theta_1+\mu_2\theta_2$$ can be written as $$\gamma(\theta_1, \theta_2)= \tilde{a}(\theta_2)\theta_1^2+\tilde{b}(\theta_2)\theta_1+\tilde{c}(\theta_2)
=a(\theta_1)\theta_2^2+b(\theta_1)\theta_2+c(\theta_1)$$ where
----------------------------------------------- -------------------------------------------------- -----------------------------------------------------------------------
$\tilde{a}(\theta_2)=\frac{1}{2}\sigma_{11}$, $\tilde{b}(\theta_2)=\sigma_{12}\theta_2+\mu_1$, $\tilde{c}(\theta_2)=\frac{1}{2}\sigma_{22}\theta_2^2+\mu_2\theta_2$,
$a(\theta_1)=\frac{1}{2}\sigma_{22}$, $b(\theta_1)=\sigma_{12}\theta_1+\mu_2$, $c(\theta_1)=\frac{1}{2}\sigma_{11}\theta_1^2+\mu_1\theta_1$.
----------------------------------------------- -------------------------------------------------- -----------------------------------------------------------------------
The equation $\gamma(\theta_1, \theta_2)\equiv 0$ defines a two-valued algebraic function $\Theta_1(\theta_2)$ such that $\gamma(\Theta_1(\theta_2),\theta_2)\equiv 0$ and a two-valued algebraic function $\Theta_2(\theta_1)$ such that $\gamma(\theta_1, \Theta_2(\theta_1)\equiv 0$. These functions have two branches:
------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------
$\Theta_1^+(\theta_2)=\frac{-\tilde{b}(\theta_2)+\sqrt{\tilde{d}(\theta_2)}}{2\tilde{a}(\theta_2)}$, $\Theta_1^-(\theta_2)=\frac{-\tilde{b}(\theta_2)-\sqrt{\tilde{d}(\theta_2)}}{2\tilde{a}(\theta_2)}$,
------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------
and
------------------------------------------------------------------------------ ------------------------------------------------------------------------------
$\Theta_2^+(\theta_1)=\frac{-b(\theta_1)+\sqrt{d(\theta_1)}}{2a(\theta_1)}$, $\Theta_2^-(\theta_1)=\frac{-b(\theta_1)-\sqrt{d(\theta_1)}}{2a(\theta_1)}$.
------------------------------------------------------------------------------ ------------------------------------------------------------------------------
where
------------------------------------------------------------------------------------------------------------------------------
$\tilde{d}(\theta_2)=\theta_2^2(\sigma_{12}^2-\sigma_{11}\sigma_{22})+2\theta_2(\mu_1\sigma_{12}-\mu_2\sigma_{11})+\mu_1^2$,
$d(\theta_1)=\theta_1^2(\sigma_{12}^2-\sigma_{11}\sigma_{22})+2\theta_1(\mu_2\sigma_{12}-\mu_1\sigma_{22})+\mu_2^2$.
------------------------------------------------------------------------------------------------------------------------------
The discriminant ${d}(\theta_1)$ (resp. $\tilde{d}(\theta_2)$) has two zeros $\theta_1^+$, $\theta_1^{-}$ (resp. $\theta_2^{+}$ and $\theta_2^{-}$) that are both real and of opposite signs:
---------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------
$\theta_1^-= \frac{(\mu_2\sigma_{12}-\mu_1\sigma_{22}) - \sqrt{D_1}}{\det{\Sigma}}<0$, $\theta_1^+= \frac{(\mu_2\sigma_{12}-\mu_1\sigma_{22}) + \sqrt{D_1}}{\det{\Sigma}}>0$,
---------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------
------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------
$\theta_2^- = \frac{(\mu_1\sigma_{12}-\mu_2\sigma_{11}) - \sqrt{D_2}}{\det{\Sigma}} <0$, $\theta_2^+ = \frac{(\mu_1\sigma_{12}-\mu_2\sigma_{11}) + \sqrt{D_2}}{\det{\Sigma}} >0$,
------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------
with notations $D_1 =(\mu_2\sigma_{12}-\mu_1\sigma_{22})^2 +\mu_2^2 \det{\Sigma}$ and $D_2 =(\mu_1\sigma_{12}-\mu_2\sigma_{11})^2 +\mu_1^2 \det{\Sigma}$. Then $\Theta_2(\theta_1)$ (resp. $\Theta_1(\theta_2)$) has two branch points: $\theta_1^-$ and $\theta_1^+$ (resp $\theta_2^{-}$ and $\theta_2^{+}$). We can compute: $$\Theta_2^\pm(\theta_1^-)=\frac{\mu_1\sigma_{12}-\mu_2\sigma_{11} + \frac{\sigma_{12}}{\sigma_{22}} \sqrt{D_1}}{\det \Sigma}, \ \
\Theta_2^\pm(\theta_1^+)=\frac{\mu_1\sigma_{12}-\mu_2\sigma_{11} - \frac{\sigma_{12}}{\sigma_{22}} \sqrt{D_1}}{\det \Sigma},$$ $$\Theta_1^\pm(\theta_2^-) =
\frac{\mu_2\sigma_{12}-\mu_1\sigma_{22} + \frac{\sigma_{12}}{\sigma_{11}} \sqrt{D_2}}{\det \Sigma},\ \
\Theta_1^\pm(\theta_2^+) =
\frac{\mu_2\sigma_{12}-\mu_1\sigma_{22} - \frac{\sigma_{12}}{\sigma_{11}} \sqrt{D_2}}{\det \Sigma}.$$ Furthermore, $d(\theta_1)$ (resp. $\tilde d(\theta_2)$) being positive on $]\theta_1^-, \theta_1^+[$ (resp. $]\theta_2^-, \theta_2^+[$) and negative on ${\bf R} \setminus [\theta_1^-, \theta_1^+]$ (resp. ${\bf R}\setminus [\theta_2^-, \theta_2^+]$) , both branches $\Theta_2^{\pm}(\theta_1)$ (resp. $\Theta_1^\pm (\theta_2)$) take real values on $[\theta_1^-, \theta_1^+]$ (resp. $[\theta_2^-, \theta_2^+]$) and complex values on ${\bf R} \setminus [\theta_1^-, \theta_1^+]$ (resp. ${\bf R}\setminus [\theta_2^-, \theta_2^+]$).
![Functions $\Theta_2^\pm (\theta_1)$ and $\Theta_1^\pm (\theta_2)$ on $[\theta_1^-,\theta_1^+]$ and $[\theta_2^-,\theta_2^+]$[]{data-label="fonctionellipse"}](ellipsefonctions2.pdf "fig:") ![Functions $\Theta_2^\pm (\theta_1)$ and $\Theta_1^\pm (\theta_2)$ on $[\theta_1^-,\theta_1^+]$ and $[\theta_2^-,\theta_2^+]$[]{data-label="fonctionellipse"}](ellipsefonctions1.pdf "fig:")
Construction of the Riemann surface S {#constructionS}
-------------------------------------
We now construct the Riemann surface $\mathbf{S}$ of the algebraic function $\Theta_2(\theta_1)$. For this purpose we take two Riemann spheres $\mathbb{C}_{\theta_1}^1\cup\{\infty\}$ and $\mathbb{C}_{\theta_1}^2\cup\{\infty^\prime\}$, say $\mathbf{S}_{\theta_1}^1$ and $\mathbf{S}_{\theta_1}^2$, cut along $([-\infty^{(\prime)},\theta_1^{-}] \cup [\theta_1^{+},\infty^{(\prime)}])$, and we glue them together along the borders of these cuts, joining the lower border of the cut on $\mathbf{S}_{\theta_1}^1$ to the upper border of the same cut on $\mathbf{S}_{\theta_1}^2$ and vice versa. This procedure can be viewed as gluing together two half-spheres, see Figure \[constrsurfa\]. The resulting surface $\mathbf{S}$ is homeomorphic to a sphere (i.e., a compact Riemann surface of genus 0) and is projected on the Riemann sphere $\mathbb{C}\cup\{\infty\}$ by a canonical covering map $h_{\theta_1}: \mathbf{S} \to \mathbb{C}\cup\{\infty\}$. In a standard way, we can lift the function $\Theta_2(\theta_1)$ to $\mathbf{S}$, by setting $\Theta_2(s)=\Theta_2^+(h_{\theta_1}(s))$ if $s\in\mathbf{S}_{\theta_1}^1 \subset \mathbf{S}$ and $\Theta_2(s)=\Theta_2^-(h_{\theta_1}(s))$ if $s\in\mathbf{S}_{\theta_1}^2 \subset \mathbf{S}$.
In a similar way one constructs the Riemann surface of the function $\Theta_1(\theta_2)$, by gluing together two copies $\mathbf{S}_{\theta_2}^1$ and $\mathbf{S}_{\theta_2}^2$ of the Riemann sphere $\mathbf{S}$ cut along $([-\infty^{(\prime)},\theta_2^{-}] \cup [\theta_2^{+},\infty^{(\prime)}])$. We obtain again a surface homeomorphic to a sphere where we lift function $\Theta_1(\theta_2)$.
![Construction of the Riemann surface ${\mathbf{S}}$[]{data-label="constrsurfa"}](constructionsurface2)
Since the Riemann surfaces of $\Theta_1(\theta_2)$ and $\Theta_2(\theta_1)$ are equivalent, we can and will work on a [*single*]{} Riemann surface $\mathbf{S}$, with two different covering maps $h_{\theta_1}, h_{\theta_2} : \mathbf{S} \to \mathbb{C}\cup\{\infty\}$. Then, for $s\in\mathbf{S}$, we set $\theta_1(s)=h_{\theta_1}(s)$ and $\theta_2(s)=h_{\theta_2}(s)$. We will often represent a point $s\in\mathbf{S}$ by the pair of its *coordinates* $(\theta_1(s),\theta_2(s))$. These coordinates are of course not independent, because the equation $\gamma(\theta_1(s),\theta_2(s))=0$ is valid for any $s\in\mathbf{S}$. One can see ${\bf S}$ with points $
s_1^\pm=(\theta_1^\pm,\frac{\mu_1\sigma_{12}-\mu_2\sigma_{11} \mp \frac{\sigma_{12}}{\sigma_{22}} \sqrt{D_1}}{\det \Sigma})$, $
s_2^\pm=(\frac{\mu_2\sigma_{12}-\mu_1\sigma_{22} \mp \frac{\sigma_{12}}{\sigma_{11}} \sqrt{D_2}}{\det \Sigma},\theta_2^\pm)$, $
s_\infty=(\infty,\infty)$, $
s_{\infty^\prime}=(\infty^\prime,\infty^\prime)$ on Figure \[realpointss\]. It is the union of ${\bf S}_{\theta_1}^1$ and ${\bf S}_{\theta_1}^2$ glued along the contour ${{\mathcal}R}_{\theta_1}=\{s : \theta_1(s) \in {\bf R} \setminus ]\theta_1^-, \theta_1^+[\}$ that goes from $s_\infty$ to $s_{\infty^\prime}$ via $s_1^-$ and back to $s_\infty$ via $s_1^+$. It is also the union of ${\bf S}_{\theta_2}^1$ and ${\bf S}_{\theta_2}^2$ glued along the contour ${{\mathcal}R}_{\theta_2}=\{s : \theta_2(s) \in {\bf R} \setminus ]\theta_2^-, \theta_2^+[\}$. This contour goes from $s_\infty$ to $s_{\infty^\prime}$ and back as well, but via $s_2^-$ and $s_2^+$. Let ${{\mathcal}E}$ be the set of points of ${\bf S}$ where both coordinates $\theta_1(s)$ and $\theta_2(s)$ are real. Then $$\mathcal{E}=\{ s\in\mathbf{S}:\theta_1(s)\in[\theta_1^-,\theta_1^{+}] \}
=
\{ s\in\mathbf{S}:\theta_2(s)\in[\theta_2^-,\theta_2^{+}] \}.$$ One can see $\mathcal{ E}$ on Figures \[fonctionellipse\] and \[realpointss\], it contains all branch points $s_1^{\pm}$ and $s_2^{\pm}$.
![Points of ${\mathbf{S}}$ with $\theta_1(s)$ or $\theta_2(s)$ real[]{data-label="realpointss"}](sphereens.png)
Galois automorphisms eta and zeta
---------------------------------
Now we need to introduce Galois automorphisms on ${\bf S}$. For any $s\in \mathbf{S} \setminus s_1^{\pm}$ there is a unique $s^\prime \ne s \in \mathbf{S} \setminus s_1^{\pm}$ such that $\theta_1(s)=\theta_1(s^\prime)$. Furthermore, if $s \in {\mathbf{S}}_{\theta_1}^1$ then $s^\prime \in {\mathbf{S}}_{\theta_1}^2$ and vice versa. On the other hand, whenever $s=s_1^-$ or $s=s_1^+$ (i.e. $\theta_1(s)=\theta_1^\pm$ is one of branch points of $\Theta_2(\theta_1)$) we have $s=s^\prime$. Also, since $\gamma(\theta_1(s),\theta_2(s))=0$, $\theta_2(s)$ and $\theta_2(s^\prime)$ represent both values of function $\Theta_2(\theta_1)$ at $\theta_1=\theta_1(s)=\theta_1(s^\prime)$. By Vieta’s theorem we obtain $\theta_2(s)\theta_2(s^\prime)=\frac{c(\theta_1(s))}{a(\theta_1(s))}$.
Similarly, for any $s\in\mathbf{S} \setminus s_2^{\pm}$, there exists a unique $s^{\prime\prime} \neq s \in\mathbf{S}
\setminus s_2^{\pm} $ such that $\theta_2(s)=\theta_2(s^{\prime\prime})$. If $s\in {\mathbf{S}}_{\theta_2}^1$ then $s^\prime \in {\mathbf{S}}_{\theta_2}^2$ and vice versa. On the other hand, if $s=s_2^-$ or $s=s_2^+$ (i.e. $\theta_2(s)=\theta_2^\pm$ is one of branch points of $\Theta_1(\theta_2)$) we have $s=s^{\prime\prime}$. Moreover, since $\gamma(\theta_1(s),\theta_2(s))=0$, $\theta_1(s)$ and $\theta_1(s^{\prime\prime})$ give both values of function $\Theta_1(\theta_2)$ at $\theta_2=\theta_2(s)=\theta_2(s^{\prime\prime})$. Again, by Vieta’s theorem $\theta_1(s)\theta_1(s^{\prime\prime})=\frac{\tilde{c}(\theta_2(s))}{\tilde{a}(\theta_2(s))}$.
With the previous notations we now define the mappings $\zeta:\mathbf{S}\to\mathbf{S}$ and $\eta:\mathbf{S}\to\mathbf{S}$ by
$\left\{
\begin{tabular}{cc}
$s = s\^$ & if $\_1(s)=\_1(s\^)$,
\\
$s = s\^$ & if $\_2(s)=\_2(s\^)$
\\
\end{tabular}
\right. $
Following [@malyshev_sluchainye_1970] we call them *Galois automorphisms* of $\mathbf{S}$. Then $\zeta^2=\eta^2=\text{Id}$, and
----------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------
$\theta_2(\zeta s)= \frac{c(\theta_1(s))}{a(\theta_1(s))} \frac{1}{\theta_2(s)}$, $\theta_1(\eta s)=\frac{\tilde{c}(\theta_2(s))}{\tilde{a}(\theta_2(s))} \frac{1}{\theta_1(s)}$.
----------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------
Points $s_1^-$ and $s_1^+$ (resp. $s_2^-$ and $s_2^+$) are fixed points for $\zeta$ (resp. $\eta$).
It is known that conformal automorphisms of a sphere (that can be identified to ${\bf C} \cup \infty$) are transformations of type $z\mapsto \frac{az+b}{cz+d}$ where $a,b,c,d$ are any complex numbers satisfying $ad-bc\neq 0$. The automorphisms $\zeta$ and $\eta$, which are conformal automorphisms of ${\mathbf{S}}$, have each two fixed points and are involutions (because $\zeta^2=\eta^2=\text{Id}$). We can deduce from it that $\zeta$ (resp. $\eta$) is a symmetry w.r.t. the axis $A_1$ (resp. $A_2$) that passes through fixed points $s_1^-$ and $s_1^+$ (resp. $s_2^-$ and $s_2^+$). In other words $\zeta$ (resp. $\eta$) is a rotation of angle $\pi$, around $D_1$ (resp. $A_2$), see Figure \[sphereaxes\]. Let us draw the axis $A$ orthogonal to the plane generated by the axes $A_1$ and $A_2$ and passing through the intersection point of $A_1$ and $A_2$. We denote by $\beta$ the angle between the axes $A_1$ and $A_2$. Automorphisms $\eta\zeta$ and $\zeta\eta$ are then [*rotations of angle $ 2\beta$ and $-2\beta$ around the axis $A$.*]{} This axis goes through points $s_\infty$ and $s_{\infty\prime}$ which are fixed points for $\eta\zeta$ and $\zeta\eta$, see Figure \[sphereaxes\].
In the particular case $\Sigma=\text{Id}$, we have $\eta\zeta=\zeta\eta$, the axes $A_1$ and $A_2$ are orthogonal. We deduce that $\beta=\frac{\pi}{2}$ and that $\eta\zeta$ and $\zeta\eta$ are symmetries w.r.t. the axis $A$.
![Axes $A_1$, $A_2$ and $A$ of Galois automorphisms $\zeta$, $\eta$ and $\zeta\eta$ respectively[]{data-label="sphereaxes"}](sphereaxe.png)
Domains of initial definition of phi1 and phi2 on S
---------------------------------------------------
We would like to lift functions $\phi_1(\theta_2)$ and $\phi_2(\theta_1)$ on ${\bf S}$ naturally as $\phi_1(s)=\phi_1(\theta_2(s))$ and $\phi_2(s)=\phi_2(\theta_1(s))$. But it can not be done for all $s \in {\bf S}$, $\phi_1(\theta_2)$ and $\phi_2(\theta_1)$ being not defined on the whole of ${\bf C}$. Nevertheless, we are able to do it for points $s$ where $\theta_2(s)$ or $\theta_1(s)$ respectively have non-positive real parts. Therefore, in this section we study the domains on ${\bf S}$ where it holds true.
For any $\theta_1 \in {\bf C}$ with ${{\mathcal{R}}}(\theta_1)=0$, $\Theta_2(\theta_1)$ takes two values $\Theta_2^{\pm}(\theta_1)$. Let us observe that under assumption that the second coordinate of the interior drift is negative, i.e $\mu_2<0$ we have ${{\mathcal{R}}} \Theta_2^-(\theta_1){\leqslant}0$ and ${{\mathcal{R}}} \Theta_2^+(\theta_1)>0$. Furthermore ${{\mathcal{R}}}\Theta_2^-(\theta_1)=0$ only at $\theta_1=0$, and then $\Theta_2^-(\theta_1)=0$. The domain $$\Delta_1=\{s \in {\mathbf{S}}: \ {{\mathcal{R}}}\theta_1(s)<0\}$$ is simply connected and bounded by the contour $\mathcal{I}_{\theta_1}=\{s : {{\mathcal{R}}}\theta_1(s)=0\}$.
The contour $\mathcal{I}_{\theta_1}$ can be represented as the union of $\mathcal{I}_{\theta_1}^-\cup \mathcal{I}_{\theta_1}^+$ , where $\mathcal{I}_{\theta_1}^-=\{s : {{\mathcal{R}}}\theta_1(s)=0, {{\mathcal{R}}} \theta_2(s) {\leqslant}0\}$, $\mathcal{I}_{\theta_1}^+=\{s : {{\mathcal{R}}}\theta_1(s)=0, {{\mathcal{R}}} \theta_2(s) > 0\}$, see Figure \[contIth1\].
The contour $\mathcal{I}_{\theta_1}^-$ goes from $s_\infty$ to $s_{\infty\prime}$ crossing the set of real points $\mathcal{ E}$ at $s_0=(0,0)$, while $\mathcal{I}_{\theta_1}^{+}$ goes from $s_\infty$ to $s_{\infty\prime}$ crossing $\mathcal{ E}$ at $s'_0=(0, -2 \frac{\mu_2}{\sigma_{22}})$ where the second coordinate is positive.
In the same way, under assumption that the first coordinate of the interior drift is negative, i.e. $\mu_1<0$, for any $\theta_2 \in {\bf C}$ with ${{\mathcal{R}}}(\theta_2)=0$, $\Theta_1(\theta_2)$ takes two values $\Theta_1^{\pm}(\theta_2)$, where ${{\mathcal{R}}} \Theta_1^-(\theta_2){\leqslant}0$ and ${{\mathcal{R}}} \Theta_1^+(\theta_2)>0$, moreover ${{\mathcal{R}}}\Theta_1^-(\theta_2)=0$ only at $\theta_2=0$, and then $\Theta_1^-(\theta_2)=0$. The domain $$\Delta_2=\{s \in {\mathbf{S}}: {{\mathcal{R}}}\theta_2(s)<0\}$$ is simply connected and bounded by the contour $\mathcal{I}_{\theta_2}=\{s : {{\mathcal{R}}}\theta_2(s)=0\}$. The contour $\mathcal{I}_{\theta_2}$ can be represented as the union of $\mathcal{I}_{\theta_2}^-\cup \mathcal{I}_{\theta_2}^+$ , where $\mathcal{I}_{\theta_2}^-=\{s : {{\mathcal{R}}}\theta_2(s)=0, {{\mathcal{R}}} \theta_1(s) {\leqslant}0\}$, $\mathcal{I}_{\theta_2}^+=\{s : {{\mathcal{R}}}\theta_2(s)=0, {{\mathcal{R}}} \theta_1(s) > 0\}$. The contour $\mathcal{I}_{\theta_2}^-$ goes from $s_\infty$ to $s_{\infty\prime}$ crossing the set of real points $\mathcal{ E}$ at $s_0=(0,0)$, while $\mathcal{I}_{\theta_2}^+$ goes from $s_\infty$ to $s_{\infty\prime}$ crossing $\mathcal{ E}$ at $s''_
0=(-2 \frac{\mu_1}{\sigma_{11}}, 0)$, see Figure \[contIth1\].
Assume now that the interior drift has both coordinates negative, i.e. (\[mumu\]). From what said above, $\mathcal{I}_{\theta_1}^-\setminus s_0 \subset \Delta_2$ and $\mathcal{I}_{\theta_2}^-\setminus s_0 \subset \Delta_1$. The intersection $\Delta_1 \cap \Delta_2$ consists of two connected components, both bounded by $\mathcal{I}_{\theta_1}^-
$ and $\mathcal{I}_{\theta_2}^-$. The union $\Delta_1 \cup \Delta_2$ is a connected domain, but not simply connected because of the point $s_0$. The domain $\Delta_1 \cup \Delta_2 \cup s_0$ is open, simply connected and bounded by $\mathcal{I}_{\theta_1}^+
$ and $\mathcal{I}_{\theta_2}^+$, see Figure \[contIth1\]. We set $\Delta=\Delta_1\cup\Delta_2$.
Note that in the cases of stationary SRBM with drift $\mu$ having one of coordinates non-negative, the location of contours $\mathcal{I}_{\theta_1}^+$, $\mathcal{I}_{\theta_1}^-$, $\mathcal{I}_{\theta_2}^+$, $\mathcal{I}_{\theta_2}^-$ on ${\bf S}$ is different. For example, assume that $\mu_2>0$. Then ${{\mathcal{R}}} \Theta_2^-(\theta_1)<0$ and ${{\mathcal{R}}} \Theta_2^+(\theta_1){\geqslant}0$, the contour $\mathcal{I}_{\theta_1}^-$ goes from $s_\infty$ to $s_{\infty\prime}$ crossing the set of real points $\mathcal{ E}$ at $s'_0=(0, -2 \frac{\mu_2}{\sigma_{22}})$ where the second coordinate is negative, while $\mathcal{I}_{\theta_1}^{+}$ goes from $s_\infty$ to $s_{\infty\prime}$ crossing $\mathcal{ E}$ at $s_0=(0, 0)$. In order to shorten the number of cases and pictures, we restrict ourselves in this paper to the case (\[mumu\]) of both coordinates of $\mu$ negative, although all our methods work in these other cases as well.
![Pure imaginary points of ${\mathbf{S}}$[]{data-label="contIth1"}](sphereima.png)
Parametrization of S {#param}
--------------------
It is difficult to visualize on three-dimensional sphere different points, contours, automorphisms and domains introduced above that will be used in future steps. For this reason we propose here an explicit and practical parametrisation of ${\mathbf{S}}$. Namely we identify ${\mathbf{S}}$ to $\mathbb{C}\cup\{\infty\}$ and in the next proposition we explicitly define $h_{\theta_1}$ and $h_{\theta_2}$ two recoveries introduced in Section \[constructionS\]. Such a parametrisation allows to visualize better in two dimensions the sphere ${\mathbf{S}}\equiv \mathbb{C}\cup\{\infty\}$ and all sets we are interested in, as we can see in Figure \[recouvrS\].
We set the following covering maps $$\begin{array}{lrcl}
h_{\theta_1}: & \mathbb{C}\cup\{\infty\}\equiv{\mathbf{S}}& \longrightarrow & \mathbb{C}\cup\{\infty\}
\\
& s & \longmapsto & h_{\theta_1}(s)=\theta_1(s):=\frac{\theta_1^{+}+\theta_1^{-}}{2}+\frac{\theta_1^{+}-\theta_1^{-}}{4} (s+\frac{1}{s})
\end{array}$$ and $$\begin{array}{lrcl}
h_{\theta_2}: & \mathbb{C}\cup\{\infty\}\equiv{\mathbf{S}}& \longrightarrow & \mathbb{C}\cup\{\infty\}
\\
& s & \longmapsto & h_{\theta_2}(s)=\theta_2(s):=\frac{\theta_2^{+}+\theta_2^{-}}{2}+\frac{\theta_2^{+}-\theta_2^{-}}{4} (\frac{s}{e^{i\beta}}+\frac{e^{i\beta}}{s}),
\end{array}$$ where $$\beta= \arccos {-\frac{\sigma_{12}}{\sqrt{\sigma_{11}\sigma_{22}}}}.$$ The equation $\gamma(\theta_1(s),\theta_2(s))=0$ is valid for any $s\in\mathbf{S}$. Galois automorphisms can be written
------------------------- ----------------------------------
$\zeta(s)=\frac{1}{s}$, $\eta(s)=\frac{e^{2i\beta}}{s}$,
------------------------- ----------------------------------
and $\eta\zeta$ (resp. $\zeta\eta$) is a rotation around $s_\infty\equiv0$ of angle $2\beta$ (resp. $-2\beta$) according to counterclockwise direction.
We set $ h_{\theta_1}(s)=\theta_1(s):=\frac{\theta_1^{+}+\theta_1^{-}}{2}+\frac{\theta_1^{+}-\theta_1^{-}}{4} (s+\frac{1}{s})$. One can notice that $h_{\theta_1}(1)=\theta_1^+$, $h_{\theta_1}(-1)=\theta_1^-$, $h_{\theta_1}'(1)=0$, $h_{\theta_1}'(-1)=0$. This parametrization is practical because it leads to a similar rational recovery $h_{\theta_2}$. In order to make the equation $\gamma(\theta_1(s),\theta_2(s))=0$ valid for any $s\in\mathbf{S}$ we naturally set $$\theta_2(s)= \Theta_2^+ (\theta_1(s)):=\frac{-b(\theta_1(s))+\sqrt{d(\theta_1(s))}}{2a(\theta_1(s))}$$ and we are going to show that $\theta_2(s)=
\frac{\theta_2^{+}+\theta_2^{-}}{2}+\frac{\theta_2^{+}-\theta_2^{-}}{4} (\frac{s}{e^{i\beta}}+\frac{e^{i\beta}}{s})$ where $\beta= \arccos {-\frac{\sigma_{12}}{\sqrt{\sigma_{11}\sigma_{22}}}}$. We note that $d(\theta_1(s))$ is the opposite of the square of a rational fraction $$\begin{aligned}
d(\theta_1(s))&=-\det \Sigma (\theta_1(s)-\theta_1^+)(\theta_1(s)-\theta_1^-)
=-\det \Sigma (\frac{\theta_1^{+}-\theta_1^{-}}{4})^2(-2+(s+\frac{1}{s}))(2+(s+\frac{1}{s}))
\\ &= -\det \Sigma
(\frac{\theta_1^+-\theta_1^-}{4})^2 (s-\frac{1}{s})^2 \leqslant 0.\end{aligned}$$ Then we have $$\label{232}
\theta_2(s)= \Theta_2^+ (\theta_1(s))
:=
\frac{-\sigma_{12}\frac{\theta_1^{+}+\theta_1^{-}}{2}+\frac{\theta_1^{+}-\theta_1^{-}}{4} (s+\frac{1}{s})-\mu_2+i \sqrt{\det \Sigma}(\frac{\theta_1^+-\theta_1^-}{4})(s-\frac{1}{s})}{\sigma_{22}}.$$ Furthermore this parametrization leads to simple expressions for Galois automorphisms $\eta$ and $\zeta$. We derive immediately that $\theta_1(s)=\theta_1(\frac{1}{s})$ and $\theta_2 (\frac{1}{s})=\Theta_2^-(\theta_1(s))$. Then we have $$\zeta (s)=\frac{1}{s}.$$ Next we search $\eta$ as an automorphism of the form $\eta s =\frac{K}{s}$. Since $\theta_2(s)$ is of the form $\theta_2(s)=us+\frac{v}{s}+w$ with constants $u,v,w$ defined by (\[232\]), then $\theta_2(s)=\theta_2(\frac{K}{s})$ with $K=\frac{u}{v}$. This leads to $$\eta (s) =\frac{K}{s} \text{ with }
K=\frac{-\sigma_{12}-i\sqrt{\det \Sigma}}{-\sigma_{12}+i\sqrt{\det \Sigma}}.$$ After setting $$K=e^{2i\beta} \text{ with }
\beta= \arccos {-\frac{\sigma_{12}}{\sqrt{\sigma_{11}\sigma_{22}}}}$$ we have
--------------------------- -----------------------------------
${\zeta}(s)=\frac{1}{s}$, ${\eta}(s)=\frac{e^{2i\beta}}{s}$
--------------------------- -----------------------------------
and then
--------------------------------- ----------------------------------
$\eta\zeta (s) = e^{2i\beta}s$, $\zeta\eta (s) = e^{-2i\beta}s$.
--------------------------------- ----------------------------------
It follows that ${\eta}{\zeta}$ and ${\zeta}{\eta}$ are just rotations for angles $2\beta$ et $-2\beta$ respectively. By symmetry considerations we can now rewrite $$\theta_2 (s)=\sqrt{uv}(\frac{s}{\sqrt{K}}+\frac{\sqrt{K}}{s})+w=\frac{\theta_1^+-\theta_1^-}{4}\sqrt{\frac{\sigma_{11}}{\sigma_{22}}}(\frac{s}{\sqrt{K}}+\frac{\sqrt{K}}{s})+\frac{-\sigma_{12}(\frac{\theta_1^++\theta_1^-}{2})-\mu_2}{\sigma_{22}}.$$ For $i=1,2$ we have $\theta_i^+-\theta_i^- = 2 \frac{\sqrt{D_i}}{\det \Sigma} $ and $\sigma_{11}D_1=\sigma_{22}D_2$. Then we obtain $\frac{\theta_1^+-\theta_1^-}{4}\sqrt{\frac{\sigma_{11}}{\sigma_{22}}}=\frac{\theta_2^+-\theta_2^-}{4}$. Moreover $\frac{-\sigma_{12}(\frac{\theta_1^++\theta_1^-}{2})-\mu_2}{\sigma_{22}}=\frac{\Theta_2^\pm(\theta_1^+)+\Theta_2^\pm(\theta_1^-)}{2}=\frac{\theta_2^++\theta_2^-}{2}$ (the last equality follows from elementary geometric properties of an ellipse). It implies $$h_{\theta_2}(s)=\theta_2(s)=
\frac{\theta_2^++\theta_2^-}{2}+
\frac{\theta_2^+-\theta_2^-}{4} (\frac{s}{\sqrt{K}}+\frac{\sqrt{K}}{s})$$ concluding the proof.
![Parametrization of ${\mathbf{S}}$[]{data-label="recouvrS"}](recouvr1.pdf)
Figure \[recouvrS\] shows different sets we are interested in according to the parametrization we have just introduced. We have $\theta_1(1)=\theta_1^+$, $\theta_1(-1)=\theta_1^-$, $\theta_2 (e^{i\beta})=\theta_2^+$ et $\theta_2 (e^{i(\pi+\beta)})=\theta_2^-$, $\theta_1(0)=\theta_2(0)=\infty$, $\theta_1(\infty)=\theta_2(\infty)=\infty$. Then we write $s_1^+=1$, $s_1^-=-1$, $s_2^+=e^{i\beta}$, $s_2^-=e^{i(\pi+\beta)}$, $s_\infty=0$, $s_{\infty'}=\infty$. It is easy to see that $$\mathcal{E}=\{s\in \mathbb{C}| \ |s|=1\},$$ and $$\mathcal{R}_{\theta_1}={\bf R}, \ \mathcal{R}_{\theta_2}=e^{i\beta}{\bf R}.$$ We can determine the equation of the analytic curves of pure imaginary points of $\theta_i$. We have ${\mathcal{I}}_{\theta_1}=\{ s \in {\mathbf{S}}| \theta_1(s) \in i{\bf R} \}$. If we write $s=e^{i\omega}$ with $\omega=a+ib\in\mathbb{C}$ we find that $\Re (\theta_1(s))=\frac{\theta_1^++\theta_1^-}{2}+\frac{\theta_1^+-\theta_1^-}{2} \cos (a) \cosh (b)$. It follows that $${\mathcal{I}}_{\theta_1}=\{s=e^{i\omega}\in{\mathbf{S}}|\omega=a+ib,
a\in {\bf R}, b \in{\bf R}, \cos (a) \cosh (b)=\frac{\theta_1^++\theta_1^-}{\theta_1^--\theta_1^+}\}.$$ Similarly we have $${\mathcal{I}}_{\theta_2}=\{s=e^{i\omega}\in{\mathbf{S}}|\omega=a+ib,
a\in {\bf R}, b \in{\bf R}, \cos (a) \cosh (b)=\frac{\theta_2^++\theta_2^-}{\theta_2^--\theta_2^+}\}.$$ We can easily notice that $${\zeta} {\mathcal{I}}_{\theta_1}^-= {\mathcal{I}}_{\theta_1}^+, \
{\zeta} {\mathcal{I}}_{\theta_1}^+= {\mathcal{I}}_{\theta_1}^-
\text{ and }
{\eta} {\mathcal{I}}_{\theta_2}^-= {\mathcal{I}}_{\theta_2}^+, \
{\eta} {\mathcal{I}}_{\theta_2}^+= {\mathcal{I}}_{\theta_2}^-.$$
Meromorphic continuation of phi1 and phi2 on S
==============================================
Lifting of phi1 and phi2 on S and their meromomorphic continuation {#meromcontinuation}
------------------------------------------------------------------
#### Lifting of $\varphi_1$ and $\varphi_2$ on ${\mathbf{S}}$.
Since the function $\theta_1 \to \varphi_2(\theta_1)$ is holomorphic on the set $\{\theta_1\in \mathbb{C} : \Re \theta_1< 0\}$ and continuous up to its boundary, we can lift it to $\bar \Delta_1 = \{s\in{\mathbf{S}}:\Re \theta_1(s) {\leqslant}0 \}$ as $$\varphi_2(s)=\varphi_2(\theta_1(s)), \ \forall s \in \bar \Delta_1.$$ In the same way we can lift $\varphi_1$ to $\bar\Delta_2$ as $$\varphi_1(s)=\varphi_1(\theta_2(s)), \ \forall s \in \bar\Delta_2.$$ Moreover, by definition of Galois automorphisms, functions $\phi_1$ and $\phi_2$ are invariant w.r.t. $\eta$ and $\zeta$ respectively: $$\label{dhf}
\varphi_2(\zeta s)=\varphi_2(\theta_1(\zeta s))=\varphi_2(\theta_1(s))=\varphi_2(s), \ \ \forall s\in\bar \Delta_1,$$
$$
\varphi_1(\eta s)=\varphi_1(\theta_2(\eta s))=\varphi_1(\theta_2(s))=\varphi_1(s), \ \ \forall s\in \bar \Delta_2.$$ Functions $\gamma_1$ and $\gamma_2$ can be lifted naturally on the whole of ${\bf S}$ as $$\gamma_1(s)=\gamma_1(\theta_1(s), \theta_2(s)), \ \ \ \ \gamma_2(s)=\gamma_2(\theta_1(s), \theta_2(s)) \ \ \ \forall s\in {\bf S}.$$ Since $\gamma(\theta_1(s), \theta_2(s))=0$, then the right-hand side in the main functional equation (\[maineq\]) equals zero for any $\theta=(\theta_1(s), \theta_2(s))$ such that $s \in \bar\Delta_1 \cap \bar\Delta_2$. Thus we have $$\label{eqoo}
\gamma_1(s)\varphi(s)+\gamma_2(s)\varphi_2(s)=0, \ \ \ \forall s\in \bar \Delta_1 \cap \bar \Delta_2.$$
#### **Continuation of $\varphi_1$ and $\varphi_2$ on $\Delta$.**
\[lempro\] Functions $\varphi_1$ and $\varphi_2$ (defined on $\bar{\Delta}_2$ and $\bar{\Delta}_1$ respectively) can be meromorphically continued on $\Delta\cup\{s_0\}$ by setting $$\_1(s)=
------------------------------------------------- --------------------
$-\frac{\gamma_2(s)}{\gamma_1(s)} \varphi_2(s)$ if $s\in\Delta_1,$
------------------------------------------------- --------------------
$$and$$\_2(s)=
------------------------------------------------- --------------------
$-\frac{\gamma_1(s)}{\gamma_2(s)} \varphi_1(s)$ if $s\in\Delta_2.$
------------------------------------------------- --------------------
$$ Furthermore, $$\label{dfg}
\gamma_1(s)\varphi_1(s)+\gamma_2(s)\varphi_2(s)=0
\ \ \forall s\in\Delta\cup\{s_0\},$$ $$\label{invf}
\varphi_1(s)=\varphi(\eta s), \ \ \ \ \varphi_2(s)=\varphi(\zeta s) \ \ \forall s\in\Delta\cup\{s_0\}.$$
The open set $\Delta_1 \cap \Delta_2$ is non-empty and bounded by the curve $\mathcal{I}_{\theta_1}^- \cup \mathcal{I}_{\theta_2}^-$. Functional equation (\[eqoo\]) is valid for $s\in \Delta_1 \cap \Delta_2$. It allows us to continue functions $\varphi_1$ and $\varphi_2$ as meromorphic on $\Delta$ as stated in this lemma. The functional equation (\[eqoo\]) is then valid on the whole of $\Delta$, as well as the invariance formulas (\[invf\]).
The function $\phi_1(s)$ is defined in a neighborhood $\mathcal{O}(s_0)$ of $s_0$ as $\phi_1(\theta_2(s))$ for any $s \in \Delta_2 \cap \mathcal O(s_0)$ and $ -\frac{\gamma_2(s)}{\gamma_1(s)} \varphi_2(\theta_1(s))$ for any $s \in \Delta_1 \cap \mathcal O(s_0)$. Furthermore, $$\lim_{s \to s_0, s\in \Delta_2} \phi_1(s)= \mathbb{E}_{\pi}(\int_0^1 \mathrm{d}L_t^1)$$ by definition of the function $\phi_1.$ It is easy to see that function $\frac{\gamma_2(s)}{\gamma_1(s)}$ has a removable singularity at $s_0$ and to compute $\lim_{s\to s_0} \frac{\gamma_2(s)}{\gamma_1(s)}=\frac{r_{12}\mu_2-r_{22}\mu_1}{r_{11}\mu_2-r_{21}\mu_1}$. Hence $$\lim_{s \to s_0, s\in \Delta_1} \phi_1(s)= \lim_{s \to s_0, s\in \Delta_1}
-\frac{\gamma_2(s)}{\gamma_1(s)} \phi_2(\theta_1(s))=
\frac{r_{22}\mu_1 -r_{12}\mu_2 }{r_{11}\mu_2-r_{21}\mu_1 } \mathbb{E}_{\pi}(\int_0^1 \mathrm{d}L_t^2).$$ For any $s \in \Delta_1 \cap \Delta_2 \cap \mathcal{O}(s_0)$, by (\[eqoo\]) $\phi_1(s)=-\frac{\gamma_2(s)}{\gamma_1(s)}\phi_2(s)$, from where $\lim_{s \to s_0, s\in \Delta_2} \phi_1(s)= \lim_{s \to s_0, s\in \Delta_1} \phi_1(s)$. Hence, function $\phi_1(s)$ has a removable singularity at $s_0$, and so is $\phi_2(s)$ by the same arguments.
Functions $\phi_1$ and $\phi_2$ can be then of course continued to $\bar \Delta$. Moreover we have the following lemma.
\[lemrev\] The domains $ \bar \Delta \cup \eta \zeta \bar \Delta$ and $\bar \Delta \cup \zeta \eta\bar \Delta$ are simply connected.
Since $\eta \zeta$ and $\zeta \eta$ are just rotations for a certain angle $2\beta$ or $-2\beta$, it suffices to check that $\eta \zeta \mathcal{I}_{\theta_1}^+ \subset \bar \Delta$ and that $\zeta \eta \mathcal{I}_{\theta_2}^+ \in \bar \Delta$. In fact, $ \zeta \mathcal{I}_{\theta_1}^+ = \mathcal I_{\theta_1}^- \subset
\bar \Delta_2.$ Since $\eta \bar \Delta_2=\bar \Delta_2$, it follows that $ \eta \mathcal I_{\theta_1}^- \subset \bar \Delta_2 \subset \bar \Delta$. By the same arguments $\zeta \eta \mathcal{I}_{\theta_2}^+\in \bar \Delta$. One can refer to Figure \[etazetadelta\].
![$\Delta$ and $\eta\zeta\Delta$[]{data-label="etazetadelta"}](recouvr2.pdf)
Now we would like to continue function $\phi_1$ (resp. $\phi_2$) on $\eta \zeta \bar \Delta$ (resp. $\zeta \eta \bar \Delta$) as $\phi_1(s)=G(s) \phi_1(\zeta \eta s)$ for all $s\in \eta \zeta \bar \Delta$, where $G(s)$ is a known function and $\phi_1 (\zeta \eta s)$ is well defined since $\zeta \eta s \in \bar \Delta$. We could then continue this procedure for $(\eta \zeta)^2 \bar \Delta$, $(\eta \zeta)^3 \bar \Delta$, (resp. $(\zeta \eta)^2 \bar \Delta$, $(\zeta \eta)^3 \bar \Delta$) etc and hence to define $\phi_1$ (resp. $\phi_2$) on the whole of ${\bf S}$. Unfortunately, the domain $\bar \Delta$ is closed, from where it will be difficult to establish that the function is meromorphic. From the other hand, neither $ \Delta \cup \eta \zeta \Delta$ nor $ (\Delta \cup s_0) \cup \eta \zeta
(\Delta \cup s_0)$ are simply connected, there is a “gap” at $s''_0$. See figure \[etazetadelta\]. To avoid this technical complication, we will first continue $\phi_1$ and $\phi_2$ on a slightly bigger open domain $\Delta^\epsilon$ defined as follows. Let $$\label{de1}
\Delta_1^\epsilon=\{s : {{\mathcal{R}}}\theta_1(s)<\epsilon\}, \ \ \ \Delta_2^\epsilon=\{s : {{\mathcal{R}}}\theta_2(s)<\epsilon\}$$ and $$\label{de}
\Delta^\epsilon=\Delta_1^\epsilon\cup \Delta_2^\epsilon$$
Let us fix any $\epsilon>0$ small enough. For any $\theta_1 \in {\bf C}$ with ${{\mathcal{R}}} \theta_1 =\epsilon$, the function $\Theta_2(\theta_1)$ takes two values $\Theta_2^{\pm} (\theta_1)$ where ${{\mathcal{R}}} (\Theta_2^{-}(\theta_1)) <0$ and ${{\mathcal{R}}}(\Theta_2(\theta_1)) >0$. The domain $\Delta_1^\epsilon$ is bounded by the contour $\mathcal{I}_{\theta_1}^{\epsilon}= \mathcal{I}_{\theta_1}^{\epsilon, -}\cup \mathcal{I}_{\theta_1}^{\epsilon,+}$ where $\mathcal{I}_{\theta_1}^{\epsilon, -}$, $\mathcal{I}_{\theta_1}^{\epsilon, +}$ both go from $s_\infty$ to $s_{\infty\prime}$, $\mathcal{I}_{\theta_1}^{\epsilon, -} \subset \Delta_2$ and $\mathcal{I}_{\theta_1}^{\epsilon,+ }\cap \Delta =\emptyset$. See Figure \[deltaepsilon\]. The same is true about the contour $\mathcal{I}_{\theta_2}^{\epsilon}= \mathcal{I}_{\theta_2}^{\epsilon, -}\cup \mathcal{I}_{\theta_2}^{\epsilon,+}$ limiting $\Delta_2^\epsilon$, namely $\mathcal{I}_{\theta_2}^{\epsilon, -} \subset \Delta_1$ and $\mathcal{I}_{\theta_2}^{\epsilon,+ }\cap \Delta =\emptyset$.
![$\mathcal{I}_{\theta_1}^{\epsilon, -}$ and $\mathcal{I}_{\theta_1}^{\epsilon, +}$[]{data-label="deltaepsilon"}](recouvr3.pdf)
\[lemproepsilon\] Functions $\phi_1(s)$ and $\phi_2(s)$ can be continued as meromorphic functions on $\Delta^\epsilon$. Moreover equation (\[dfg\]) and the invariance formulas (\[invf\]) remain valid.
For any $s\in \Delta_1^\epsilon\setminus \Delta$, we have $\zeta s \in \Delta_2 \subset \Delta$, except for $s =s'_0$, for which $\zeta s'_0=s_0$. Anyway, function $\phi_2(s)$ can be continued as meromorphic function on $\Delta_1^\epsilon/\Delta$ as : $$\phi_2(s)=\phi_2(\zeta s), \ \ \ \forall s\in \Delta_1^\epsilon \setminus \Delta.$$ Then $\phi_1(s)$ can be continued on the same domain by (\[dfg\]): $$\phi_1(s)=-\frac{\gamma_2(s)}{\gamma_1(s)} \varphi_2(s) \ \ \ \forall s \in \Delta_1^\epsilon\setminus\Delta.$$ Similarly, the formulas $$\phi_1(s)=\phi_1(\eta s), \ \ \phi_2(s)=-\frac{\gamma_1(s)}{\gamma_2(s)} \varphi_1(s) \ \ \ \forall s \in \Delta_2^\epsilon\setminus\Delta$$ determine the meromorphic continuation of $\phi_1(s)$ and $\phi_2(s)$ on $\Delta_2^\epsilon\setminus\Delta$.
\[lemconteps\] The domains $\Delta^\epsilon \cap \eta\zeta \Delta^\epsilon$ and $\Delta^\epsilon \cap \zeta\eta \Delta^\epsilon$ are open simply connected domains. Function $\phi_1(s)$ can be continued as meromorphic on $ \Delta^\epsilon \cup \eta\zeta \Delta^\epsilon$ by the formula : $$\varphi_1(s)=
\frac{\gamma_1( \zeta \eta s)\gamma_2( \eta s)}{\gamma_2( \zeta \eta s)\gamma_1( \eta s)} \varphi_1( \zeta \eta s), \ \ \forall s \in \eta\zeta \Delta^\epsilon\ \hbox{ continuation by rotation of } 2\beta. \label{pro+trans}$$ Function $\phi_2(s)$ can be continued as meromorphic on $ \Delta^\epsilon \cup \zeta\eta \Delta^\epsilon$ by the formula : $$\varphi_2(s)=
\frac{\gamma_2( \eta \zeta s)\gamma_1( \zeta s)}{\gamma_1( \eta \zeta s)\gamma_2( \zeta s)} \varphi_2( \eta \zeta s), \ \ \forall s\in
\zeta\eta \Delta^\epsilon \hbox{ continuation by rotation of} -2\beta. \label{pro-trans1}$$
We have shown in the proof of Lemma \[lemrev\] that $\eta\zeta \mathcal I_{\theta_1}^{+} \subset \bar \Delta \subset \Delta^\epsilon$, and that $\zeta\eta \mathcal I_{\theta_2}^{+} \subset \bar \Delta \subset \Delta^\epsilon$. Since $\zeta \eta$ and $\eta \zeta$ are just rotations, this implies that $\Delta^\epsilon \cap \eta\zeta \Delta^\epsilon$ and $\Delta^\epsilon \cap \zeta\eta \Delta^\epsilon$ are non-empty open simply connected domains, and that $\Delta^\epsilon \cup \eta\zeta \Delta^\epsilon$ and $\Delta^\epsilon \cup \zeta\eta \Delta^\epsilon$ are simply connected.
Let us take $s \in \Delta^\epsilon \cap \eta\zeta \Delta^\epsilon$. Then $\zeta \eta s \in \Delta^\epsilon \cap \zeta\eta \Delta^\epsilon $ and we can write by (\[dfg\]) $$\label{efty}
\gamma_1(\zeta \eta s ) \phi_1(\zeta \eta s )+ \gamma_2( \zeta \eta s ) \phi_2( \zeta \eta s )=0.$$ Furthermore, we have shown in the proof of Lemma \[lemrev\] that $\zeta \eta \mathcal{I}_{\theta_2}^+ \in \bar \Delta_1$. It follows that for all $\epsilon$ small enough $\zeta \eta \mathcal{I}_{\theta_2}^{ \epsilon, +} \in \Delta_1$, and hence $ \Delta^\epsilon \cap \zeta\eta \Delta^\epsilon \subset \Delta_1^\epsilon $. Since $\zeta \Delta_1^\epsilon=\Delta_1^\epsilon$, then $ \zeta ( \Delta^\epsilon \cap \zeta\eta \Delta^\epsilon) \subset \Delta_1^\epsilon \subset \Delta_\epsilon$. Then $\zeta(\zeta \eta s)=\eta s \in \Delta^\epsilon$ and we can write (\[dfg\]) and (\[invf\]) at this point as well: $$\label{smh}
\gamma_1( \eta s ) \phi_1 (\eta s )+ \gamma_2( \eta s ) \phi_2( \eta s )=0.$$ $$\label{smhh}
\phi_1 (\eta s )=\phi_1(s)$$ $$\label{dooo}
\phi_2 (\eta s)=\phi_2(\zeta \eta s).$$ Combining (\[smhh\]) and (\[smh\]) we get $\phi_1(s)=-\gamma_2(\eta s) \phi_2(\eta s)/\gamma_1(\eta s)$ from where by (\[dooo\]) $$\label{sdnew}
\phi_1(s)= -\frac{\gamma_2(\eta s) }{\gamma_1(\eta s)} \phi_2(\zeta \eta s).$$ Due to (\[efty\]) $$\label{sdnew1}
\phi_2(\zeta \eta s)=-\frac{\gamma_1(\zeta\eta s)}{\gamma_2(\zeta\eta s)} \phi_1(\zeta \eta s).$$ Substituting (\[sdnew1\]) into (\[sdnew\]), we obtain the formula (\[pro+trans\]) valid for any $s \in \Delta^\epsilon \cap \eta\zeta \Delta^\epsilon$. By principle of analytic continuation this allows to continue $\phi_1$ on $\eta\zeta \Delta^\epsilon$ as meromorphic function. The proof is completely analogous for $\varphi_2$.
We may now in the same way, using formulas (\[pro+trans\]) and (\[pro-trans1\]), continue function $\phi_1(s)$ (resp. $\phi_2(s)$) as meromorphic on $(\eta\zeta)^2 \Delta^\epsilon$, $(\eta\zeta)^3 \Delta^\epsilon$ (resp. $(\zeta\eta)^2 \Delta^\epsilon$, $(\zeta\eta)^3 \Delta^\epsilon$) etc proceeding each time by rotation for the angle $2\beta$ \[resp. $-2\beta$\]. Namely we have the following lemma.
\[lemcontepse\] For any $n {\geqslant}1$ the domains $\Delta^\epsilon \cup \eta\zeta \Delta^\epsilon \cup \cdots \cup (\eta \zeta)^{n}\Delta^\epsilon$ and $\Delta^\epsilon \cup \zeta\eta \Delta^\epsilon \cup \cdots (\zeta \eta)^n \Delta^\epsilon $ are open simply connected domains. Function $\phi_1(s)$ can be continued as meromorphic subsequently on $\eta \zeta \Delta^\epsilon, (\eta \zeta)^2 \Delta^\epsilon, \cdots
(\eta \zeta) ^n \Delta^\epsilon $ by the formulas : $$\varphi_1(s)=
\frac{\gamma_1( \zeta \eta s)\gamma_2( \eta s)}{\gamma_2( \zeta \eta s)\gamma_1( \eta s)} \varphi_1( \zeta \eta s), \ \ \forall s \in (\eta\zeta)^k \Delta^\epsilon, k =1,2\ldots, n,\ \hbox{ continuation by rotation of } 2\beta. \label{pro+transe}$$ Function $\phi_2(s)$ can be continued as meromorphic on $\zeta \eta \Delta^\epsilon, (\zeta \eta)^2 \Delta^\epsilon, \cdots
(\zeta \eta) ^n \Delta^\epsilon $ by the formulas : $$\varphi_2(s)=
\frac{\gamma_2( \eta \zeta s)\gamma_1( \zeta s)}{\gamma_1( \eta \zeta s)\gamma_2( \zeta s)} \varphi_2( \eta \zeta s), \ \ \forall s\in
(\zeta\eta)^k \Delta^\epsilon, k=1,2,\ldots n, \hbox{ continuation by rotation of} -2\beta. \label{pro-trans1e}$$
We proceed by induction on $k=1,2,\ldots n$. For $k=1$, this is the subject of the previous lemma. For any $k=2,\ldots,n$, assume the formula (\[pro+transe\]) for any $s\in (\eta \zeta)^{k-1} \Delta$. The domain $(\eta \zeta)^{k-1} \Delta^\epsilon \cap (\eta \zeta)^k \Delta^\epsilon=(\eta \zeta)^{k-1}(\Delta^\epsilon \cap \eta \zeta \Delta^\epsilon)$ is a non empty open domain by Lemma \[lemconteps\], $(\eta \zeta)^{k-1}$ being just the rotation for the angle $2(k-1)\beta$. The formula (\[pro+transe\]) is valid for any $s \in (\eta \zeta)^{k-1} \Delta^\epsilon \cap (\eta \zeta)^k \Delta^\epsilon$ by induction assumption. Hence, by the principle of meromorphic continuation it is valid for any $s\in (\eta \zeta)^k \Delta^\epsilon$. The same is true for the formula (\[pro-trans1e\]).
Proceeding as in Lemma \[lemcontepse\] by rotations, we will continue $\phi_1$ soon on the first half of ${\bf S}$, that is ${\bf S}_{\theta_2}^1$, then the whole of ${\bf S}$ and go further, turning around ${\bf S}$ for the second time, for the third, etc up to infinity. In fact, each time we complete this procedure on one of two halves of ${\bf S}$, we recover a new branch of the function $\phi_1$ as function of $\theta_2 \in {\bf C}$. So, going back to the complex plane, we continue this function as multivalued and determine all its branches. The same is true for $\phi_2$ if we proceed by rotations in the opposite direction. This procedure could be presented better on the universal covering of ${\bf S}$, but for the purpose of the present paper it is enough to complete it only on one-half of ${\bf S}$, that is to study just the first (main) branch of $\phi_1$ and $\phi_2$. We summarize this result in the following theorem. We recall that ${\bf S} ={\bf S}_{\theta_1}^1 \cup {\bf S}_{\theta_1}^2$ and we denote by ${\bf S}_{\theta_1}^1$ the half that contains $s'_0$ (and not $s_0$, as $\zeta s_0=s'_0$). In the same way ${\bf S} ={\bf S}_{\theta_2}^1 \cup {\bf S}_{\theta_2}^2$ and we denote by ${\bf S}_{\theta_2}^1$ the half that contains $s''_0$ (and not $s_0$, as $\eta s_0=s''_0$), see Figure \[recouvrS\].
\[thmpro\] For any $s\in {\bf S}_{\theta_2}^1$ there exists $n{\geqslant}0$ such that $(\zeta \eta)^n s \in \bar \Delta$. Let us define $$\label{qs}
\phi_1(s)= \frac{\gamma_1( (\zeta \eta)^n s)\dots \gamma_1( \zeta \eta s)}{\gamma_2( (\zeta \eta)^n s)\dots \gamma_2( \zeta \eta s)}
\frac{\gamma_2( \eta (\zeta \eta)^{n-1} s)\dots\gamma_2( \eta s)}{\gamma_1( \eta \zeta \eta)^{n-1} s)\dots \gamma_1( \eta s)}
\phi_1((\zeta\eta)^n s)$$ Then the function $\phi_1(s)$ is meromorphic on ${\bf S}_{\theta_2}^1$. For any $s\in {\bf S}_{\theta_1}^1$, there exists $n{\geqslant}0$ such that $(\eta \zeta)^n s \in \bar \Delta$. Let us define $$\label{qse}
\phi_2(s)=
\frac{\gamma_2( (\eta \zeta)^n s)\dots \gamma_2( \eta \zeta s)}{\gamma_1( (\eta \zeta)^n s)\dots \gamma_1( \eta \zeta s)}
\frac{\gamma_1( \zeta (\eta \zeta)^{n-1} s)\dots\gamma_1( \eta s)}{\gamma_2( \zeta \eta \zeta)^{n-1} s)\dots \gamma_2( \zeta s)}
\phi_1((\eta\zeta)^n s)$$ Then the function $\phi_2(s)$ is meromorphic on ${\bf S}_{\theta_1}^1$.
It is a direct corollary of Lemma \[lemrev\] and Lemma \[lemcontepse\].
Poles of functions phi1 and phi2 on S {#mainpole}
-------------------------------------
It follows from meromorphic continuation procedure that all poles of $\phi_1(s)$ and $\phi_2(s)$ on ${\bf S}$ are located on the ellipse ${{\mathcal}E}$, they are images of zeros of $\gamma_1$ and $\gamma_2$ by automorphisms $\eta$ and $\zeta$ applied several times. Then all poles of all branches of $\phi_2(s)$ (resp. $\phi_2(s)$) on ${\bf C}_{\theta_1}$ (resp. ${\bf C}_{\theta_2}$) are on the real segment $[\theta _1^-, \theta_1^+]$ (resp. $ [\theta _2^-, \theta_2^+]$).
[**Notations of arcs on $\mathcal{E}$.**]{} Let us remind that we denote by $\{s_1, s_2\}$ an arc of the ellipse ${{\mathcal}E}$ with ends at $s_1$ and $s_2$ passing through the origin, see Theorem \[thmmain\]. From now on, we will denote in square brackets $]s_1, s_2[$ or $[s_1,s_2]$ an arc of ${{\mathcal}E}$ going in the anticlockwise direction from $s_1$ to $s_2$.
In order to compute the asymptotic expansion of stationary distribution density, we are interested in poles of $\phi_1$ on the arc $]s''_0, s_2^+[$ and in those of $\phi_2$ on the arc $]s_1^+, s'_0[$. To determine the main asymptotic term, we are particularly interested in the pole of $\phi_1(\theta_2(s))$ on $]s''_0, s_2^+[$ closest to $s''_0$ and in the one of $\phi_2(\theta_1(s))$ on $]s_1^+, s'_0[$ closest to $s'_0$. We identify them in this section.
We remind that $\theta^{*}$ is a zero of $\gamma_1(s)$ on $\mathcal{E}$ different from $s_0$ and that $\theta^{**}$ is a zero of $\gamma_2(s)$ on $\mathcal{E}$ different from $s_0$. Their coordinates are $$\begin{aligned}
\theta^*&=&
2 \frac{r_{21}\mu_1-r_{11}\mu_2}{r_{21}^2\sigma_{11}-2r_{11}r_{21}\sigma_{12}+r_{11}^2\sigma_{22}}
\Big(-r_{21} , r_{11} \Big),\nonumber \\
\theta^{**}&=&
2 \frac{r_{12}\mu_2-r_{22}\mu_1}{r_{22}^2\sigma_{11}-2r_{22}r_{12}\sigma_{12}+r_{12}^2\sigma_{22}}
\Big(r_{22} , -r_{12} \Big)
\label{zerostheta}\end{aligned}$$ Their images by automorphisms $\eta$ and $\zeta$ have the following coordinates: $$\begin{aligned}
\eta \theta^*&=&\Big( -\frac{r_{11}}{\sigma_{11}r_{21}}(\sigma_{22}\theta_2^{*}+2\mu_2) , \theta_2^* \Big),\nonumber\\
\zeta \theta^{**}&=& \Big(\theta_1^{**} ,-\frac{r_{22}}{\sigma_{22}r_{12}}(\sigma_{11}\theta_1^{**}+2\mu_1) \Big).\label{zerosthetaauto}\end{aligned}$$
\[lem3\]
- If $\theta^{**}\in ]s_0,s_1^+[$, then $\zeta \theta^{**}$ is a pole of $\phi_2(\theta_1(s))$ on $]s_1^+, s'_0[$.
- If $\theta^{*}\in ]s_2^+, s_0[$, then $\eta \theta^{*}$ is a pole of $\phi_1(\theta_2(s))$ on $]s_0'',s_2^+[$.
By meromorphic continuation procedure $$\label{mpo}
\phi_2(\zeta\theta^{**})=\frac{\gamma_2(\eta \theta^{**})\gamma_1(\theta^{**})\phi_2(\eta\theta^{**})}{\gamma_2(\theta^{**})
\gamma_1(\eta\theta^{**})}.$$
Let us check that the numerator in (\[mpo\]) is non zero, this will prove the statement (1) the lemma.
It is clear that $\gamma_1(\theta^{**})\ne 0$ due to stability conditions (\[u\]) and (\[v\]).
Suppose that $\gamma_2(\eta\theta^{**})=0$. This could be only if $\eta \theta^{**}=\theta^{**} \in ]s_0, s_1^+[$, thus $\theta^{**}=s_2^-$ where $\theta_2(s_2^-) <0$ and consequently $\phi_1( \eta\theta^{**})<\infty$. But by meromorphic continuation of $\phi_2(\theta_1(s))$ to the arc $\{s \in {{\mathcal}E} : \theta_2(s)<0\}$ we have: $\phi_2(\eta \theta^{**}_1)=-\frac{\gamma_1(\eta \theta^{**})\phi_1(\eta \theta^{**}_2)}{\gamma_2( \eta\theta^{**})}$, from where by (\[mpo\]) $$\phi_2(\zeta \theta^{**})=-\frac{\gamma_1(\theta^{**})} { \gamma_2(\theta^{**})}\phi_1(\eta \theta_2^{**}).$$ Then $\zeta \theta^{**}$ is clearly a pole of $\phi_2$, this finishes the proof of statement (1) of the lemma in this particular case. Otherwise $\gamma_2(\eta \theta^{**}) \ne 0$.
Let us finally check that $\phi_2(\eta \theta^{**})\ne 0$. Let us first observe that $\phi_2(\theta_1(s))\ne 0$ for any $s \in {{\mathcal}E}$ with one of two coordinates non-positive. In fact, if the first coordinate $\theta_1(s)$ of $s$ is non-positive, then $\phi_2(\theta_1(s))\ne 0$ by its definition. If $s$ has the second coordinate $\theta_2(s)$ non-positive, then $\phi_2(\theta_1(s))=-\frac{\gamma_1(s)}{\gamma_2( s)}\phi_1(\theta_2(s))$ where $\gamma_1(s)$ can not have zeros with the second coordinate non-positive by stability conditions and neither $\phi_1(\theta_2(s))$ by its definition. Hence, $\phi_2(\theta_1(s))\ne 0$ on the arc $\{s \in {{\mathcal}E} : \theta_1(s){\leqslant}0 \hbox{ or } \theta_2(s) {\leqslant}0\}$.
It remains to consider the case where both coordinates of $\eta \theta^{**}$ are positive, i.e. $\theta^{**} \in ]\eta s'_0, s_1^+[$ where the parameters are such that $s^{1,+}_2>\theta_2(\eta s'_0)>0$ and to show that $\phi_2(\eta \theta^{**}_1)\ne 0$. Suppose the contrary, that $\phi_2(\eta \theta^{**})=0$. Then there are zeros of $\phi_2$ on $] \eta s_1^+, s'_0 [$ and among these zeros there exists $\theta^0$ the closest one to $s'_0$. By meromorphic continuation $$\label{zez}
\phi_2(\theta^{0}_1)=\frac{\gamma_2( \eta \zeta \theta^{0})\gamma_1(\zeta \theta^{0})\phi_2(\eta\zeta \theta^{0})}{\gamma_2(\zeta \theta^{0}_1)
\gamma_1(\eta\zeta \theta^{0})},$$ where $\eta\zeta \theta^0 \in ]\theta_0, s''_0]$. First of all, we note that $\phi_2(\eta\zeta \theta^{0})\ne 0$ if $\eta \zeta \theta^0 \in [ \theta_0, s'_0[$, since $\theta^0$ is the closest zero to $s'_0$, and $\phi_2(\eta\zeta \theta^{0})\ne 0$ if $\eta \zeta \theta^0 \in [s'_0, s''_0]$, because one of coordinates of $\eta \zeta \theta^0$ is non-positive within this segment. Hence, $\phi_2(\eta\zeta \theta^{0})\ne 0$ for any point $\eta\zeta \theta^0 \in ]\theta_0, s''_0[$.
Furthermore, since $\eta\zeta \theta^0 \in ]\theta_0, s''_0[$, then $\eta \zeta \theta^0 \ne \theta^{**}$ and thus $\gamma_2(\eta \zeta \theta^0)\ne 0$ except for $\eta \zeta \theta^0= s_0$. As for this particular case $\eta \zeta \theta^0= s_0$, we would have $\phi_2(\theta^0)=-\gamma_1(s''_0)\phi_1(s_0)\gamma_2^{-1}(s''_0)\ne 0$, so that $\theta^0=\zeta \eta s_0$ can not be a zero of $\phi_2$.
The point $\zeta \theta^0 \in \zeta [ \eta \theta^{**}, s'_0]=\zeta \eta [\eta s'_0, \theta^{**}]$ that is the segment $[\eta s'_0, \theta^{**}]$ rotated for the angle $-2\beta$. Hence $\zeta \theta^0$ is located on ${{\mathcal}E}$ below $\theta^{**}$. Then $\gamma_1(\zeta \theta^0)=0$ combined with $\gamma_2(\theta^{**})=0$ is impossible by stability conditions (\[u\]) and (\[v\]). Thus $\gamma_1(\zeta \theta^0)\ne 0$. It follows from (\[zez\]) that $\phi_2(\theta^0_1)\ne 0$. Thus there exist no zeros of $\phi_2 $ on $] \eta s^{1,+}, s'_0 [$ and finally $\phi_2(\eta \theta^{**}) \ne 0$. Therefore the numerator in (\[mpo\]) is non zero, hence $\zeta \theta^{**}$ is a pole of $\phi_2$.
The reasoning for $\theta^*$ is the same.
\[lem1\]
- Assume that $\theta^p \in ]s_1^+, s'_0[$ is a pole of $\phi_2(\theta_1)$ and it is the closest pole to $s'_0$ .
If the parameters $(\Sigma, \mu)$ are such that $\theta_2(s_1^+){\leqslant}0$, or the parameters $(\Sigma, \mu, R)$ are such that $\theta_2(s_1^+) >0$ but $\eta \zeta \theta^p \not\in ]\eta s_1^+, s_0[$, then $\gamma_2( \zeta \theta^p)=0$ where $\zeta \theta^p \in ]s_1^+, s_0[$ and $\theta^p$ is a pole of the first oder.
If the parameters $(\Sigma, \mu, R)$ are such that $\theta_2(s_1^+)>0$ and $\eta \zeta \theta^p \in ]\eta s_1^+, s_0[$, then either $\gamma_2( \zeta \theta^p)=0$ where $\zeta \theta^p \in ]s_0, s_1^+[$ or $\gamma_1 (\eta\zeta \theta^p)=0$. Furthermore, in this case, if $\gamma_2( \zeta \theta^p)$ and $\gamma_1 (\eta\zeta \theta^p)$ do not equal zero simultaneously, then $\theta^p$ is a pole of the first order.
- Assume that $\theta^p \in ]s_0'',s_2^+[$ is a pole of $\phi_1(\theta_2)$ and it is the closest pole to $s_0''$.
If the parameters $(\Sigma, \mu)$ are such that $\theta_1(s_2^+){\leqslant}0$, or the parameters $(\Sigma, \mu, R)$ are such that $\theta_1(s_2^+) >0$ but $\zeta \eta \theta^p \not\in ]s_0,\zeta s_2^+[$, then $\gamma_1( \eta \theta^p)=0$ where $\eta \theta^p \in ]s_0, s_2^+[$ and $\theta^p$ is a pole of the first oder.
If the parameters $(\Sigma, \mu, R)$ are such that $\theta_1(s_2^+)>0$ and $\zeta \eta \theta^p \in ]s_0, \eta s_1^+, [$, then either $\gamma_1( \eta \theta^p)=0$ where $\eta \theta^p \in ] s_2^+,s_0[$ or $\gamma_2 (\zeta\eta \theta^p)=0$. Furthermore, in this case, if $\gamma_1( \eta \theta^p)$ and $\gamma_2 (\zeta\eta \theta^p)$ do not equal zero simultaneously, then $\theta^p$ is a pole of the first order.
Due to meromorphic continuation procedure we have $$\label{pri}
\phi_2(\theta^p_1)=\frac{\gamma_2(\eta\zeta \theta^p) \gamma_1(\zeta\theta^p) \phi_2(\eta\zeta\theta^p_1)}{\gamma_2(\zeta\theta^p)\gamma_1(\eta\zeta\theta^p)}$$ where $\eta\zeta \theta^p \in ]\theta^p, s''_0]$.
Assume that $\eta\zeta\theta^p \in ]\theta^p, s_0[$. In this case point $\eta\zeta \theta^p$ has the second coordinate positive and so does $\zeta\theta^p \in [s_0, s_1^+[$. It follows that $\theta_2(s_1^+)>0$ and $\eta\zeta \theta^p \in \eta ]s_0, s_1^+[ \cap \{\theta : \theta_2>0\}=]\eta s_1^+, s_0[$.
Thus, if the parameters $(\Sigma, \mu)$ are such that $\theta_2(s_1^+){\leqslant}0$, or the parameters $(\Sigma, \mu, R)$ are such that $\theta_2(s_1^+)>0$ but $\eta \zeta \theta^p \not\in ]\eta s_1^+, s_0[$, then the second coordinate of $\eta\zeta \theta^p$ is non-positive, i.e. $\eta\zeta\theta^p \in [s_0, s''_0]$. In this case $$\phi_2(\eta\zeta\theta^p_1)=-\frac{ \gamma_1(\eta\zeta\theta^p) }{ \gamma_2(\eta\zeta\theta^p) }\phi_1(\eta\zeta\theta^p)$$ from where by (\[pri\]) $$\label{dp}
\phi_2(\theta^p_1)=-\frac{ \gamma_1(\zeta\theta^p) \phi_1(\eta\zeta\theta^p_2)}{\gamma_2(\zeta\theta^p)}.$$ Since $\phi_1(\eta\zeta\theta^p_2)$ is finite for for any $\eta\zeta\theta^p \in [s_0, s''_0]$ by its initial definition, the formula (\[dp\]) implies that $\gamma_2(\zeta\theta^p)=0$ and the pole $\theta^p$ is of the first order.
If parameters $(\Sigma, \mu, R)$ are such that $\theta_2(s_1^+)>0$ and $\eta \zeta \theta^p \in ]\eta s_1^+, s_0[$, then either $\eta\zeta\theta^p \in [s''_0, s_0]$ or $\eta\zeta\theta^p \in ]s_0, \theta^p[$. In the first case we have (\[dp\]) as previously from where $\gamma_2(\zeta\theta^p)=0$ and the pole $\zeta \theta^p$ is of the first order. Let us turn to the second case $\eta\zeta\theta^p \in ]\theta^p, s_0[$ for which we will use the formula (\[pri\]). The pole $\theta^p$ being the closest to $s'_0$, then $\eta\zeta\theta^p$ can not be a pole of $\phi_2$ on $]\theta^p, s'_0[$. It can neither be a pole of $\phi_2$ on $[s'_0, s_0]$, since this function is initially well defined on this segment. Hence in formula (\[pri\]) $\phi_2(\eta\zeta\theta^p_1)\ne \infty$ for $\eta\zeta \theta^p \in ]\theta^p, s'_0[$. It follows from (\[pri\]) that either $\gamma_2( \zeta \theta^p)=0$ or $\gamma_1 (\eta\zeta \theta^p)=0$ and if these two equalities do not hold simultaneously, then pole $\theta^p$ must be of the first order.
The proof in the case (ii) is symmetric.
Figure \[fi12\] gives two illustrations of Lemmas \[lem3\] and \[lem1\].
On the left figure the parameters are such that $\theta_1(s_2^+)>0$ and $\theta_2(s_1^+)>0$. Let us look at zeros $\theta^*$ of $\gamma_1$ and $\theta^ {**}$ of $\gamma_2$ different from $s_0$. We see $\theta^{*} \in ]s_2^+, s_0[$, then $\eta \theta^*$ is the first candidate for the closest pole of $\phi_1$ to $s''_0$ on $]s''_0 ,s_2^+[$. We also see $\theta^{**} \not\in [s_0, \zeta s_2^+]$, then there are no other candidates. Hence the closest pole of $\phi_1$ to $s''_0$ on $]s''_0 ,s_2^+[$ is $\eta \theta^{*}$. Since $\theta^{**} \in ]s_0, s_1^+[$, then $\zeta \theta^{**}$ is the first candidate for the closest pole of $\phi_2$ to $s'_0$ on $]s_1^+, s'_0[$. Furthermore, $\theta^* \in ] \eta s_1^+, s_0[$, so that $\zeta \eta \theta^*$ is the second candidate to be the closest pole of $\phi_2$ to $s'_0$ on $]s_1^+, s_0[$. We see at the picture that $\zeta \eta \theta^*$ is closer to $s'_0$ than $\zeta \eta^{**}$.
On the right figure the parameters are such that $\theta_1(s_2^+)<0$ and $\theta_2(s_1^+)<0$. We see $\theta^{*} \in ]s_2^+, s_0[$, then $\eta \theta^*$ is immediately the closest pole of $\phi_1$ to $s''_0$ on $]s''_0 ,s_2^+[$. Since $\theta^{**} \not \in ]s_0, s_1^+[$, then there are no poles of $\phi_2$ on $]s_1^+, s'_0[$.
![On the left figure : $\eta \theta^*$ is the closest pole of $\phi_1$ to $s''_0$ on $]s''_0 ,s_2^+[$, $\zeta \eta \theta^*$ is the closest pole of $\phi_2$ to $s'_0$ on $]s_1^+, s_0[$. On the right figure: $\eta \theta^*$ is the closest pole of $\phi_1$ to $s''_0$ on $]s''_0 ,s_2^+[$, there are no poles of $\phi_2$ on $]s_1^+, s_0[$ []{data-label="fi12"}](ellipsepole.pdf "fig:") ![On the left figure : $\eta \theta^*$ is the closest pole of $\phi_1$ to $s''_0$ on $]s''_0 ,s_2^+[$, $\zeta \eta \theta^*$ is the closest pole of $\phi_2$ to $s'_0$ on $]s_1^+, s_0[$. On the right figure: $\eta \theta^*$ is the closest pole of $\phi_1$ to $s''_0$ on $]s''_0 ,s_2^+[$, there are no poles of $\phi_2$ on $]s_1^+, s_0[$ []{data-label="fi12"}](ellipsepole2.pdf "fig:")
We will also need the following two lemmas.
\[lem2\]
- Assume that $\theta_2(s_1^+)>0$. Then for any $s \in ]\eta s_1^+, s_0[$ we have $\theta_2(\zeta \eta s)> \theta_2(\eta s)$.
- Assume that $\theta_1(s_2^+)>0$. Then for any $s \in ]s_0,\zeta s_2^+ [$ we have $\theta_1(\eta \zeta s)> \theta_1(\zeta s)$.
Since $\theta_2(s_1^+)>0$, then $\theta_2(\zeta \eta s_0) - \theta_2(\eta s_0)>0$. Consider the function $f(s)=\theta_2(\zeta s)-\theta_2(s)$ for $s \in [s''_0, s_1^+]$. It depends continuously on $s$ on this arc. We note that $f(s''_0)= \theta_2(\zeta \eta s_0)-\theta_2(\eta s_0)>0$, $f(s_1^+)=0$. Furthermore, since $s_1^-\not\in [s''_0, s_1^+]$, then $f(s)\ne 0$ for all $s \in ]s''_0, s^{1,+}]$. Hence $f(s)>0$ for all $s \in ]s''_0, s_1^+]$, from where $\theta_2(\zeta \eta s)- \theta_2(\eta s) =f(\eta s)>0$ for any $s \in \eta ]s''_0, s_1^+]= [\eta s_1^+, s_0[$. The proof in the other case is symmetric.
\[lem4\]
Assume that $\gamma_2(s)$ has a zero $\theta^{**} \in ]s_0, s_1^+[$ and $\gamma_1(s)$ has a zero $\theta^* \in]s_2^+, s_0[$. Then one of the following three assertions holds true :
- The closest pole of $\phi_2(\theta_1(s))$ to $s'_0$ on $]s_1^+, s'_0[$ is $\zeta \theta^{**}$, the closest pole of $\phi_1(\theta_2(s))$ to $s''_0$ on $]s''_0, s_2^+[$ is $\eta \theta^*$, both of them are of the first order.
- The closest pole of $\phi_2(\theta_1(s))$ to $s'_0$ on $]s_1^+, s'_0[$ is $\zeta \theta^{**}$, it is of the first order. The closest pole of $\phi_1(\theta_2(s))$ to $s''_0$ on $]s''_0, s_2^+[$ is $\eta \zeta \theta^{**}$ where $ \theta_1(\eta \zeta \theta^{**})>
\theta_1(\zeta \theta^{**})$.
- The closest pole of $\phi_2(\theta_1(s))$ to $s'_0$ on $]s_1^+, s'_0[$ is $\zeta \eta \theta^{*}$ where $
\theta_2(\zeta \eta \theta^{*})> \theta_2(\eta \theta^{*})$. The closest pole of $\phi_1(\theta_2(s))$ to $s''_0$ on $]s''_0, s_2^+[$ is $\eta \theta^{*}$, it is of the first order.
The case (ii) is illustrated on Figure \[fi12\].
By Lemma \[lem3\] there exist poles of the function $\phi_1(\theta_2(s))$ on $]s''_0, s_2^+[$. By Lemma \[lem1\] under parameters such that $\theta_1(s_2^+){\leqslant}0$ or $\theta_1(s_2^+)> 0$ and $\theta^{**} \not \in ]s_0, \zeta s_2^+[$, $\eta \theta^*$ is the closest pole to $s''_0$ and it is of the first order. By the same lemma under parameters such that $\theta_1(s_2^+)> 0$ and $\theta^{**} \in ]s_0, \zeta s_2^+[$, either $\eta \theta^{*}$ or $\eta \zeta \theta^{**}$ is the closest pole to $s'_0$. By Lemma \[lem1\], if $\gamma_1(\zeta \theta^{*})\ne 0$, pole $\eta \theta^*$ is of the first order. Condition $\gamma_1(\zeta\theta^*) \ne 0$ is equivalent to $\zeta \theta^* \ne \theta^{**}$. This means just that pole $\eta \theta^{*}$ is different from $\eta \zeta \theta^{**}$ which is another candidate for the closest pole to $s''_0$. By Lemma \[lem2\] $\theta_1(\eta\zeta \theta^{**})>\theta_1(\zeta \theta^{**})$ To summarize, one of two following statements holds true:
- Point $\eta \theta^*$ is the closest pole of $\phi_1(\theta_2(s))$ to $s''_0$ on $]s''_0, s_2^+[$ and it is of the first order;
- The parameters are such that $\theta_1(s_2^+)> 0$ and $\theta^{**} \in ]s_0, \zeta s_2^+[$. Point $\eta \zeta \theta^{**}$ is the closest pole of $\phi_1(\theta_2(s))$ to $s''_0$ on $]s''_0, s_2^+[$ and $ \theta_1(\eta\zeta \theta^{**})> \theta_1(\zeta \theta^{**})$.
By Lemmas \[lem3\], \[lem1\], \[lem2\] and the same considerations, one of the following statements about $\phi_2(\theta_1)$ holds true:
- Pole $\zeta \theta^{**}$ is the closest pole of $\phi_2(\theta_1(s))$ to $s'_0$ on $]s_1^+, s'_0[$ and it is of the first order.
- The parameters are such that $\theta_2(s_1^+)>0$, $\theta^* \in ]\eta s_1^+, s_0[$, point $\zeta \eta \theta^{*}$ is the closest pole of $\phi_2(\theta_1(s))$ to $s'_0$ on $]s_1^+, s'_0[$ and $\theta_2(\zeta\eta \theta^{*})>\theta_2(\eta \theta^{*})$
Let us finally prove that (b1) and (b2) can not hold true simultaneously. Assume that $\theta_2(s_1^+)>0$, $\theta^{*} \in ]\eta s_1^+, s_0[$, $\theta_1(s_2^+)>0$, $\theta^{**} \in ]s_0, \zeta s_2^+[$ and e.g. (b2), that is $\zeta \eta \theta^{*}$ is the closest pole to $s_0$. Note that in this case $\zeta \theta^{**} \in ]s_2^+, s'_0[$. Then $\zeta \eta \theta^{*}$ is closer to $s'_0$ than the pole $\zeta \theta^{**}$ on this segment or coincides with it. Hence $\theta_1(\zeta \eta \theta^{*}) {\leqslant}\theta_1(\zeta \theta^{**})$ and $\theta_2(\zeta \eta \theta^{*}) {\leqslant}\theta_2 (\zeta \theta^{**})$. By Lemma \[lem2\] $\theta_1(\eta \theta^*) =\theta_1 (\zeta \eta \theta^{*})$ and $\theta_1(\zeta \theta^{**})<\theta_1(\eta \zeta \theta^{**})$, $\theta_2(\eta \theta^*) < \theta_2(\zeta \eta \theta^{*})$ and $\theta_2(\zeta \theta^{**})=\theta_2(\eta \zeta \theta^{**})$. Then $\theta_1(\eta \theta^*)<\theta_1(\eta \zeta \theta^{**})$, $\theta_2(\eta \theta^*) < \theta_2(\eta \zeta \theta^{**})$. This means that that $\eta \theta^*$ is the closest pole of $\phi_1(\theta_2(s))$ to $s''_0$, $\eta\theta^* \ne \eta \zeta \theta^{**}$ , so that (b1) is impossible for $\phi_1(\theta_2(s))$, then we have (a1).
In the same way assumption (b1) leads to (a2). Thus (b1), (b2) can not hold true simultaneously, the lemma is proved.
Contribution of the saddle-point and of the poles to the asymptotic expansion
=============================================================================
Stationary distribution density as a sum of integrals on S {#invlaplacetransf}
-----------------------------------------------------------
By the functional equation (\[maineq\]) and the inversion formula of Laplace transform (we refer to [@doetsch_introduction_1974] and [@brychkov_multidimensional_1992]), the density $\pi(x_1,x_2)$ can be represented as a double integral $$\begin{aligned}
\pi(x_1,x_2)
&=
\frac{1}{(2\pi i)^2}
\int_{-i\infty}^{i\infty} \int_{-i\infty}^{i\infty}
e^{-x_1\theta_1-x_2\theta_2} \varphi(\theta_1,\theta_2)
\mathrm{d} \theta_1 \mathrm{d} \theta_2\nonumber
\\
&=\frac{-1}{(2\pi i)^2} \int_{-i\infty}^{i\infty} \int_{-i\infty}^{i\infty}
e^{-x_1\theta_1-x_2\theta_2}
\frac{\gamma_1(\theta) \varphi_1(\theta_2)+\gamma_2(\theta)\varphi_2(\theta_1)}{\gamma(\theta)}
\mathrm{d} \theta_1 \mathrm{d} \theta_2.\label{vbn}\end{aligned}$$ We now reduce it to a sum of single integrals.
For any $(x_1, x_2) \in {\bf R}^2_+$ $$\pi(x_1,x_2)= I_1(x_1,x_2)+I_2(x_1,x_2)$$ where $$\begin{aligned}
I_1 (x_1, x_2) &=
\frac{1}{2\pi i} \int_{-i\infty}^{i\infty}
\varphi_2(\theta_1) \gamma_2(\theta_1, \Theta_2^+(\theta_1)) e^{-x_1\theta_1-x_2\Theta_2^+(\theta_1)}
\frac{\mathrm{d} \theta_1}{\sqrt{ d(\theta_1) }} \label{I1}
$$ and $$\begin{aligned}
I_2 (x_1, x_2) &=
\frac{1}{2\pi i} \int_{-i\infty}^{i\infty}
\varphi_1(\theta_2) \gamma_1(\Theta_1^+(\theta_2), \theta_2) e^{-x_1\Theta_1^+(\theta_2)-x_2\theta_2}
\frac{\mathrm{d} \theta_2}{\sqrt{ \tilde d (\theta_2) }}. \label{I2}
$$ \[propI\]
By inversion formula (\[vbn\]) $$\begin{aligned}
\pi(x_1,x_2)
&=
\frac{-1}{2\pi i} \int_{-i\infty}^{i\infty}
\varphi_2(\theta_1) e^{-x_1\theta_1}
\Big(\frac{1}{2\pi i} \int_{-i\infty}^{i\infty} \frac{\gamma_2(\theta)}{\gamma(\theta)} e^{-x_2\theta_2} \mathrm{d} \theta_2\Big) \mathrm{d} \theta_1
\\ &+
\frac{-1}{2\pi i} \int_{-i\infty}^{i\infty}
\varphi_1(\theta_2) e^{-x_2\theta_2}
\Big(\frac{1}{2\pi i} \int_{-i\infty}^{i\infty} \frac{\gamma_1(\theta)}{\gamma(\theta)} e^{-x_1\theta_1} \mathrm{d} \theta_1\Big) \mathrm{d} \theta_2.\\\end{aligned}$$ Now it suffices to show the following formulas $$\label{xxxp}
\frac{-1}{2\pi i} \int_{-i\infty}^{i\infty} \frac{\gamma_2(\theta)e^{-x_2\theta_2}}{\gamma(\theta)} \mathrm{d} \theta_2
=
\frac{\gamma_2(\theta_1, \Theta_2^+(\theta_1))}{\sqrt{ d(\theta_1) }} e^{-x_2\Theta_2^+(\theta_1)},$$ $$\label{yyyp}
\frac{-1}{2\pi i} \int_{-i\infty}^{i\infty} \frac{\gamma_1(\theta)e^{-x_1\theta_1}}{\gamma(\theta)} \mathrm{d} \theta_1
=
\frac{\gamma_1(\Theta_1^+(\theta_2), \theta_2)}{\sqrt{\tilde d(\theta_2) }} e^{-x_1\Theta_1^+(\theta_2)}.$$ Let us prove (\[xxxp\]). For any $\theta_1 \in i{\bf R} \setminus \{0\}$, the function $\gamma(\theta)=\frac{\sigma_{22}}{2}(\theta_2-\Theta_2^{+}(\theta_1))(\theta_2-\Theta_2^{-}(\theta_1))$ has two zeros $\Theta_2^{+}(\theta_1)$ and $\Theta_2^{-}(\theta_1)$. Their real parts are of opposite signs: $\Re(\Theta_2^-(\theta_1))< 0$ and $\Re(\Theta_2^+(\theta_1))>0$. Thus for any fixed $\theta_1 \in i{\bf R}\setminus\{0\}$, function $\frac{\gamma_2(\theta) e^{-x_2\theta_2}}{\gamma(\theta)}$ of the argument $\theta_2$ has two poles on the complex plane ${\bf C}_{\theta_2}$, one at $\Theta^{-}_2(\theta_1)$ with negative real part and another one at $\Theta^{+}_2(\theta_1)$ with positive real part. Let us construct a contour $\mathcal{C}_R =[-iR,iR] \cup \{R e^{it} \mid t \in ]-\pi/2, \pi/2[\}$ composed of the purely imaginary segment $[-iR,iR]$ and the half of the circle with radius $R$ and center $0$ on ${\bf C}_{\theta_2}$, see Figure \[contourintresidu\]. For $R$ large enough $\Theta^{+}_2(\theta_1)$ is inside the contour. The integral of $\frac{\gamma_2(\theta) e^{-x_2\theta_2}}{\gamma(\theta)}$ over this contour taken in the counter-clockwise direction equals the residue at the unique pole of the integrand: $$\frac{1}{2\pi i}
\int_{\mathcal{C}_R} \frac{\gamma_2(\theta) e^{-x_2\theta_2}}{\gamma(\theta)} d\theta_2 = {\rm res}_{\theta_2=\Theta^{+}_2(\theta_1)}
\frac{\gamma_2(\theta) e^{-x_2\theta_2}}{\gamma(\theta)}= \frac{\gamma_2(\theta_1 ,\Theta^{+}_2(\theta_1))}{(\sigma_{22}/2) (
\Theta^{+}_2(\theta_1)-\Theta^{-}_2(\theta_1) )}e^{-x_2 \Theta^{+}_2(\theta_1)}$$ $$\label{jmmcb}
=
\frac{\gamma_2(\theta_1, \Theta^{+}_2(\theta_1))}{
\sqrt{d(\theta_1)}} e^{-x_2 \Theta^{+}_2(\theta_1)} \ \ \ \hbox{ for all large enough }R{\geqslant}0.$$
![Contour $\mathcal{C}_R$ on ${\bf C}_{\theta_2}$.[]{data-label="contourintresidu"}](contourint1.pdf)
Let us take the limit of this integral as $R \to \infty$: $$\lim_{R \to \infty} \int_{\mathcal{C}_R} \frac{\gamma_2(\theta) e^{-x_2\theta_2}}{\gamma(\theta)} d\theta_2 =
- \lim_{R \to \infty} \int_{-i R}^{i R} \frac{\gamma_2(\theta) e^{-x_2\theta_2}}{\gamma(\theta)} d\theta_2 +
\lim_{R \to \infty} \int_{\{Re^{it}|t\in]-\frac{\pi}{2},\frac{\pi}{2}[\}} \frac{\gamma_2(\theta) e^{-x_2\theta_2}}{\gamma(\theta)} d\theta_2.
\label{eqax}$$ The last term equals $$\label{56}
\lim_{R \to \infty} \int_{\{Re^{it}|t\in]-\frac{\pi}{2},\frac{\pi}{2}[\}} \frac{\gamma_2(\theta) e^{-x_2\theta_2}}{\gamma(\theta)} d\theta_2=\lim_{ R \to \infty }
\int_{-\frac{\pi}{2}}^{\frac{\pi}{2}} \frac{\gamma_2(\theta_1,Re^{it}) }{\gamma(\theta_1,Re^{it})} e^{-x_2 Re^{it}} iR e^{it}d t.$$ We note that $\sup_{R>0}\sup_{t\in]-\frac{\pi}{2},\frac{\pi}{2}[} |iRe^{it} \frac{\gamma_2(\theta_1,Re^{it}) }{\gamma(\theta_1,Re^{it})}|<\infty $ and $\sup_{R>0}\sup_{t\in]-\frac{\pi}{2},\frac{\pi}{2}[}|e^{-x_2 R e^{it}}|{\leqslant}1$. Furthermore $|e^{-x_2 Re^{it}}|=e^{-x_2 R\cos{t}}\to 0$ as $R\to \infty$ for all $t\in]-\frac{\pi}{2},\frac{\pi}{2}[$. Then by dominated convergence theorem the limit (\[56\]) equals 0 as $R \to \infty$. Hence, due to (\[jmmcb\]) and (\[eqax\]) $$\frac{\gamma_2(\theta_1 ,\Theta^{+}_2(\theta_1))}{(\sigma_{22}/2) (
\Theta^{+}_2(\theta_1)-\Theta^{-}_2(\theta_1) )}e^{-x_2 \Theta^{+}_2(\theta_1)} =
\lim_{R \to \infty} \int_{\mathcal{C}_R} \frac{\gamma_2(\theta) e^{-x_2\theta_2}}{\gamma(\theta)} d\theta_2 =
\int_{-i \infty}^{i \infty} \frac{\gamma_2(\theta) e^{-x_2\theta_2}}{\gamma(\theta)} d\theta_2,$$ that proves (\[xxxp\]) for any $\theta_1 \in i{\bf R} \setminus \{0\}$. The proof of is analogous.
Note also that the integral $$\frac{1}{2\pi i} \int_{-i\infty}^{i\infty}
\varphi_2(\theta_1) \gamma_2(\theta_1, \Theta_2^+(\theta_1)) e^{-x_1\theta_1-x_2\Theta_2^+(\theta_1)}
\frac{\mathrm{d} \theta_1}{\sqrt{ d(\theta_1) }}$$ is absolutely convergent. In fact $\sup_{\theta_1 \in i{\bf R}} |\phi_2(\theta_1)| {\leqslant}\nu_2({\bf R}^2_+)$ by definition of $\phi_2$. It is elementary to see that $\sup_{\theta_1 \in i{\bf R}} |\gamma_2(\theta_1, \Theta^+_2(\theta_1)) d^{-1/2} (\theta_1)|<\infty$. Furthermore, for any $\theta_1 \in i{\bf R}$, $\Re \Theta_2^+(\theta_1) =\sigma_{22}^{-1} (-\mu_2 +\Re \sqrt{d(\theta_1)})$, thus for some constant $c>0$ we have $\Re \Theta_2^+(\theta_1) > c |\Im \theta_1|$. Then the integral is absolutely convergent. This concludes the proof of formula (\[I1\]). The proof of (\[I2\]) is completely analogous.
#### **Remark**
These integrals are equal to those on the Riemann surface ${\bf S}$ along properly oriented contours ${{\mathcal}I}_{\theta_1}^+$ and ${{\mathcal}I}_{\theta_2}^+$ respectively. Thanks to the parametrization of Section \[param\] we have $$\label{sss}
\frac{\mathrm{d}\theta_1}{\sqrt{d(\theta_1)}}=\frac{\mathrm{d}\theta_2}{\sqrt{\tilde d(\theta_2)}}=\frac{i \mathrm{d}s}{s\sqrt{\det \Sigma }}.$$ Then we can write for $x=(x_1, x_2) \in{\bf R}_+^2$ the density $\pi(x_1, x_2)$ as a sum of two integrals on ${\bf S}$ : $$I_1+I_2=
\frac{1}{2\pi \sqrt{\det \Sigma}} \int_{\mathcal{I}_{\theta_1}^+}
\frac{\varphi_2(s) \gamma_2(\theta(s))}{s} e^{-\langle \theta(s) \mid x \rangle }
{\mathrm{d}s} +
\frac{1}{2\pi \sqrt{\det \Sigma}} \int_{\mathcal{I}_{\theta_2}^+}
\frac{\varphi_1(s) \gamma_1(\theta(s))}{s} e^{-\langle \theta(s) \mid x \rangle }
\mathrm{d}s.$$
Saddle-point {#subsec:saddle}
------------
Let us fix $\alpha \in ]0, \pi/2[$ and put $(x_1,x_2)=r e_\alpha= r(\cos (\alpha), \sin (\alpha))$ where $\alpha \in ]0, \pi/2[$, Our aim now is to find the asymptotic expansion of $\pi(r\cos (\alpha), r\sin (\alpha))$, that is the one of the sum $$\label{iwi}
I_1+I_2=
\frac{1}{2\pi \sqrt{\det \Sigma}} \int_{\mathcal{I}_{\theta_1}^+}
\frac{\varphi_2(s) \gamma_2(\theta(s))}{s} e^{- r \langle \theta(s) \mid e_\alpha \rangle }
{\mathrm{d}s}
+
\frac{1}{2\pi \sqrt{\det \Sigma}} \int_{\mathcal{I}_{\theta_2}^+}
\frac{\varphi_1(s) \gamma_1(\theta(s))}{s} e^{- r \langle \theta(s) \mid e_\alpha \rangle }
\mathrm{d}s$$ as $r \to \infty$ and to prove that for any $\alpha_0 \in ]0, \pi/2[$ this asymptotic expansion is uniform in a small neighborhood ${{\mathcal}O}(\alpha_0)\in ]0, \pi/2[$.
These integrals are typical to apply the saddle-point method, see [@fedoryuk_asymptotic_1989] or [@pemantle_analytic_2013]. Let us study the function $\langle \theta(s) \mid e_\alpha \rangle $ on ${\bf S}$ and its critical points.
\[sp\]
- For any $\alpha \in ]0, \pi/2[$ function $\langle \theta(s) \mid e_\alpha \rangle $ has two critical points on ${\bf S}$ denoted by $\theta^{+}(\alpha)$ and $\theta^{-}(\alpha)$. Both of them are on ellipse $\mathcal{E}$, $\theta^{+}(\alpha) \in ]s_1^+, s_2^+[$, $\theta^{-}(\alpha) \in ]s_1^-, s_2^-[$. Both of them are non-degenerate.
- The coordinates of $\theta^{+}(\alpha)=(\theta_1^+(\alpha), \theta_2^+(\alpha))$ are given by formulas : $$\begin{aligned}
\theta_1^\pm(\alpha)&=& \frac{\mu_2\sigma_{12}-\mu_1\sigma_{22}}{\det \Sigma}
\pm\frac{1}{\det \Sigma}
\sqrt{\frac{D_1}{1+\frac{\tan(\alpha)^2}{(\sigma_{22}-\tan(\alpha)\sigma_{12})^2} \det\Sigma}
} \nonumber\\
\theta_2^\pm(\alpha)&=& \frac{\mu_1\sigma_{12}-\mu_2\sigma_{11}}{\det \Sigma}
\pm\frac{1}{\det \Sigma}
\sqrt{\frac{D_2}{1+\frac{\tan(\alpha)^2}{(\sigma_{11}-\tan(\alpha)\sigma_{12})^2} \det\Sigma}
} \label{cdpt}\end{aligned}$$ where notations $D_1 =(\mu_2\sigma_{12}-\mu_1\sigma_{22})^2 +\mu_2^2 \det{\Sigma}$ and $D_2 =(\mu_1\sigma_{12}-\mu_2\sigma_{11})^2 +\mu_1^2 \det{\Sigma}$ are used. With the parametrization of Section \[param\] the corresponding points on $\bf S$ are such that: $$s_\pm(\alpha)^2=\frac{\cos \alpha (\theta_1^+ -\theta_1^-)+\sin \alpha (\theta_2^+ -\theta_2^-) e^{i\beta}}{\cos \alpha (\theta_1^+ -\theta_1^-)+\sin \alpha (\theta_2^+ -\theta_2^-) e^{-i\beta}}.$$
- Function $\theta^{+}(\alpha)$ is an isomorphism between $[0, \pi/2]$ and ${{\mathcal}A}= [s^{1,+}, s^{2,+}]$, $\lim_{\alpha \to 0} \theta^{+}(\alpha)=s_1^+$, $\lim_{\alpha \to \pi/2} \theta^{+}(\alpha)=s_2^+$. Function $\theta^{-}(\alpha)$ is an isomorphism between $[0, \pi/2]$ and $ [s^{1,-}, s^{2,-}]$, $\lim_{\alpha \to 0} \theta^{-}(\alpha)=s_1^-$, $\lim_{\alpha \to \pi/2} \theta^{-}(\alpha)=s_2^-$.
- Function $\langle \theta(s) \mid e_\alpha \rangle$ is strictly [*increasing* ]{} on the arc $[\theta^{-}(\alpha), \theta^{+}(\alpha)]$ of $\mathcal{E}$ and strictly decreasing on the arc $[\theta^{+}(\alpha), \theta^{-}(\alpha)]$. Namely, $\theta^{+}(\alpha)$ is its maximum on ${\mathcal}{ E}$ and $\theta^{-}(\alpha)$ is its minimum: $$\theta^{+}(\alpha)= {\rm argmax}_{s \in {{\mathcal}E}} \langle \theta(s) \mid e_\alpha \rangle \ \ \
\theta^{-} (\alpha)= {\rm argmin}_{s \in {{\mathcal}E}} \langle \theta(s) \mid e_\alpha \rangle.$$
Let us look for critical points with coordinates $(\theta_1, \theta_2)$ of $\langle \theta(s) \mid e_\alpha \rangle $ on ${\bf S}$. Equation $ (\theta_1 \cos (\alpha) + \theta_2(\theta_1) \sin (\alpha) )'_{\theta_1}=0$ implies $\tan(\alpha) \frac{d \theta_2}{d \theta_1}=-1$. Substituting it into equation $\gamma(\theta_1, \theta_2(\theta_1))'_{\theta_1} \equiv 0$ and writing also $\gamma(\theta_1,\theta_2) \equiv 0$ we get the system of two equations $$\left\{
\begin{array}{l}
-\sigma_{11}\theta_1 \tan(\alpha) +\sigma_{22}\theta_2 +\sigma_{12}\theta_1 -\sigma_{12}\theta_2\tan(\alpha)
-\mu_1\tan(\alpha) +\mu_2 =0\\
\sigma_{11}\theta_1^2+\sigma_{22} \theta_2^2 +2\sigma_{12}\theta_1 \theta_2 +\mu_1 \theta_1 +\mu_2 \theta_2=0
\end{array}
\right.$$ from where we compute $\theta^{-}(\alpha)=(\theta^{-}_1(\alpha), \theta^{-}_2(\alpha))$ and $\theta^{+}(\alpha)=(\theta^{+}_1(\alpha), \theta^{+}_2(\alpha))$ explicitly as announced in (\[cdpt\]). We check directly that $\frac{d^2 \theta_2}{d \theta_1}\ne 0$ at these points, so they are non-degenerate critical points. It is also easy to see from (\[cdpt\]) that $\theta^{-}_1(\alpha)$ is strictly increasing from branch point $\theta_1^-$ to $\theta_1(\theta_2^-)$ and that $\theta^{+}_1(\alpha)$ is strictly decreasing from branch point $\theta_1^+$ to $\theta_1(\theta_2^+)$ when $\alpha$ runs the segment $[0, \pi/2]$. In the same way $\theta^{-}_2(\alpha)$ is strictly decreasing from $\theta_2(\theta_1^-)$ to $\theta_2^-$ and $\theta^{+}_2(\alpha)$ is strictly increasing from $\theta_2(\theta_1^+)$ to $\theta_2^+$ when $\alpha$ runs the segment $[0, \pi/2]$. This proves assertions (i)–(iii).
Finally, since there are no critical points on ${\mathcal}{ E}$ except for $\theta^+(\alpha)$ and $\theta^-(\alpha)$, function $\langle \theta(s) \mid e_\alpha \rangle$ is monotonous on the arcs $[\theta^{-}(\alpha), \theta^{+}(\alpha)]$ and $[\theta^{+}(\alpha), \theta^{-}(\alpha)]$. In view of the inequality $\langle \theta^+(\alpha) \mid e_\alpha \rangle >
\langle \theta^-(\alpha) \mid e_\alpha \rangle$, assertion (iv) follows.
[**Notation of the saddle-point.**]{} From now one we are interested in point $\theta^{+}(\alpha)$ that we denote by $\theta(\alpha)$ for shortness.
[**The steepest-descent contour $\gamma_\alpha$.**]{} The level curves $\{s : \Re \langle \theta(s) \mid e_\alpha\rangle= \langle \theta(\alpha) \mid e_\alpha\rangle \}$ are orthogonal at $\theta(\alpha)$ and subdivide its neighborhood into four sections. The curves of steepest descent $\{s : \Im \langle \theta(s) \mid e_\alpha\rangle=0 \}$ on ${\bf S}$ are orthogonal at $\theta(\alpha)$ as well, see Lemma 1.3, Chapter IV in [@fedoryuk_saddle-point_1977]. One of them coincides with ${\mathcal}{E}$. We denote the other one by $\gamma_\alpha$. The real part $\Re \langle \theta(s) \mid \alpha\rangle$ is strictly increasing on $\gamma_\alpha$ as $s$ goes far away from $\theta(\alpha)$, see [@fedoryuk_asymptotic_1989 Section 4.2]. The level curves of functions $\Re \langle \theta(s) \mid e_\alpha\rangle$ and $\Im \langle \theta(s) \mid e_\alpha\rangle$ are pictured in Figure \[ligneniveau\].
![Level sets of $\Im \langle \theta(s) \mid e_\alpha \rangle $ and $\Re \langle \theta(s) \mid e_\alpha \rangle$[]{data-label="ligneniveau"}](niveauimaginaire.pdf "fig:")![Level sets of $\Im \langle \theta(s) \mid e_\alpha \rangle $ and $\Re \langle \theta(s) \mid e_\alpha \rangle$[]{data-label="ligneniveau"}](niveau.pdf "fig:")
Let $z_{\alpha,+}= (\theta_1( z_{\alpha,+} ), \theta_2(z_{\alpha,+} )) $ and $z_{\alpha,-}=(\theta_1(
z_{\alpha,-} ), \theta_2 (z_{\alpha,-} )) $ be the end points of $\gamma_\alpha$ where $\Im \theta_1(
z_{\alpha,-} )>0$ and $\Im \theta_1(
z_{\alpha,-} )<0$. We can fix end points $z_{\alpha , -}$ and $z_{\alpha , +}$ in such a way that $\forall \alpha \in \mathcal{O}(\alpha_0)$ and some small $\epsilon>0$ $$\Re \langle z_{\alpha,\pm} \mid e_\alpha \rangle = \langle \theta(\alpha) \mid e_\alpha \rangle +\epsilon.$$ For technical reasons we choose $\epsilon$ small enough such that $\Re \theta_1(z_{\alpha,\pm}) \in ]\theta_1^{-}, \theta_1^+[$ and $\Re \theta_2(z_{\alpha,\pm}) \in ]\theta_2^{-}, \theta_2^+[$.
Shifting the integration contours {#subsec:shiftingcontours}
---------------------------------
Our aim now is to shift the integration contours $\mathcal{I}_{\theta_1}^+$ and $\mathcal{I}_{\theta_2}^+$ in (\[iwi\]) up to new contours $\Gamma_{\theta_1,\alpha}$ and $\Gamma_{\theta_2,\alpha}$ respectively which coincide with $\gamma_\alpha$ in a neighborhood of $\theta(\alpha)$ on ${\bf S}$ and are “higher” than $\theta(\alpha)$ in the sense of level curves of the function $\Re \langle \theta(s) \mid e_\alpha \rangle$, that is $\Re \langle \theta(s) \mid e_\alpha \rangle > \Re \langle \theta(\alpha) \mid e_\alpha \rangle+\epsilon$ for any $s \in \Gamma_{\theta_i,\alpha}\setminus \gamma_{\alpha}$ with $i=1,2$. When shifting the contours we should of course take into account the poles of the integrands and the residues at them.
Let us construct $\Gamma_{\theta_1,\alpha}$ and $\Gamma_{\theta_2,\alpha}$. We set $$\Gamma_{\theta_1, \alpha}^{1,+}=\{s : \Re \theta_1(s)= \Re \theta_1(z_{\alpha,+}), \Im \theta_1(z_{\alpha,+} ) {\leqslant}\Im \theta_1(s) {\leqslant}V(\alpha) \}$$ where $V(\alpha)>0$ will be defined later. Then the end points of $\Gamma_{\theta_1, \alpha}^{1,+}$ are $z_{\alpha,+}$ and $Z_{\alpha,+}$ where $\Re \theta_1(z_{\alpha,+})= \Re \theta_1(Z_{\alpha,+})$, $\Im \theta_1(Z_{\alpha,+})= V(\alpha)$. Next $$\Gamma_{\theta_1,\alpha}^{2,+}=\{s : \Im \theta_1(s)=V(\alpha), 0{\leqslant}\Re \theta_1(s) {\leqslant}\Re \theta_1(z_{\alpha,+})\}$$ if $ \Re \theta_1(z_{\alpha,+}) >0$ and $$\Gamma_{\theta_1,\alpha}^{2,+}=\{s : \Im \theta_1(s)=V(\alpha), 0{\geqslant}\Re \theta_1(s) {\geqslant}\Re \theta_1(z_{\alpha,+})\}$$ if $ \Re \theta_1(z_{\alpha,+}) <0$. This contour goes from $Z_{\alpha,+}$ up to $Z^0_{\alpha, +}$ on ${\mathcal I}_{\theta_1}$ with $\Re (\theta_1(s))= 0$, $\Im (\theta_1(s))=V(\alpha)$. Finally $\Gamma_{\theta_1, \alpha}^{3,+}$ coincides with $\mathcal{I}_{\theta_1}^+$ from $Z_{\alpha, +}^0$ up to infinity : $$\Gamma_{\theta_1, \alpha}^{3,+} = \{s : \Re \theta_1(s)=0, \Im \theta_1(s){\geqslant}V(\alpha)\}.$$ We define in the same way $\Gamma_{\theta_1, \alpha}^{1,-}=\{s : \Re \theta_1(s)= \Re \theta_1(z_{\alpha,-}), -V(\alpha) {\leqslant}\Im s {\leqslant}\Im \theta_1(z_{\alpha,-} ) \}$. The end points of $\Gamma_{\theta_1, \alpha}^{1,-}$ are $z_{\alpha,-}$ and $Z_{\alpha,-}$ where $\Re \theta_1(z_{\alpha,-})= \Re \theta_1(Z_{\alpha,-})$, $\Im \theta_1(Z_{\alpha,-})= -V(\alpha)$. Next $\Gamma_{\theta_2,\alpha}^{2,-}=\{s : \Im \theta_1(s)=-V(\alpha), 0{\leqslant}\Re \theta_1(s) {\leqslant}\Re \theta_1(z_{\alpha,-})\}$ or $\Gamma_{\theta_2,\alpha}^{2-}=\{s : \Im \theta_1(s)=-V(\alpha), 0{\geqslant}\Re \theta_1(s) {\geqslant}\Re \theta_1(z_{\alpha,-})\}$ according to the sign of $\Re \theta_1(z_{\alpha, -})$. It goes from $Z_{\alpha,-}$ to $Z^0_{\alpha, -}$ on ${\bf S}_{\theta_1}$ with $\Re (\theta_1( Z^0_{\alpha, -} ))= 0$, $\Im (\theta_1(Z^0_{\alpha, -} ))= - V(\alpha)$. Finally $\Gamma_{\theta_1, \alpha}^{3,+}$ coincides with ${\mathcal I}_{\theta_1}^+$ from $Z_{\alpha, -}^0$ up to infinity. Then contour $\Gamma_{\theta_1,\alpha}= \Gamma_{\theta_1,\alpha}^{3,-}\cup \Gamma_{\theta_1,\alpha}^{2,-} \cup \Gamma_{\theta_1,\alpha}^{1,-}\cup \gamma_\alpha \cup \Gamma_{\theta_1,\alpha}^{1,+} \cup \Gamma_{\theta_1,\alpha}^{2,+} \cup
\Gamma_{\theta_1,\alpha}^{3,+} \subset {\bf S}_{\theta_1}^1$. One can visualize this contour on Figure \[contint\] : in the left picture it is drawn on parametrized ${\bf S}$, in the right picture it is projected on the complex plane ${\bf C}_{\theta_1}$ .
![Contour $\Gamma_{\theta_1, \alpha}$ on parametrized ${\bf S}$ and projected on ${\bf C}_{\theta_1}$.[]{data-label="contint"}](cheminirecouvr.pdf "fig:") ![Contour $\Gamma_{\theta_1, \alpha}$ on parametrized ${\bf S}$ and projected on ${\bf C}_{\theta_1}$.[]{data-label="contint"}](chemini.pdf "fig:")
The contour $\Gamma_{\theta_2, \alpha}$ is constructed analogously with respect to $\theta_2$-coordinate, $\Gamma_{\theta_2,\alpha}= \Gamma_{\theta_2,\alpha}^{3,-}\cup \Gamma_{\theta_2,\alpha}^{2,-} \cup \Gamma_{\theta_2,\alpha}^{1,-}\cup \gamma_\alpha \cup \Gamma_{\theta_2,\alpha}^{1,+} \cup \Gamma_{\theta_2,\alpha}^{2,+} \cup
\Gamma_{\theta_2,\alpha}^{3,+} \subset {\bf S}_{\theta_2}^1$. The curve of steepest descent $\gamma_\alpha$ is common for $\Gamma_{\theta_1, \alpha}$ and $\Gamma_{\theta_2, \alpha}$.
Let us recall that poles of $\phi_1(s)$ and $\phi_2(s)$ on ${\bf S}$ may occur only at ${\mathcal}{E}$. Let us also recall the convention that an arc $\}a, b\{$ on ${\mathcal}{E}$ is the one with ends $a$ and $b$ which does not include $s_0=(0,0)$.
[**Notation of the sets of poles $\mathcal{P}'_\alpha$ and $\mathcal{P}''_\alpha$.**]{} Let ${{\mathcal}P'}_\alpha$ be the set of poles of the first order of the function $\phi_2(\theta_1(s))$ on the arc $\}\theta(\alpha), s'_0\{$. Let ${{\mathcal}P''}_\alpha$ be the set of poles of the first order of the function $\phi_1(\theta_2(s))$ on the arc $\}\theta(\alpha), s''_0\{$.
Then the following lemma holds true.
\[ppmm\] Let $\alpha_0 \in ]0, \pi/2[$ be such that $\theta(\alpha_0)$ is not a pole of $\phi_1(\theta_2(s))$ neither of $\phi_2(\theta_1(s))$. If ${{\mathcal}P'}_\alpha \cup {{\mathcal}P''}_\alpha $ is not empty, then for any $\alpha \in \mathcal{O}(\alpha_0)$ $$\begin{aligned}
\lefteqn{ \pi (r e_\alpha) =
\sum_{p \in {{\mathcal}P'}_\alpha} {\rm res}_p \phi_2(\theta_1(s))
\frac{\gamma_2(p)}{\sqrt{d(\theta_1(p))}} e^{-r \langle \theta(p) \mid e_\alpha \rangle } +
\sum_{p \in {{\mathcal}P''}_\alpha} {\rm res}_p \phi_1(\theta_2(s))
\frac{\gamma_1(p)}{\sqrt{\tilde d(\theta_2(p))}} e^{-r \langle \theta(p) \mid e_\alpha \rangle } }\nonumber \\
&&{} + \frac{1}{2\pi \sqrt{\det \Sigma}}
\big( \int_{\Gamma_{\theta_1, \alpha} }
\frac{\varphi_2(s) \gamma_2(\theta(s))}{s} e^{- r \langle \theta(s) \mid e_\alpha \rangle }
{\mathrm{d}s}
+
\int_{\Gamma_{\theta_2, \alpha} }
\frac{\varphi_1(s) \gamma_1(\theta(s))}{s} e^{- r \langle \theta(s) \mid e_\alpha \rangle }
\big)
\mathrm{d}s.\label{prolongam}\end{aligned}$$ If $\mathcal{ P'}_\alpha \cup
\mathcal{ P''}_\alpha$ is empty, representation (\[prolongam\]) stays valid where the corresponding sums over $ p \in {{\mathcal}P'}_\alpha$ and $p \in {{\mathcal}P''}_\alpha$ are omitted.
It follows from the assumption of the lemma that $\theta(\alpha)$ is not a pole of $\phi_1(\theta_2(s))$ neither of $\phi_2(\theta_1(s))$ for any $\alpha$ in a small enough neighborhood $\mathcal{O}(\alpha_0)$. Then we use the representation of the density (\[iwi\]) and apply Cauchy theorem shifting the contours to $\Gamma_{\theta_1, \alpha}$ and $\Gamma_{\theta_2, \alpha}$.
In order to find the asymptotic expansion of the density $\pi(r\cos(\alpha), r \sin(\alpha)) $, we have to evaluate now the contribution of the residues at poles in (\[prolongam\]) and the one of integrals along shifted contours $\Gamma_{\theta_1,\alpha}$ and $\Gamma_{\theta_2, \alpha}$. This is the subject of the next two sections.
Asymptotics of integrals along shifted contours Gtheta1 and Gtheta2 {#subsec:asymptintalongshift}
-------------------------------------------------------------------
To finish the construction of $\Gamma_{\theta_1,\alpha}$ and $\Gamma_{\theta_2, \alpha}$, it remains to specify $V(\alpha)$. For that purpose we consider closer the function $$f_\alpha(s)=\langle \theta(s) \mid e_\alpha \rangle = \theta_1(s) \cos \alpha + \theta_2(s) \sin \alpha.$$ Let us define the projection of this function on ${\bf C}_{\theta_1}$ : $$f_\alpha(\theta_1)= \theta_1 \cos \alpha +\Theta_2^{+}(\theta_1) \sin \alpha , \ \ \theta_1 \in {\bf C}_{\theta_1}.$$ Clearly $f_\alpha (s)= f_\alpha (\theta_1(s))= \langle \theta(s) \mid e_\alpha \rangle $ on ${\bf S}_{\theta_1}^1$.
\[uv\]
- For any fixed $u \in [\theta_1^-,\theta_1^+]$ the function $ v \to \Re (f_\alpha (u+iv))$ is increasing on $[0, \infty[$ and decreasing on $]-\infty, 0]$.
- There exist constants $d_1{\leqslant}0$, $d_2>0$ and $V>0$ such that: $$\label{tra}
\inf_{u\in [\theta_1^-,\theta_1^+]} \Re (f_\alpha (u+iv)) {\geqslant}d_1+d_2 \sin (\alpha) |v| \ \ \forall v{\geqslant}V \ \hbox{and } \forall v{\leqslant}-V, \ \ \forall \alpha \in ]0, \pi/2[.$$
We compute : $$\Re (f_\alpha (u+iv))=\cos(\alpha) u+\frac{\sin(\alpha)}{\sigma_{22}} (-\sigma_{12} u-\mu_2 +\Re \sqrt{d(u+iv)})$$ with the discriminant $d(u+iv)=(\det \Sigma) (u+iv-\theta_1^-)(\theta_1^+ -u-iv)$. Then $$\Re \sqrt{d(u+iv)}=\sqrt{\det\Sigma}\sqrt{|(u+iv-\theta_1^-)(\theta_1^+ -u-iv)|} \cos (\frac{\omega_-(u+iv) +\omega_+(u+iv)}{2})$$ where $\omega_-(u+iv)$ et $\omega_+(u+iv)$ are defined as $\omega_-(u+iv)=\arg (\theta_1^+ -u-iv)$ and $\omega_+(u+iv)=\arg (u+iv-\theta_1^-)$, see Figure \[angleomega\]. We have $$\cos (\frac{\omega_-(u+iv) +\omega_+(u+iv)}{2})=
\sqrt{ \frac{1}{2} \cos ( \omega_-(u+iv) +\omega_+(u+iv) ) +\frac{1}{2} }$$ $$\begin{aligned}
&=\sqrt{ \frac{1}{2} \cos ( \omega_-(u+iv) ) \cos( \omega_+(u+iv) ) - \frac{1}{2} \sin ( \omega_-(u+iv) ) \sin ( \omega_+(u+iv) ) +
\frac{1}{2} }
\\
&=\sqrt{ \frac{(u-\theta_1^-)(\theta_1^+-u)- v (-v)}{ 2 |(u+iv-\theta_1^-)(\theta_1^+ -u-iv)|} +\frac{1}{2} }.\end{aligned}$$ Thus $$\begin{gathered}
\Re (f_\alpha (u+iv))=\cos(\alpha) u
\\ +\frac{\sin(\alpha)}{\sigma_{22}} \Big(-\sigma_{12} u-\mu_2
+\sqrt{ \frac{1}{2} } \sqrt{ (u-\theta_1^-)(\theta_1^+-u)+v^2 + |(u+iv-\theta_1^-)(\theta_1^+ -u-iv)|}\Big)
\end{gathered}$$ Both statements of the lemma follow directly from this representation.
![$\omega_- (u+iv)$ et $\omega_+ (u+iv)$[]{data-label="angleomega"}](angles.pdf)
We may now choose $V(\alpha)$ and such that $$\label{ddddd}
V(\alpha) = \max \Big( V, \ \frac{\langle \theta(\alpha) \mid e_\alpha \rangle +\epsilon-d_1}{d_2 \sin (\alpha)}\Big)$$ in accordance with notations of Lemma \[uv\]. This concludes the construction of $\Gamma_{\theta_1,\alpha}$ and $\Gamma_{\theta_2, \alpha}$.
The asymptotic expansion of integrals along these contours is given in the following lemma. The main contribution comes from the integrals along $\gamma_\alpha$, while all other parts of integrals are proved to be exponentially negligible by construction.
\[saddleas\] Let $\alpha_0 \in ]0, \pi/2[$ and $\mathcal{O}(\alpha_0)$ a small enough neighborhood of $\alpha_0$. Then when $r \to \infty$ uniformly for $\alpha \in \mathcal{O}(\alpha_0)$ we have $$\label{assh1}
\frac{1}{2\pi \sqrt{\det \Sigma}} \int_{\Gamma_{\theta_1, \alpha} }
\frac{\varphi_2(s) \gamma_2(\theta(s))}{s} e^{- r \langle \theta(s) \mid e_\alpha \rangle }
{\mathrm{d}s} \sim \sum_{l=0}^{k} \frac{c^l_{\theta_1} (\alpha)}{r^l\sqrt{ r}}
e^{-r \langle \theta(\alpha) \mid e_\alpha \rangle },$$ $$\label{assh2}
\frac{1}{2\pi \sqrt{\det \Sigma}} \int_{\Gamma_{\theta_2, \alpha} }
\frac{\varphi_1(s) \gamma_1(\theta(s))}{s} e^{- r \langle \theta(s) \mid e_\alpha \rangle }
\mathrm{d}s \sim
\sum_{l=0}^{k} \frac{c^l_{\theta_2} (\alpha)}{r^l\sqrt{ r}}
e^{-r \langle \theta(\alpha) \mid e_\alpha \rangle }.$$ The constants $c^l_{\theta_1}(\alpha)$, $c^l_{\theta_2}(\alpha)$, $l=0,1,2, \ldots$ depend continuously of $\alpha$ and can be made explicit in terms of functions $\phi_1$ and $\phi_2$ and their derivatives at $\theta(\alpha)$. Namely $$\begin{aligned}
c^0_{\theta_1}(\alpha) &=& \frac{1}{\sqrt{2\pi \det\Sigma}} \frac{\varphi_2(s(\alpha))\gamma_2(\theta(\alpha))}{s(\alpha)\sqrt{ f_\alpha ''(s(\alpha))}}, \nonumber\\
c^0_{\theta_2}(\alpha)&=&
\frac{1}{\sqrt{2\pi \det\Sigma}} \frac{\varphi_1(s(\alpha))\gamma_1(\theta(\alpha))}{s(\alpha)\sqrt{ f_\alpha ''(s(\alpha))}}. \nonumber \end{aligned}$$
By Lemma \[uv\] (i) and by (\[sss\]) for any $r>0$. $$\label{z}
\Big|\int_{\Gamma_{\theta_1,\alpha}^{1,\pm} }
\frac{ \phi_2(\theta_1(s)) \gamma_2(s)}{ s\sqrt{\det \Sigma}} \exp^{-r\langle \theta(s) \mid e_\alpha \rangle }{\rm d}\,s \Big| {\leqslant}2 V(\alpha) \sup_{s \in \Gamma_{\theta_1,\alpha}^{1,\pm} } \Big|\frac{ \phi_2(\theta_1(s)) \gamma_2(s) }{\sqrt{d(\theta_1(s))} }\Big| e^{-r \langle \theta(\alpha) \mid e_\alpha \rangle -r\epsilon }.$$
The length of $\Gamma_{\theta_2,\alpha}^{\pm}$ being smaller than $(\theta_1^+-\theta_1^-)$, by Lemma \[uv\] (ii) and by (\[sss\]) for any $r>0$ $$\label{zz}
\Big|\int_{\Gamma_{\theta_1,\alpha}^{2,\pm} }
\frac{ \phi_2(\theta_1(s)) \gamma_2(s) }{ s \sqrt{\det\Sigma} } \exp^{-r\langle \theta(s) \mid e_\alpha \rangle }{\rm d}\, s \Big| {\leqslant}(\theta_1^+ -\theta_1^-) \sup_{s \in \Gamma_{\theta_2,\alpha}^{2,\pm} } \Big|\frac{ \phi_2(\theta_1(s)) \gamma_2(s) }{\sqrt{d(\theta_1(s))} } \Big| e^{-r(d_1 +d_2 \sin (\alpha) V(\alpha))
}$$ where due to the choice of $V(\alpha)$ $$\label{zzz}
e^{-r (d_1+d_2 \sin(\alpha) V(\alpha)) } {\leqslant}e^{-r \langle \theta(\alpha) \mid e_\alpha \rangle -r\epsilon }.$$ Finally note that for any $s \in \Gamma_{\theta_1,\alpha}^{3,\pm}$ $$\frac{\gamma_2(s)}{\sqrt{d(\theta_1(s) )} } =
r_{1,2} \frac{ \theta_1(s)}{ \sqrt{d(\theta_1(s) )} } + r_{22} \frac{-b(\theta_1(s)) + \sqrt{d(\theta_1(s))} }{ 2 a(\theta_1(s)) \sqrt{d(\theta_1(s) )} }$$ where $\Re \theta_1(s)=0$, $\Im \theta_1(s) {\geqslant}V$. Then there exists a constant $D>0$ such that $| \gamma_2(s) d^{-1/2}(\theta_1(s))|{\leqslant}D$ for any $s \in \Gamma_{\theta_1,\alpha}^{3,\pm}$ and any $\alpha \in ]0, \pi/2[$. Moreover $|\phi_2(\theta_1(s))|{\leqslant}\nu_1({\bf R_+}) $ for any $s \in {\mathcal I}_{\theta_1}$. Thus by Lemma \[uv\] (ii) and by (\[sss\]) $$\label{zzzz}
\Big|\int_{\Gamma_{\theta_1,\alpha}^{3,\pm} }
\frac{ \phi_2(\theta_1(s)) \gamma_2(s) }{s \sqrt{ \det \Sigma } } \exp^{-r\langle \theta(s) \mid e_\alpha \rangle } {\rm d}\,s \Big| {\leqslant}2 D \nu_1({\bf R_+}) \int\limits_{V(\alpha)}^\infty e^{- r (d_1+d_2 \sin (\alpha) v) }dv $$
$$ {\leqslant}2 D \nu_1({\bf R_+}) \frac{1}{ c \sin (\alpha) V(\alpha)}
e^{-r(d_1+d_2 \sin (\alpha) V(\alpha) ) }
{\leqslant}2 D \nu_1({\bf R_+}) \frac{1}{ c \sin (\alpha) V(\alpha)} e^{-r \langle \theta(\alpha) \mid e_\alpha \rangle -r\epsilon }.$$
The contours $\Gamma_{\theta_1, \alpha}^{i,\pm}$ for $i=1,2$ being far away from poles of $\phi_2$ and zeros of $d(\theta_1(s))$ for all $\alpha \in {{\mathcal}O}(\alpha_0)$, $\sup_{\alpha \in {{\mathcal}O}(\alpha_0) } \sup_{s \in \Gamma_{\theta_1,\alpha}^{i,\pm} } \Big|\frac{ \phi_2(\theta_1(s)) \gamma_2(s) }{\sqrt{d(\theta_1(s))} }\Big|<\infty$ for $i=1,2$, and of course $ \sup_{\alpha \in {{\mathcal}O}(\alpha_0) } (\sin(\alpha) V (\alpha))^{-1}$ and $\sup _{\alpha \in {{\mathcal}O}(\alpha_0)} V(\alpha)$ are finite as well. It follows that for some constant $C>0$ , any $r>0$ and any $\alpha \in {{\mathcal}O}(\alpha_0)$ $$\label{bbbbb}
\Big|\int_{\Gamma_{\theta_1,\alpha}^{1,\pm} \cup \Gamma_{\theta_1,\alpha}^{2,\pm} \cup \Gamma_{\theta_1,\alpha}^{3,\pm} }
\frac{ \phi_2(\theta_1(s)) \gamma_2(s)}{s \sqrt{\det \Sigma } } e^{-r\langle \theta(s) \mid e_\alpha \rangle }{\rm d}\, s \Big|
{\leqslant}C e^{-r \langle \theta(\alpha) \mid e_\alpha \rangle -r\epsilon }.$$ As for the contour $\gamma_\alpha$ of the steepest descent of the function $\langle \theta (s) \mid e_\alpha \rangle$, we apply the standard saddle-point method, see e.g. Theorem 1.7, Chapter IV in [@fedoryuk_saddle-point_1977]: for any $k>0$ when $r \to \infty$, uniformly $\forall \alpha \in {\mathcal O}(\alpha_0)$, $$\label{sspp}
\frac{1}{2\pi\sqrt{\det \Sigma}}\int_{\gamma_\alpha} \frac{\phi_2(s) \gamma_2(\theta(s))}{s} e^{-r\langle \theta(s) \mid e_\alpha \rangle }{\rm d}\,s
\sim \sum_{l=0}^{k} \frac{c^l_{\theta_1}(\alpha) }{r^l\sqrt{ r}} e^{-r \langle \theta(\alpha) \mid e_\alpha \rangle },$$ where $c^0_{\theta_1}(\alpha)$ is given explicitly in the statement of the lemma and all other constants $c^l_{\theta_1}(\alpha)$ can be written in terms of the same functions and their derivatives at $\theta(\alpha)$. Thus (\[assh1\]) is proved and the proof of (\[assh2\]) for the integral over $\Gamma_{\theta_2, \alpha}$ is absolutely analogous.
Contribution of poles into the asymptotics of pi r alpha {#subsec:contributionpole}
--------------------------------------------------------
Once Lemma \[saddleas\] established the asymptotics of integrals along shifted contours $\Gamma_{\theta_1, \alpha}$ and $\Gamma_{\theta_2, \alpha}$, let us come back to Lemma \[ppmm\] and evaluate the contribution to the density of residues at poles over $\mathcal{P}'_\alpha \cup \mathcal{P}''_\alpha$. There are two possibilities:
- $\mathcal{P}'_\alpha \cup \mathcal{P}''_\alpha$ is empty, then the asymptotics of the density is determined by the saddle-point via Lemma \[saddleas\].
- $\mathcal{P}'_\alpha \cup \mathcal{P}''_\alpha$ is not empty. Then due to monotonicity of the function $\langle \theta(s) \mid e_\alpha \rangle $ on $\mathcal{E}$, see Lemma \[sp\] (iv), for any $p \in \mathcal{P}'_\alpha \cup \mathcal{P}''_\alpha$ we have $\langle \theta(p) \mid e_\alpha \rangle < \langle \theta(\alpha) \mid e_\alpha \rangle$. Hence [*all*]{} residues at poles $p \in \mathcal{P}'_\alpha \cup \mathcal{P}''_\alpha$ bring more important contribution to the asymptotic expansion as $r \to \infty$ than integrals over $\Gamma_{\theta_1, \alpha}$ and $\Gamma_{\theta_2, \alpha}$.
First of all, we would like to distinguish the set of parameters $(\Sigma, \mu , R)$ under which (i) or (ii) hold true. Secondly, under (ii), we would like to find the most important pole from the asymptotic point of view. Let us look closer at the arc $\{s'_0, \theta(\alpha) \}$. Under parameters such that $\theta_1(s_2^+)<0$ we have $s'_0 \in ]s_1^+, s_2^+[$, see Figure \[fipoles\], the left picture. Then for some $\alpha' \in ]0, \pi/2[$ $\theta(\alpha')=s'_0$. This arc written in square brackets in the anticlockwise direction is $]s'_0, \theta(\alpha)[$ for any $\alpha \in ]\alpha', \pi/2[$ and the function $\langle \theta(s) \mid e_\alpha \rangle$ is increasing when $s$ runs from $s'_0$ to $\theta(\alpha)$. For any $\alpha \in [0, \alpha'[$ this arc is written $]\theta(\alpha), s'_0[$ and the function $\langle \theta(s) \mid e_\alpha \rangle$ is decreasing when $s$ runs from $\theta(\alpha)$ so $s'_0$. Under parameters such that $\theta_1(s_2^+){\geqslant}0$, we have $s'_0 \not\in ]s_1^+, s_2^+[$, see Figure \[fipoles\] the right picture, from where this arc is written $]\theta(\alpha), s'_0[ $ for any $\alpha \in ]0, \pi/2[$. The function $\langle \theta(s) \mid e_\alpha \rangle$ is decreasing when $s$ runs from $\theta(\alpha)$ to $s'_0$.
![The arc $\mathcal{A}= [s_1^+,s_2+]$ if $\theta_1(s_2^+)<0$, $\theta_2(s_1^+)<0$ on the left picture, if $\theta_1(s_2^+)>0$, $\theta_2(s_1^+)>0$ on the right picture[]{data-label="fipoles"}](fipoles1.pdf "fig:") ![The arc $\mathcal{A}= [s_1^+,s_2+]$ if $\theta_1(s_2^+)<0$, $\theta_2(s_1^+)<0$ on the left picture, if $\theta_1(s_2^+)>0$, $\theta_2(s_1^+)>0$ on the right picture[]{data-label="fipoles"}](fipoles2.pdf "fig:")
The important conclusion is that in all cases, the pole $p$ of $\phi_2$ on the arc $\{ s'_0, \theta(\alpha)\}$ with the smallest $\langle \theta(p) \mid e_\alpha \rangle$ is the to $s'_0$. In the same way we can consider the arc $\{s''_0, \theta(\alpha)\}$ and find out, due to monotonicity of the function $\langle \theta(s) \mid e_\alpha \rangle$, that the pole of $\phi_1$ with the smallest $\langle \theta(p) \mid e_\alpha \rangle$ is the to $s''_0$. We know from Lemmas \[lem3\]–\[lem4\] the way that these poles are related to zeros of $\gamma_1$ and $\gamma_2$. Now we summarize this information in the following theorem.
\[thmpoles\]
- Let $\zeta \theta^{**} \not \in \{\theta(\alpha), s'_0\}$, $\eta \theta^* \not\in \{\theta(\alpha), s''_0)\}$. Then ${{\mathcal}P'}_\alpha$ and ${{\mathcal}P''}_\alpha$ are both empty, $\theta(\alpha)$ is not a pole of $\phi_1$ and neither of $\phi_2$.
- Let $\zeta \theta^{**} \in \}\theta(\alpha), s'_0\}$ and $\eta \theta^* \not\in \} \theta(\alpha), s''_0\}$. Then $$\label{ki}
\min_{p \in {{\mathcal}P'}_\alpha \cup {{\mathcal}P''}_\alpha } \langle \theta(p) \mid e_\alpha \rangle = \langle \zeta \theta^{**} \mid e_\alpha \rangle$$ and this minimum over $ {{\mathcal}P'}_\alpha \cup {{\mathcal}P''}_\alpha $ is achieved at the unique element $p=\zeta \theta^{**}$ which is a pole of the first order of $\phi_2$.
- Let $\zeta \theta^{**} \not\in \}\theta(\alpha), s'_0 \}$ and $\eta \theta^* \in \{s''_ 0, \theta(\alpha) \{$. Then $$\label{kii}
\min_{ p \in {{\mathcal}P'}_\alpha \cup {{\mathcal}P''}_\alpha} \langle \theta(p) \mid e_\alpha \rangle = \langle \eta \theta^{*} \mid e_\alpha \rangle$$ and this minimum over $ {{\mathcal}P'}_\alpha \cup {{\mathcal}P''}_\alpha $ is achieved at the unique element $p=\eta \theta^{*}$ which is a pole of the first order of $\phi_1$.
- Let $\zeta \theta^{**} \in \}\theta(\alpha), s'_0 \}$ and $\eta \theta^* \in \{s''_0, \theta(\alpha) \{$.
If $\langle \zeta \theta^{**} \mid e_\alpha \rangle < \langle \eta \theta^{*} \mid e_\alpha \rangle$, then (\[ki\]) is valid. If $\langle \zeta \theta^{*} \mid e_\alpha \rangle > \langle \eta \theta^{**} \mid e_\alpha \rangle$, then (\[kii\]) is valid. In both cases the minimum over ${{\mathcal}P'}_\alpha \cup {{\mathcal}P''}_\alpha$ is achieved at the unique element which is the pole of the first order $p=\zeta \theta^{**}$ of $\phi_2$ or the pole of the first order $p=\eta \theta^*$ of $\phi_1$ respectively.
If $\langle \zeta \theta^{**} \mid e_\alpha \rangle = \langle \eta \theta^{*} \mid e_\alpha \rangle$, then $$\label{kit}
\min_{p \in {{\mathcal}P'}_\alpha \cup {{\mathcal}P''}_\alpha } \langle \theta(p) \mid e_\alpha \rangle = \langle \zeta \theta^{**} \mid e_\alpha \rangle=
\langle \eta \theta^{*} \mid e_\alpha \rangle.$$ This minimum over ${{\mathcal}P'}_\alpha \cup {{\mathcal}P''}_\alpha $ is achieved at exactly two elements $p=\zeta \theta^{**}$ and $p=\eta \theta^*$ which are poles of the first order of $\phi_1$ and $\phi_2$ respectively.
\(a) Let $\theta_1(s_2^+)<0$ and let $\alpha >\alpha'$ defined above. Then $\theta_1(\alpha)<0$ and all points of the arc $\{\theta(\alpha), s'_0\{$ have the first coordinate negative, so that function $\phi_2(\theta_1(s))$ is initially well defined at them and holomorphic. Let now $\theta_1(s_2^+)<0$ and $\alpha \in ]0, \alpha'[$ or $\theta_1(s_2^+){\geqslant}0$. Then $\theta_1(\alpha)>0$ and the arc $\{\theta(\alpha), s'_0\}$ written in the anticlockwise direction is $[\theta(\alpha), s'_0]$. Assume that $\phi_2(\theta_1(s))$ has poles on $[\theta(\alpha), s'_0[$ and $\theta^p$ is the closest to $s'_0$. Then by Lemma \[lem1\] either $\gamma_2(\zeta \theta^p)=0$ or parameters are such that $\theta_2(s_1^+)>0$, $\eta \zeta \theta^p \in ]\eta s_1^+, s_0[$ and $\gamma_1(\eta\zeta \theta^p)=0$. In the first case $\zeta \theta^p=\theta^{**}$ is a zero of $\gamma_2$ different from $s_0$. This implies $\theta^p =\zeta \theta^{**} \in [\theta(\alpha) s'_0[$ which is impossible by assumptions. In the second case $\eta \zeta \theta^p=\theta^*$ is a zero of $\gamma_1$ different from $s_0$. This implies $\zeta \theta^p=\eta \theta^* \in \eta ]\eta s_1^+, s_0[=]s''_0, s_1^+[ \subset ]s''_0, \theta(\alpha)[ =\}\theta(\alpha), s''_0\{$ that contradicts the assumptions as well. Hence $\phi_2(\theta_1(s))$ has no poles on the open arc $\}\theta(\alpha), s'_0\{$ and neither at $\theta(\alpha)$, ${{\mathcal}P'}_\alpha$ is empty, The reasoning for ${{\mathcal}P''}_\alpha$ is the same.
\(b) By stability conditions (\[u\]) and (\[v\]) $\theta^{**}_1>0$, then $\zeta\theta^{**}_1>0$. Thus $\theta_1(\alpha)>0$, in the case $\theta_1(s_2^+)<0$ the angle $\alpha$ must be smaller than $\alpha'$ and the arc $\}\theta(\alpha), s'_0\{$ should be written $]\theta(\alpha), s'_0[$. By Lemma \[lem3\] there exist poles of function $\phi_2(\theta_1(s))$ on this arc and $\zeta \theta^{**}$ is one among them. By Lemma \[lem1\] $\zeta \theta^{**}$ can not be the closest pole to $s'_0$ only if the parameters are such that $\theta_2(s_1^+)>0$ and for some $\theta^p \in ]\theta(\alpha), s'_0[$ such that $\eta \zeta \theta^p \in ]\eta s_1^+, s_0[$ $\gamma_1(\eta\zeta \theta^p)=0$. But then $\eta \zeta \theta^p=\theta^*$ is a zero of $\gamma_1$ different from $s_0$. It follows $\zeta \theta^p=\eta \theta^* \in \eta ]\eta s_1^+, s_0[=]s''_0, s_1^+[ \subset ]s''_0, \theta(\alpha)[= \}\theta(\alpha), s''_0\{$ that is impossible by assumptions. Hence by Lemma \[lem1\] $\zeta \theta^{**}$ is the closest pole to $s'_0$ of $\phi_2(\theta_1(s))$ and it is of the first order. The function $\langle \theta(s) \mid e_\alpha \rangle $ being decreasing on $]\theta(\alpha), s'_0[$ when $s$ runs the arc in the anticlockwise direction, thus $$\label{pp}
\min_{p \in {{\mathcal}P'}_\alpha } \langle \theta(p) \mid e_\alpha \rangle = \langle \zeta \theta^{**} \mid e_\alpha \rangle,$$ and the minimum is achieved on the unique element $\zeta \theta^{**}$.
If ${{\mathcal}P}''_\alpha$ is empty then the statement (b) is proved.
Assume that ${{\mathcal}P}''_\alpha$ is not empty. Then there exist poles of $\phi_1(\theta_2(s))$ on the arc $\}\theta(\alpha), s''_0\{$. Since function $\phi_1(\theta_2(s))$ is initially well defined and holomorphic at all points with the second coordinate negative, then $\theta_2(\alpha)>0$ and the arc is $]s''_0, \theta(\alpha)[$ when written in the anticlockwise direction. Let $\theta^p$ be a pole of $\phi_1(\theta_2(s))$ which is the closest to $s''_0$. Then by Lemma \[lem1\] either $\gamma_1(\eta \theta^p)=0$ or parameters are such that $\theta_1(s_2^+)>0$, $\zeta \eta \theta^p \in ]s_0, \zeta s_2^+[$ and $\gamma_2(\zeta\eta \theta^p)=0$. In the first case $\eta \theta^p=\theta^{*}$ is a zero of $\gamma_1$ different from $s_0$. This implies $\theta^p =\eta \theta^{*} \in ]s''_0, \theta(\alpha)[$ which is impossible by assumptions. In the second case $\zeta \eta \theta^p=\theta^{**}$ where $\eta \theta^p=\zeta \theta^{**} \in \zeta ]s_0, \zeta s_2^+[=]s_2^+, s'_0[ \subset ]\theta(\alpha), s'_0[$. Thus $\theta^p=\eta \zeta \theta^{**}$ is the closest pole to $s''_0$. Hence, the closest pole of the first order coincides with it or is further away from $s''_0$. Since the function $<\theta(s) \mid e_\alpha>$ is increasing on $]s''_0, \theta(\alpha)[$ when $s$ is running from $s''_0$ to $\theta(\alpha)$, we derive $$\min_{p \in {{\mathcal}P''}_\alpha } \langle \theta(p) \mid e_\alpha \rangle {\geqslant}\langle \eta \zeta \theta^{**} \mid e_\alpha \rangle .$$ But by Lemma \[lem2\] $$\theta_1(\eta \zeta \theta^{**})> \theta_1(\zeta \theta^{**}), \ \ \theta_2(\eta \zeta \theta^{**})= \theta_2(\zeta \theta^{**})$$ from where $$\langle \eta \zeta \theta^{**} \mid e_\alpha \rangle > \langle \zeta \theta^{**} \mid e_\alpha \rangle .$$ Thus, whenever ${{\mathcal}P''}_\alpha$ is non empty, $$\min_{p \in {{\mathcal}P''}_\alpha } \langle \theta(p) \mid e_\alpha \rangle > \langle \zeta \theta^{**} \mid e_\alpha \rangle .$$ This inequality combined with (\[pp\]) finishes the proof of (b).
The proof of (c) is symmetric.
\(d) Since $\theta^{*}_2=\eta \theta^{*}_2>0$ and $\theta^{**}_1=\zeta \theta^{**}_1>0$ by stability conditions (\[u\]) and (\[v\]), then $\theta(\alpha)$ has both coordinates positive. The corresponding arcs written in the anticlockwise direction are $]\theta(\alpha), s'_0[\subset ]s_1^+ , s'_0[$ and $]s''_0, \theta(\alpha)[ \subset ]s''_0, s_2^+[$. By Lemma \[lem1\] $\zeta \theta^{**}$ is a pole of $\phi_2(\theta_1(s))$ on the first of these arcs while $\eta \theta^*$ is a pole of $\phi_1(\theta_2(s))$ on the second one. Then one of the statements of Lemma \[lem4\] (i), (ii) or (iii) holds true.
Under the statement (i), taking into account the monotonicity of the function $\langle \theta(s) \mid e_\alpha \rangle$ on the arcs, we derive immediately that $ \min_{p \in {{\mathcal}P'}_\alpha } \langle \theta(p) \mid e_\alpha \rangle = \langle \zeta \theta^{**} \mid e_\alpha \rangle$, and this minimum is achieved on the unique element $p=\zeta \theta^{**}$ . We derive also that $\min_{p \in {{\mathcal}P''}_\alpha } \langle \theta(p) \mid e_\alpha \rangle = \langle \eta \theta^{*} \mid e_\alpha \rangle $ and this minimum is achieved on the unique element $p=\eta \theta^*$. Thus, under the statement (i) of Lemma \[lem4\], the theorem is immediate.
Assume now (ii) of Lemma \[lem4\]. Again by monotonicity of $\langle \theta(s) \mid e_\alpha \rangle$ we deduce $ \min_{p \in {{\mathcal}P'}_\alpha } \langle \theta(p) \mid e_\alpha \rangle = \langle \zeta \theta^{**} \mid e_\alpha \rangle$ where the minimum is achieved at the unique element $\zeta \theta^{**}$. Under (ii) all poles of $\phi_1(\theta_2(s))$ on $]s''_0, \theta(\alpha)[$ are not closer to $s''_0$ than $ \eta\zeta \theta^{**}$ , so that either ${{\mathcal}P''}_\alpha$ is empty or $$\min_{p \in {{\mathcal}P''}_\alpha } \langle \theta(p) \mid e_\alpha \rangle {\geqslant}\langle \eta\zeta \theta^{**} \mid e_\alpha \rangle.$$ By Lemma \[lem2\] $\theta_1( \eta\zeta \theta^{**})>\theta_1(\zeta \theta^{**}),$ $ \theta_2 ( \eta\zeta \theta^{**})= \theta_2(\zeta \theta^{**})$ from where $\langle \eta\zeta \theta^{**} \mid e_\alpha \rangle > \langle \zeta \theta^{**} \mid e_\alpha \rangle $. Hence $$\min_{p \in {{\mathcal}P''}_\alpha } \langle \theta(p) \mid e_\alpha \rangle > \langle \zeta \theta^{**} \mid e_\alpha \rangle,$$ and finally $$\label{ll}
\min_{p \in {{\mathcal}P'}_\alpha \cup {{\mathcal}P''}_\alpha } \langle \theta(p) \mid e_\alpha \rangle =
\langle \zeta \theta^{**} \mid e_\alpha \rangle$$ where the minimum is achieved on the unique element $\zeta \theta^{**}$. From the other hand, the pole $\eta \theta^* \in ]s''_0, \theta(\alpha)[$ of $\phi_1(\theta_2(s))$ in this case is not closer to $s''_0$ than $ \eta\zeta \theta^{**}$. Then the inequality $$\label{mm}
\langle \eta \theta^* \mid e_\alpha \rangle {\geqslant}\langle \eta \zeta \theta^{**} \mid e_\alpha \rangle >
\langle \zeta \theta^{**} \mid e_\alpha \rangle$$ is valid.
Under the statement (iii) of Lemma \[lem4\], by symmetric arguments, $\min_{p \in {{\mathcal}P'}_\alpha \cup {{\mathcal}P''}_\alpha } \langle \theta(p) \mid e_\alpha \rangle =
\langle \eta \theta^{*} \mid e_\alpha \rangle$ where the minimum is achieved on the unique element $\eta \theta^{*}$, while $\langle \eta \theta^* \mid e_\alpha \rangle< \langle \zeta \theta^{**} \mid e_\alpha \rangle$. The concludes the proof of the lemma.
Asymptotic expansion of the density pi r alpha, {#sec:asymptexpansion}
================================================
Given angle alpha0, asymptotic expansion of the density as a function of parameters Sigma mu R {#subsec:givenangle}
----------------------------------------------------------------------------------------------
We are now ready to formulate and prove the results. In this section we fix an angle $\alpha_0 \in ]0, \pi/2[$ and give the asymptotic expansion of the density of stationary distribution depending on parameters $(\Sigma, \mu, R)$, and more precisely on the position of zeros of $\gamma_1$ and $\gamma_2$ on ellipse $\mathcal{E}$.
In the first theorem parameters $(\Sigma, \mu, R)$ are such that the asymptotic expansion is determined by the saddle-point.
\[thmresults1\] Let $\alpha_0 \in ]0, \pi/2[$, $\mathcal{O}(\alpha_0)$ is a small enough neighborhood of $\alpha_0$. Assume that $\zeta \theta^{**} \not \in \{\theta(\alpha_0), s'_0\}$, $\eta \theta^* \not\in \{\theta(\alpha_0), s''_0)\}$. Then there exist constants $c^l(\alpha)$, $l=0,1,2,\ldots $, such that for any $k>0$: $$\label{as1}
\pi(r \cos (\alpha), r \sin (\alpha) ) \sim \sum_{l=0}^k \frac{c^l(\alpha)}{ r^l \sqrt{r}} e^{-r
\langle \theta (\alpha) \mid e_\alpha \rangle },\ \hbox{ as }r \to \infty, \ \hbox{ uniformly for }\alpha \in \mathcal{O}(\alpha_0).$$ Constants $c^l(\alpha)$ $l=0,1,2,\ldots $ depend continuously on $\alpha$ and can be expressed in terms of functions $\phi_1$ and $\phi_2$ and their derivatives at $\theta(\alpha)$. Namely $$c^0(\alpha)= c_{\theta_1}^0(\alpha) + c_{\theta_2}^0(\alpha)$$ where $c_{\theta_1}^0(\alpha)$ and $c_{\theta_2}^0(\alpha)$ are defined in Lemma \[saddleas\].
By Lemma \[sp\] (iii) $\theta(\alpha)$ depends continuously on $\alpha$, then $\zeta \theta^{**} \not \in \{\theta(\alpha), s'_0\}$, $\eta \theta^* \not\in \{\theta(\alpha), s''_0\}$ for all $\alpha \in \mathcal{O}(\alpha_0)$. By Theorem \[thmpoles\] (a) the sets ${{\mathcal}P'}_\alpha$ and ${{\mathcal}P''}_\alpha$ are both empty, furthermore, $\theta(\alpha)$ is not a pole of $\phi_1$ and neither of $\phi_2$. Then by Lemma \[ppmm\] the density equals the sum of integrals along shifted contours $\Gamma_{\theta_1, \alpha}$ and $\Gamma_{\theta_2, \alpha}$ the asymptotics of which is found in Lemma \[saddleas\], $c^l(\alpha)=c^l_{\theta_1}(\alpha)+ c^l_{\theta_2}(\alpha)$, $l=0,1,2,\ldots$.
In the second theorem parameters $(\Sigma, \mu, R)$ are such that the most important terms of the asymptotic expansion come from the poles of $\phi_1$ or $\phi_2$ and the smaller ones come from the saddle-point.
\[thmresults2\] Let $\alpha_0 \in ]0, \pi/2[$, $\mathcal{O}(\alpha_0)$ is a small enough neighborhood of $\alpha_0$. Assume that $\zeta \theta^{**} \in \}\theta(\alpha_0), s'_0\}$ or $\eta \theta^* \in \}\theta(\alpha_0), s''_0\}$. Assume also that $\theta(\alpha_0)$ is a pole of $\phi_1(\theta_2(s))$ neither of $\phi_2(\theta_1(s))$. Then for any $k>0$ when $r \to \infty$, uniformly for $\alpha \in \mathcal{O}(\alpha_0)$ we have $$\begin{aligned}
\pi(r\cos(\alpha), r \sin(\alpha)) & \sim &
\sum_{p \in {{\mathcal}P'}_{\alpha_0}} {\rm res}_p \phi_2(\theta_1(s))
\frac{\gamma_2(p)}{\sqrt{d(\theta_1(p))}} e^{-r \langle \theta(p) \mid e_\alpha \rangle } \nonumber \\
&& { }+
\sum_{p \in {{\mathcal}P''}_{\alpha_0} } {\rm res}_p \phi_1(\theta_2(s))
\frac{\gamma_1(p)}{\sqrt{\tilde d(\theta_2(p))}} e^{-r \langle \theta(p) \mid e_\alpha \rangle } \nonumber \\
&& {} +
\sum_{l=0}^k \frac{c^l(\alpha)}{ r^l \sqrt{r}} e^{-r
\langle \theta (\alpha) \mid e_\alpha \rangle }. \label{th23} \end{aligned}$$ Constants $c^l(\alpha)$ $l=0,1,2,\ldots$ are the same as in Theorem \[thmresults1\]. Furthermore
- If $\zeta \theta^{**} \in \}\theta(\alpha_0), s'_0\}$ and $\eta \theta^* \not\in \}\theta(\alpha_0), s''_0\}$, then the main term in the expansion (\[th23\]) is at $p=\zeta \theta^{**}$.
- If If $\zeta \theta^{**} \not\in \}\theta(\alpha_0), s'_0\}$ and $\eta \theta^* \in \}\theta(\alpha_0), s''_0\}$, then the main term in (\[th23\]) is at $p=\eta \theta^{*}$.
- Let $\zeta \theta^{**} \in \}\theta(\alpha_0), s'_0\}$ and $\eta \theta^* \in \}\theta(\alpha_0), s''_0\}$. If $\langle \zeta \theta^{**} \mid e_{\alpha_0 }\rangle < \langle \eta \theta^{*} \mid e_{\alpha_0 }\rangle $, then the main term in (\[th23\]) is at $p=\zeta \theta^{**}$.
If $\langle \zeta \theta^{**} \mid e_{\alpha_0 }\rangle > \langle \eta \theta^{*} \mid e_{\alpha_0 }\rangle $, then main term in (\[th23\]) is at $p=\eta \theta^{*}$.
If $\langle \zeta \theta^{**} \mid e_{\alpha_0 }\rangle = \langle \eta \theta^{*} \mid e_{\alpha_0 }\rangle $, then two the most important terms in the expansion (\[th23\]) are at $p=\zeta \theta^{**}$ and at $p=\eta \theta^*$.
Point $\theta(\alpha_0)$ being not a pole of $\phi_1$ neither of $\phi_2$, one can choose $\mathcal{O}(\alpha_0)$ small enough such that $\theta(\alpha)$ is not a pole of no one of these functions and $\mathcal{P}'_{\alpha} \cup \mathcal{P}''_\alpha = \mathcal{P}'_{\alpha_0} \cup \mathcal{P}''_{\alpha_0} $ for all $ \alpha \in \mathcal{O}(\alpha_0)$. By assumptions $\zeta \theta^{**} \in \}\theta(\alpha_0), s'_0\}$ or $\eta \theta^* \in \}\theta(\alpha_0), s''_0\}$, then by Theorem \[thmpoles\] (b), (c) or (d) $\mathcal{P}'_{\alpha_0} \cup \mathcal{P}''_{\alpha_0} $ is not empty. Finally by virtue of Lemma \[ppmm\] and Lemma \[saddleas\] the representation (\[th23\]) holds true.
Let us study the main asymptotic term. Statements (i), (ii) and (iii) for $\alpha=\alpha_0$ follow directly from Theorem \[thmpoles\] (b), (c) and (d). They remain valid for any $\alpha \in \mathcal{O}(\alpha_0)$ due to the continuity of the functions $\alpha \to \langle \theta(p) \mid e_\alpha \rangle $ for any $p \in \mathcal{P}'_{\alpha_0}
\cup \mathcal{P}''_{\alpha_0} $.
[**Remark.**]{} Under parameters such that $\zeta \theta^{**} \in \}\theta(\alpha_0), s'_0\}$, $\eta \theta^* \in \}\theta(\alpha_0), s''_0\}$ and $\langle \zeta \theta^{**} \mid e_{\alpha_0 }\rangle = \langle \eta \theta^{*} \mid e_{\alpha_0 }\rangle $ (case (iii)), for any fixed angle $\alpha<\alpha_0$, the main asymptotic term is at $\eta \theta^{*}$ and the second one is at $\zeta \theta^{**}$; for any fixed angle $\alpha>\alpha_0$ the pole $\zeta\theta^{**}$ provides the main asymptotic term and $\eta \theta^*$ gives the second one. If $r\to \infty$ and $\alpha \to \alpha_0$, both of these terms should be taken into account.
In Theorem \[thmresults2\] $\theta(\alpha_0)$ is assumed not to be a pole of $\phi_1$ and neither of $\phi_2$, that is why Lemma \[ppmm\] applies. Nevertheless, it may happen (for a very few angles and under some sets of parameters) that $\theta(\alpha_0)$ is a pole of one of these functions. In this case the following theorem holds true.
\[thmresults3\] Let $\alpha_0 \in ]0, \pi/2[$. Assume that $\zeta \theta^{**} \in \}\theta(\alpha_0), s'_0\}$ or $\eta \theta^* \in \}\theta(\alpha_0), s''_0)\}$.
Assume also that $\theta(\alpha_0)$ is a pole of $\phi_1(\theta_2(s))$ or of $\phi_2(\theta_1(s))$.
Then for any $\delta>0$ there exists a small enough neighborhood $\mathcal{O}(\alpha_0)$ such that $$\begin{aligned}
\pi(r\cos(\alpha), r \sin(\alpha)) & \sim &
\sum_{p \in {{\mathcal}P'}_{\alpha_0}} {\rm res}_p \phi_2(\theta_1(s))
\frac{\gamma_2(p)}{\sqrt{d(\theta_1(p))}} e^{-r \langle \theta(p) \mid e_\alpha \rangle } \nonumber \\
&& { }+
\sum_{p \in {{\mathcal}P''}_{\alpha_0} } {\rm res}_p \phi_1(\theta_2(s))
\frac{\gamma_1(p)}{\sqrt{\tilde d(\theta_2(p))}} e^{-r \langle \theta(p) \mid e_\alpha \rangle } \nonumber \\
&&{} + o(e^{-r (\langle \theta(\alpha) \mid e_\alpha \rangle -\delta ) } ) \ \ \ r \to \infty, \hbox{ uniformly } \forall \alpha \in \mathcal{O}(\alpha_0) \label{th233} \end{aligned}$$
Furthermore, the main term in this expansion is the same as in Theorem \[thmresults2\], cases (i), (ii) and (iii).
For any $\delta>0$ one can choose $\tau' \in \}s'_0, \theta(\alpha_0)\{$ and $\tau'' \in \}s''_0, \theta(\alpha_0)\{$ close enough to $\theta(\alpha_0)$ so that $\mathcal{P}'_{\alpha_0} \subset \}s'_0, \tau'\{$ and $\mathcal{P}''_{\alpha_0} \subset \}s''_0, \tau"\{$. Furthermore $\tau'$ and $\tau''$ can be chosen close enough to $\alpha_0$ so that $ \langle \theta(\alpha_0) \mid e_{\alpha_0} \rangle - \langle \tau' \mid e_{\alpha_0} \rangle <\delta/4$ and $ \langle \theta(\alpha_0) \mid e_{\alpha_0} \rangle - \langle \tau'' \mid e_{\alpha_0} \rangle <\delta/4$. Then by continuity of the functions $\alpha \to \langle \theta(\alpha) \mid e_\alpha \rangle $, $\alpha \to \langle \tau '\mid e_\alpha \rangle $, $\alpha \to \langle \tau'' \mid e_\alpha \rangle $ one can fix a small enough neighborhood $\mathcal{O}(\alpha_0)$ such that $$\label{at}
\langle \theta(\alpha) \mid e_{\alpha} \rangle - \langle \tau' \mid e_{\alpha} \rangle <\delta/2,\ \ \
\langle \theta(\alpha) \mid e_{\alpha} \rangle - \langle \tau'' \mid e_{\alpha} \rangle <\delta/2 ,\ \ \forall \alpha \in \mathcal{O}(\alpha_0).$$ Next, we shift the integration contours in (\[iwi\]) $\mathcal{I}_{\theta_1}^+$ and $\mathcal{I}_{\theta_2}^+$ to the new ones $\Gamma'_{\theta_1, \alpha}$ and $\Gamma''_{\theta_2, \alpha}$ going through $\tau'$ and $\tau''$ respectively that we construct as follows: $\Gamma'_{\theta_1, \alpha} =\Gamma'^{1}_{\theta_1, \alpha} \cup \Gamma'^{2, \pm }_{\theta_1, \alpha}\cup \Gamma'^{3, \pm}_{\theta_1, \alpha}$ where $\Gamma'^{1}_{\theta_1}=\{ s : \Re \theta_1(s)= \Re \theta_1(\tau'), -V(\alpha){\leqslant}\Im \theta_1(s) {\leqslant}V(\alpha) \}$, $\Gamma'^{2,\pm }_{\theta_1, \alpha}= \{s : \Im \theta_1(s)= \pm V(\alpha), 0{\leqslant}\Re \theta_1(s) {\leqslant}\Re \theta_1(\tau')\}
$ if $\Re \theta_1(\tau')>0$ and $\Gamma'^{2, \pm }_{\theta_1, \alpha}= \{s : \Im \theta_1(s)= \pm V(\alpha), 0{\geqslant}\Re \theta_1(s) {\geqslant}\Re \theta_1(\tau')\}$ if $\Re \theta_1(\tau')<0$ , finally $\Gamma'^{3, +}_{\theta_1, \alpha}= \{s : \Re \theta_1(s)=0, \Im \theta_1(s) {\geqslant}V(\alpha)\}$, $\Gamma'^{3, -}_{\theta_1, \alpha}= \{s : \Re \theta_1(s)=0, \Im \theta_1(s) {\leqslant}-V(\alpha)\}$. The construction of $\Gamma''_{\theta_2, \alpha}$ is analogous. The value $V(\alpha)$ is fixed as: $$\label{va}
V(\alpha)=\max\Big(V, \frac{\langle \tau' \mid e_\alpha \rangle -d_1}{d_2 \sin (\alpha)}, \frac{\langle \tau'' \mid e_\alpha \rangle -d_1}{d_2 \sin (\alpha)}
\Big)$$ with notations from Lemma \[uv\]. Thanks to the representation (\[iwi\]) and Cauchy theorem $$\label{lre}
\pi(re_\alpha) =
\sum_{p \in {{\mathcal}P'}_{\alpha_0}} {\rm res}_p \phi_2(\theta_1(s))
\frac{\gamma_2(p)}{\sqrt{d(\theta_1(p))}} e^{-r \langle \theta(p) \mid e_\alpha \rangle } +
\sum_{p \in {{\mathcal}P''}_{\alpha_0}} {\rm res}_p \phi_1(\theta_2(s))
\frac{\gamma_1(p)}{\sqrt{\tilde d(\theta_2(p))}} e^{-r \langle \theta(p) \mid e_\alpha \rangle }$$
$${}+ \frac{1}{2\pi \sqrt{\det \Sigma}} \int_{\Gamma'_{\theta_1, \alpha} }
\frac{\varphi_2(s) \gamma_2(\theta(s))}{s} e^{- r \langle \theta(s) \mid e_\alpha \rangle }
{\mathrm{d}s}
+
\frac{1}{2\pi \sqrt{\det \Sigma}} \int_{\Gamma''_{\theta_2, \alpha} }
\frac{\varphi_1(s) \gamma_1(\theta(s))}{s} e^{- r \langle \theta(s) \mid e_\alpha \rangle }
\mathrm{d}s.$$ Applying Lemma \[uv\] (i) for the estimation of integrals along $\Gamma'^{1}_{\theta_1,\alpha}$ and $\Gamma''^{1}_{\theta_1,\alpha}$, and the same lemma (ii) for the estimation of those along $\Gamma'^{\pm 2 }_{\theta_1,\alpha}$ $\Gamma''^{\pm 2}_{\theta_2,\alpha}$, $\Gamma'^{\pm 3}_{\theta_1,\alpha}$ and $\Gamma''^{\pm 3}_{\theta_2,\alpha}$ exactly as in Lemma \[saddleas\] and in view of (\[va\]) we can show that with some constant $C>0$ $$\begin{aligned}
\label{in11}
\Big| \int_{\Gamma'_{\theta_1, \alpha} }
\frac{\varphi_2(s) \gamma_2(\theta(s))}{s} e^{- r \langle \theta(s) \mid e_\alpha \rangle }
{\mathrm{d}s}\Big| & {\leqslant}& C e ^{- r \langle \tau' \mid e_\alpha \rangle},\nonumber \\
\Big| \int_{\Gamma''_{\theta_1, \alpha} }
\frac{\varphi_2(s) \gamma_2(\theta(s))}{s} e^{- r \langle \theta(s) \mid e_\alpha \rangle }
{\mathrm{d}s}\Big| &{\leqslant}& C e ^{- r \langle \tau'' \mid e_\alpha \rangle} \ \ \forall r>0, \forall \alpha \in \mathcal{O}(\alpha_0).
\nonumber \end{aligned}$$ Hence, by (\[at\]) $$\int_{\Gamma'_{\theta_1, \alpha} }
\frac{\varphi_2(s) \gamma_2(\theta(s))}{s} e^{- r \langle \theta(s) \mid e_\alpha \rangle }
{\mathrm{d}s}
+
\int_{\Gamma''_{\theta_2, \alpha} }
\frac{\varphi_1(s) \gamma_1(\theta(s))}{s} e^{- r \langle \theta(s) \mid e_\alpha \rangle }
\mathrm{d}s = o(e^{- r(\langle \theta(\alpha) \mid e_\alpha \rangle -\delta ) })$$ as $r \to \infty$ uniformly $\forall \alpha \in \mathcal{O}(\alpha_0)$. This finishes the proof of the representation (\[th233\]). The analysis of the main term is the same as in Theorem \[thmresults2\].
It remains to study the cases of parameters such that
- $\zeta \theta^{**}=\theta(\alpha_0) $ and $\eta \theta^{*} \not\in \{s''_0, \theta(\alpha_0)\{$
- $\eta \theta^*=\theta(\alpha_0)$ and $\zeta \theta^{**} \not\in \{s''_0, \theta(\alpha_0)\{$.
By Lemma \[lem3\] this means that $\theta(\alpha_0)$ is a pole of one of functions $\phi_1$ or $\phi_2$. Since in both cases $ \eta \theta^{*} \not\in \{s''_0, \theta(\alpha)\{$, $\zeta \theta^{**} \not\in \{s''_0, \theta(\alpha)\{$, we derive by the same reasoning as in Theorem \[thmpoles\] (a) that $\mathcal{P}'_{\alpha_0} \cup \mathcal{P}''_{\alpha_0}$ is empty. The following theorem is valid.
\[thmresults4\] Assume that $\alpha_0$ is such that the assumptions on parameters (O1) or (O2) are valid. Then for any $\delta>0$ there exists a small enough neighborhood $\mathcal{O}(\alpha_0)$ such that $$\label{th2334}
\pi(r\cos(\alpha), r \sin(\alpha)) = o(e^{-r (\langle \theta(\alpha) \mid e_\alpha \rangle -\delta ) } ) \ \ \ r \to \infty, \hbox{ uniformly } \forall \alpha \in \mathcal{O}(\alpha_0).$$
We choose $\tau'$ and $\tau''$ according to (\[at\]) and proceed exactly as in the proof of Theorem \[thmresults3\].
[**Remark.**]{} In Theorems \[thmresults3\] and \[thmresults4\] $\theta(\alpha_0)$ is a pole of one of the functions $\phi_1$ or $\phi_2$, hence at least one of the integrals (\[iwi\]) can not be shifted to $\Gamma_{\theta_1, \alpha_0}$ or $\Gamma_{\theta_2, \alpha_0}$ going through $\theta(\alpha_0)$. Furthermore, although for any $\alpha \in \mathcal{O}(\alpha_0)$, $\alpha\ne \alpha_0$, this shift is possible, the uniform asymptotic expansion by the saddle-point method as in Lemma \[saddleas\] does not stay valid, that is why we are not able to specify small asymptotic terms in Theorem \[thmresults3\] neither to obtain a more precise result in Theorem \[thmresults4\]. This should be possible if we consider the double asymptotics $r\to \infty$ and $\alpha \to \alpha_0$ and apply the (more advanced) saddle-point method in the special case when the saddle-point is approaching a pole of the integrand. We do not do it in the present paper.
[**Remark.**]{} Assumptions of theorems \[thmresults1\] — \[thmresults4\] are expressed in terms of positions on ellipse $\mathcal{E}$ of points $\zeta \theta^{**}$ and $\eta \theta^*$ that are images of zeros of $\gamma_1$ and $\gamma_2$ on $\mathcal{E}$ by Galois automorphisms. They can be also expressed in terms of the following simple inequalities.
Under parameters such that $\theta_1(\alpha_0)>0$, we have $\zeta\theta^{**} \ne \{s'_0, \theta(\alpha_0)\}$ iff $\theta^{**} \ne \{s_0, \zeta \theta(\alpha_0)\}$ that is equivalent to $\gamma_2(\zeta \theta(\alpha))<0$. Under parameters such that $\theta_1(\alpha_0){\leqslant}0$, we have always $\zeta \theta^{**} \ne \{s'_0, \theta(\alpha_0)\}$ because $\theta_1(\zeta \theta^{**})>0$ by stability conditions, in this case we have also $\gamma_2(\zeta \theta(\alpha_0)){\geqslant}0$. We come to the following conclusions.
- Assumption $\zeta\theta^{**} \ne \{s'_0, \theta(\alpha_0)\}$ is equivalent to the one that $\gamma_2(\zeta \theta(\alpha_0))<0$ or $\theta_1(\alpha_0){\leqslant}0$.
Assumption $\zeta\theta^{**} \in \{s'_0, \theta(\alpha_0)\{$ is equivalent to the one that $\gamma_2(\zeta \theta(\alpha_0))>0$ and $\theta_1(\alpha_0)> 0$.
- Assumption $\eta\theta^{*} \ne \{s''_0, \theta(\alpha_0)\}$ is equivalent to the one that $\gamma_1(\eta \theta(\alpha_0))<0$ or $\theta_2(\alpha_0){\leqslant}0$.
Assumption $\eta\theta^{**} \in \{s''_0, \theta(\alpha_0)\{$ is equivalent to the one that $\gamma_1(\eta \theta(\alpha_0))>0$ and $\theta_2(\alpha_0)> 0$.
Given parameters Sigma mu R, density asymptotics for all angles alpha0 {#subsec:givenparameters}
-----------------------------------------------------------------------
In this section we state the asymptotics of the density for all angles $\alpha_0 \in ]0, \pi/2[$ once parameters $(\Sigma, \mu, R)$ are fixed. Theorems \[thmresultsnew1\] – \[thmresultsnew3\] below are direct corollaries of Theorems \[thmresults1\] – \[thmresults4\] and elementary geometric properties of ellipse $\mathcal{E}$ and straight lines $\gamma_1(\theta)=0$ and $\gamma_2(\theta)=0$, therefore we do not give their proofs. To shorten the presentation, we restrict ourselves to the main term in the formulations of the results, although of course further terms of the expansions could be written. The different cases of Theorem \[thmresultsnew1\] are illustrated by Figures \[firstfig\]–\[lastfig\].
Let $\theta_1(s_2^+)>0$, $\theta_2(s_1^+)>0$. \[thmresultsnew1\]
- Let $\gamma_2(s_1^+){\leqslant}0$ and $\gamma_1(s_2^+){\leqslant}0$ Then for any $\alpha_0 \in ]0, \pi/2[$ we have :
$$\label{assadlep}
\pi(r \cos \alpha, r \sin \alpha) \sim \frac{c(\alpha_0)}{\sqrt{r}} \exp(-r\langle \theta(\alpha) \mid e_\alpha \rangle), \ \ r \to \infty,
\alpha \to \alpha_0,$$
where the constant $c(\alpha_0)$ depends continuously on $\alpha_0 \in ]0, \pi/2[$ and $\lim_{\alpha_0 \to 0} c(\alpha_0)=\lim_{\alpha_0 \to \pi/2} c(\alpha_0)=0$.
- Let $\gamma_2(s_1^+)>0$ and $\gamma_1(s_2^+){\leqslant}0$.
- Let $\gamma_2(\zeta s_2^+){\geqslant}0$ or equivalently $\frac{ d \Theta_2^{+}(\theta_1) }{ d \theta_1 } \Bigm|_{\theta_1^{**} }{\geqslant}0$. Then for any $\alpha_0 \in ]0, \pi/2[$ we have $$\label{p**}
\pi(r \cos \alpha, r \sin \alpha) \sim d_1 \exp(-r\langle \zeta \theta^{**} \mid e_\alpha \rangle), \ \ r \to \infty,
\alpha \to \alpha_0,$$ with some constant $d_1>0$.
- Let $\gamma_2(\zeta s_2^+)<0$ or equivalently $A^{**}\equiv \frac{ d \Theta_2^{+}(\theta_1) }{ d \theta_1 } \Bigm|_{\theta_1^{**} }
< 0$. Define $\alpha_1= \arctan ( -1/ A^{**}) \in ]0, \pi/2[$. Then for any $\alpha_0 \in ]0, \alpha_1[$ we have (\[p\*\*\]) and for any $\alpha \in ]\alpha_1, \pi/2[$ we have (\[assadlep\]).
- Let $\gamma_2(s_1^+)<0$ and $\gamma_1(s_2^+){\geqslant}0$.
- Let $\gamma_1(\eta s_1^+){\geqslant}0$ or equivalently $\frac{ d \Theta_1^{+}(\theta_2) }{ d \theta_2 } \Bigm|_{\theta_2^{*} } {\geqslant}0$. Then for any $\alpha_0 \in ]0, \pi/2[$ we have $$\label{p*}
\pi(r \cos \alpha, r \sin \alpha) \sim d_2 \exp(- r\langle \eta \theta^{*} \mid e_\alpha \rangle), \ \ r \to \infty,
\alpha \to \alpha_0,$$ with some constant $d_2>0$.
- Let $\gamma_1(\eta s_1^+)<0$ or equivalently $ A^*\equiv \frac{ d \Theta_1^{+}(\theta_2) }{ d \theta_2 } \Bigm|_{\theta_2^{*} } < 0$. Define $\alpha_2= \arctan (- A^*) \in ]0, \pi/2[$. Then for any $\alpha_0 \in ]0, \alpha_2[$ we have (\[assadlep\]) and for any $\alpha \in ]\alpha_2, \pi/2[$ we have (\[p\*\]).
- Let $\gamma_2(s_1^+)>0$ and $\gamma_1(s_2^+)>0$.
- Let $\theta_1(\zeta \theta^{**}){\leqslant}\theta_1(\eta \theta^*)$ and $\theta_2(\zeta \theta^{**}){\leqslant}\theta_2(\eta \theta^*)$ where at least one of inequalities is strict. Then for any $\alpha_0 \in ]0, \pi/2[$ we have (\[p\*\*\]).
- Let $\theta_1(\zeta \theta^{**}){\geqslant}\theta_1(\eta \theta^*)$ and $\theta_2(\zeta \theta^{**}){\geqslant}\theta_2(\eta \theta^*)$ where at least one of inequalities is strict. Then for any $\alpha \in ]0, \pi/2[$ we have (\[p\*\]).
- Let $\theta_1(\zeta \theta^{**}){\leqslant}\theta_1(\eta \theta^*)$ and $\theta_2(\zeta \theta^{**}) {\geqslant}\theta_2(\eta \theta^*)$ where at least one of the inequalities is strict. Let us define $\beta_0 =\arctan \frac{ \theta_1(\zeta\theta^{**})-\theta_1(\eta \theta^*) }{ \theta_2(\eta\theta^{*}) -\theta_2(\zeta \theta^{**}) }$ Then for any $\alpha_0 \in ]0, \beta_0[$ we have (\[p\*\*\]), for any $\alpha_0 \in ]\beta_0, \pi/2[$ we have (\[p\*\]) and for $\alpha_0=\beta_0$ we have $$\label{p***}
\pi(r \cos \alpha, r \sin \alpha) \sim
d_1 \exp(-r\langle \zeta \theta^{**} \mid e_\alpha \rangle)+ d_2 \exp(-r\langle \eta \theta^{*} \mid e_\alpha \rangle), \ \ r \to \infty,
\alpha \to \alpha_0.$$
- Let $\theta_1(\zeta \theta^{**}){\geqslant}\theta_1(\eta \theta^*)$ and $\theta_2(\zeta \theta^{**}) {\leqslant}\theta_2(\eta \theta^*)$. Let us define angles $\alpha_1$ and $\alpha_2$ as in (ii) and (iii). Then $0<\alpha_1 {\leqslant}\alpha_2 <\pi/2$; for any $\alpha_0 \in ]0, \alpha_1[$ we have (\[p\*\*\]) , for any $\alpha_0 \in ]\alpha_1, \alpha_2[$ we have (\[assadlep\]) and for any $\alpha_0 \in ]\alpha_2, \pi/2[$ we have (\[p\*\]).
Let $\theta_1(s_2^+)\leqslant 0$, $\theta_2(s_1^+) \leqslant 0$. \[thmresultsnew2\]
- Let $\gamma_2(s_1^+){\leqslant}0$ and $\gamma_1(s_2^+){\leqslant}0$ Then for any $\alpha_0 \in ]0, \pi/2[$ the asymptotics (\[assadlep\]) is valid.
- Let $\gamma_2(s_1^+)>0$ and $\gamma_1(s_2^+){\leqslant}0$. Let $A^{**}\equiv \frac{ d \Theta_2^{+}(\theta_1) }{ d \theta_1 } \Bigm|_{\theta_1^{**} }$. Then $\alpha_1= \arctan ( -1/ A^{**}) \in ]0, \pi/2[$. For any $\alpha_0 \in ]0, \alpha_1[$ the asymptotics (\[p\*\*\]) is valid and and for any $\alpha \in ]\alpha_1, \pi/2[$ the asymptotics (\[assadlep\]) holds true.
- Let $\gamma_2(s_1^+)<0$ and $\gamma_1(s_2^+){\geqslant}0$. Let $ A^*\equiv \frac{ d \Theta_1^{+}(\theta_2) }{ d \theta_2 } \Bigm|_{\theta_2^{*} } $. Then $\alpha_2= \arctan (- A^*) \in ]0, \pi/2[$. For any $\alpha_0 \in ]0, \alpha_2[$ the asymptotics (\[assadlep\]) is valid and for any $\alpha \in ]\alpha_2, \pi/2[$ the asymptotics (\[p\*\]) holds true.
- Let $\gamma_2(s_1^+)>0$ and $\gamma_1(s_2^+)>0$. Then either $\theta_1(\zeta \theta^{**})< \theta_1(\eta \theta^*)$ and $\theta_2(\zeta \theta^{**}) > \theta_2(\eta \theta^*)$, or $\theta_1(\zeta \theta^{**})> \theta_1(\eta \theta^*)$ and $\theta_2(\zeta \theta^{**}) < \theta_2(\eta \theta^*)$, or finally $\theta_1(\zeta \theta^{**})=\theta_1(\eta \theta^*)$ and $\theta_2(\zeta \theta^{**}) = \theta_2(\eta \theta^*)$.
- Let $\theta_1(\zeta \theta^{**})< \theta_1(\eta \theta^*)$ and $\theta_2(\zeta \theta^{**}) > \theta_2(\eta \theta^*)$. Let us define $\beta_0 =\arctan \frac{ \theta_1(\zeta\theta^{**})-\theta_1(\eta \theta^*) }{ \theta_2(\eta\theta^{*}) -\theta_2(\zeta \theta^{**}) }$ Then for any $\alpha_0 \in ]0, \beta_0[$ we have (\[p\*\*\]), for any $\alpha_0 \in ]\beta_0, \pi/2[$ we have (\[p\*\]) and for $\alpha_0=\beta_0$ we have (\[p\*\*\*\]).
- Let $\theta_1(\zeta \theta^{**})> \theta_1(\eta \theta^*)$ and $\theta_2(\zeta \theta^{**}) < \theta_2(\eta \theta^*)$ or $\theta_1(\zeta \theta^{**})= \theta_1(\eta \theta^*)$ and $\theta_2(\zeta \theta^{**}) = \theta_2(\eta \theta^*)$ Let us define angles $\alpha_1$ and $\alpha_2$ as in (ii) and (iii). Then $0<\alpha_1 {\leqslant}\alpha_2 <\pi/2$; for any $\alpha_0 \in ]0, \alpha_1[$ we have (\[p\*\*\]) , for any $\alpha_0 \in ]\alpha_1, \alpha_2[$ we have (\[assadlep\]) and for any $\alpha_0 \in ]\alpha_2, \pi/2[$ we have (\[p\*\]).
\[thmresultsnew3\] Let $\theta_1(s_2^+)> 0$, $\theta_2(s_1^+) \leqslant 0$.
- Let $\gamma_2(s_1^+){\leqslant}0$ and $\gamma_1(s_2^+){\leqslant}0$ Then for any $\alpha_0 \in ]0, \pi/2[$ the asymptotics (\[assadlep\]) is valid.
- Let $\gamma_2(s_1^+)>0$ and $\gamma_1(s_2^+){\leqslant}0$.
- Let $\gamma_2(\zeta s_2^+){\geqslant}0$ or equivalently $\frac{ d \Theta_2^{+}(\theta_1) }{ d \theta_1 } \Bigm|_{\theta_1^{**} }{\geqslant}0$. Then for any $\alpha_0 \in ]0, \pi/2[$ the asymptotics is valid.
- Let $\gamma_2(\zeta s_2^+)<0$ or equivalently $A^{**}\equiv \frac{ d \Theta_2^{+}(\theta_1) }{ d \theta_1 } \Bigm|_{\theta_1^{**} }
< 0$. Define $\alpha_1= \arctan ( -1/ A^{**}) \in ]0, \pi/2[$. Then for any $\alpha_0 \in ]0, \alpha_1[$ we have (\[p\*\*\]) and for any $\alpha \in ]\alpha_1, \pi/2[$ we have (\[assadlep\]).
- Let $\gamma_2(s_1^+)<0$ and $\gamma_1(s_2^+){\geqslant}0$. Let $ A^*\equiv \frac{ d \Theta_1^{+}(\theta_2) }{ d \theta_2 } \Bigm|_{\theta_2^{*} } $. Then $\alpha_2= \arctan (- A^*) \in ]0, \pi/2[$. For any $\alpha_0 \in ]0, \alpha_2[$ the asymptotics (\[assadlep\]) is valid and for any $\alpha \in ]\alpha_2, \pi/2[$ the asymptotics (\[p\*\]) holds true.
- Let $\gamma_2(s_1^+)>0$ and $\gamma_1(s_2^+)>0$. Then either $\theta_1(\zeta \theta^{**}){\leqslant}\theta_1(\eta \theta^*)$ and $\theta_2(\zeta \theta^{**}){\leqslant}\theta_2(\eta \theta^*)$, or $\theta_1(\zeta \theta^{**})< \theta_1(\eta \theta^*)$ and $\theta_2(\zeta \theta^{**}) > \theta_2(\eta \theta^*)$, or $\theta_1(\zeta \theta^{**})> \theta_1(\eta \theta^*)$ and $\theta_2(\zeta \theta^{**}) < \theta_2(\eta \theta^*)$, or finally $\theta_1(\zeta \theta^{**})=\theta_1(\eta \theta^*)$ and $\theta_2(\zeta \theta^{**}) = \theta_2(\eta \theta^*)$.
- Let $\theta_1(\zeta \theta^{**}){\leqslant}\theta_1(\eta \theta^*)$ and $\theta_2(\zeta \theta^{**}){\leqslant}\theta_2(\eta \theta^*)$ where at least one of inequalities is strict. Then for any $\alpha_0 \in ]0, \pi/2[$ we have (\[p\*\*\]).
- Let $\theta_1(\zeta \theta^{**})< \theta_1(\eta \theta^*)$ and $\theta_2(\zeta \theta^{**}) > \theta_2(\eta \theta^*)$. Let us define $\beta_0 =\arctan \frac{ \theta_1(\zeta\theta^{**})-\theta_1(\eta \theta^*) }{ \theta_2(\eta\theta^{*}) -\theta_2(\zeta \theta^{**}) }$ Then for any $\alpha_0 \in ]0, \beta_0[$ we have (\[p\*\*\]), for any $\alpha_0 \in ]\beta_0, \pi/2[$ we have (\[p\*\]) and for $\alpha_0=\beta_0$ we have (\[p\*\*\*\]).
- Let $\theta_1(\zeta \theta^{**})> \theta_1(\eta \theta^*)$ and $\theta_2(\zeta \theta^{**}) < \theta_2(\eta \theta^*)$ or $\theta_1(\zeta \theta^{**})= \theta_1(\eta \theta^*)$ and $\theta_2(\zeta \theta^{**}) = \theta_2(\eta \theta^*)$ Let us define angles $\alpha_1$ and $\alpha_2$ as in (ii) and (iii). Then $0<\alpha_1 {\leqslant}\alpha_2 <\pi/2$; for any $\alpha_0 \in ]0, \alpha_1[$ we have (\[p\*\*\]) , for any $\alpha_0 \in ]\alpha_1, \alpha_2[$ we have (\[assadlep\]) and for any $\alpha_0 \in ]\alpha_2, \pi/2[$ we have (\[p\*\]).
The symmetric theorem for the case $\theta_1(s_2^+)\leqslant 0$, $\theta_2(s_1^+) > 0$ holds.
Concluding remarks {#subsec:concludingremarks}
------------------
Let us remark that the approach of this article applies to the SRBM in . Thanks to a linear transformation $T\in {\bf R}^{2\times 2}$, it is easy to transform $Z(t)$, a reflected Brownian motion of parameters $(\Sigma,\mu,R)$ in a cone into $T Z(t)$ a reflected Brownian motion of parameters $(T\Sigma T^t,T\mu,TR)$ in the quarter plane. For example if the initial cone is the set $\{(x,y) | x\geqslant0 \text{ and } y\leqslant ax \}$ for some $a>0$, we may just take $T=
\begin{pmatrix}
1 & -\frac{1}{a} \\
0 & 1
\end{pmatrix}$. The process $T Z(t)$ lives in a quarter plane. Then the approach of this article applies and its results can be converted to the initial cone by the inverse linear transformation. The analytic approach for discrete random walks is essentially restricted to those with jumps to the nearest neighbors in the interior of the quarter plane. Since a linear transformation can not generally keep the length of jumps, this procedure does not work in the discrete case. That is why the analytic approach in ${\bf R}^2$ has a more general scope of applications.
To conclude this article, we sketch the way of recovering the asymptotic results of Dai and Miyazawa [@dai_reflecting_2011] via the approach of this article. Given a directional vector $c=(c_1,c_2) \in {\bf R}_+^2$, thanks to the representation of Lemma \[propI\] we obtain $$\begin{aligned}
\lefteqn{
\mathbf{P}( \langle c \mid Z(\infty) \rangle {\geqslant}R) =\int_{x_1 \geqslant 0, \ x_2 \geqslant 0 \atop c_1x_1+c_2x_2 \geqslant R} \pi(x_1,x_2) \mathrm{d}x_1 \mathrm{d}x_2}\nonumber\\
&=&
\int_{x_1 \geqslant 0, \ x_2 \geqslant 0 \atop c_1x_1+c_2x_2 \geqslant R}
I_1(x_1, x_2) \mathrm{d}x_1 \mathrm{d}x_2
+
\int_{x_1 \geqslant 0, \ x_2 \geqslant 0 \atop c_1x_1+c_2x_2 \geqslant R}
I_2(x_1,x_2) \mathrm{d}x_1 \mathrm{d}x_2\nonumber\\
&=&\int\limits_{\mathcal{I}_{\theta_1}^{\epsilon,+}} g_1(\theta_1) \frac{1}{\theta_1}\frac{1}{\Theta_2^+(\theta_1)-\theta_1 \frac{c_2}{c_1}} e^{-\frac{ R}{c_1}\theta_1} \mathrm{d}\theta_1+
\int\limits_{\mathcal{I}_{\theta_1}^{\epsilon,+}} g_1(\theta_1) \frac{-c_2/c_1}{\Theta_2^+(\theta_1)(\Theta_2^+(\theta_1)-\theta_1 \frac{c_2}{c_1})} e^{-\frac{ R}{c_2}\Theta_2^+(\theta_1)} \mathrm{d}\theta_1\label{gg}\\
&&{}+\int\limits_{\mathcal{I}_{\theta_2}^{\epsilon,+}} g_2(\theta_2) \frac{1}{\theta_2}\frac{1}{\Theta_1^+(\theta_2)-\theta_2 \frac{c_1}{c_2}} e^{-\frac{ R}{c_2}\theta_2} \mathrm{d}\theta_2+
\int\limits_{\mathcal{I}_{\theta_2}^{\epsilon,+}} g_2(\theta_2) \frac{-c_1/c_2}{\Theta_1^+(\theta_2)(\Theta_1^+(\theta_2)-\theta_2 \frac{c_1}{c_2})} e^{-\frac{ R}{c_1}\Theta_1^+(\theta_2)} \mathrm{d}\theta_2,\nonumber\\
\label{ggg}\end{aligned}$$ where $$g_1(\theta_1)= \frac{\varphi_2(\theta_1) \gamma_2(\theta_1, \Theta_2^+(\theta_1)) }{\sqrt{ d(\theta_1) }}, \ \ \ \
g_2(\theta_2)= \frac{\varphi_1(\theta_2) \gamma_1(\Theta_1^+(\theta_2), \theta_2) }{\sqrt{ d(\theta_2) }}.$$ The first term in (\[gg\]) is just the Laplace transform of the function $h_1(\theta_1)=g_1(\theta_1) \frac{1}{\theta_1}\frac{1}{\Theta_2^+(\theta_1)-\theta_1 \frac{c_2}{c_1}}$, its asymptotics is determined by the smallest real singularity of $h_1(\theta_1)$, see e.g. [@doetsch_introduction_1974]. This may be either the branch point $\theta_1^+$ of $\phi_2(\theta_1)$, or the smallest pole of $h_1(\theta_1)$ on $]0, \theta_1^+[$ whenever it exists, the natural candidates are $\zeta \theta^{**}$, $\zeta \eta \theta^*$ due to Lemmas \[lem3\]– \[lem4\] or a point $\theta^c=(\theta_1^c, \theta_2^c)$ such that $\theta_2^c=\Theta_2^+(\theta_1^c)=
\theta_1^c \frac{c_2}{c_1}$. To determine the asymptotics of the second integral in (\[gg\]), we shift the integration contour to the new one passing through the saddle-point $\Theta_1(\theta_2^+)$ and take into account the poles of the integrand we encounter, the most important of these poles are those listed above. The asymptotics of two terms in (\[ggg\]) is determined in the same way. Combining all these results together we derive the main asymptotic term depending on the parameters that can be either $e^{-\frac{R}{c_1}\theta_1^{+}}$, $e^{-\frac{R}{c_2}\theta_2^+}$ preceding by $R^{-1/2}$ or $R^{-3/2}$ with some constant, or $e^{-\frac{R}{c_1}\theta_1^c}=e^{\frac{R}{c_2}\theta_2^c}$, $e^{-\frac{R}{c_i} (\zeta\theta^{**})_i}$, $e^{-\frac{R}{c_i} (\eta\theta^{*})_i}$, $i=1,2$ preceding by some constant and the factor $R$ in some critical cases. This analysis leads to the results of [@dai_reflecting_2011].
[**Acknowledgement**]{} We are grateful to Kilian Raschel and Sandrine Péché for helpful discussions and suggestions.
![Theorem \[thmresultsnew1\] case (i)[]{data-label="firstfig"}](cas1.pdf "fig:") ![Theorem \[thmresultsnew1\] case (i)[]{data-label="firstfig"}](casa1.pdf "fig:")
![Theorem \[thmresultsnew1\] case (iia) and (iva)](cas2a.pdf "fig:") ![Theorem \[thmresultsnew1\] case (iia) and (iva)](cas2aa.pdf "fig:") ![Theorem \[thmresultsnew1\] case (iia) and (iva)](casa2a.pdf "fig:")
![Theorem \[thmresultsnew1\] case (iiia) and (ivb)](cas2b.pdf "fig:") ![Theorem \[thmresultsnew1\] case (iiia) and (ivb)](cas2bb.pdf "fig:") ![Theorem \[thmresultsnew1\] case (iiia) and (ivb)](casa2b.pdf "fig:")
![Theorem \[thmresultsnew1\] case (iib)](cas3a.pdf "fig:") ![Theorem \[thmresultsnew1\] case (iib)](casa3a.pdf "fig:")
![Theorem \[thmresultsnew1\] case (iiib)](cas3b.pdf "fig:") ![Theorem \[thmresultsnew1\] case (iiib)](casa3b.pdf "fig:")
![Theorem \[thmresultsnew1\] case (ivc)](cas4a.pdf "fig:") ![Theorem \[thmresultsnew1\] case (ivc)](casa4.pdf "fig:")
![Theorem \[thmresultsnew1\] case (ivd)[]{data-label="lastfig"}](cas5.pdf "fig:") ![Theorem \[thmresultsnew1\] case (ivd)[]{data-label="lastfig"}](casa5.pdf "fig:")
[^1]: Version of
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Joris Raeymaekers\
Institute of Physics of the ASCR,\
Na Slovance 2, 182 21 Prague 8, Czech Republic.\
email: [`[email protected]`]{}
title: Quantization of conical spaces in 3D gravity
---
Introduction
============
Pure gravity in 2+1 dimensions with negative cosmological constant has proven to be an interesting laboratory to test ideas in quantum gravity and holography. Despite having no local degrees of freedom, it possesses an asymptotic Virasoro symmetry [@Brown:1986nw] which can be viewed as the symmetry of a dual 1+1 dimensional CFT [@Maldacena:1997re]. Furthermore the gravity theory contains black hole solutions [@Banados:1992wn] whose entropy can be understood from modular invariance of the dual CFT [@Strominger:1997eq]. Despite these qualitative successes, it is still unclear whether pure gravity at weak coupling (large central charge $c$) exists as a quantum theory and, if so, what are the properties of the dual CFT [@Witten:2007kt],[@Gaberdiel:2007ve],[@Maloney:2007ud].
An important step in identifying a candidate dual CFT is to understand the spectrum of smooth, asymptotically AdS, classical solutions, which are believed to represent semiclassical states in the spectrum of the dual CFT. In this note, we will illustrate a subtlety in the concept of smoothness related to the formulation of the theory in terms of Chern-Simons gauge fields [@Achucarro:1987vz],[@Witten:1988hc]. In this formulation the metric is a derived quantity and in particular there is no natural way to impose its invertibility. As a consequence, many observables that exist in the metric formulation and which involve the inverse metric, are not natural observables in the Chern-Simons formulation. Instead, the natural observables in the Chern-Simons formulation are based on holonomies of the gauge field, and we will see that these provide a somewhat cruder measure of smoothness.
The solutions we will study are labelled by an integer $s$ greater than one and correspond to metrics with a conical singularity of excess angle $2 \p (s-1)$. Conical defects in 2+1 dimensional gravity have been extensively studied following the pioneering work [@Deser:1983tn],[@Deser:1983nh]. The conical excess solutions were, as far as we know, first explored in [@Izquierdo:1994jz] as BPS solutions of 3-dimensional supergravity, and were subsequently studied in [@Mansson:2000sj],[@Balasubramanian:2000rt]. Despite being metrically singular, we will argue that the conical excess solutions appear smooth to the natural observables in the Chern-Simons formulation. One new insight coming from our analysis is that requiring the solutions to appear smooth to the point particle probes considered in [@Witten:1989sx],[@Ammon:2013hba] picks out the conical spaces without angular momentum.
Aside from the singularity in the metric, another reason that conical excesses are usually discarded is that they have energies below that of the global AdS solution, implying that they cannot be part of the spectrum of a unitary dual CFT. Nevertheless, many consistent nonunitary 2-dimensional CFTs are known to exist, some of which have gravity duals[^1], and one might expect the conical spaces to play a role in this context. Indeed, the conical excesses are the pure gravity avatars of similar solutions in higher spin gravity theories[^2], which were argued to have a consistent dual interpretation in a nonunitary semiclassical limit of the dual CFT[@Castro:2011iw],[@Perlmutter:2012ds],[@Hijano:2013fja],[@Campoleoni:2013iha]. The arguments leading to this identification were so far made in the classical approximation in the bulk, and the main motivation of this work was to develop a framework to include bulk quantum corrections in the simplified setting of pure gravity.
We will see that the conical solutions in pure gravity are topological solitons characterized by a winding number. To discuss some of their quantum aspects we will follow the standard approach for quantizing solitons, namely to quantize the fluctuations around the classical solution (see [@Coleman] for a review and further references). In the case of pure three dimensional gravity, the fluctuations around a given solution come from acting on it with asymptotic Virasoro generators, or in other words from dressing the solution with boundary graviton excitations. Exponentiating these infinitesimal variations to yield an action of the Virasoro group on the solution, one obtains what is called a coadjoint orbit of the Virasoro group. Each orbit carries a natural Poisson bracket, which upon quantization should lead to a Virasoro representation. The Virasoro coadjoint orbits were classified in [@Lazutkin],[@Segal:1981ap] and further analyzed by Witten [@Witten:1987ty]. Subsequent work includes [@Alekseev:1988ce],[@Bershadsky:1989mf],[@Balog:1997zz], see also [@Garbarz:2014kaa],[@Barnich:2014zoa] for recent discussions of coadjoint orbits in the context of three dimensional gravity.
It turns out that the conical spaces give rise to rather special orbits: they are the ‘exceptional’ orbits which possess an $SO(2,1)^{(s)} \times SO(2,1)^{(s)}$ symmetry. Here, the superscript $s$ indicates that the group is an $s$-fold cover of $SO(2,1)$. The generators of this symmetry are embedded in the Virasoro algebra in a different manner for each conical space and, in particular, are different from the symmetry generators of the global AdS solution. By analyzing the exceptional coadjoint orbits one finds that the conical spaces possess boundary graviton fluctuations which lower the Virasoro energy, which is unbounded below [@Witten:1987ty]. For this reason they cannot be quantized in a way that leads to a unitary highest weight Virasoro representation, and it has so far proved impossible to quantize the exceptional orbits by standard methods.
In this work we will propose a framework to quantize the exceptional orbits which leads to a nonunitary highest weight representation of the Virasoro algebra, constructed as a semiclassical perturbation expansion in the inverse central charge, and proceed to compute the 1-loop correction to the energy of the conical spaces. Furthermore, as anticipated in [@Witten:1987ty], the quantized exceptional orbits contain a null vector at level $s$. Hence the conical spaces constitute an appearance of the most interesting representations of the Virasoro algebra, the degenerate ones, in a gravity context. Our results for the energy correction and the null vector lead us to identify the quantized conical spaces with the degenerate representations of the type $(1,s)$ in Kac’s classification [@Kac:1978ge], which are well-known to be nonunitary at large values of the central charge $c$. It is fascinating that these representations become unitary, and belong of the spectrum of the Virasoro minimal models, at small values of the central charge $0<c<1$, which represents the strong coupling regime of the gravity theory.
Chern-Simons formulation of 3D AdS gravity {#secCS}
==========================================
Let us briefly review the Chern-Simons formulation of gravity and state our conventions. Pure gravity in 2+1 dimensions with negative cosmological constant can be reformulated as a Chern-Simons theory with gauge group $SO(1,2)\times {SO(1,2)}$ with opposite levels for the two factors [@Achucarro:1987vz],[@Witten:1988hc]: S &=& S\_[CS]{}\[A\] - S\_[CS]{}\[A\]\
S\_[CS]{}\[A\] &=& [k 4 ]{} \_\_3 ( Ad A + [ 23]{} AA A ).\[SCS\] The gauge potentials $A, \tilde A$ take values in the $so(1,2)$ Lie algebra whose generators $J_a, a = 0,1,2$ satisfy the commutation relations[^3] = \_[ab]{}\^[ c]{} J\_c\[Js\] Viewing $SO(1,2)$ as the 2+1-dimensional Lorentz group, the generator $J_0$ is compact and corresponds to spatial rotations, while $J_{1,2}$ generate boosts. In our conventions the trace in (\[SCS\]) is taken in the defining three dimensional representation[^4]. The vielbein and spin connection are obtained from A &=& ( ø\^a + [e\^a l]{}) J\_aA &=& ( ø\^a - [e\^a l]{}) J\_a where $\o^a=\half \e^a_{\ bc} \o^{bc}$ and $l$ is the AdS radius.
The Chern-Simons field strengths are related to the torsion $\calt^a = d e^a +\e^a_{\ bc} \o^b \wedge e^c $ and curvature $\calr^a= d \o^a + \half \e^a_{\ bc} \o^b \wedge \o^c$ two-forms as follows: (F- F ) &=&\
( F + F) &=&+ [1l\^2]{} e. Hence the equations of motion, which impose the flatness of $A$ and $\tilde A$, imply that the connection $\o$ is torsionless and that Einstein’s equations hold with negative cosmological constant $\L = -{2\over l^2}$.
Writing the action in terms of $e, \o$ gives S &=& [k l]{}\_d\^3 x ( R + [ 2l\^2]{} ) +[k l]{}\_ e\^a ø\_a . This allows us to obtain Newton’s constant from k = [l 16 G]{}. The Brown-Henneaux central charge [@Brown:1986nw] is the combination c = [3 l 2 G]{}= 24 k. The large $c$ (weak coupling) limit is the semiclassical regime, where the path integral is dominated by classical gravity solutions, while for small $c$ (strong coupling) quantum corrections are significant.
Gauge transformations act as A Ł\^[-1]{} A Ł+ Ł\^[-1]{} dŁA Ł\^[-1]{} A Ł+ Ł\^[-1]{} dŁ\[gaugetransf\] and in order for $e^{i S}$ to be gauge-invariant, $k$ must be quantized in integer units [@Witten:2007kt], which implies that $c$ is a multiple of 24[^5]. By contrast, if we consider Euclidean gravity, the relevant Chern-Simons gauge group is replaced by $SL(2,{\mathbb{C}})$ and there is no quantization condition on $k$ and $c$ [@Witten:1989ip] (this is one of the reasons why the relation between Lorentzian and Euclidean Chern-Simons gravity is poorly understood). For most of this paper we will consider the Lorentzian theory but will occasionally comment on the Euclidean case.
Asymptotic symmetries and coadjoint orbits {#secas}
==========================================
In this section we will review the analysis of the asymptotic symmetries in the Chern-Simons formulation [@Banados:1998gg] and rephrase these standard results in the language of coadjoint orbits of the Virasoro group, which we will use in section \[secquant\]. We start by using the isomorphism $so(1,2) \sim sl(2,{\mathbb{R}})$ to introduce a new Lie algebra basis $\{ V_0 , V_{\pm 1} \}$ V\_0 &=& - J\_2V\_[1]{} &=& J\_0 J\_1. which obey $sl(2,{\mathbb{R}})$ commutation relations = (m-n) V\_[m+n]{}. Note that in the new basis, the compact generator is $\half ( V_1 + V_{-1} )$. We will take $\calm$ to have the same topology as the global AdS$_3$ manifold, namely that of the solid cylinder ${\mathbb{R}}\times D$. The time coordinate $T$ runs along the length of the cylinder while on the disk $D$ we choose polar coordinates $\r, \f$, where $\f$ has period $2\p$.
Gauge connections $A, \tilde A$ satisfying asymptotically AdS boundary conditions can be gauge-fixed to the form [@Banados:1998gg] A &=& g\^[-1]{} a(x\^+) g dx\^+ + g\^[-1]{} d g, g = e\^[V\_0]{}A &=& g a(x\^-) g\^[-1]{} dx\^- + g d g\^[-1]{}\[partialgf\] where $x^\pm = \f \pm T$ and $a(x^+), \tilde a(x^-)$ are Lie algebra valued functions which are assumed to be in the so-called highest weight gauge a &=& V\_1 - [12c]{} t(x\^+) V\_[-1]{}a &=& V\_[-1]{} - [12c]{} t(x\^-) V\_[1]{}.\[hwgauge\] The functions $t (x^+), \tilde t (x^- )$ parameterize the phase space of flat connections obeying asymptotically AdS boundary conditions. As we will see below, they are to be identified with the left- and right-moving boundary stress tensors. The global AdS$_3$ solution corresponds to taking t\_[AdS]{}= t\_[AdS]{}= - [c 48]{}. \[adscov\]
At fixed time, say $T=0$, the form (\[partialgf\]) is preserved by gauge parameters of the form ł= g\^[-1]{} () g, ł= g () g\^[-1]{}. under which $a, \tilde a$ transform as a = ’ + \[a, \], a = ’ + \[a, \]. Decomposing $\e, \tilde \e$ as () &=& j ()V\_1 + \_0 () V\_0+\_[-1]{} () V\_[-1]{}() &=& j () V\_[-1]{} + \_0 () V\_0+\_[1]{} () V\_[1]{}\[asgauge\] the requirement that the gauge transformations preserve the highest weight gauge (\[hwgauge\]) fixes $\e_0, \e_{-1}$ and $\tilde \e_0, \tilde \e_{1}$ in terms of $j$ and $\tilde j$ respectively. Such transformations should be viewed as infinitesimal asymptotic symmetries, which are therefore parameterized by two functions on the circle $j(\f )$ and $\tilde j (\f )$. Under the action of the infinitesimal asymptotic symmetries $t(\f)$ and $\tilde t(\f)$ transform as \_j t &=& 2 t j ’ + j t’ - [c 24 ]{} j”’\_[j]{} t &=& 2t j ’ + j t’ - [c 24 ]{} j”’ \[transfo\] The conserved charge $q_j$ corresponding to the asymptotic symmetry $j(\f)$ is q\_j(t) &=& \_0\^[2]{} d j () t ()\[Noethercharges\] and similarly for the right-moving charges $\tilde q_{\tilde j}$ (for the rest of this section, we will display only the formulas in the left-moving sector, the right-moving side proceeding analogously). The conserved charges are linear functionals on the phase space and we would like to compute their Poisson brackets. These can be deduced from the fact for every function $\calo$ on the phase space, its variation under $j$ arises from the Poisson bracket with $q_j$: $ \{q_j, \calo \}_{PB}= \d_j \calo$. Taking $\calo$ to be the conserved charge $\calo = q_k$ we find the Poisson bracket { q\_j, q\_k }\_[PB]{} (t) &=& \_0\^[2 ]{} dk \_j t\
&=& q\_[(j’k - j k’)]{}(t) - [c 48 ]{} \_0\^[2 ]{} d(j”’ k - j k”’)\[PBVir\]
Let us review the group theoretic meaning of this expression. If $c$ were zero, this Poisson bracket would realize the Lie algebra $diff (S^1 )$ of reparametrizations of the circle in the sense that { q\_j, q\_k }\_[PB]{} = q\_[\[k,j\]]{} where $[k,j] = j'k - jk'$ is the commutator in $diff (S^1 )$. Therefore $j(\f), k(\f) $ should be thought of as components of tangent vectors $j(\f) \pa_\f,\ k(\f) \pa_\f$. As for $t(\f)$, the expression for the charge (\[Noethercharges\]) allows us to identify $t(\f)$ as an element of the vector space dual to $diff (S^1 )$. The transformation law (\[transfo\]) at $c=0$ defines the coadjoint representation of $diff(S^1 )$ and shows that $t ( \f)$ should be seen as the component of a quadratic differential $t(\f) d\f^2$.
When $c$ is nonzero, the Poisson bracket (\[PBVir\]) instead realizes a central extension of $diff (S^1 )$, the Virasoro algebra, as we shall presently review. Extending $diff (S^1 )$ with a central generator $\hat c$, elements of the Virasoro algebra can be represented as pairs (j(), n) j () \_+ n c where $n$ is a real number. The commutation relations are = ( (j k’ - j’ k ), [1 48 ]{} \_0\^[2]{} d(j”’ k - j k”’ )). Similarly, we extend the dual vector space by including the constant parameter $c$ as an extra coordinate. The dual vector space to the Virasoro algebra consists of pairs of the form $(t(\f ),c)$, and the pairing between adjoint and coadjoint vectors is given by the generalization of the expression (\[Noethercharges\]) for the charge: q\_[(j,n)]{} (t,c) = \_0\^[2]{} d j t + c n (t,c), (j,n)\[pairing\] The transformation law (\[transfo\]) can be extended in such a way that the pairing is Virasoro-invariant, meaning \_[(k,m)]{} (t,c), (j,n)+ (t,c), \[(k,m), (j,n)\]=0 .This leads to \_[(j,n)]{}( t, c) = ( 2 t j ’ + j t’ - [c 24 ]{} j”’,0).\[coadjinf\] This transformation law defines the coadjoint representation of the Virasoro algebra. Using these definitions one finds that the Poisson bracket (\[PBVir\]) can be written as { q\_[(j,n)]{}, q\_[(k,m)]{} }\_[PB]{} (t,c) = q\_[\[(k,m),(j,n)\]]{} (t,c)\[KK\] from which we see that the Poisson bracket (\[PBVir\]) indeed realizes the Virasoro algebra.
The Poisson bracket can be written in a more standard form by choosing the following basis for the charges generating $diff(S^1 )$: l\_m = q\_[(e\^[i m ]{},0)]{}. In this basis the Poisson brackets take the form -i { l\_m, l\_n }\_[PB]{} =(m-n)l\_[m+n]{} +[c m\^3 12]{} \_[m,-n]{}.\[VirPB\] Note that our Virasoro energies $l_0, \tilde l_0$ are naturally defined on the boundary cylinder, and on the global AdS solution (which will turn out to correspond to the $SL(2,{\mathbb{R}})$ invariant vacuum in the dual CFT) they take the values l\_0 ((t\_[AdS]{},c)) = l\_0 ((t\_[AdS]{},c)) = -[c 24]{}. This corresponds to the Casimir energy of a cylinder of circumference $2 \p$. To make contact with the more standard conventions where $l_0$ and $\tilde l_0$ act on the plane one should shift them by ${c \over 24}$.
Extending this analysis to include the right-moving sector, we conclude that the asymptotic charges generate two copies of the Virasoro algebra through Poisson brackets. The combinations $(t(\f), c)$ and $(\tilde t(\f), c)$ transform in the coadjoint representation of the respective Virasoro algebras. At fixed central charge $c$, asymptotically AdS solutions come in families obtained by acting on a given solution with the Virasoro symmetries, which are are referred to as coadjoint orbits of the Virasoro group. Physically, moving around on a coadjoint orbit can be seen as dressing (or undressing) a given solution with boundary graviton excitations. A standard result, which we will use to quantize coadjoint orbits in section \[secquant\], is that each orbit possesses a natural symplectic form and hence a Poisson bracket (the Kirillov-Kostant bracket) such that the restrictions of the charges to the orbit obey (\[KK\]).
Of course, not all coadjoint orbits are physically acceptable, as some of them may correspond to singular gravity solutions, and the issue of regularity will be the subject of the next section. The regular coadjoint orbits are expected to correspond to semiclassical states in the quantum theory, and upon quantizing them we expect to obtain Virasoro representations belonging to the spectrum of the dual CFT.
Conical spaces as topological Chern-Simons solitons {#seccon}
===================================================
In this section we will revisit, in the pure gravity context, the conical solutions which were recently studied in higher spin gravity [@Castro:2011iw]. We will emphasize their similarity to solitons in the sense that they carry a topological winding number, and discuss their smoothness as seen by observables both in the metric and Chern-Simons formulations.
Winding numbers
---------------
Let us study the space of asymptotically AdS solutions on the solid cylinder in more detail. At any fixed time $T$, the flat connections $a, \tilde a$ (see (\[hwgauge\])) on the boundary cylinder can be expressed as a () d = h\^[-1]{} d h, a () d = h\^[-1]{} d h\[puregauge\]where $h(\f), \tilde h (\f)$ are maps from the boundary circle into the gauge group $SO(1,2)$. These are classified by the first homotopy group of $SO(1,2)$, and since $SO(1,2)$ is contractible to its maximal compact subgroup $U(1)$, the first homotopy group is $\p_1 \left(SO(1,2)\right)= {\mathbb{Z}}$. Hence the space of flat connections on the boundary consists of topological sectors labelled by the two winding numbers $s, \tilde s$ of the maps $h (\f ), \tilde h (\f )$, which measure how many times the $U(1)$ directions in $SO(1,2) \times SO(1,2)$ are traversed when we go around the boundary circle. The following are representative maps with winding numbers $s,\tilde s$ m\_s () &=& e\^[[s2]{} (V\_[1]{} + V\_[-1]{})]{} m\_[s]{} () &=& e\^[[s2]{} (V\_[-1]{} + V\_[1]{})]{}. These however don’t give rise, upon substituting in (\[puregauge\]), to connections in the highest weight gauge (\[hwgauge\]). This can be remedied by choosing different representatives related by a constant gauge transformation h\_s () &=& b m\_s () b\^[-1]{} = e\^[ ( V\_1 +[s\^2 4]{} V\_[-1]{})]{}, b = e\^[ V\_0]{}h\_[s]{} () &=& b m\_[s]{} () b\^[-1]{} = e\^[ ( V\_[-1]{} +[s\^2 4]{} V\_[1]{})]{},b = e\^[- V\_0]{}.\[hs\] Substituting in (\[puregauge\]) gives the following highest weight gauge connections[^6] a\_s &=& V\_1 +[s\^2 4]{} V\_[-1]{}a\_[s]{} &=& V\_[-1]{} +[s\^2 4]{} V\_[1]{}. We will label these solutions by their winding numbers $(s, \tilde s)$ in what follows. The $(s, \tilde s)$ solution is characterized by constant covectors $(t_s,c)$ and $(\tilde t_{\tilde s},c)$ with t\_s = - [c s\^2 48 ]{}, t\_[s]{} = - [c s\^2 48 ]{}.From (\[pairing\]) we read off the Virasoro energies l\_0 ((t\_s,c)) = - [c s\^2 24]{}, l\_0 ((t\_[s]{}),c) = - [c s\^2 24]{}. The other Virasoro charges vanish on these solutions. Comparing to (\[adscov\]), we see that the case $s = \tilde s =1$ corresponds to global AdS: it is a somewhat peculiar feature of the Chern-Simons description that the natural classical vacuum of the theory appears in a winding sector. We note that the winding states with $s,\tilde s>1$ have energies below the AdS vacuum energy. Since in a unitary CFT all primaries have conformal weights above the $SL(2,{\mathbb{R}})$ invariant vacuum, this is already an indication that these winding states can only play a role in nonunitary versions of holography, as we shall see in more detail below.
Metric-like observables
-----------------------
Let us now discuss whether the $(s, \tilde s)$ winding solutions are smooth. Since the $\f$-circles (i.e. the curves of constant $T, \r$) are by assumption contractible, our coordinate system is singular at some value of $\r$ (which we call the ‘origin’), where the $\f$ coordinate is ill-defined. Our solutions satisfy the equations of motion everywhere except possibly in the origin. To decide whether the solution is singular in the origin, we will look at suitable gauge-invariant observables which could measure such a singularity. As we already mentioned in the Introduction, the analysis depends on whether we work in the Chern-Simons formulation, where the natural observables come from holonomies of the gauge field, or the metric formulation, where we have at our disposal the standard curvature invariants which are constructed using the inverse metric. Let’s start by addressing smoothness in the metric formulation.
For simplicity we restrict our attention to the solutions where $\tilde s =s$ which have vanishing angular momentum[^7]. The corresponding metrics are, after a shift $ \r \to \r + \ln { s\ \over 2}$, ds\^2\_[(s, s)]{} = l\^2 \[staticdefs\] As anticipated in (\[adscov\]), for $s =1$ we recover the standard global AdS metric. In the limit $s \to 0$ (where we cannot perform the shift of the $\r$-variable) the metric is that of the zero mass, zero angular momentum BTZ black hole ds\^2\_[(0,0)]{} = l\^2 which has energy above that of global AdS.
For $s>1$ the metric (\[staticdefs\]) can be seen to have a curvature singularity in $\r=0$, corresponding to a conical singularity with an excess angle of $2 \p (s-1)$ [@Deser:1983tn],[@Deser:1983nh]. Let’s rederive this fact in a way that emphasizes the difference between the metric and Chern-Simons formulations. In the metric formulation we postulate that the vielbein is invertible. If this is the case, we can construct many gauge-invariant observables besides the eigenvalues of the holonomies (\[hols\]). For example, for every (non-null) two-surface we can define a surface observable () = \_(F\^a + F\^a) e\_a\^n\_\[surfaceobs\] where $n_\m$ is the unit normal to the surface. Now let’s evaluate $\calo (\cald)$ with $\cald$ a disc $0 \leq \r \leq \r_0$ at constant $T$, in the background of the $(s,s)$ solution. For any solution with constant $t = \tilde t$ we can rewrite $\calo (\cald) $ as () = \_(R + [2l\^2]{}) dd\[integr\] with $\tilde g_{\m\n}$ the induced metric on $\cald$ and $\tilde R$ its scalar curvature. Due to the equations of motion the integrand vanishes everywhere, except possibly in the origin, and we can replace (\[integr\]) by () = \_[\_]{} R ddwhere ${\cald_\e} $ is a tiny disc around the origin. The spatial metric near the origin is d s\^2 \~l\^2 which can be written as the flat metric $l^2 du d\bar u$ in terms of the complex coordinate $u = \r e^{i s \f}$. However, the transformation from $(\r,\f)$ and $(u, \bar u)$ is not one-to-one, rather it is one-to-one for the coordinate $v = u^{1/s}$, in terms of which the metric is only conformally flat, with a conformal factor which is singular in the origin: d s\^2 \~l\^2 |v|\^[2(s-1)]{} dvd|v. Hence a useful way to picture the geometry near the origin is as the Riemann surface of the function $u^{1/s}$, i.e. as an $s$-sheeted branched covering over the complex plane. We then evaluate R = (1- s)\_v\_[|v]{} |v|\^2 = 4(1-s) \^[(2)]{}(v,|v) from which we find () = 4(1-s). Hence from the metric point of view only the global AdS solution with $s=1$ is regular while the remaining winding states are not. The surface observable $\calo (\cald)$ which sees the singularity is not a natural observable in the Chern-Simons formulation of the theory as it explicitly involves the inverse vielbein. From the Chern-Simons point of view, the natural observables are rather based on the holonomies of the gauge field which we shall now discuss.
Holonomies
----------
The natural observables in a three-dimensional Chern-Simons theory are related to the holonomies of the gauge fields around a closed curve $\calc$: H ( ) = e\^[\_A]{}, H ( ) = e\^[\_A]{}\[hols\] The eigenvalues of $H ( \calc ) , \tilde H ( \calc ) $ are gauge-invariant and, when the connections $A, \tilde A$ are flat, depend only on the homotopy class of $\calc$. However, they do depend on the representation of the gauge group used to evaluate them. Since we are considering the pure gravity theory, which contains only the gauge fields in the adjoint representation, it’s natural to evaluate the holonomies (\[hols\]) in the adjoint representation, in our case the 3-dimensional representation of $SO(1,2)$. For a smooth flat gauge connection, the holonomy around a contractible loop should be unity. Here we shall take $\calc$ to be a $\f$-circle, which is by assumption contractible, in the $(s, \tilde s)$ winding solution. We find that H &\~& e\^[2 a]{} \~e\^[- 2 i s V\_0]{}H &\~& e\^[2 a]{} \~e\^[- 2 i s V\_0]{}\[holcon\] where in the last step we have used ( V\_1 + V\_[-1]{} ) = M\^[-1]{} ( - i V\_0 ) M with $M = e^{- i \p/4 ( V_1 - V_{-1} )}$. Since the eigenvalues of $V_0$ in the adjoint representation are $1,0$ and $-1$, we see that the holonomy is indeed trivial and all the $(s, \tilde s)$ winding solutions appear to be smooth [@Mansson:2000sj]. Note that, if we replace the adjoint representation with any other finite-dimensional representation of $SO(2,1)$, the holonomy remains trivial as the eigenvalues of $V_0$ are integer in these representations. Hence from the point of view of these observables, all the $(s, \tilde s)$ solutions appear to be regular.
More generally, we can ask if the $(s, \tilde s)$ winding states still appear smooth if we probe them with external matter, for example a spinning point particle probe of mass $M$ and spin $J$. It was argued in [@Witten:1989sx] that such probes are related to holonomies (\[hols\]) evaluated in infinite-dimensional unitary representations of the gauge group. We shall follow here the recent discussion[^8] [@Ammon:2013hba] (see [@Castro:2014tta] for the extension to spinning particles) to which we refer for more details. We will focus on the gauge-invariant Wilson loop operator W\_R () = [tr]{}\_R ( e\^[\_( A + A ) ]{}).\[Wline\] where $\calc$ is a closed loop and $R$ is a representation of the gauge group. We take $R$ to be the infinite-dimensional highest weight representation of $SO(1,2)\times SO(1,2)$ built on a primary state $| h, \tilde h\rangle$: L\_1 | h, h&=& L\_1 | h, h=0L\_0 | h, h&=& h | h, h, L\_0 | h, h= h | h, hThe quantum numbers $h, \tilde h$ are related to the mass and spin of the probe as l M = h + h, J = h - h. The physical meaning of this Wilson loop was found in [@Castro:2014mza] to be as follows. When $\calc$ is noncontractible, the Wilson loop is related to the proper distance along the curve (or a generalization thereof for spinning particles). For example, for $J=0$ and $\calc$ the $\f$-circle in the BTZ black hole background, $W_R (\calc )$ measures the Bekenstein-Hawking entropy. When $\calc$ is contractible, it measures rather the phase picked up by the wavefunction of the test particle[^9] when going around $\calc$, and when this phase is nontrivial it measures a singularity in the gauge field as seen by the probe. For example, for the $\f$-circle in the global AdS background one finds a trivial phase, while for conical defect solutions with $-c/48 \p< t, \tilde t <0$ the phase is nontrivial, signaling a singularity.
Let us now compute the Wilson loop in the background of our conical excess solutions labelled by $(s, \tilde s)$, where we take $\calc$ to be a $\f$-circle. The Wilson loop can be evaluated by representing the trace as a path integral over auxiliary variables, see Appendix E in [@Castro:2014tta] for details in the case of spinning particles. For large $h, \tilde h$, where the point particle approximation is valid, the auxiliary path integral can be approximated by a saddle point contribution with the result W\_R () = e\^[-\_3( (h ł\_- h ł\_) V\_0 ) ]{}\[Wloopspin\] where the trace is taken in the 3-dimensional representation of $SO(1,2)$. The matrices $\l_\f, \tilde \l_\f$ are the eigenvalue matrices of $a_\f$ and $\tilde a_\f$ respectively, which we already determined in (\[holcon\]): ł\_= - i s V\_0, ł\_= - i s V\_0. Substituting in (\[Wloopspin\]) we obtain Let us discuss this result first for the case where the probe particles are bosons, so that $J$ is an integer, with arbitrary mass $M$. From (\[phase\]) we see that their wavefunctions are single-valued only when s = s i.e. the left- and right winding numbers must be equal, which restricts to the non-spinning defects with $l_0 = \tilde l_0$. For fermionic probe particles, $J$ is half-integer and in a regular background the wavefunction should pick up a phase $-1$ when going around the origin. From (\[phase\]) we see that only the conical spaces with odd $s$ appear regular to fermionic probes, which obey the familiar Neveu-Schwarz boundary conditions on the boundary cylinder. The solutions with even $s$ do appear singular to fermionic probes, which experience the insertion of a worldline defect which causes them to obey Ramond boundary conditions on the boundary cylinder. For $s=0$ this reduces to the familiar result that the zero-mass BTZ black hole lives in the Ramond sector of the dual CFT. Since we are in 2+1 dimensions we can also consider probes with fractional spin, $J \in {\mathbb{N}}+ 1/n$, with $n$ a positive integer, whose wavefunctions should pick up a phase $e^{2 \p i\over n}$. From (\[phase\]) we see that only the conical spaces with $s = 1 {\rm \ mod \ } n$ appear regular to such probes.
To summarize, we have found that from the point of view of natural Chern-Simons observables all $(s, \tilde s)$ conical solutions appear regular in pure gravity, while adding bosonic probes selects the non-spinning solutions with $s = \tilde s$. Adding fermionic or fractional spin matter further restricts the allowed values of $s$.
We end this section by commenting on the fate of the conical spaces when working in Euclidean rather than Lorentzian signature. In this case, the Chern-Simons gauge group is to be replaced by $SL(2,{\mathbb{C}})$, which has the property that all closed loops are contractible, $\p_1 (SL(2,{\mathbb{C}}))=0$. In this case, the analytic continuation of the conical solutions still yields smooth Euclidean solutions which however don’t carry any topological winding charge. In fact, independent of the signature, the search for smooth solutions which lie on the orbits of a constant covector yields precisely the $(s,\tilde s)$ solutions and nothing else [@Castro:2011iw]. We should note that there also exist coadjoint orbits which do not contain a constant covector [@Witten:1987ty], whose role in 3D gravity remains to be fully understood (see [@Garbarz:2014kaa] for a further discussion of these solutions).
Quantizing the conical spaces {#secquant}
=============================
In the previous section we have provided some plausibility arguments that the conical spaces are soliton-like solutions which appear smooth to observables in the Chern-Simons formulation, and we would now like to quantize them and identify them with states in a dual CFT. The standard method to quantize solitons is to make a perturbative expansion of the action in small fluctuations around the soliton and proceed to quantize them. This gives the correction to the energy of the solution and the spectrum of excited bound states in an expansion in $\hbar$ (or, in our case, $1/c$) [@Coleman].
Since the Chern-Simons theory is topological all on-shell fluctuations arise from gauge transformations (\[gaugetransf\]) of $A, \tilde A$. However as we reviewed in section \[secas\], after imposing asymptotically AdS boundary conditions the subset of transformations (\[asgauge\]) should rather be seen as generators of global symmetries which change the solution. Therefore the fluctuations we are to quantize come from acting on our solutions with asymptotic symmetries, i.e. from displacements on the coadjoint orbits of the solutions. In order to quantize them we use the fact that each coadjoint orbit carries a well-defined Poisson bracket, the Kirillov-Kostant bracket, under which the Poisson brackets of the Virasoro charges on the orbit satisfy the classical Virasoro algebra (\[VirPB\]).
Upon quantizing this Poisson bracket we expect to obtain representations of the Virasoro group[^10] which we would like to determine[^11] . As we shall see, the main obstacle in carrying out this programme is that the conical orbits have negative directions for the energy $l_0$. For this reason it has so far proved impossible to quantize these orbits by standard methods. We will here propose a semiclassical quantization of the conical orbits which leads to non-unitary highest weight Virasoro representations. The most important property of the resulting representation is that it will turn out to have a null vector at level $s$, and we will make the connection with Kac’s classification of degenerate representations.
Symplectic form on coadjoint orbits
-----------------------------------
We start by briefly reviewing the Kirillov-Kostant symplectic structure on the Virasoro coadjoint orbits, referring to [@Witten:1987ty], [@Alekseev:1988ce] for more details. It will be convenient to parameterize points on the coadjoint orbit of some fixed covector $(t(\f), c)$ by the finite group element which maps $(t(\f), c)$ to the desired point[^12]. In our case the group which acts on the orbits is $Diff (S^1 )$, the group of diffeomorphisms of the circle, i.e. periodic maps F(),F(+ 2) = F() + 2 . The infinitesimal coadjoint action (\[coadjinf\]) integrates to the well-known finite transformation which involves the Schwarzian derivative $S(F)$: (t(), c) ) & & ( t\_F() , c )\
t\_F()&=& t( F()) (F’)\^2 - [c24 ]{} S (F)\
S(F) && [F”’ F’]{} - [32]{} ([F” F’]{})\^2 Combining this with (\[pairing\]) we find the Virasoro charges along a coadjoint orbit l\_n ((t\_F ,c )) = \_0\^[2]{} e\^[i n]{} ( t(F()) (F’)\^2 - [c24 ]{} S (F) ). \[chargesorb\] The Kirillov-Kostant symplectic form at the point $( t_F , c )$ on the orbit of the covector $(t,c)$ can be written as [@Alekseev:1988ce] Ø&=& -(t\_F, c) ,\
&=&\_0\^[2 ]{} d\[sympl1\] where the brackets in the first line denote the pairing introduced in (\[pairing\]). We have here introduced the notation $\d$ to denote the de Rham operator acting on phase space variables, as oppposed to the spacetime de Rham operator $d$. The main property of this symplectic form is that the Poisson brackets of the charges along the orbit (\[chargesorb\]) are guaranteed to satisfy the classical Virasoro algebra (\[VirPB\]). Following [@Alekseev:1988ce], (\[sympl1\]) can be rewritten in a useful simplified form as a total derivative Ø=\_0\^[2 ]{}d( t(F()) F’ F -[c 48 ]{} [F F’]{} ( [F”’ F’]{} - 2 ([F” F’]{})\^2) ).\[sympl\]
Conical spaces and exceptional orbits {#secflucts}
-------------------------------------
Next we want to apply these general formulas to the coadjoint orbits of the conical solutions. We recall that the conical orbits are those of the constant covectors $(t_s,c)$, where t\_s = - [c s\^2 48 ]{}. and $s > 1$.
Let us first analyze the symmetries of the conical orbits. Any covector is trivially left invariant by $\hat c = (0,1)$, and from (\[coadjinf\]) we see that constant covectors are also invariant under the action of $l_0$. The conical covectors are special in that they are left invariant by two more generators; indeed one easily checks from (\[coadjinf\]) that \_[l\_s]{} ((t\_s,c)) =\_[l\_[-s]{}]{} ((t\_s,c))=0. The infinitesimal transformations generated by the charges $l_0, l_{\pm s}$[^13] integrate to finite reparameterizations of the circle of the form [@Balog:1997zz] F() = [1i s]{} , ||\^2 - ||\^2 =1. These form an $s$-fold cover of $SO(2,1)$ which we denote as $SO(2,1)^{(s)}$. Hence the conical orbits have the structure of the coset spaces $ Diff (S^1 ) / SO(2,1)^{(s)}$ and are often referred to as exceptional coadjoint orbits.
Next we would like to compute the Virasoro charges along the conical orbits. We parameterize the elements of $Diff(S^1)$ as F() = + \_[n ]{} f\_n e\^[- in ]{}, and reality of $F(\f)$ implies that f\_m\^\* = f\_[-m]{}.\[Freal\] In view of the symmetries discussed above, the $f_n$ for $n \neq 0, s, -s$ are coordinates on the orbit of $(t_s,c)$ (it will be a good check to see that the Virasoro charges are independent of $f_0, f_s$ and $f_{-s}$). We start with the expression for the energy $l_0$, which using (\[chargesorb\]) reads l\_0 ((t\_[s,F]{},c)) = -[c s\^2 24]{} + [c 12]{} \_[m ]{} m\^2 (m\^2 - s\^2) |f\_m|\^2+ (f\^3 )\[l0orb\]\
It’s immediately clear that there is a major difference between the $s=1$ and $s>1$ cases: from (\[l0orb\]) we see that for $s=1$ the covector $(t_s,c)$ around which we are expanding is a minimum of the energy, while for $s>1$ it is only a saddle point. The fluctuations $f_m$ with $|m|<s$ represent unstable directions of the conical spaces. Moreover, it can be shown that the energy on the conical orbits is unbounded below [@Witten:1987ty].
The other Virasoro charges along the conical orbits are[^14] l\_m ((t\_[s,F]{},c)) &=& [ i c 12]{} m(s\^2-m\^2) f\_m + (f\^2 ); m 0,s,-s\
l\_[s]{} ((t\_[s,F]{},c)) &=& [c 24]{} \_[m ]{} (m\^2 - s\^2) ( ( ms)\^2-s\^2) f\_m f\_[s - m]{} + (f\^3 ).\[Lsorb\] Next we evaluate the expression (\[sympl\]) for the symplectic form: Ø= - [i c 12]{}\_[m ]{} m (m\^2-s\^2) f\_m f\_[-m]{} + ….\[symplpert\] where the dots mean that we have omitted terms of higher order in the $f_n$.
Semiclassical expansion
-----------------------
Now we would like to use the symplectic form (\[symplpert\]) to quantize the conical orbits. We start by introducing new coordinates $a_m, m \neq 0, s, -s,$ on the orbit: a\_m= f\_m + (f\^2)\[alphas\] in terms of which the symplectic form (\[sympl\]) is Ø= - i \_[m ]{} m(m\^2- s\^2) a\_ma\_[-m]{}\[symplorb\]. Here, the $\calo (f^2)$ terms in (\[alphas\]) are chosen such that (\[symplorb\]) is exact without further higher order corrections; it follows from Darboux’s theorem[^15] that this is indeed possible. Note that the reality of $\O$ implies that a\_m\^\* = a\_[-m]{} ,\[realas\] which at leading order order follows from (\[Freal\]). The $a_m$ are related to canonical coordinates $x_m, p_m$, in terms of which $\O = \sum_{m \in {\mathbb{N}}_0 \backslash{ \{ s \} }} dp_m \wedge dx_m$, as follows: a\_[m]{} &=& [1 ]{} ( x\_m i p\_m), 0<m<s\
a\_[m]{} &=& [1 ]{} ( x\_m i p\_m), m>s. The Poisson brackets among the $a_m$ following from (\[symplorb\]) are essentially those of harmonic oscillators, {a\_m, a\_n }\_[PB]{} = [i \_[m,-n]{} m( m\^2 - s\^2)]{} m, n, 0, s\[canPB\] If the Darboux-like coordinates (\[alphas\]) and the expression for the Virasoro charges (\[Lsorb\]) are known, we can in principle express the Virasoro charges in terms of the $a_m$, yielding a perturbation series in powers of $1/c$.
Now we turn to the issue of quantizing the conical coadjoint orbits. The $s=1$ case, which is the orbit of global AdS, can be quantized using standard methods [@Witten:1987ty](it can be given a Kähler structure) and gives the unitary representation based on the $SL(2,{\mathbb{R}})$ invariant vacuum. Hence we recover the standard result that global AdS is dual to the $SL(2,{\mathbb{R}})$ invariant vacuum in the CFT.
On the other hand, the $s>1$ orbits, corresponding to the conical excesses, have so far defied quantization using standard methods [@Witten:1987ty]. This is most likely related to their being only saddle points of the energy: we don’t expect there to be any way of quantizing such solutions in a way that leads to unitary highest weight Virasoro representations. Nevertheless, we will argue that they do give rise to a nonunitary highest weight Virasoro representation acting on a Fock space, which we will construct as a perturbation expansion in $1/c$. This, then, is what we will mean by ‘quantization’ of these orbits.
We start by replacing the classical coordinates $a_m$ on the orbit by quantum operators $A_m$ satisfying commutation relations obtained from (\[canPB\])by sending $-i \{\cdot, \cdot \}_{PB} \to [\cdot, \cdot ]$: = [ \_[m,-n]{} m (m\^2- s\^2)]{}\[cancomm\]. From the classical expressions (\[Lsorb\]) we can expect that the quantum Virasoro generators, which we will denote by $L_m$, can be represented as composite operators in terms of the oscillators $A_m$, with some as yet unknown ordering prescription. Let’s start with the Virasoro energy $L_0$, which to our current level of accuracy can be written as L\_0 = -[c s\^2 24]{}+ n\_0 + \_[m ]{} m\^2 (m\^2 - s\^2) A\_[-m]{} A\_m + ( c\^[-1/2]{} ) \[L0orb2\] where we have introduced an ordering constant $n_0$ of order $c^0$ reflecting the ordering ambiguity in the the quadratic term. From its commutator with $A_{\pm m}$ \[L\_0, A\_[m]{}\] = m A\_[m]{} + ( c\^[-1/2]{} ) we see that, for large $c$, the $A_m$ with $m>0$ lower the $L_0$ eigenvalue, while the $A_m$ with $m<0$ act as raising operators. Note that the oscillators with modes $|m| < s$ have a nonstandard minus sign in both the energy (\[L0orb2\]) and the commutation relations (\[cancomm\]) as a consequence of expanding around a saddle point. These modes are similar to those arising in the matter CFT of string theory from the timelike field $X^0$ with negative kinetic energy. As is the case for that theory, we need to make some choice for the ‘vacuum’ state by stating which of the $A_m$ with $|m| < s$ annihilate it. No choice is free from unpleasant features: we must either give up having an $L_0$ which is bounded below or having only positive norm states. We will here make the choice to keep the energy levels positive by taking our the Fock vacuum $|0\rangle_s$ to be the state annihilated by the lowering operators, A\_m |0\_s =0 m>0 \[defvac\]. By acting on this vacuum with the raising operators, we build up a Fock space in the usual way. We require the inner product on this Fock space to satisfy A\^\_m= A\_[-m]{}reflecting the reality condition (\[realas\]) and assume the ground state to be normalized as $\,_s\langle 0| 0 \rangle_s =1$. For $s>1$, our Fock space contains negative norm states, since we must have $\| A_{-1} |0\rangle_s \|^2=1/(1-s^2)<0$. The problems with normalizability are also evident from the expression for the position space wavefunction \^[0]{}\_[s]{} (x) x| 0\_s \~e\^[- \_[m \_0]{} [sgn ]{} (m-s) x\_m\^2]{}\[vacwav\] i.e. the modes with $m<s$ come with a wrong sign in the Gaussian wavefunction.
Let’s discuss some properties of the expression for the Virasoro generators $L_m$ in terms of the oscillators $A_n$. We will choose to represent the $L_m$ as creation-annihilation normal ordered expressions in the oscillators satisfying $L_m^\dagger = L_{-m}$, introducing unknown normal ordering constants where necessary. These constants are then to be fixed by requiring that the $L_m$ obey the quantum Virasoro algebra \[L\_m,L\_n\] = (m-n) L\_[m+n]{} + [c 12]{} m\^3 \_[m,-n]{}. We will comment on a systematic procedure to achieve this below. Since the $L_m$ are normal ordered and of level $m$ in the oscillators, it follows that we will automatically have L\_m |0\_s =0 m>0 i.e. our Fock space will furnish a highest weight (i.e. primary) representation of the Virasoro algebra, which is however nonunitary because of the negative norm states mentioned above. In the following we will determine precisely which nonunitary representations the conical spaces correspond to.
To the order at which we have been working, only $L_0$ requires the introduction of a normal ordering constant ($n_0$ in (\[L0orb2\])), and the remaining Virasoro generators are given by L\_m &=& - i m(m\^2 - s\^2) A\_m + ( c\^[0]{} ) m, 0, s\[Lmorb2\]\
L\_[s]{} &=& \_[m+ \_s]{} 2\^[-\_[m,0]{}]{} ( m\^2 - [s\^2 4]{} ) ( m\^2 - [9 s\^2 4]{}) A\_[-m ]{} A\_[m ]{} + ( c\^[-1/2]{} )\[Lsorb2\] where we have introduced a constant $\n_s$ which is 0 for $s$ even and $1/2$ for $s$ odd: \_s .
Before turning to the determination of the normal ordering constant $n_0$, we outline our proposed strategy for quantizing the conical orbits in a semiclassical expansion in $1/c$. We have seen that Virasoro generators admit an expansion in powers of $1/c$ in terms of the oscillators $A_m$ (given by our formulas (\[L0orb2\],\[Lmorb2\],\[Lsorb2\]) plus corrections), where we consider the oscillators $A_m$ to be of ‘order $c^0$’ as the commutation relations (\[cancomm\]) are $c$-independent. We introduce the operators $\calf_{m,n}$ measuring the failure of the Virasoro algebra to hold: \_[m,n]{} = \[L\_m,L\_n\] - (m-n) L\_[m+n]{} - [c 12]{} m\^3 \_[m,-n]{}. When expressed in terms of the oscillators $A_m$ these admit an expansion in powers of $1/c$: \_[m,n]{} = \_[/2]{} [ \_[m,n]{}\^[( -2)]{}c\^[- 2]{}]{} . with $ \calf_{m,n}^{(\a)}$ independent of $c$. We expect that, by adjusting a finite number of normal ordering constants up to some order $(1/c)^\b$, it will possible to make the Virasoro commutation relations hold to order $(1/c)^\b$ in the sense that \_[m,n]{}\^[()]{} =0 -2 . Proceeding in this way we build up an oscillator realization of the Virasoro algebra, order by order in $1/c$. Although we don’t have a general proof that this procedure is consistent, we will see that it gives sensible and satisfying results at the first nontrivial order.
It is instructive to check the vanishing of the first few $\calf_{m,n}^{(\a)}$ using (\[L0orb2\],\[Lmorb2\],\[Lsorb2\]) and the commutation relations (\[cancomm\]). The operators $\calf_{m,n}^{(-2)}$ and $\calf_{m,n}^{(-3/2)}$ vanish trivially as they are given by commutators with the c-number term in $L_0$. For the operators $\calf_{m,n}^{(-1)}$, $\calf_{m,0}^{(-1/2)}$ and $\calf_{m,s}^{(-1/2)}$ one finds \_[m,n]{}\^[(-1)]{} &=& -[1 12]{} m (m\^2 -s\^2) n (n\^2 -s\^2) \[A\_m, A\_n\] - [m 12]{} (m\^2 - s\^2)\_[m,-n]{} \_[m,0]{}\^[(-1/2)]{} &=& -[i ]{} m (m\^2 -s\^2) \_[n ]{} n\^2 (n\^2-s\^2) \[A\_m, A\_[-n]{} A\_[n]{}\] + [i m\^2(m\^2-s\^2) ]{} A\_m\_[m,s]{}\^[(-1/2)]{} &=& -[i ]{} m (m\^2 -s\^2) \_[n + \_s]{} 2\^[-\_[n,0]{}]{} ( n\^2 - [s\^2 4]{} ) ( n\^2 - [9 s\^2 4]{}) \[A\_m, A\_[-n + [s 2]{}]{} A\_[n + [s 2]{}]{} \] &&+[i ]{}(m\^2-s\^2)((m+s)\^2 - s\^2) A\_[m+s]{} . These operators do not involve any normal ordering constants and hence their vanishing is guaranteed because the classical Virasoro algebra holds by construction. It is straightforward to verify that they indeed vanish upon using (\[cancomm\]) and the commutator = [\_[m,n- [s 2]{}]{} + \_[m,- n- [s 2]{}]{}m(m\^2-s\^2)]{} A\_[m+ s]{} which follows from (\[cancomm\]).
One-loop energy correction
--------------------------
The first nontrivial check on our proposed method to quantize the conical orbits comes at order $c^0$, where the operator $\calf_{s,-s}^{(0)}$ involves the normal ordering constant $n_0$ in (\[L0orb2\]). Since $c$ multiplies the action (\[SCS\]), $1/c$ is the loop-counting parameter in our theory and the constant $n_0$, which is of order $c^0$, can be interpreted as the 1-loop correction to the energy of the conical solutions. The quickest way to determine $n_0$ is to impose that the expectation value of $\calf_{s,-s}^{(0)}$ in the Fock vacuum vanishes: \_s0 | \_[s,-s]{}\^[(0)]{} |0\_s =0. This leads to n\_0 = [1 2 s]{} \_s0 | \[ L\_s’, L\_[-s]{}’ \]|0\_s.\[novac\] where $ L_s'$ ( $L_{-s}'$ ) is the piece of $L_s$ ($L_{-s}$) in (\[Lsorb2\]) which involves two annihilation (creation) operators: L\_[s]{}’= \_[m= \_s]{}\^[s/2 -1]{} 2\^[-\_m,0]{} ( m\^2 - [s\^2 4]{} ) ( m\^2 - [9 s\^2 4]{}) A\_[-m ]{} A\_[m ]{}.\[Lprime\] From this expression we can evaluate (\[novac\]) using (\[cancomm\]) and obtain n\_0 &=& \_[m= \_s]{}\^[s/2 -1]{} 2\^[-\_m,0]{} ( [9 s 8]{} - [m\^2 2 s]{})\
&=& [1 24]{} (s-1) (1 + 13 s)\[n0\].
It remains to check that, with the value (\[n0\]) of $n_0$, the operator $\calf_{s,-s}^{(0)}$ indeed vanishes. It is given by \_[s,-s]{}\^[(0)]{} &=& \_[m,n + \_s]{}( 2\^[-\_[m,0]{}-\_[n,0]{}]{} ( m\^2 - [s\^2 4]{} ) ( m\^2 - [9 s\^2 4]{}) ( n\^2 - [s\^2 4]{} ) ( n\^2 - [9 s\^2 4]{}).&&.\[ A\_[-m + [s 2]{}]{} A\_[m + [s 2]{}]{}, A\_[-n - [s 2]{}]{} A\_[n - [s 2]{}]{}\]) -2 s n\_0 - 2 s \_[m ]{} m\^2 (m\^2 - s\^2) A\_[-m]{} A\_m and can be shown to vanish upon using $$\begin{gathered}
[ A_{-m + {s \over 2}} A_{m + {s \over 2}}, A_{-n - {s \over 2}} A_{n - {s \over 2}}]=
\left( { A_{-m+ {s\over 2}} A_{m- {s\over 2}} \over (m + {s \over 2}) \left( (m+{s \over 2})^2 - s^2 \right) }
+ { A_{-m- {s\over 2}} A_{m+ {s\over 2}} \over (-m + {s \over 2}) \left( (m-{s \over 2})^2 - s^2 \right) } \right) \d_{m,n}\nonu+ \left({ A_{m+ {s\over 2}} A_{-m- {s\over 2}} \over (-m + {s \over 2}) \left( (m-{s \over 2})^2 - s^2 \right) }+
{ A_{m- {s\over 2}} A_{-m+ {s\over 2}} \over (m + {s \over 2}) \left( (m+{s \over 2})^2 - s^2 \right) } \right)\d_{m,-n}.\end{gathered}$$ Similarly one verifies that $\calf_{s,0}^{(0)}$ is zero as well. To check the vanishing of the remaining order $c^0$ operators $\calf_{m,n}^{(0)}$ would require going to the next order in our expansion (\[Lmorb2\]) and would involve determining further integration constants, which we will leave for further study.
Summarizing, we have computed the Virasoro energy of the conical spaces to 1-loop order to be \[1loopen\]
Null states and Kac’s degenerate representations
------------------------------------------------
We have seen in section \[secflucts\] that the coadjoint orbits of the conical solutions are rather special in that they have a three parameter isotropy group $SO(2,1)^{(s)}$. As a consequence, the classical Virasoro charges $l_0,l_s,l_{-s}$ on the coadjoint orbit can be expressed in terms of the other $l_m$, see (\[Lsorb\]). It is natural to expect that also on the quantum level the corresponding Virasoro representation is such that $L_0,L_s,L_{-s}$ can be expressed in terms of the remaining $L_m$ [@Witten:1987ty]. This property is the hallmark of a Virasoro representation which contains a null vector at level $s$: acting with the vanishing combination of $L_{-s}$ and the other generators on the Fock vacuum we obtain a null vector, and conversely, in a highest weight representation with a null vector at level $s$ it is possible to express $L_0, L_{\pm s}$ as a function of the remaining $L_m$ (see [@Witten:1987ty] for a proof).
For definiteness, from (\[Lsorb2\]) we see that, to our level of accuracy, we can express $L_0,L_s,L_{-s}$ as L\_0 &=& - [c s\^2 24]{} + n\_0+ [12 c]{}\_[m\\{0,s}]{} [L\_[-m]{} L\_m m\^2 - s\^2]{} + ( c\^[-3/2]{} )\
L\_[s]{} &=& [12c]{} \_
[cc]{}\
[2\^[-\_[m,0]{}]{} L\_[-ms/2]{} L\_[ms/2]{} m\^2- [s\^2 4]{} ]{} + (c\^[-3/2]{} )\[LsitoLm\] The null vector obtained by acting with (\[LsitoLm\]) on the Fock vacuum is, after a shift in the summation variable and a reordering of terms: This expression is the operator version of a classical formula derived in [@Witten:1987ty]. The second term is again (minus) the operator $L_{-s}'$ defined in (\[Lprime\]).
Virasoro representations containing a null vector are referred to as degenerate and were classified by Kac [@Kac:1978ge] (see also [@Feigin:1981st]). We will now identify precisely which Kac representations the quantized conical spaces correspond to. Kac’s result can be summarized as follows: for any value of the central charge $c$, there is a degenerate representation for every pair of nonzero natural numbers[^16] $(r,s)$. It contains a null vector at level $r s$ and is based on a primary of weight h\_[(r,s)]{} = - -+ (13-c) (r\^2+s\^2) + (r\^2-s\^2). The degenerate representations are usually only considered at small values of the central charge, $0<c<1$, since only in that regime they have a conformal weights above the vacuum and stand a chance of being unitary. However, they exist as nonunitary representations also for $c$ outside this range. Note that, for $r\neq s$, $h_{(r,s)}$ becomes complex in the range $1<c<25$ and is real again for $c\geq 25$. Comparing the large $c$ expansion h\_[(r,s)]{} \~-+ (-12 r s+13 s\^2-1)+ (c\^[-1]{} ) with (\[1loopen\]) leads to the unambiguous identification of the quantized orbit of the conical space labelled by $s$ with the degenerate representation of type $(1,s)$ of the left-moving Virasoro algebra. The conformal weight of the representations $h_{(1,s)}$ is plotted as a function of $c$ in figure \[1sfig\].
(100,150) (0,0)[![The conformal weights $h_{(1,s)}$ for $1 \leq s \leq 6$ as a function of the central charge $c$. The red lines correspond to $c=1$ and $c=25$, between which the conformal weights become complex.[]{data-label="1sfig"}](1sreps.pdf "fig:"){height="150pt"}]{} (45,150)[$c$]{} (90,100)[Re $h$]{} (-10,152)[Im $h$]{}
Similarly, the right-moving winding number $\tilde s$ corresponds to the representation of type $(1,\tilde s)$ of the right-moving Virasoro algebra.
A further check comes from comparing the null vectors: the explicit form of the null vector at level $s$ of the $(1,s)$ degenerate representation was first derived by Benoit and Saint-Aubin [@Benoit:1988aw]. Their expression (see (2.7) in [@Benoit:1988aw]) is written as a sum over partitions of $s$, and the leading contributions at large $c$ come from the partitions with one and two elements. It is straightforward to check that these contributions coincide with our expression (\[nullvectors\]). This identification between the conical states and $(1,s)$ degenerate representations was already proposed in [@Perlmutter:2012ds] based on classical considerations in the bulk, and we have hereby confirmed agreement also on the 1-loop level.
Discussion
==========
We end with some comments and directions for future work.
- We argued that the conical spaces play a role in nonunitary versions of holography and are dual to primaries of degenerate Virasoro representations. The origin of the nonunitary behaviour came from expanding around a saddle point of the Virasoro energy. This seems to be the generic situation in the known consistent nonunitary quantum field theories. For example in Vafa’s nonunitary holographic theories [@Vafa:2014iua], which are based on large $N$ gauge theories where the gauge group is a supergroup, the classical ground state in the super matrix model description also has negative energy directions. We also want to remark that our Fock vacua $|0\rangle_s$ seem rather similar to the Kodama state in 4D quantum gravity [@Kodama:1990sc], which is a saddle point of the Hamiltonian as was pointed out in [@Witten:2003mb]. Our results suggest that this state may yet play a role in examples of nonunitary holography.
- We computed the 1-loop energy correction for conical spaces from the bulk perspective using operator methods, by quantizing the natural Poisson bracket on coadjoint orbits. It would be instructive to rederive this result from a 1-loop determinant in the path integral formalism. Here the challenge is to derive the action governing the boundary graviton fluctuations. A natural guess for this action is the so-called geometric action [@Alekseev:1988ce] which reproduces the Poisson bracket on the coadjoint orbits by construction, and which should be obtainable from the original Chern-Simons action (some evidence for this appears in [@Alekseev:1988ce],[@Bershadsky:1989mf]).
- The method outlined in section \[secquant\] allows one to derive a perturbative expression for the null vector in the degenerate representation in a large $c$-expansion. The expression for the exact null vectors [@Benoit:1988aw] can similarly be derived from a large $c$ expansion [@Bauer:1991ai], but the relation between the two approaches is as yet unclear. Clarifying this link would be especially interesting since the knowledge of the exact null vector implies knowledge of the energy correction to all orders.
- We provided evidence that the Chern-Simons formulation of 3D gravity accommodates the degenerate Virasoro representations of the type $(1,s)$. While a CFT containing only these primaries does have a spectrum which is closed under OPE’s, our experience with minimal model CFTs at $c<1$ suggests that such a theory cannot be modular invariant and that this requires the inclusion of the more general $(r,s)$ representations. It therefore natural to ask what type of matter has to be added to pure gravity in order to get all $(r,s)$ representations in the spectrum. There is evidence [@Perlmutter:2012ds] that the theory that accomplishes this is the Prokushkin-Vasiliev theory [@Prokushkin:1998bq] at the value $\l =2$ of the vacuum parameter, where it reduces to gravity coupled to a somewhat unusual scalar field, whose excitations around the conical spaces give the $(r,s)$ representations. In fact, our computation of the energy shift (\[1loopen\]) was the missing ingredient in the matching of the bulk 1-loop partition function to the partition function of such a CFT at large $c$. It is an interesting open question whether the full CFT partition function including $1/c$ corrections is (presumably in some formal sense) modular invariant, and whether it can be reproduced from the bulk side.
- The degenerate Virasoro representations display various interesting features at small values of the central charge ($0<c<1$), which would be very interesting to understand physically as strong coupling phenomena in the bulk[^17]. Examples are the fact that some of the $(r,s)$ representations can become equivalent to each other and become unitary at special values of the central charge in the strong coupling regime $0<c<1$. For those special values of $c$ a truncation of the spectrum of $(r,s)$ representations gives a unitary modular invariant theory, leading to the Virasoro minimal models, some of which may be interpretable as strongly coupled gravity theories [@Castro:2011zq].
- As already mentioned in the Introduction, the conical solutions in pure gravity which we studied in this paper are the more tractable cousins of similar solutions[^18] [@Castro:2011iw] in higher spin theories with asymptotic $W$-symmetry [@Henneaux:2010xg],[@Campoleoni:2010zq]. It would be interesting to extend the regularity arguments of section \[seccon\] using point particle probes to the higher spin case, as well to understand whether they, too, are singular in the metric-like formulation of the theory [@Campoleoni:2012hp]. For the higher spin conical solutions, a similar holographic interpretation as degenerate nonunitary representations of the $W$-algebra was proposed in [@Perlmutter:2012ds]. The arguments for this identifications were purely classical, and it would be interesting to generalize our method to compute quantum corrections to these solutions. While the concepts of finite $W$-symmetry transformations and coadjoint orbits are certainly much less understood than their Virasoro counterparts, we do feel that it should be possible to extend the perturbative setup of the current work the higher spin case.
Acknowledgements {#acknowledgements .unnumbered}
================
I would like to thank Andrea Campoleoni, Frederik Denef, Rajesh Gopakumar, Gustavo Lucena Gómez and Tomáš Procházka for useful discussions. This research was supported by the Grant Agency of the Czech Republic under the grant 14-31689S. I also want to acknowledge support from the collaborative grant WBI/14-1 for a visit to ULB Brussels which sparked some of the questions that led to this work.
[10]{}
J. D. Brown and M. Henneaux, “Central Charges in the Canonical Realization of Asymptotic Symmetries: An Example from Three-Dimensional Gravity,” Commun. Math. Phys. [**104**]{} (1986) 207. J. M. Maldacena, “The Large N limit of superconformal field theories and supergravity,” Int. J. Theor. Phys. [**38**]{}, 1113 (1999) \[Adv. Theor. Math. Phys. [**2**]{}, 231 (1998)\] \[hep-th/9711200\]. M. Banados, C. Teitelboim and J. Zanelli, “The Black hole in three-dimensional space-time,” Phys. Rev. Lett. [**69**]{}, 1849 (1992) \[hep-th/9204099\]. A. Strominger, “Black hole entropy from near horizon microstates,” JHEP [**9802**]{}, 009 (1998) \[hep-th/9712251\]. E. Witten, “Three-Dimensional Gravity Revisited,” arXiv:0706.3359 \[hep-th\]. M. R. Gaberdiel, “Constraints on extremal self-dual CFTs,” JHEP [**0711**]{}, 087 (2007) \[arXiv:0707.4073 \[hep-th\]\]. A. Maloney and E. Witten, “Quantum Gravity Partition Functions in Three Dimensions,” JHEP [**1002**]{}, 029 (2010) \[arXiv:0712.0155 \[hep-th\]\]. A. Achucarro and P. K. Townsend, “A Chern-Simons Action for Three-Dimensional anti-De Sitter Supergravity Theories,” Phys. Lett. B [**180**]{}, 89 (1986). E. Witten, “(2+1)-Dimensional Gravity as an Exactly Soluble System,” Nucl. Phys. B [**311**]{}, 46 (1988). S. Deser, R. Jackiw and G. ’t Hooft, “Three-Dimensional Einstein Gravity: Dynamics of Flat Space,” Annals Phys. [**152**]{}, 220 (1984). S. Deser and R. Jackiw, “Three-Dimensional Cosmological Gravity: Dynamics of Constant Curvature,” Annals Phys. [**153**]{} (1984) 405. J. M. Izquierdo and P. K. Townsend, “Supersymmetric space-times in (2+1) adS supergravity models,” Class. Quant. Grav. [**12**]{}, 895 (1995) \[gr-qc/9501018\]. T. Mansson and B. Sundborg, “Multi - black hole sectors of AdS(3) gravity,” Phys. Rev. D [**65**]{}, 024025 (2002) \[hep-th/0010083\]. V. Balasubramanian, J. de Boer, E. Keski-Vakkuri and S. F. Ross, “Supersymmetric conical defects: Towards a string theoretic description of black hole formation,” Phys. Rev. D [**64**]{}, 064011 (2001) \[hep-th/0011217\].
E. Witten, “Topology Changing Amplitudes in (2+1)-Dimensional Gravity,” Nucl. Phys. B [**323**]{}, 113 (1989).
M. Ammon, A. Castro and N. Iqbal, “Wilson Lines and Entanglement Entropy in Higher Spin Gravity,” JHEP [**1310**]{}, 110 (2013) \[arXiv:1306.4338 \[hep-th\]\]. D. Grumiller, W. Riedler, J. Rosseel and T. Zojer, “Holographic applications of logarithmic conformal field theories,” J. Phys. A [**46**]{}, 494002 (2013) \[arXiv:1302.0280 \[hep-th\]\]. C. Vafa, “Non-Unitary Holography,” arXiv:1409.1603 \[hep-th\].
A. Castro, R. Gopakumar, M. Gutperle and J. Raeymaekers, “Conical Defects in Higher Spin Theories,” JHEP [**1202**]{}, 096 (2012) \[arXiv:1111.3381 \[hep-th\]\]. E. Perlmutter, T. Prochazka and J. Raeymaekers, “The semiclassical limit of W$_N$ CFTs and Vasiliev theory,” JHEP [**1305**]{}, 007 (2013) \[arXiv:1210.8452 \[hep-th\]\]. E. Hijano, P. Kraus and E. Perlmutter, “Matching four-point functions in higher spin $AdS_3/CFT_2$,” JHEP [**1305**]{}, 163 (2013) \[arXiv:1302.6113 \[hep-th\]\]. A. Campoleoni and S. Fredenhagen, “On the higher-spin charges of conical defects,” Phys. Lett. B [**726**]{}, 387 (2013) \[arXiv:1307.3745\]. S. Coleman, “Classical lumps and their quantum descendants,” in Aspects of Symmetry, Cambridge University Press 1985.
V.F. Lazutkin,T.F. Pankratova, “Normal forms and versal deformations for Hill’s equation” Funkts. Anal. Prilozh. **9**, 41 (1975)
G. Segal, “Unitarity Representations of Some Infinite Dimensional Groups,” Commun. Math. Phys. [**80**]{}, 301 (1981).
E. Witten, “Coadjoint Orbits of the Virasoro Group,” Commun. Math. Phys. [**114**]{}, 1 (1988). A. Alekseev and S. L. Shatashvili, “Path Integral Quantization of the Coadjoint Orbits of the Virasoro Group and 2D Gravity,” Nucl. Phys. B [**323**]{}, 719 (1989). M. Bershadsky and H. Ooguri, “Hidden SL(n) Symmetry in Conformal Field Theories,” Commun. Math. Phys. [**126**]{}, 49 (1989). J. Balog, L. Feher and L. Palla, “Coadjoint orbits of the Virasoro algebra and the global Liouville equation,” Int. J. Mod. Phys. A [**13**]{}, 315 (1998) \[hep-th/9703045\]. A. Garbarz and M. Leston, “Classification of Boundary Gravitons in AdS$_3$ Gravity,” JHEP [**1405**]{}, 141 (2014) \[arXiv:1403.3367 \[hep-th\]\]. G. Barnich and B. Oblak, “Holographic positive energy theorems in three-dimensional gravity,” Class. Quant. Grav. [**31**]{}, 152001 (2014) \[arXiv:1403.3835 \[hep-th\]\].
V. G. Kac, “Contravariant Form for Infinite Dimensional Lie Algebras and Superalgebras,” In \*Austin 1978, Proceedings, Group Theoretical Methods In Physics\*, Berlin 1979, 441-445
E. Witten, “Quantization of [Chern-Simons]{} Gauge Theory With Complex Gauge Group,” Commun. Math. Phys. [**137**]{}, 29 (1991).
M. Banados, “Three-dimensional quantum geometry and black holes,” AIP Conf. Proc. [**484**]{}, 147 (1999) \[hep-th/9901148\].
J. de Boer and J. I. Jottar, “Entanglement Entropy and Higher Spin Holography in AdS$_3$,” JHEP [**1404**]{}, 089 (2014) \[arXiv:1306.4347 \[hep-th\]\].
A. Castro and E. Llabrés, “Unravelling Holographic Entanglement Entropy in Higher Spin Theories,” arXiv:1410.2870 \[hep-th\].
A. Castro, S. Detournay, N. Iqbal and E. Perlmutter, “Holographic entanglement entropy and gravitational anomalies,” JHEP [**1407**]{}, 114 (2014) \[arXiv:1405.2792 \[hep-th\]\]. A.A. Kirillov, “Lectures on the Orbit Method,” Graduate Studies in Mathematics Vol. 64, American Mathematical Society 2004.
B. L. Feigin and D. B. Fuks, “Invariant skew symmetric differential operators on the line and verma modules over the Virasoro algebra,” Funct. Anal. Appl. [**16**]{}, 114 (1982) \[Funkt. Anal. Pril. [**16**]{}, 47 (1982)\]. L. Benoit and Y. Saint-Aubin, “Degenerate Conformal Field Theories and Explicit Expression for Some Null Vectors,” Phys. Lett. B [**215**]{}, 517 (1988). H. Kodama, “Holomorphic Wave Function of the Universe,” Phys. Rev. D [**42**]{}, 2548 (1990). E. Witten, “A Note on the Chern-Simons and Kodama wave functions,” gr-qc/0306083.
M. Bauer, P. Di Francesco, C. Itzykson and J. B. Zuber, “Covariant differential equations and singular vectors in Virasoro representations,” Nucl. Phys. B [**362**]{}, 515 (1991). S. F. Prokushkin and M. A. Vasiliev, “Higher spin gauge interactions for massive matter fields in 3-D AdS space-time,” Nucl. Phys. B [**545**]{}, 385 (1999) \[hep-th/9806236\]. A. Castro, M. R. Gaberdiel, T. Hartman, A. Maloney and R. Volpato, “The Gravity Dual of the Ising Model,” Phys. Rev. D [**85**]{}, 024032 (2012) \[arXiv:1111.1987 \[hep-th\]\]. M. Henneaux and S. J. Rey, “Nonlinear $W_{infinity}$ as Asymptotic Symmetry of Three-Dimensional Higher Spin Anti-de Sitter Gravity,” JHEP [**1012**]{}, 007 (2010) \[arXiv:1008.4579 \[hep-th\]\].
A. Campoleoni, S. Fredenhagen, S. Pfenninger and S. Theisen, “Asymptotic symmetries of three-dimensional gravity coupled to higher-spin fields,” JHEP [**1011**]{}, 007 (2010) \[arXiv:1008.4744 \[hep-th\]\].
A. Campoleoni, S. Fredenhagen, S. Pfenninger and S. Theisen, “Towards metric-like higher-spin gauge theories in three dimensions,” J. Phys. A [**46**]{}, 214017 (2013) \[arXiv:1208.1851 \[hep-th\]\].
[^1]: Known nonunitary examples of holographic duality include the gravity/logarithmic CFT duality (see [@Grumiller:2013at] for a review and further references), and the large $N$ gauge theories based on a supergroup which were recently discussed in [@Vafa:2014iua].
[^2]: At first sight, the pure gravity conical spaces may seem metrically more singular than their cousins in higher spin theories, since for the latter one can always apply a higher spin gauge gauge transformation such that the metric has no curvature singularity [@Castro:2011iw]. However we should keep in mind that metric curvature invariants are not good observables in higher spin theories; for example one can construct higher spin solutions where the metric appears in some gauges to be regular yet the Chern-Simons gauge field describing them is singular.
[^3]: Indices are raised with the metric $\h_{ab} = diag (-1,1,1)$ and $\e_{012} \equiv 1$.
[^4]: Some useful properties are $\tr_3 (J_a J_b) = 2 \h_{ab},
\tr_3 (J_a J_b J_c) = \e_{abc}$.
[^5]: As explained in [@Witten:2007kt], if we replace the gauge group by an $n$-fold cover of $SO(1,2)$, $k$ is quantized in units of $n^{-2}$ and $c$ is a multiple of $24/n^2$. For example taking the gauge group to be $SL(2, {\mathbb{R}})$, $c$ should be a multiple of $6$.
[^6]: Here we encounter the subtlety that (\[hs\]) is a good gauge transformation only when both $s$ and $\tilde s$ are positive. Since we can flip the sign of both $s$ and $\tilde s$ by sending $\f \to - \f$, we can only transform the maps with $s \tilde s >0$ to the highest weight gauge. We will assume from now on that $s$ and $\tilde s$ are positive.
[^7]: When $s \neq \tilde s$, the metric is that of a spinning conical defect: $$ds^2_{(s,\tilde s)} = l^2 \left[ d\r^2 + {s \tilde s\over 2} \cosh 2\r dx^+ dx^- - {s^2 \over 4 } (dx^+)^2 - {\tilde s^2 \over 4} (dx^-)^2\right]$$ which apart from a curvature singularity also contains closed timelike curves.
[^8]: See also [@deBoer:2013vca] for a closely related description of point particles in the Chern-Simons formulation, and [@Castro:2014mza] for the precise relation between the two descriptions.
[^9]: An argument for this goes as follows. Let’s take $\calc$ to be a $\f$-circle and consider the wavefunction of a position eigenstate, i.e. the propagator $\langle \r,\f, T|, \r',\f', T'\rangle$, which can be represented as a sum over paths weighted by (\[Wline\]). When sending $\f \to \f + 2 \p$, each contributing path $\calp$ gets deformed to a new path homotopic to $\calp + \calc$. Hence the phase picked up by the propagator is given by the Wilson loop around $\calc$.
[^10]: See [@Kirillov] for a general perspective on coadjoint orbits and representation theory.
[^11]: To be more precise, since our setup will be perturbative around the conical solutions, we will obtain representations of the Virasoro algebra rather than the Virasoro group.
[^12]: Once again we focus in this section on the left-moving sector, the right-moving sector being analogous.
[^13]: This symmetry algebra was referred to as a ‘twisted $sl(2,{\mathbb{R}})$’ in [@Perlmutter:2012ds].
[^14]: The second expression requires some explanation. Applying (\[chargesorb\]) one obtains l\_[s]{} ((t\_[s,F]{},c)) = [c 24]{} \_[m ]{} m (m s) ( m\^2 ms - 3 s\^2) f\_m f\_[s - m]{}+ (f\^3 ). One can check this expression does not depend on $f_0, f_s$ and $f_{-s}$, which can be made more explicit by adding zero in the form 0 = \_[m ]{} m (m s) ( m ) s f\_m f\_[s -m]{} after which one obtains (\[Lsorb\]).
[^15]: Or rather an infinite-dimensional generalization thereof explained in [@Witten:1987ty], section 4.
[^16]: For small values of the central charge some of those representations can become equivalent, but this doesn’t play a role in the large $c$ regime we are considering.
[^17]: As we reviewed in section \[secCS\], with our choice of gauge group $c$ is quantized in units of 24, so to consider small values of $c$ we should either consider a covering group or work in Euclidean signature, where $c$ is not quantized.
[^18]: The interpretation as topological solitons characterized by a winding number seems to be special to the pure gravity case, since the higher spin theories are based on the gauge groups $SL(N,{\mathbb{R}})$ (for the Lorenzian theory) or $SL(N,{\mathbb{C}})$ (for the Euclidean theory), which have $\p_1\left(SL(N,{\mathbb{R}})\right) = {\mathbb{Z}}_2$ and $\p_1\left(SL(N,{\mathbb{C}})\right) = 1$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The large two-pion exchange amplitudes are calculated in HB$\chi$PT and their net contribution to the reaction cross section is large.'
---
[^1]
[*Department of Physics and Astronomy\
University of South Carolina, Columbia, SC 29208* ]{}
Refs. [@pmmmk96; @cfmv96] evaluated this reaction at threshold in HB$\chi$PT. They found that the impulse approximation (I.A.) and one-pion-exchange (Resc) diagrams interfere destructively resulting in a very small reaction cross section. According to Lee and Riska [@lr93] a model explanation of the measured reaction cross section near threshold requires the contributions from heavy meson ($\sigma$ and $\omega$) exchange in addition to the one-pion-exchange. The $\sigma$-meson-exchange is more properly described by correlated two-pion-exchange. This knowledge prompted a HB$\chi$PT study of two-pion-exchange (TPE) contributions to the $pp\to pp\pi^0$ reaction amplitude [@dkms99].
In Weinberg chiral counting the TPE contributions are of higher chiral order in HB$\chi$PT. However, it was shown by Ref. [@dkms99] that some TPE amplitudes are as large or larger than the lower chiral order Resc contribution. At threshold the typical momentum is $p \sim \sqrt{m_\pi m_N}$. This large momentum prompted Cohen [*et al.*]{} [@cfmv96] to propose a momentum counting rule, reviewed in Ref. [@hanhart04]. According to this counting, one finds that the Resc diagram is higher order in $p/\Lambda$ compared to some “dominant" TPE diagrams, and this counting agrees with the numerical evaluations of the TPE diagrams of Ref. [@dkms99]. One drawback with the momentum counting is that the sum of the diagrams in each “momentum order" no longer is independent of the definition of the pion field. Hanhart and Kaiser (HK) [@hk02] used momentum counting to evaluate the “leading" momentum behavior of the dominant TPE diagrams. HK also found that diagram II in Ref. [@dkms99] should have opposite sign (which we confirmed), and they found that the sum of the leading momentum behavior of the three dominant TPE diagrams cancel. We will present results from Ref. [@dkms99] and a recent calculations [@kkms06] which show that the “sub-leading" parts of a dominant TPE diagram gives a contribution comparable to the Resc amplitude.
The TPE transition operators (TO) were evaluated analytically by Ref. [@dkms99] in HB$\chi$PT. When these operators are sandwiched between phenomenologically determined distorted $NN$ wave functions, the momentum integrals converge slowly [@kkms06]. This slow convergence can be understood when we adopt the threshold fixed kinematics approximation (FKA). Imposing FKA on the analytic expressions for the TO given in Ref. [@dkms99], we make an asymptotic expansion in the two-nucleon momentum transfer ($ |\vec{k} |
= |\vec{p}-\vec{p}^{\; \prime}|
\to \infty $). The TO matrix $T$ of the TPE diagrams is of the form $$\begin{aligned}
T &=& \left( \frac{g_A}{f_\pi} \right)
\left( \vec{\Sigma}\cdot \vec{k} \right)
t(p,p^\prime , x)\end{aligned}$$ where $x=\hat{p}\cdot\hat{p}^\prime$. The [*asymptotic*]{} momentum behavior for $ t(p,p^\prime , x)$ is $
t(p,p^\prime , x) \sim
t_1 \; | \vec{k} | + t_2 \;
{\rm ln}[\Lambda^2/ |\vec{k} |^2 ]
+ t_3 + \delta t(p,p^\prime , x) ,
$ where $\delta t(p,p^\prime , x) $ is ${\cal O}(|\vec{k} |^{-1}) $, and the amplitudes $t_i$, $i=1, 2, 3$ are known analytic expressions for each diagram. The $t_1$ amplitude is the dominant TPE amplitude of HK [@hk02]. $$\begin{array}{|l| |r| r|r|r|r|r|r|}
\hline
{\rm Amplitude \; K=} & I&II&III&IV&V&VI&VII \\
\hline
R_K & -.70& -6.70& -6.70 & 9.50 &0.18&0.14&2.65 \\
\hline
t_1 \; \; \; \; \; \propto & - & -2 & -1
& +3 &- &- & - \\
\hline
\end{array}$$ In the table the row marked $R_K$, gives the values of the ratio of TO to the Resc amplitude in the plane wave approximation for the seven amplitudes of Ref. [@dkms99]. As indicated in the last row of the table, marked $t_1$, the leading momentum terms of the TO from diagrams II, III and IV sum to zero, confirming HK’s result [@hk02]. The non-cancellation of the dominant amplitudes can however be inferred from the $R_K$ row since the ratios II:III:IV are not -2:-1:3 but roughly -2:-2:3. When we $t_1$, we find that the sum of the two-pion-exchange amplitudes is larger than the Resc amplitude [@kkms06].
This research is supported in part by a grant from NSF.
[99]{}
B.-Y. Park [*et al.*]{}, Phys. Rev. [**C 53**]{}, 1519 (1996)
T.D. Cohen [*et al.*]{}, Phys. Rev. [**C 53**]{}, 2661 (1996)
T.-S.H. Lee and D.O. Riska, Phys. Rev. Lett. [**70**]{}, 2237 (1993)
V. Dmitrašinović [*et al.*]{}, Phys. Lett. [**B 465**]{}, 43 (1999).
C. Hanhart, Phys. Rep. [**397**]{}, 155 (2004)
C. Hanhart and N. Kaiser, Phys. Rev. [**C 66**]{}, 054005 (2002).
Y. Kim, K. Kubodera, F. Myhrer and T. Sato, in preparation.
[^1]: E-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Entanglement of formation is a fundamental measure that quantifies the entanglement of bipartite quantum states. This measure has recently been extended into multipartite states taking the name $\alpha$-entanglement of formation. In this work, we follow an analogous multipartite extension for the Gaussian version of entanglement of formation, and focusing on the the finest partition of a multipartite Gaussian state we show this measure is fully additive and computable for 3-mode Gaussian states.'
author:
- Sho Onoe
- Spyros Tserkis
- 'Austin P. Lund'
- 'Timothy C. Ralph'
bibliography:
- 'bibliography.bib'
title: Multipartite Gaussian Entanglement of Formation
---
Introduction
============
Entanglement is a property of quantum mechanics that allows correlations beyond the classical limit. As such, it is considered a crucial resource that allows certain quantum protocols to be more efficient than their classical counterpart [@horodecki2009quantum]. Several entanglement measures have been defined in the literature [@horodecki2009quantum; @plenio2014introduction], however in general the quantification of their values is a challenging task.
Bipartite entanglement of formation (EoF) [@bennett1996mixed] is defined as the least expected amount of bipartite entropy of entanglement (EoE) required to create a state. In general, the quantification of bipartite EoF involves a minimization over infinite degrees of freedom, making it hard to compute. Initial research focused on simple systems such as the 2-qubit system [@hill1997entanglement; @wootters1998entanglement], which led to analytical expressions for the measure.
An analogous measure, called Gaussian EoF (GEoF), focusing only on Gaussian states and operations, was defined by Wolf et al. [@wolf2004gaussian]. A few years later, this measure was proven to be equal to EoF in the case of 2-mode Gaussian states [@akbari2015entanglement; @Wilde.PRA.18]. For these types of states, several efficient numerical methods and analytical expressions have been derived [@marian2008entanglement; @wolf2004gaussian; @Tserkis.Ralph.PRA.17; @tserkis2019quantifying]. Recently, in Ref. [@szalay2015multipartite], Szalay introduced a measure referred to as $\alpha$-EoF, which is the multipartite extension of bipartite EoF. In this paper, we follow Wolf’s approach and apply the notion of $\alpha$-EoF onto the Gaussian regime. We show that $\alpha$-GEoF is a computable multipartite entanglement measure. We utilize a special case of $\alpha$-GEoF, which we refer to as N-mode GEoF, to quantify the *total entanglement* in a 3-mode Gaussian system, in the sense that it is the sum of the entanglement of all internal partitions of the state.
Our paper is set out in the following way. In section II, we introduce the conventions adopted in this paper. In section III we review bipartite entanglement measures. In section IV, we review $\alpha$-entanglement measures [@szalay2015multipartite] and introduce a special subset, referring to it as N-mode EoF. In section V, we apply $\alpha$-entanglement measures to the Gaussian regime and prove N-mode EoF is fully additive. In section VI, we consider the tripartite case and compute the total entanglement for simple cases. We summarize and conclude our results in section VII.
Preliminaries
=============
Modes, Partitions and Sub-Systems
---------------------------------
In the discrete variable case, the smallest sub-systems are referred to as qudits (or qubits for 2-level systems). In the continuous variable case, the smallest sub-systems are referred to as modes. For simplicity, this paper will be utilizing the terminology mode, but in this context it can be used interchangeably with qudits if we are not considering the case of Gaussian states.
Let us consider an $N$-mode state $\hat{\rho}$. The state of the $n$th mode $\hat{\rho}_n$ can be found via the partial trace over all other modes: $$\hat{\rho}_{n}\equiv {\mbox{Tr}}_{\forall i\neq n}(\hat{\rho}) \,.$$ $\hat{\rho}$ can be split into $M$ partitions, via assigning each mode into one of the $M$ partitions (where $N \geqslant M$). By doing this, we introduce $M$ sub-systems, denoted $\{s_1,s_2,...,s_M\}$. This defines the M-partitioning, $\alpha=s_1|s_2|...|s_M$. Each sub-system $s_j$ is defined as the reduced state, achieved through the partial trace over all other sub-system, i.e., $$\hat{\rho}_{s_n}\equiv {\mbox{Tr}}_{\forall s_i\neq s_n}(\hat{\rho}) \,.$$
von Neumann Entropy
-------------------
Before we get into the quantification of entanglement, we need to first define a function that a broad family of entanglement measures are based on, i.e., quantum entropy [@Neumann1927; @bravyi2003entanglement; @ohya2004quantum]. In particular, we focus on the von Neumann entropy, which for a state $\hat{\rho}$ is defined as $$S(\hat{\rho}) \equiv - {\mbox{Tr}}(\hat{\rho} \ln \hat{\rho}) \,.$$ $S(\hat{\rho})$ is a symmetric, basis-independent function, which vanishes for pure states. Also, note that it is fully additive for non-correlated states (although sub-additive in general), i.e., $$S\left(\hat{\rho}_{s_1} \otimes \hat{\rho}_{s_2} \right) =S(\hat{\rho}_{s_1})+S(\hat{\rho}_{s_2}) \,,$$ and convex $$S \left(\sum_j p_j \hat{\rho}_j \right) \geqslant \sum_j p_j S(\hat{\rho}_j) \,.$$
Gaussian States
---------------
In the later part of this paper, we will be considering quantum systems comprised of bosonic Gaussian [modes]{}, $\hat{a}_n$ [@welsch1999ii; @weedbrook2012gaussian; @adesso2014continuous; @serafini2017quantum]. These bosonic annihilation operators satisfy the bosonic commutation relations $[\hat{a}_n,\hat{a}_m^{\dag}]=\delta^n_m$, where $\delta$ is the Kronecker delta. For Gaussian states, the analysis of first and second moment [@weedbrook2012gaussian] is sufficient to characterize the Wigner function of a particular output mode [@scully1999quantum]. The first moment of an $N$-mode Gaussian state is fully characterized by its $2N$-dimensional displacement vector, $\vec{D}$. The second order moment is described by its $2N \times 2N$ real symmetric covariance matrix [@wang2007quantum], $\boldsymbol{\sigma}$. As a result, all Gaussian states can be written as $\hat{\rho}_{\boldsymbol{\sigma},\vec{D}}$.
The $i$th element of the displacement vector is defined in the following way: $$\begin{gathered}
\vec{D}_{i} \equiv {\mbox{Tr}}(\rho \hat{R}_i) \,,\end{gathered}$$ where $$\begin{gathered}
\vec{R} \equiv(\hat{q}_{1},...,\hat{q}_{N},\hat{p}_{1},...,\hat{p}_{N})^{T} \,,\end{gathered}$$ and we have defined $\hat{q}_n\equiv \hat{a}_n+\hat{a}_n^{\dag}$ and $\hat{p}_n\equiv \hat{a}_n-\hat{a}_n^{\dag}$. The $\{i,i'\}$th element of the covariance matrix $\boldsymbol{\sigma}$ is defined in the following way: $$\sigma_{ii'} \equiv {\mbox{Tr}}\left[\hat{\rho}(\hat{R}_i\hat{R}_{i'}+\hat{R}_{i'}\hat{R}_i)\right]-2{\mbox{Tr}}\left[\hat{\rho} \hat{R}_i\right] {\mbox{Tr}}\left[\hat{\rho} \hat{R}_{i'}\right] \,.$$
Bipartite Entanglement Measures
===============================
Bipartite Entropy of Entanglement
---------------------------------
EoE, $E_{s_1|s_2}$, is the typical way to quantify bipartite entanglement in pure states, $\hat{\psi} := {\mbox{$| \psi \rangle$}}{\mbox{$\langle \psi |$}}$ [@bennett1996concentrating]. This measure is given by the von Neumann entropy of the reduced state: $$E_{s_1|s_2}(\hat{\psi}) \equiv S\left[{\mbox{Tr}}_{s_2}(\hat{\psi})\right]\,.$$ As $\hat{\psi}$ is a pure state, EoE is invariant under permutations, i.e., $E_{s_1|s_2}(\hat{\psi}) = E_{s_2|s_1}(\hat{\psi})$. This is a reliable bipartite entanglement measure as it satisfies the following postulates [@horodecki2009quantum; @plenio2014introduction]:
1. $E_{s_1|s_2}$ is an indicator function for separability between the subsystem $s_1$ and $s_2$; $$E_{s_1|s_2}(\hat{\psi})=0 \Leftrightarrow \hat{\psi}=\hat{\psi}_{s_1} \otimes \hat{\psi}_{s_2} \,.$$
2. $E_{s_1|s_2}$ is non-increasing on average under local operations and classical communications (LOCC), $\hat{\Lambda}_{s_1|s_2}$, where the locality is defined in terms of the sub-system $s_1$ and $s_2$ [@bennett1996mixed; @vidal2000entanglement; @horodecki2001separability; @chitambar2014everything; @eltschka2014quantifying; @plenio2014introduction]; $$E_{s_1|s_2}(\hat{\psi}) \geqslant \sum_j p_j E_{s_1|s_2}\left[\hat{\Lambda}_{j,s_1|s_2} (\hat{\psi})\right],$$ where $$\hat{\Lambda}_{s_1|s_2} (\hat{\psi}) = \sum_j p_j \hat{\Lambda}_{j,s_1|s_2} (\hat{\psi}) \, ,$$ are pure LOCC sub-operations [@horodecki2005simplifying; @szalay2015multipartite].
Bipartite Entanglement of Formation
-----------------------------------
A natural way to extend an entanglement measure to mixed states is via the convex-roof extension [@uhlmann2010roofs; @bennett1996mixed; @uhlmann2000fidelity; @rothlisberger2009numerical]. EoF is defined as the convex-roof extension of EoE: $$\mathcal{E}_{\text{F},s_1|s_2}(\hat{\rho}) \equiv \inf_{\hat{\rho}=\sum_j p_j \hat{\psi}_{j}}\left\{\sum_j p_j E_{s_1|s_2}(\hat{\psi}_{j}) \right\} \,,$$ where “$\inf$” becomes a “$\min$” for discrete variable states, and the sum can be replaced with an integral when considering a continuum of pure states.
This is a reliable bipartite entanglement measure as it satisfies the mixed state extension of the aforementioned postulates [@szalay2015multipartite] and an extra one, i.e.,
3. for pure states $\mathcal{E}_{\text{F},s_1|s_2}$ reduces to the entropy of entanglement, i.e., $$\mathcal{E}_{\text{F},s_1|s_2}(\hat{\psi})={E}_{s_1|s_2}(\hat{\psi}) \,.$$
As von Neumann entropy is convex, postulate 2 implies that bipartite EoF is also non-increasing under LOCC; $\mathcal{E}_{\text{F},s_1|s_2}(\hat{\rho})\geq \mathcal{E}_{\text{F},s_1|s_2}\left[\hat{\Lambda}_{s_1|s_2}(\hat{\rho})\right]$.
M-Partite Entanglement Measures
===============================
$\alpha$ - Separability
-----------------------
Entanglement can also exist among several partitions. There are several ways to divide an $N$-mode system into $M$ partitions. To make a distinction between the partitioning, Szalay [@szalay2015multipartite] introduced a hierarchy of separability classes. A pure state, ${\mbox{$| \psi \rangle$}}_{\alpha}$, is called “$\alpha$-separable” when $${\mbox{$| \psi \rangle$}}_{\alpha} \equiv \bigotimes_{s_i \in \alpha}{\mbox{$| \psi_{s_i} \rangle$}} \,.$$
For example, a pure five-mode state is $1|23|45$-separable if and only if the state can be written in the following way $${\mbox{$| \psi \rangle$}}_{1|23|45} = {\mbox{$| \psi_{1} \rangle$}} \otimes {\mbox{$| \psi_{23} \rangle$}}\otimes {\mbox{$| \psi_{45} \rangle$}} \,.$$
Then an $\alpha$-separable mixed state can be written in the following way $$\hat{\rho}_{\alpha}=\sum_j p_j {\mbox{$| \psi_j \rangle$}}_{\alpha}{\mbox{$\langle \psi_j |$}}_{\alpha} \,.$$ We can then make a hierarchy for separability as follows: $\alpha$ precedes or equals $\beta$, if all sub-system in $\beta$ can be written as a subset or equal to a subsystem in $\alpha$, i.e., $$\alpha \preceq \beta \Leftrightarrow \forall s_i \in \beta , \; \exists s_{i'} \in \alpha : s_i \subseteq s_{i'} \,.$$ If $\alpha$ has a finer partition than $\beta$ (i.e. $\alpha \preceq \beta$), then a state which is $\beta$ separable must also be $\alpha$ separable.
$\alpha$-Entropy of Entanglement and $\alpha$-Entanglement of Formation
-----------------------------------------------------------------------
### $\alpha$-Von Neumann Entropy
Let us define $\alpha$-von Neumann entropy in the following way: $$S_{\alpha}(\hat{\rho})\equiv \frac{1}{2}\sum_{s_i\in \alpha}S(\hat{\rho}_{s_i}) \,.$$
This is a measure that is well-defined for all states $\hat{\rho}$. Due to the fully additivity of $S$, $S_{\alpha}$ is also fully additive: $$S_{\alpha}(\hat{\rho}_A\otimes \hat{\rho}_B)= S_{\alpha_A}(\hat{\rho}_A) +S_{\alpha_B}(\hat{\rho}_B) \,,$$ where $\alpha_C$, $C \in \{A,B\}$, is the subset of $\alpha$ which includes the part that overlaps with the system $C$.
### $\alpha$-Entropy of Entanglement and Entanglement of Formation
In the multipartite case, Szalay [@szalay2015multipartite] defined the $\alpha$-EoE of a pure state $\hat{\psi}$ to be: $$E_{\alpha}(\hat{\psi})= S_{\alpha}(\hat{\psi}) \,.$$ This measure can be interpreted as the sum of entanglement between the partitions.
$\alpha$-EoF is defined as the convex-roof extension to $\alpha$-EoE[@szalay2015multipartite]: $$\mathcal{E}_{\text{F},\alpha}(\hat{\rho}) \equiv \inf_{\hat{\rho}=\sum_j p_j \hat{\psi}_{j}}\left\{\sum_j p_j E_{\alpha}(\hat{\psi}_{j}) \right\} \,.$$
$\alpha$-EoE and EoF are reliable $\alpha$-entanglement measure as they satisfy the same postulates as the bipartite case, except we must replace $s_1|s_2$ with $\alpha$. $\alpha$-entanglement measures also satisfy an extra postulate:
4. $E_\alpha$ and $\mathcal{E}_\alpha$ must satisfy the multipartite monotonicity; $$\begin{gathered}
{E}_\alpha(\hat{\rho})\leqslant {E}_\beta(\hat{\rho}), \; \forall \alpha \preceq \beta\,,
\\
\mathcal{E}_{\text{F},\alpha}(\hat{\rho})\leqslant \mathcal{E}_{\text{F},\beta}(\hat{\rho}), \; \forall \alpha \preceq \beta \,.\end{gathered}$$
This means that an entanglement measure of finer partition is sensitive to more entanglement within the system, hence giving a larger value.
N-Mode Entropy of Entanglement and N-Mode Entanglement of Formation
-------------------------------------------------------------------
### N-Mode Entropy of Entanglement
In this section, we consider the finest partitioning of $\alpha$-entanglement measure and refer to it as the $N$-mode entropy of entanglement (NEoE) and formation (NEoF). We notice that we have replaced the term partition with mode, as we are no longer interested in the entanglement between the partition that we assign, but with every mode that exists within the system, i.e. $N=M$. NEoE and NEoF satisfy the same postulates as $\alpha$-entanglement measures with the finest partitioning.
For a pure $N$-mode state, $\hat{\psi}$, NEoE is defined in the following way: $$\tilde{E}(\hat{\psi})= \tilde{S}(\hat{\psi}) \equiv \frac{1}{2}\sum_{n=1}^{\mathrm{N}} \; S\left[Tr_{\forall i\neq n}(\hat{\psi})\right] \,.$$ NEoE is the sum of all entanglement between each mode and the rest of the system. Due to multipartite monotonicity, this measure is also the largest pure entanglement measure out of the $\alpha$-EoF. For this reason, we refer to this quantity as the total of entanglement within the system.
To highlight a feature of this measure, let us consider a 2-mode entangled state, with a vacuum input in the 3rd mode. In this case, this measure will reduce down to the bipartite entanglement between the 2-mode entangled state, giving the total entanglement within the system. In comparison, a genuine tripartite entanglement measure [@adesso2006multipartite; @schneeloch2020quantifying] would be zero in this case, as there is only bipartite entanglement.
### N-Mode Entanglement of Formation
For an $N$-mode mixed state, $\hat{\rho}$, NEoF is defined in the following way: $$\tilde{\mathcal{E}}_{\text{F}}(\hat{\rho}) \equiv \inf_{\hat{\rho}=\sum_j p_j \hat{\psi}_{j}}\left\{\sum_{j} {p_j} \tilde{E}(\hat{\psi}_{j}) \right\} \,.$$
This measure quantifies the least expected total entanglement that is required to create the mixed state. Even though this is a well-defined measure it is hard to compute as there are infinite degrees of freedom for the set $\{p_j,\hat{\psi}_{j}\}$. In this paper, we limit ourselves to a Gaussian convex roof-extension to overcome this problem.
$\alpha$-Gaussian Entanglement of Formation
===========================================
Von Neumann Entropy and $\alpha$-EoE for Gaussian States
--------------------------------------------------------
For Gaussian states, the von Neumann entropy of a state, $\hat{\rho}_{\boldsymbol{\sigma},\vec{D}}$, is fully characterized by its covariance matrix. The von Neumann entropy of an $N$-mode Gaussian state with covariance matrix $\boldsymbol{\sigma}$ can be calculated as follows [@Agarwak1971]: $$S(\boldsymbol{\sigma})=\frac{1}{2}\sum_{n=1}^{N} \; h(\nu_n) \,,$$ where $\nu_n$ is the $n$th symplectic eigenvalue of $\boldsymbol{\sigma}$, and $$h(x) \equiv \frac{x_+}{2}\log_2(\frac{x_+}{2})-\frac{x_-}{2}\log_2(\frac{x_-}{2}) \,,$$ with $x_{\pm} \equiv x\pm 1$ an auxiliary function.
As the von Neumann entropy is fully characterized by its covariance matrix, $\alpha$-EoE of a pure state, $\hat{\psi}_{\boldsymbol{\pi},\vec{D}}$, is also fully characterized by its covariance matrix. The $\alpha$-EoE of a pure state with covariance matrix $\boldsymbol{\pi}$ is calculated as follows: $$E_{{\alpha}}(\boldsymbol{\pi}) =\frac{1}{2} \sum_{s_i\in \alpha} S\left[{\mbox{Tr}}_{s_i}(\boldsymbol{\pi})\right] \,.$$ A covariance matrix is pure if and only if $\text{det}(\boldsymbol{\pi})=1$.
$\alpha$-Gaussian Entanglement of Formation
-------------------------------------------
A mixed Gaussian state $\rho_{\boldsymbol{\sigma},\vec{D}}$ can be decomposed into a mixture of pure Gaussian states in the following way: $$\hat{\rho}_{\boldsymbol{\sigma},\vec{D}}=\int \mathrm{d} \boldsymbol{\pi} \mathrm{d}\vec{D}' \; \mu (\boldsymbol{\pi},\vec{D}') \hat{\psi}_{\boldsymbol{\pi},\vec{D}'} \,,$$ where $\mu$ is the probability density of $\hat{\rho}_{\boldsymbol{\pi},\vec{D}'}$. In Ref. [@wolf2004gaussian] the authors defined the bipartite Gaussian entanglement of formation (GEoF), and analogously we define the $\alpha$-GEoF as follows $$\label{GEoF}
\begin{aligned}
\mathcal{E}_{\text{G},{\alpha}}(\hat{\rho}_{\boldsymbol{\sigma},\vec{D}}) \equiv \underset{\mu}{\text{inf}}\left\{ \int\right. & \mathrm{d} \boldsymbol{\pi} \mathrm{d}\vec{D}' \; \mu (\boldsymbol{\pi},\vec{D}')E_\alpha(\boldsymbol{\pi})
\\
& \left. | \hat{\rho}_{\boldsymbol{\sigma},\vec{D}}=\int \mathrm{d} \boldsymbol{\pi} \mathrm{d}\vec{D}' \; \mu (\boldsymbol{\pi},\vec{D}') \hat{\psi}_{\boldsymbol{\pi},\vec{D}'} \right\} \, .
\end{aligned}$$ This definition involves a minimization over infinite degrees of freedom, however by following the analysis of Ref. [@wolf2004gaussian], we find that Eqn. (\[GEoF\]) reduces to the following expression $$\label{Wolf}
\mathcal{E}_{\text{G},{\alpha}}(\boldsymbol{\sigma})=\underset{\pi}{\text{inf}}\left\{ E_{\alpha}(\boldsymbol{\pi})| \boldsymbol{\sigma}=\boldsymbol{\pi} + \boldsymbol{\varphi} \right\} \,,$$ where $\boldsymbol{\varphi}$ is a positive semi-definite matrix. This equation has finite free parameters, and therefore is a computable entanglement measure. In the App. \[ApAdd\] we utilize Eqn. (\[Wolf\]) to prove the additivity of NGEoF.
N-mode Gaussian Entanglement of Formation for 3 mode states
===========================================================
Mixed 3-mode Gaussian states
----------------------------
For mixed 3-modes states, we can utilize Gaussian local unitary operations (GLUO; refer to App. \[ApGLUO\]) to reduce the state into the standard form [@adesso2006multipartite; @ferraro2005gaussian]: $$\boldsymbol{\sigma}_{\text{sf}}=
\begin{bmatrix}
a_1 &e_{1} &e_{3} &0 & 0 & e_{4}\\
e_{1}& a_2 &e_{6} &0 &0 & e_{7}\\
e_{3}& e_{6} & a_{3} & 0 & e_8 &0 \\
0 &0 &0 & a_1 & e_{2} & e_{5}\\
0& 0 & e_{8}& e_{2} & a_2 & e_{9} \\
e_4& e_7&0 &e_{5}& e_{9} & a_{3}
\end{bmatrix} \,.$$
As GLUO do not affect the entanglement, we can reduce Eqn. (\[Wolf\]) to the following: $$\label{EqnGEoF}
\mathcal{E}_{\text{G},\alpha}(\boldsymbol{\sigma})=\underset{\boldsymbol{\pi}}{\text{inf}}\left\{ E(\boldsymbol{\pi})| \boldsymbol{\sigma}_{\text{sf}}-\boldsymbol{\pi}\geqslant 0 \right\} \,.$$ In the next subsection, we fully parametrize $\boldsymbol{\pi}$.
Pure 3-mode Gaussian States
---------------------------
By utilizing GLUO, $\textbf{L}$, we can reduce any $\boldsymbol{\pi}$ to the standard form [@duan2000inseparability; @simon2000peres]. For the 3-mode pure state, the standard form is [@adesso2005gaussian]: $$\boldsymbol{\pi}_{\text{sf}}
=\begin{bmatrix}
a_1 & e_{12}^{+}& e_{13}^{+}&0 &0 &0\\
e_{12}^{+}& a_2& e_{23}^{+} & 0 &0 & 0\\
e_{13}^{+}& e_{23}^{+}& a_{3} & 0 &0 & 0 \\
0&0&0 & a_1 & e_{12}^{-} & e_{13}^{-}\\
0&0&0 &e_{12}^{-} & a_2 & e_{23}^{-} \\
0&0&0 & e_{13}^{-} & e_{23}^{-} & a_{3}
\end{bmatrix} \,,$$ where $e_{ij}^{\pm}$ are a function of $a_1$, $a_2$ and $a_3$. For $\boldsymbol{\pi}_{\text{sf}}$ to be a physical covariance matrix the inequality $|a_i-a_j|\leqslant a_k-1$ must be satisfied [@adesso2006multipartite]. All pure states can then be decomposed in the following way: $$\begin{aligned}
\boldsymbol{\pi} = \boldsymbol{L} \boldsymbol{\pi}_{\text{sf}}(a_1,a_2,a_3) \boldsymbol{L}^T \,.
\end{aligned}$$
In general, $\boldsymbol{L}$ has 9 free parameters, and hence the minimization of Eqn. (\[EqnGEoF\]) can be conducted over 12 free parameters. A numerical code which scans over all possible $\boldsymbol{\pi}$ with finite size step for these 12 free parameters can be created. The condition $\left(\boldsymbol{\sigma}_{\text{sf},n}-\boldsymbol{\pi}_n \right) \geqslant 0 $, gives a finite range for all local squeezing operations, $a_1$, $a_2$ and $a_3$. The phase parameters are limited to $0 \geqslant \phi \geqslant 2 \pi $.
q-p states
----------
In this section we consider a special class of states where we can reduce the number of free parameters to 6. In special cases, the standard form of the mixed state reduces to the following form: $$\boldsymbol{\sigma}_{\text{qp}}=
\begin{bmatrix}
a_1 & e_{1}& e_{3}&0 &0 & 0\\
e_{1}& a_2& e_{6} & 0 &0 & 0\\
e_{3}& e_{6}&a_{3} & 0 & 0 &0 \\
0 &0 &0 & a_1 & e_{2} & e_{5}\\
0& 0& 0&e_{2} & a_2 & e_{9} \\
0& 0&0& e_{5} & e_{9} & a_{3}
\end{bmatrix} \,.$$ We will refer to these states as q-p states. q-p states have the property that the $\hat{q}$-quadrature is completely uncorrelated to the $\hat{p}$-quadrature. This means that we can write the following: $$\boldsymbol{\sigma}_{\text{qp}}=\boldsymbol{\sigma}_{\hat{q}} \oplus \boldsymbol{\sigma}_{\hat{p}} \,.$$ Following the analysis in Ref. [@wolf2004gaussian], we prove in App. \[ApQp\] that the optimum pure state to create such a state must also be a q-p state: $$\boldsymbol{\pi}_{\text{qp}}=\boldsymbol{\pi}_{\hat{q}} \oplus \boldsymbol{\pi}_{\hat{p}} \,.$$ These states only have 6 free parameters, which greatly reduces the complexity of the problem.
Numerical Results
-----------------
Consider a two-mode Gaussian state, where one of the modes is thermal, while the others are vacuum. When a two-mode squeezer is applied to such a state, the bipartite GEoF is constant regardless of the number of photon in the thermal mode [@adesso2004extremal; @adesso2005gaussian; @giovannetti2014ultimate].
\[FigNGEoF\] [{width="\columnwidth"}]{}
We aim to replicate an analogous result in the tripartite case, utilizing NGEoF. We consider a case where a 3-mode squeezer, $\hat{S}_3$ (details of this operation can be found in App. \[ApS3\]), is applied to an input with with all three modes which are thermal with an average of $\bar{n}$ particles. Since the output state is a q-p state, we conduct an numerical optimization over q-p state to obtain Fig. 1. We repeat this process in the case where $\hat{S}_3$ is applied to an input with 1 mode which is thermal and the rest being a vacuum. NGEoF is constant in when there is only one thermal input, which is an analogous result to the 2-mode case.
Conclusion
==========
In this paper, we utilized the analysis of Ref. [@szalay2015multipartite] on multipartite entanglement measures, and applied it to the Gaussian regime. We successfully demonstrated that the degree of freedom for this measure reduces down to a finite one for all Gaussian states. In particular, we were interested in a special case of $\alpha$-GEoF; NGEoF which quantifies the least expected total entanglement that is required to create the state. We proved that this measure is fully additive. In the last section we quantified its value for simple 3-mode Gaussian states and demonstrated that this measure displayed analogous features to the 2-mode case.
An interesting future research direction would be to compare NGEoF and NEoF. For the 2-mode case, it has been proven that NGEoF and NEoF coincides with each other for Gaussian states [@akbari2015entanglement; @Wilde.PRA.18]. It would be beneficial to prove that this can be extended to the N-mode case. Combined with the result that NGEoF is additive, as proven by this paper, the additivity of NEoF would then be proven for Gaussian states in general.
In this paper, we were particularly interested in NEoF, however there are other interesting $\alpha$-EoF measures. In particular, there is a $\alpha$-EoF which quantifies the genuine tripartite entanglement within a three mode system [@szalay2015multipartite; @schneeloch2020quantifying]. We refer to this measure as a genuine tripartite entanglement measure, as it vanishes for all states which are not genuinely tripartite entangled states. A recent paper [@schneeloch2020quantifying] looked into finding an upper bound to this measure for the DV case. It would be interesting to apply this to the Gaussian regime, and investigate how useful the measure is.
Acknowledgements
================
This work is supported by the Australian Research Council (ARC) under the Centre of Excellence for Quantum Computation and Communication Technology (Grant No. CE170100012).
Additivity of N-mode Gaussian Entanglement of Formation {#ApAdd}
=======================================================
NGEoF for Gaussian states $\boldsymbol{\sigma}=\boldsymbol{\sigma}_{A}\oplus \boldsymbol{\sigma}_B$ is fully additive, i.e., $$\tilde{\mathcal{E}}_{\mathrm{G}}(\boldsymbol{\sigma}_A \oplus \boldsymbol{\sigma}_B)=\tilde{\mathcal{E}}_{\mathrm{G}}(\boldsymbol{\sigma}_A)+\tilde{\mathcal{E}}_{\mathrm{G}}(\boldsymbol{\sigma}_B) \,.$$ where $\boldsymbol{\sigma}_A$ and $\boldsymbol{\sigma}_B$ is an $N$-mode and $N'$-mode Gaussian state, respectively.
NGEoF is by construction sub-additive, i.e., $$\tilde{\mathcal{E}}_{G}(\boldsymbol{\sigma}_A \oplus \boldsymbol{\sigma}_B) \leqslant \tilde{\mathcal{E}}_{\mathrm{G}}(\boldsymbol{\sigma}_A)+\tilde{\mathcal{E}}_{\mathrm{G}}(\boldsymbol{\sigma}_B)\,,$$ and thus its additivity can be shown by proving that NGEoF is super-additive too, i.e., $$\tilde{\mathcal{E}}_{\mathrm{G}}(\boldsymbol{\sigma}_A \oplus \boldsymbol{\sigma}_B) \geqslant \tilde{\mathcal{E}}_{\mathrm{G}}(\boldsymbol{\sigma}_A)+\tilde{\mathcal{E}}_{\mathrm{G}}(\boldsymbol{\sigma}_B) \,.
\label{superadditive}$$ The Gaussian state $\boldsymbol{\sigma}=\boldsymbol{\sigma}_A \oplus \boldsymbol{\sigma}_B$ can decomposed as $$\boldsymbol{\sigma}=\boldsymbol{\sigma}_A \oplus \boldsymbol{\sigma}_B=\boldsymbol{\pi}+\boldsymbol{\varphi} \,,
\label{decomposition}$$ where $\boldsymbol{\pi}$ is a pure Gaussian state and $\boldsymbol{\varphi}$ is a positive semidefinite matrix. For any $\boldsymbol{\varphi} \geqslant 0$, the NGEoF for the states $\boldsymbol{\sigma}_A$ and $\boldsymbol{\sigma}_B$ satisfies
$$\begin{gathered}
\tilde{\mathcal{E}}_{\mathrm{G}}[{\mbox{Tr}}_B(\boldsymbol{\pi})] \geqslant \tilde{\mathcal{E}}_{\mathrm{G}}[{\mbox{Tr}}_B (\boldsymbol{\pi}+\boldsymbol{\varphi}) ] = \tilde{\mathcal{E}}_{\mathrm{G}}(\boldsymbol{\sigma}_A) \,, \\
\tilde{\mathcal{E}}_{\mathrm{G}}[{\mbox{Tr}}_A (\boldsymbol{\pi})] \geqslant \tilde{\mathcal{E}}_{\mathrm{G}}[{\mbox{Tr}}_A (\boldsymbol{\pi}+\boldsymbol{\varphi})] = \tilde{\mathcal{E}}_{\mathrm{G}}(\boldsymbol{\sigma}_B) \,,\end{gathered}$$
so we have $$\tilde{\mathcal{E}}_{\mathrm{G}}[{\mbox{Tr}}_A(\boldsymbol{\pi})] + \tilde{\mathcal{E}}_{\mathrm{G}}[{\mbox{Tr}}_B (\boldsymbol{\pi})] \geqslant \tilde{\mathcal{E}}_{\mathrm{G}}(\boldsymbol{\sigma}_A)+\tilde{\mathcal{E}}_{\mathrm{G}}(\boldsymbol{\sigma}_B) \,.
\label{ineq1}$$ The $N'$-mode state ${\mbox{Tr}}_A (\boldsymbol{\pi})$ and $N$-mode state ${\mbox{Tr}}_B (\boldsymbol{\pi})$ in the above inequality can also decomposed as follows
$$\begin{gathered}
{\mbox{Tr}}_A (\boldsymbol{\pi}) = \boldsymbol{\pi}_{B} + \boldsymbol{\varphi}_B \,, \\
{\mbox{Tr}}_B (\boldsymbol{\pi}) = \boldsymbol{\pi}_{A} + \boldsymbol{\varphi}_A \,,\end{gathered}$$
\[eqredstates\]
and again for arbitrary $\boldsymbol{\varphi}_A \geqslant 0$ and $\boldsymbol{\varphi}_B \geqslant 0$ we have
$$\begin{gathered}
\tilde{\mathcal{E}}_{\mathrm{G}} (\boldsymbol{\pi}_{A}) \geqslant \tilde{\mathcal{E}}_{\mathrm{G}} (\boldsymbol{\pi}_{A} + \boldsymbol{\varphi}_A) = \tilde{\mathcal{E}}_{\mathrm{G}} [ {\mbox{Tr}}_B( \boldsymbol{\pi})] \,, \\
\tilde{\mathcal{E}}_{\mathrm{G}} (\boldsymbol{\pi}_{B}) \geqslant \tilde{\mathcal{E}}_{\mathrm{G}} (\boldsymbol{\pi}_{B} + \boldsymbol{\varphi}_B) = \tilde{\mathcal{E}}_{\mathrm{G}} [ {\mbox{Tr}}_A (\boldsymbol{\pi})]\,,\end{gathered}$$
which implies $$\tilde{\mathcal{E}}_{\mathrm{G}} (\boldsymbol{\pi}_{A}) + \tilde{\mathcal{E}}_{\mathrm{G}} (\boldsymbol{\pi}_{B}) \geqslant \tilde{\mathcal{E}}_{\mathrm{G}} [ {\mbox{Tr}}_A (\boldsymbol{\pi})] + \tilde{\mathcal{E}}_{\mathrm{G}} [ {\mbox{Tr}}_B (\boldsymbol{\pi})] \,.
\label{ineq2}$$ Since $\boldsymbol{\pi}_{A}$ and $\boldsymbol{\pi}_{B}$ are pure states, their NGEoF is equivalent to their entropy of entanglement, i.e.,
$$\begin{gathered}
{}\tilde{\mathcal{E}}_{\mathrm{G}} (\boldsymbol{\pi}_{A}) = \tilde{E} (\boldsymbol{\pi}_{A}) = \tilde{S} (\boldsymbol{\pi}_{A})\,,
\\
{}\tilde{\mathcal{E}}_{\mathrm{G}} (\boldsymbol{\pi}_{B}) = \tilde{E} (\boldsymbol{\pi}_{B}) = \tilde{S} (\boldsymbol{\pi}_{B})\,,\end{gathered}$$
and for arbitrary $\boldsymbol{\varphi}_A \geqslant 0$ and $\boldsymbol{\varphi}_B \geqslant 0$ we get
$$\begin{gathered}
\tilde{S} [{\mbox{Tr}}_B (\boldsymbol{\pi})] = \tilde{S}(\boldsymbol{\pi}_{A} + \boldsymbol{\varphi}_A ) \geqslant \tilde{S}(\boldsymbol{\pi}_{A}) \,, \\
\tilde{S} [{\mbox{Tr}}_A (\boldsymbol{\pi})]= \tilde{S}(\boldsymbol{\pi}_{B} + \boldsymbol{\varphi}_B ) \geqslant \tilde{S}(\boldsymbol{\pi}_{B}) \,\end{gathered}$$
which combined with the inequality (\[ineq1\]) and (\[ineq2\]) turns into $$\tilde{S} [{\mbox{Tr}}_A (\boldsymbol{\pi})] + \tilde{S} [{\mbox{Tr}}_B(\boldsymbol{\pi})] \geqslant \tilde{\mathcal{E}}_{\mathrm{G}}(\boldsymbol{\sigma}_A)+\tilde{\mathcal{E}}_{\mathrm{G}}(\boldsymbol{\sigma}_B) \,.
\label{ineq3}$$ We now notice that for any $(N+N')$-mode state $\boldsymbol{\sigma}$ we have $$\tilde{S}(\boldsymbol{\sigma}_{s_1s_2})= \tilde{S}[{\mbox{Tr}}_{s_1}(\boldsymbol{\sigma}_{s_1s_2})]+\tilde{S}[{\mbox{Tr}}_{s_2}(\boldsymbol{\sigma}_{s_1s_2})]$$ and thus the left-hand side of the inequality (\[ineq3\]) becomes $$\tilde{S} [{\mbox{Tr}}_B (\boldsymbol{\pi})] + \tilde{S} [{\mbox{Tr}}_A (\boldsymbol{\pi})] = \tilde{S}(\boldsymbol{\pi})= \tilde{E}(\boldsymbol{\pi}) \,.
\label{eq1}$$
Given that the above equality is true for every $\boldsymbol{\pi}$, it should be also true for the “optimal” $\boldsymbol{\pi}_{o}$ that gives the NGEoF of the global state $\boldsymbol{\sigma}=\boldsymbol{\sigma}_A \oplus \boldsymbol{\sigma}_B$ in Eqn. (\[decomposition\]), i.e., $$\tilde{\mathcal{E}}_{\mathrm{G}}(\boldsymbol{\sigma}) = \tilde{\mathcal{E}}_{\mathrm{G}}(\boldsymbol{\sigma}_A \oplus \boldsymbol{\sigma}_B) = \tilde{E}(\boldsymbol{\pi}_{o})\, .
\label{eq2}$$
Combining the above Eqns. (\[eq1\]) and (\[eq2\]) with the inequality (\[ineq3\]), we get
$$\tilde{\mathcal{E}}_{\mathrm{G}}(\boldsymbol{\sigma}_A \oplus \boldsymbol{\sigma}_B) \geqslant \tilde{\mathcal{E}}_{\mathrm{G}}(\boldsymbol{\sigma}_A) + \tilde{\mathcal{E}}_{\mathrm{G}}(\boldsymbol{\sigma}_B)\, ,$$
which completes the proof.
Gaussian Local Unitary Operations {#ApGLUO}
=================================
In this section we introduce a useful class of operations Gaussian local unitary operations (GLUO). GLUO are operations which do not increase or decrease the amount of entanglement. By definition, these operations are a subset of LOCC (here, locality is defined with respect to each mode), which means that they cannot increase the entanglement. As these operations are locally reversible (i.e. unitary in terms of the Heisenberg picture), they cannot decrease the entanglement.
\[FigGLUO\] ![A schematic decomposition of all GLUO operations.[]{data-label="fig: Unitary"}](GLUO.jpeg "fig:"){width="60.00000%"}
We introduce the GLUO of a $N$-mode state as follows: $$\boldsymbol{L}\equiv \bigoplus_{n=1}^N \; \boldsymbol{L}_n \,,$$ where $\boldsymbol{L}_n$ is the GLUO in each mode. Each GLUO can be decomposed through the Bloch Messiah decomposition [@bloch1962canonical; @Braunstein2005Squeezing] as $$\boldsymbol{L}_n= \boldsymbol{L}(\phi_n')\boldsymbol{L}(r_n)\boldsymbol{L}(\phi_n) \,,$$ where $$\boldsymbol{L}(\phi) \equiv \begin{bmatrix}
\cos(\phi) & \sin(\phi) \\
\sin(\phi) & \cos(\phi) \\
\end{bmatrix} \,,$$ corresponds to phase rotations, and $$\boldsymbol{L}(r) \equiv \begin{bmatrix}
\cosh(r) & 0 \\
0 & \sinh(r) \\
\end{bmatrix} \,,$$ corresponds squeezing operations. A schematic diagram of this decomposition for GLUO is shown in Fig \[FigGLUO\].
Optimization of NGEoF for q-p states {#ApQp}
====================================
Consider a q-p state $\boldsymbol{\sigma}_{\text{qp}}$. For every pure state, $\boldsymbol{\pi} \leqslant \boldsymbol{\sigma}$, there exists a q-p pure state $\boldsymbol{\pi}_{\text{qp}}'\leqslant \boldsymbol{\sigma}$ which satisfies the following: $$\tilde{E}(\boldsymbol{\pi})\geqslant \tilde{E}(\boldsymbol{\pi}'_{\text{qp}})$$
Any Gaussian pure state $\boldsymbol{\pi}$ can be written in the following way [@wolf2004gaussian]: $$\boldsymbol{\pi}(\boldsymbol{X},\boldsymbol{Y})=\begin{bmatrix}
\boldsymbol{X} & \boldsymbol{XY} \\
\boldsymbol{YX} & \boldsymbol{YXY}+\boldsymbol{X}^{-1}\\
\end{bmatrix} \;,$$ where $\boldsymbol{X}>0$ and $\boldsymbol{Y}$ are real symmetric $N \times N$ matrix with $\boldsymbol{X}>0$. For q-p states, $\boldsymbol{Y}=0$. For every $\boldsymbol{\sigma}_{\text{qp}}\geqslant \boldsymbol{\pi}(\boldsymbol{X},\boldsymbol{Y})$, we have the following [@wolf2004gaussian]: $$\label{Eqnpq1}
\boldsymbol{\sigma}_{\text{qp}} \geqslant \boldsymbol{\pi}(\boldsymbol{X},\boldsymbol{Y}) \Rightarrow \boldsymbol{\sigma}_{\text{qp}} \geqslant \boldsymbol{\pi}(\boldsymbol{X},0)$$ We also have that the determinant of the single mode $\boldsymbol{\pi}_n(\boldsymbol{X},\boldsymbol{Y})$ is always larger than $\boldsymbol{\pi}_n(\boldsymbol{X},\boldsymbol{0})$: $$\label{EqDetXY}
\det\left[\boldsymbol{\pi}_n(\boldsymbol{X},\boldsymbol{0})\right]\leqslant \det\left[\boldsymbol{\pi}_n(\boldsymbol{X},\boldsymbol{Y})\right]$$ The entropy of a single mode state is computed to be; $$\label{EqnSn}
S(\boldsymbol{\sigma}_n)=h\left[\sqrt{\det(\boldsymbol{\sigma}_n)}\right] \, ,$$ As this is true for every mode, combining Eqn. (\[EqnSn\]) and (\[EqDetXY\]) gives the following: $$\label{Eqnpq2}
\tilde{E}[\boldsymbol{\pi}_n(\boldsymbol{X},\boldsymbol{0})] \leqslant \tilde{E}\left[\boldsymbol{\pi}_n[\boldsymbol{X},\boldsymbol{Y})\right]$$ Eqn. (\[Eqnpq1\]) and (\[Eqnpq2\]) completes the proof.
Symmetric 3-Mode Squeezing Operation {#ApS3}
====================================
The Heisenberg evolution of a three mode squeezing operation is as follows [@Wu2005Continuous]: $$\hat{S}_3^{\dag} \hat{a}_i\hat{S}_3= \cosh(r)\hat{a}_i+\sinh(r)\left[-\frac{1}{3}\hat{a}_i^{\dag}+\frac{2}{3}(\hat{a}_j^{\dag}+\hat{a}_k^{\dag})\right] \,.$$ The covariance matrix representation of a three mode squeezer is given by $$\boldsymbol{{S}}_3(r_3)=\begin{bmatrix}
{\alpha}_+& {\beta}_+& {\beta}_+&0&0&0\\
{\beta}_+& {\alpha}_+& {\beta}_+&0&0&0\\
{\beta}_+& {\beta}_+& {\alpha}_+&0&0&0\\
0&0&0&{\alpha}_-& {\beta}_-& {\beta}_-\\
0&0&0&{\beta}_-& {\alpha}_-& {\beta}_-\\
0&0&0&{\beta}_-& {\beta}_-& {\alpha}_-\\
\end{bmatrix} \,,$$ where we have defined the following: $$\alpha_{\pm} \equiv
\cosh(r_3)\mp\frac{\sinh(r_3)}{3}\,, \quad
\beta_{\pm} \equiv
\pm\frac{2\sinh(r_3)}{3}\,.$$
We obtain the GhZ/W state [@adesso2006multipartite] when we apply this operator onto the vacuum state. In the standard form [@adesso2005gaussian], this state can be written in the following way: $$\boldsymbol{\pi}_{GhZ/W,\text{sf}}(r_3)\equiv(\boldsymbol{S}_3 \boldsymbol{S}_3^T)_{\text{sf}}= \;\begin{bmatrix}
{\alpha}'& {\beta}'_+& {\beta}'_+&0&0&0\\
{\beta}'_+& {\alpha}'& {\beta}'_+&0&0&0\\
{\beta}'_+& {\beta}'_+& {\alpha}'&0&0&0\\
0&0&0&{\alpha}'& {\beta}'_-& {\beta}'_-\\
0&0&0&{\beta}'_-& {\alpha}'& {\beta}'_-\\
0&0&0&{\beta}'_-& {\beta}'_-& {\alpha}'\\
\end{bmatrix} \,,$$ where $$\alpha' \equiv \frac{1}{3}\sqrt{9\cosh(2r_3)^2-\sinh(2r_3)^2} \,, \quad
\beta_{\pm} \equiv \pm \frac{|2\sinh(2r_3)|}{3} \sqrt{\frac{3\cosh(2r_3)\pm|\sinh(2r_3)|}{3\cosh(2r_3)\mp|\sinh(2r_3)|}} \,.$$ The Bloch-Messiah decomposition [@bloch1962canonical; @Braunstein2005Squeezing] of this operator can be found in a straightforward fashion by setting the local squeezers to be equal with $2 \pi/3$ phase differences.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
This article describes the implementation of the joint motion estimation and image reconstruction framework presented by Burger, Dirks and Schönlieb and extends this framework to large-scale motion between consecutive image frames.\
The variational framework uses displacements between consecutive frames based on the optical flow approach to improve the image reconstruction quality on the one hand and the motion estimation quality on the other. The energy functional consists of a data-fidelity term with a general operator that connects the input sequence to the solution, it has a total variation term for the image sequence and is connected to the underlying flow using an optical flow term. Additional spatial regularity for the flow is modeled by a total variation regularizer for both components of the flow.\
The numerical minimization is performed in an alternating manner using primal-dual techniques. The resulting schemes are presented as pseudo-code together with a short numerical evaluation.
author:
- 'Hendrik Dirks[^1]'
bibliography:
- 'citations.bib'
title: 'Joint Large-Scale Motion Estimation and Image Reconstruction'
---
Introduction
============
Image reconstruction and motion estimation are important tasks in image processing. Such problems arise for example in modern medicine, biology, chemistry or physics, where even the smallest objects are observed by high resolution microscopes. To characterize the dynamics involved in such data, velocity fields between consecutive image frames are calculated. This is challenging, since the recorded images often suffer from low resolution, low contrast, different gray levels and noise. Methods that simultaneously denoise the recorded image sequence and calculate the underlying velocity field offer new opportunities, since both tasks may endorse each other.\
The ansatz from [@dirks] aims at reconstructing a given sequence $u$ of images and calculating flow fields $\boldsymbol{v}$ between subsequent images at the same time. For given measurements $f=Ku$ this can be achieved by minimizing the variational model $$\begin{aligned}
\label{equation:generalModelEQ}
&\int_0^T \frac{1}{2}\left\|K_tu(\cdot,t)-f(\cdot,t)\right\|_2^2 + \alpha \mathcal{R}(u(\cdot,t)) + \beta \mathcal{S}(\boldsymbol{v}(\cdot,t)) + \|\rho(u,\boldsymbol{v})\|_1 dt\end{aligned}$$ with respect to $u$ and $\boldsymbol{v}$ simultaneously. The denoising part is based on the ROF model [@rudin1992nonlinear]. The first part $\left\|Ku-f\right\|_2^2$ connects the input data $f$ with the image sequence $u$ via a linear operator $K$. Depending on the application, $K$ may model the cutting out of a subset $\Sigma\subset\Omega$ for inpainting, a subsampling for super resolution, a blur for deconvolution or a Radon transform for computed tomography. Additional a-priori information about the structure of $u$ respectively $\boldsymbol{v}$ can be incorporated into each frame via the regularization terms $\mathcal{R}(u(\cdot,t))$ and $\mathcal{S}(\boldsymbol{v}(\cdot,t))$, while their significance is weighted using $\alpha$ and $\beta$. Finally, flow field and images are coupled by a model for the underlying dynamics, e.g. the optical flow approach $\rho(u,\boldsymbol{v}) = u_t+\nabla u\cdot\boldsymbol{v}$ (see Section \[subsection:noiseSensitivity\]).\
Notations
---------
Let us for the following assume a discrete rectangular domain $\Omega\subset{\mathbb{R}}^2$ and sequences of recorded (noisy) images $f^1,\ldots,f^n : \Omega\rightarrow{\mathbb{R}}$, clean images $u^1,\ldots,u^n : \Omega\rightarrow{\mathbb{R}}$ and motion fields between subsequent frames $\boldsymbol{v}^1,\ldots,\boldsymbol{v}^{n-1} : \Omega\rightarrow{\mathbb{R}}^2$.
Optical Flow {#subsection:noiseSensitivity}
------------
In the optical flow problem we want to estimate the flow field $\boldsymbol{v}^i = (v^{i,1},v^{i,2})^T$ that describes the displacement between each of the subsequent images $u^{i}$ and $u^{i+1}$. For this sake it is common to use the brightness constancy assumption $$u^{i+1}(x+\boldsymbol{v}^i(x)) - u^i(x) = 0,\quad i=1,\ldots,n-1,x\in\Omega
\label{opticalFlowKomplett}$$ to derive a connection between image intensities and the underlying flow. Due to the non-linearity (in terms of $\boldsymbol{v}^i(x)$) of this formulation, one often linearizes the first term and arrives at $$\boldsymbol{v}^i\cdot\nabla u^{i+1}(x) + u^{i+1}(x) - u^i(x) = 0.
\label{opticalFlowKlassisch}$$ Unfortunately, due to the Taylor expansion, equation is only valid for small displacements. Another way, which in theory may handle displacements of arbitrary magnitude, is to use a given flow field $\boldsymbol{\tilde{v}}^i$ and use a Taylor expansion of $u^{i+1}(x+\boldsymbol{v}^i(x))$ with respect to a known flow field $\boldsymbol{\tilde{v}}^i$. The result is again a linear equation for $\boldsymbol{v}^i$ which requires evaluations of the input images at a shifted domain $x+\boldsymbol{\tilde{v}^i}$ $$(\boldsymbol{v}^i-\boldsymbol{\tilde{v}}^i)\cdot\nabla u^{i+1}(x+\boldsymbol{\tilde{v}}^i) + u^{i+1}(x+\boldsymbol{\tilde{v}}^i) - u^i(x) = 0.
\label{opticalFlowBesser}$$ By defining $\tilde{u}^{i+1} := u^{i+1}(x+\boldsymbol{\tilde{v}}^i)$ the left side of equation becomes $$\rho(\boldsymbol{v}^i,u^{i},u^{i+1}) := (\boldsymbol{v}^i-\boldsymbol{\tilde{v}}^i)\cdot \nabla\tilde{u}^{i+1} + \tilde{u}^{i+1} - u^i.
\label{opticalFlowBesser2}$$ Both, formulation and , state only one equation per point for the two unknown components of $\boldsymbol{v}$ and, consequently, the problem is underdetermined. To overcome this, the optical flow formulation can be used as a data fidelity in a variational model together with an isotropic total variation term on each of the two flow components to ensure spatial regularity. This leads to the following variational problem for calculating the optical flow of a video sequence $u^1,\ldots,u^n$: $$\operatorname*{arg\,min}_{\boldsymbol{v} = \boldsymbol{v}^1,\ldots, \boldsymbol{v}^{n-1}} \sum_{i=1}^{n-1} \| \rho(\boldsymbol{v}^i,u^{i},u^{i+1}) \|_1 + \alpha \sum_{j=1}^2\left\|\nabla v^{i,j}\right\|_{1,2}.
\label{variationalMotionModelGeneral}$$ This variational approach to optical flow was first introduced by Horn and Schunck [@horn1981determining] and continuously improved [@aubert1999computing; @papenberg2006highly]. Further details about the so-called L$^1$-TV optical flow model above can be found e.g. in [@wedel2009improved; @zach2007duality].
Image Reconstruction
--------------------
Nowadays, variational models for image reconstruction have become very popular. One of the most famous models, introduced by Rudin, Osher and Fatemi in 1992 [@rudin1992nonlinear], is the total variation (TV) model, where the authors couple a L$^2$ data fidelity term with a total variation regularization. Data-term and regularizer in the ROF model match with the first two terms model . The TV-regularization results in a denoised image with cartoon-like features. This model has also been adapted to image deblurring [@Wang07afast], inpainting [@shen2002mathematical], super resolution [@mitzel2009video; @unger2010convex] and tomographic reconstruction [@sawatzky2008accurate; @kosters2011emrecon]. We collectively call these image reconstruction models.\
From our point of view, the problem of motion estimation is directly connected to the underlying image sequence and, hence, requires accurate input images $u^i$. Unfortunately, for many practical applications, only noisy variants $f^i$ of $u^i$ can be recorded. Similar to the optical flow problem, a variational formulation can be used to reconstruct $u^i$ from $f^i$. Let us assume that $f^i$ is a degraded version of $u^i$ corrupted by Gaussian noise. Then, the ROF model $$\operatorname*{arg\,min}_{u = u^1,\ldots, u^{n}} \sum_{i=1}^{n} \frac{1}{2}\|A^iu^i-f^i \|_2^2 + \alpha \|\nabla u^i\|_{1,2}
\label{variationalDenoisingModelGeneral}$$ can be used to reconstruct each $u^i$ from $f^i$. Here, $A^i$ represents some linear operator and could be the identity (denoising), subsampling (zooming) or the Radon transform (computed tomography).
Joint Model
-----------
As already mentioned, motion estimation should be done on noise-free images, so generally one first denoises the image sequence using and afterwards estimates the underlying velocity fields using . In [@dirks] it has been shown that a joint model that simultaneously recovers an image sequence and estimates motion offers a significant advantage towards subsequently applying both methods. The following joint model $$\operatorname*{arg\,min}_{u,\boldsymbol{v}} \int_0^T \frac{1}{2}\left\|Au-f\right\|_2^2 + \alpha \left\|\nabla u\right\|_{1,2} + \gamma \left\|u_t+\nabla u\cdot\boldsymbol{v}\right\|_1 + \beta\sum_{j=1}^2\left\|\nabla v_j\right\|_{1,2} dt,
\label{oldJointModel}$$ respectively its time-discrete counterpart $$\operatorname*{arg\,min}_{\substack{u = u^1,\ldots, u^{n}\\\boldsymbol{v}=\boldsymbol{v}^1,\ldots\boldsymbol{v}^{n-1}}} \sum_{i=1}^{n} \frac{1}{2}\left\|A^iu^i-f^i\right\|_2^2 + \alpha \left\|\nabla u^i\right\|_{1,2} + \sum_{i=1}^{n-1}\gamma \left\|u^i_t+\nabla u^i\cdot\boldsymbol{v}^i\right\|_1 + \beta\sum_{j=1}^2\left\|\nabla v^{i,j}\right\|_{1,2},
\label{oldJointModelTimeDiscrete}$$ was proposed. For both, image sequence and velocity field, the respective total variation is used as a regularizer and the classical optical flow formulation from connects image sequence and velocity field. From the perspective of image reconstruction the optical flow constraint acts as an additional temporal regularizer along the calculated motion fields $\boldsymbol{v}$. In [@burger2016variational] the existence of a minimizer for model under certain regularity assumptions for $\boldsymbol{v}$ and $\nabla\boldsymbol{v}$ has been shown. Despite the existence of a minimizer, calculating it is numerically challenging. Problems arise from the non-convexity and non-linearity of the optical flow term, from the non-differentiability of the L$^1$–norm and finally due to several linear operators acting on $u$ and $\boldsymbol{v}$.\
The main drawback in terms of practical applications for is the restriction to displacements of small magnitude, which cannot be expected in practical applications. By exchanging the classical optical flow constraint with the beneficial linearization from we create a model capable of handling large-scale displacements. The joint large-scale motion estimation and image reconstruction model then reads $$\operatorname*{arg\,min}_{\substack{u = u^1,\ldots, u^{n}\\\boldsymbol{v}=\boldsymbol{v}^1,\ldots\boldsymbol{v}^{n-1}}} \sum_{i=1}^{n} \frac{1}{2}\|A^iu^i-f^i \|_2^2 + \alpha \left\|\nabla u^i\right\|_{1,2} + \sum_{i=1}^{n-1} \| \rho(\boldsymbol{v}^i,u^{i},u^{i+1}) \|_1 + \beta\sum_{j=1}^2\left\|\nabla v^{i,j}\right\|_{1,2}.
\label{jointLargeScaleModel}$$ By setting $\boldsymbol{\tilde{v}}^i=0$ in the equation $\rho(\boldsymbol{v}^i,u^{i},u^{i+1})$ we obtain the original energy.
Numerical Scheme and Implementation
===================================
We propose, referring to [@dirks], a minimization scheme which alternatingly fixes $u$ and $\boldsymbol{v}$ and minimizes the energy for the other variable. The corresponding problems for static $u^1_k,\ldots,u^n_k$ in equation resp. $\boldsymbol{v}_{k+1}^1,\ldots\boldsymbol{v}_{k+1}^{n-1}$ in equation read $$\begin{aligned}
\boldsymbol{v}_{k+1} &= \operatorname*{arg\,min}_{\boldsymbol{v}=\boldsymbol{v}^1,\ldots\boldsymbol{v}^{n-1}} \sum_{i=1}^{n-1} \| \rho(\boldsymbol{v}^i,u^{i}_{k},u^{i+1}_{k}) \|_1 + \beta\sum_{j=1}^2\left\|\nabla v^{i,j}\right\|_{1,2}, \label{subproblemV} \\
u_{k+1} &= \operatorname*{arg\,min}_{u = u^1,\ldots, u^{n}} \sum_{i=1}^{n} \frac{1}{2}\|A^iu^i-f^i \|_2^2 + \alpha \| \nabla{u}^{i}\|_1 + \sum_{i=1}^{n-1} \| \rho(\boldsymbol{v}^i_{k+1},u^{i},u^{i+1}) \|_1,
\label{subproblemU}\end{aligned}$$ where the letter $k$ shall denote the iteration number. To start with the minimization problem in , let us denote that in this formulation the evaluation of $u^{i+1}_k$ on intermediate grid points $x+\boldsymbol{\tilde{v}}$ is required. These evaluations can be generated by an interpolation scheme (linear, cubic etc.). Moreover, the linearization of the brightness constancy assumption requires a flow field $\boldsymbol{\tilde{v}}$ close to $\boldsymbol{v}$. This can be generated by an iterative coarse-to-fine approach, which solves the problem on subsampled versions of $u^i$ first and using the upscaled result as initial value for the next finer version. Moreover, on each level several so-called warpings are performed, where the problem is solved for some initial $\boldsymbol{\tilde{v}}$ and the solution is used in the next step (see Section \[coarseToFinePyramid\] and [@wedel2009improved; @zach2007duality] for details).\
Treating the minimization problem in equation next, this problem incorporates the term $$\begin{aligned}
\rho(\boldsymbol{v}^i_{k+1},u^{i},u^{i+1}) = (\boldsymbol{v}^i_{k+1}-\boldsymbol{\tilde{v}}^i_{k+1})\cdot \nabla\tilde{u}^{i+1} + \tilde{u}^{i+1} - u^i
\label{rhoSubU}\end{aligned}$$ Let us recall that in the minimization process for $\boldsymbol{v}^i$, a warping scheme is applied, which creates subsequent versions of $\boldsymbol{v}$ and $\boldsymbol{\tilde{v}}$ which can be assumed to converge to a static quantity. Consequently, the difference $\boldsymbol{v}^i_{k+1}-\boldsymbol{\tilde{v}}^i_{k+1}$ becomes arbitrarily small, so it can be neglected in and the term simplifies to $\|\tilde{u}^{i+1} - u^i(x)\|_1$. We want to underline that for evaluating $\tilde{u}^{i+1}$ the same interpolation scheme as in the $\boldsymbol{v}$-problem has to be used to ensure numerical consistency.
Problem in $\boldsymbol{v}$
---------------------------
The optical flow problem has no time-correspondence in $\boldsymbol{v}$ and thus reduces to $(n-1)$ subproblems of the form $$\begin{aligned}
\operatorname*{arg\,min}_{\boldsymbol{v}} \| \boldsymbol{v}\cdot \nabla\tilde{u} + \tilde{u}_t \|_1 + \beta\sum_{j=1}^2\left\|\nabla v^{j}\right\|_{1,2},\end{aligned}$$ where $\tilde{u}_t := -\boldsymbol{\tilde{v}}^i\cdot\nabla\tilde{u}^{i+1} + \tilde{u}^{i+1} - u^i$ and $\tilde{u} := \tilde{u}^{i+1}$. Applying a primal-dual algorithm [@pock2009algorithm; @chambolle2011first] yields the following iterative scheme: $$\begin{aligned}
\boldsymbol{y}_1^{k+1} &= \frac{\boldsymbol{y}_1^k + \sigma^v_1 \nabla \bar{v}^k_1}{\max(1,\frac{\|\boldsymbol{y}_1^k + \sigma^v_1 \nabla \bar{v}^k_1\|_2}{\alpha})},\\
\boldsymbol{y}_2^{k+1} &= \frac{\boldsymbol{y}_2^k + \sigma^v_2 \nabla \bar{v}^k_2}{\max(1,\frac{\|\boldsymbol{y}_2^k + \sigma^v_2 \nabla \bar{v}^k_2\|_2}{\alpha})},\\
y_3^{k+1} &= \max(-1,\min(1,y_3^k + \sigma^v_3 (\nabla \tilde{u} \cdot \bar{\boldsymbol{v}}^k + \tilde{u}_t))),\\
v_1^{k+1} &= v_1^k - \tau^v_1(\nabla^T\cdot\boldsymbol{y}_1^{k+1} + \tilde{u}_xy_3^{k+1}),\\
v_2^{k+1} &= v_2^k - \tau^v_2(\nabla^T\cdot\boldsymbol{y}_2^{k+1} + \tilde{u}_yy_3^{k+1}),\\
\bar{\boldsymbol{v}}^{k+1} &= 2\boldsymbol{v}^{k+1} - \boldsymbol{v}^{k},\\\end{aligned}$$ where vector-valued dual variables $\boldsymbol{y}_1,\boldsymbol{y}_2$, a regular dual variable $y_3$ and step sizes $\sigma^v_1,\sigma^v_2,\sigma^v_3,\tau^v_1$ and $\tau^v_2$ were introduced. As a convergence criterion we use the primal-dual residual $r^{pd}_v$ from [@goldstein2013adaptive]. Applied to this problem, the residual can be calculated as $$\begin{aligned}
p_1 &= \|\frac{v_1^{k} - v_1^{k+1}}{\tau^v_1} - \nabla^T\cdot(\boldsymbol{y}_1^{k+1}-\boldsymbol{y}_1^{k}) - \tilde{u}_x(y_3^{k+1}-y_3^{k})\|_1,\\
p_2 &= \|\frac{v_2^{k} - v_2^{k+1}}{\tau^v_2} - \nabla^T\cdot(\boldsymbol{y}_2^{k+1}-\boldsymbol{y}_2^{k}) - \tilde{u}_y(y_3^{k+1}-y_3^{k})\|_1,\\
d_1 &= \|\frac{\boldsymbol{y}_1^{k} - \boldsymbol{y}_1^{k+1}|}{\sigma_1} - \nabla (v_1^{k} - v_1^{k+1})\|_1,\\
d_2 &= \|\frac{\boldsymbol{y}_2^{k} - \boldsymbol{y}_2^{k+1}|}{\sigma_2} - \nabla (v_2^{k} - v_2^{k+1})\|_1,\\
d_3 &= \|\frac{y_3^{k} - y_3^{k+1}|}{\sigma_3} - \nabla \tilde{u} \cdot (\boldsymbol{v}^{k} - \boldsymbol{v}^{k+1})\|_1,\\
r^{pd}_v &= p_1 + p_2 + d_1 + d_2 + d_3.\end{aligned}$$
### Coarse-to-fine pyramid {#coarseToFinePyramid}
The optical flow calculation is incorporated into a coarse-to-fine pyramid with intermediate warping steps. This is done by creating a set with $nScales$ versions of the input images with a downsampling factor $\eta\in (0,1)$. To create these images, the input image is convolved with a Gaussian kernel with standard variation $\sigma_d$ and then extrapolated using a bicubic interpolation scheme.\
The optical flow is calculated on the coarsest scale with initial $\tilde{\boldsymbol{v}}=0$ first. The result $\boldsymbol{v}$ is filtered using a two-dimensional median-filter of size $sizeMed$. Afterwards, a bicubic interpolation scheme is taken to calculate both $\tilde{\boldsymbol{v}}$ and $\nabla\tilde{\boldsymbol{v}}$. Then we solve the problem again on the same scale. This procedure is repeated $nWarps$ times, until the result $\boldsymbol{v}$ is upscaled to the next finer level using bicubic interpolation. To improve the convergence speed, it is beneficial to also upscale the dual variables $\boldsymbol{y}_1,\boldsymbol{y}_2$ and $y_3$.
Problem u
---------
The problem in $u$ does not simplify to a series of time-independent problems, since individual frames are correlated by the flow. Consequently, the problem has to be solved in the whole space/time domain. First, we want to deduce that can be rewritten in the form $$\begin{aligned}
\operatorname*{arg\,min}_{u} \frac{1}{2}\|\mathcal{A}u-f\|_2^2 + \alpha \| \bar{\nabla}u\|_{1,2} + \gamma\|\mathcal{W}u\|_1,\label{subproblemUmatrix}\end{aligned}$$ where the identities $u=(u^1,\ldots,u^n),f=(f^1,\ldots,f^n),\mathcal{A}=diag(A^1,\ldots,A^n),\bar{\nabla} = diag(\nabla,\ldots,\nabla)$ hold obviously. For the last term, note that the optical flow term from Equation consists of $(n-1)$ parts of the form $\tilde{u}^{i+1} - u^i$, where $\tilde{u}^{i+1}$ is a shifted version of $u^{i+1}$ by a velocity field $\boldsymbol{v}^i$ that can be evaluated using an interpolation scheme. This quantity can be approximated using a so-called interpolation or warping operator $W^{i+1}$ in which each row has weights at the columns according to the interpolation method (linear, cubic, spline etc.). Thus, we have $\tilde{u}^{i+1}\approx W^{i+1}u^{i+1}$ and the last part in $\eqref{subproblemU}$ reduces to $\|\mathcal{W}u\|_1$, where $$\begin{aligned}
\mathcal{W} = \begin{bmatrix}
-I & W^2 & 0 & \ldots & \ldots & 0 \\
0 & -I & W^3& 0 & \ldots & 0 \\
\vdots & \ddots & \ddots & \ddots & \ddots & \vdots \\
\vdots & \ddots & 0 & -I & W^{n-2}&0\\
0 & \ldots & \ldots & 0 & -I & W^{n-1}
\end{bmatrix}.\end{aligned}$$ Applying the primal-dual algorithm [@pock2009algorithm; @chambolle2011first] to problem we obtain the iterative scheme $$\begin{aligned}
y_1^{k+1} &= \frac{1}{\sigma^u_1 + 1} (y_1^k + \sigma^u_1 (A\bar{u}^k - f)),\\
\boldsymbol{y}_2^{k+1} &= \frac{\boldsymbol{y}_2^k + \sigma^u_2 \bar{\nabla} \bar{u}^k}{\max(1,\frac{\|\boldsymbol{y}_2^k + \sigma^u_2 \bar{\nabla} \bar{u}^k\|_2}{\alpha})},\\
y_3^{k+1} &= \max(-\gamma,\min(\gamma,y_3^k + \sigma^u_3 \mathcal{W} \bar{u}^k )),\\
u^{k+1} &= u^k - \tau^u(A^Ty_1^{k+1} + \bar{\nabla}^T\cdot\boldsymbol{y}_2^{k+1} + \mathcal{W}^Ty_3^{k+1}),\\
\bar{u}^{k+1} &= 2u^{k+1} - u^{k},\\\end{aligned}$$ where, similar to the problem for $\boldsymbol{v}$, regular dual variables $y_1,y_3$, a vector-valued dual variable $\boldsymbol{y}_2$ and step-sizes $\sigma^u_1,\sigma^u_2,\sigma^u_3$ and $\tau^u$ were introduced. Once again, we use the primal-dual residual [@goldstein2013adaptive] as a convergence criterion that can be calculated as follows: $$\begin{aligned}
p &= \|\frac{u^{k} - u^{k+1}}{\tau} - \mathcal{A}^T(y_1^{k+1}-y_1^{k}) - \bar{\nabla}^T\cdot(\boldsymbol{y}_2^{k+1}-\boldsymbol{y}_2^{k}) - \mathcal{W}^T(y_3^{k+1}-y_3^{k})\|_1\\
d_1 &= \|\frac{{y}_1^{k} - {y}_1^{k+1}|}{\sigma} - A (u^{k} - u^{k+1})\|_1\\
d_2 &= \|\frac{\boldsymbol{y}_2^{k} - \boldsymbol{y}_2^{k+1}|}{\sigma} - \bar{\nabla} (u^{k} - u^{k+1})\|_1\\
d_3 &= \|\frac{y_3^{k} - y_3^{k+1}|}{\sigma} - \mathcal{W} ({u}^{k} - {u}^{k+1})\|_1\\
r^{pd}_u &= p + d_1 + d_2 + d_3\end{aligned}$$
Initial Data
------------
Due to the fact that the initial formulation of the problem is highly non-linear, our alternating minimization scheme only solves the linear subproblems and will most likely end up in a local minimum. In our numerical evaluation, we found that initializing the image series $u$ with reasonable values we usually end up in a lower energy. This can for example be done by solving an ROF-problem for each image $u_i$, i.e. setting the weight $\gamma$ in to zero.\
Another possibility to initialize the image sequence is to neglect the flow fields $\boldsymbol{v}$ in the whole problem. One could for example expect a smooth time-derivative $u_t$ and solve to following modified problem $$\begin{aligned}
\operatorname*{arg\,min}_{u} \frac{1}{2}\|\mathcal{A}u-f\|_2^2 + \alpha \| \bar{\nabla}u\|_{1,2} + \frac{\epsilon}{2}\|u_t\|_2^2,\label{initialDataSmooth},\end{aligned}$$ where $\epsilon$ is a small weight However, depending on the application, for large displacements this may introduce smearing effects and lead to worse results than the first method.
Operator Discretization
-----------------------
We assume the underlying space-time grid to consist of the following set of discrete points: $$\begin{aligned}
\left\lbrace (i,j,t) : i=0,\ldots,n_x,j=0,\ldots,n_y,t=0,\ldots,n_t \right\rbrace
\end{aligned}$$ For the sake of simplicity we write down the discrete operators applied to an example variable $u$ respectively its adjoint $\boldsymbol{y}$. The required generalization to the appearing primal and dual variables is left to the reader. The discrete gradient is calculated using forward differences and Neumann boundary conditions. The corresponding adjoint operator consists of backward differences with Dirichlet boundary conditions. The resulting scheme reads: $$\begin{aligned}
u_x(i,j) &= \begin{cases} u(i+1,j)-u(i,j) &\mbox{if } i<n_x \\
0 & \mbox{if } i=n_x \end{cases}\\
u_y(i,j) &= \begin{cases} u(i,j+1)-u(i,j) &\mbox{if } j<n_y \\
0 & \mbox{if } j=n_y \end{cases}\\
\nabla\cdot \boldsymbol{y}(i,j) &=
\begin{cases}
y_1(i,j)-y_1(i-1,j) &\mbox{if } i>0 \\
y_1(i,j) & \mbox{if } i=0\\
-y_1(i-1,j) & \mbox{if } i=n_x
\end{cases}\\
&+
\begin{cases}
y_2(i,j)-y_2(i,j-1) &\mbox{if } j>0 \\
y_2(i,j) & \mbox{if } j=0 \\
-y_2(i,j-1) & \mbox{if } j=n_y .
\end{cases}
\end{aligned}$$
### Warping Operator
The structure of the warping operator $W^i$ depends on the interpolation scheme used to evaluate $\tilde{u}(i,j) = u(i+v_1(i,j),j+v_2(i,j))$. In our observation, a cubic interpolation scheme creates by far the best results.\
The cubic interpolation scheme incorporates a 16-point neighborhood of the target position. This is given by all pairwise combinations $p_{kl} = (i_k,j_l)^T$ of $$\begin{aligned}
i_0 &= floor (i+v_1(i,j))-1, i_1 = i_0 + 1, i_2 = i_1 + 1, i_3 = i_2 + 1,\\
j_0 &= floor (j+v_2(i,j))-1, j_1 = i_0 + 1, j_2 = j_1 + 1, j_3 = j_2 + 1.
\end{aligned}$$ Let us further denote $$\begin{aligned}
\mathcal{I}_{1D}(p_0,p_1,p_2,p_3,x) :=& (-\frac{1}{2}p_0 + \frac{3}{2}p_1-\frac{3}{2}p_2 + \frac{1}{2}p_3)x^3 + (p_0-\frac{5}{2}p_1+2p_2-\frac{1}{2}p_3)x^2 \\
&+ (-\frac{1}{2}p_0 + \frac{1}{2}p_2)x + p_1
\end{aligned}$$ as the one-dimensional cubic interpolation at position $x$ using values at position $p_0,\ldots,p_3$. Then, we obtain the bicubic interpolation of $\tilde{u}(i,j)$ as a combination of two one-dimensional problems: $$\begin{aligned}
\tilde{p}_0 &:= \mathcal{I}_{1D}(p_{00},p_{01},p_{02},p_{03},j+v_2(i,j)),\\
\tilde{p}_1 &:= \mathcal{I}_{1D}(p_{10},p_{11},p_{12},p_{13},j+v_2(i,j)),\\
\tilde{p}_2 &:= \mathcal{I}_{1D}(p_{20},p_{21},p_{22},p_{23},j+v_2(i,j)),\\
\tilde{p}_3 &:= \mathcal{I}_{1D}(p_{30},p_{31},p_{32},p_{33},j+v_2(i,j)),\\
u(i+v_1(i,j),j+v_2(i,j)) &\approx \mathcal{I}_{1D}(\tilde{p}_0,\tilde{p}_1,\tilde{p}_2,\tilde{p}_3,i+v_1(i,j))
\end{aligned}$$ Using this, one can build up the matrix $W^i$. For the row corresponding to the pair $(i,j)$ the quantities $v_1(i,j)$ and $v_2(i,j)$ are static, thus we can calculate $p_{kl}$. Plugging this into the interpolation formula and regrouping it with respect to $p_{kl}$ leads to an equation similar to $\sum_{k,l} \alpha_{kl} p_{kl}$. The weight $\alpha_{kl}$ is written at the column index corresponding to $p_{kl}$. In the boundary region we have to ensure that $i_0,j_0\geq 0, i_3\leq n_x$ and $j_3\leq n_y$. If one of these conditions is violated, the row is set to zero. The adjoint operator to $W^i$ is then simply the transposed matrix.
Parameters {#sec:parameters}
==========
In this section we give a short explanation of parameters and default choices for them, that are used in our experimental setup in Section \[section:experiments\]. For parameters concerning the flow estimation we follow [@sun2014quantitative], where various factors in the optical flow calculation are analyzed and a robust baseline approach is deduced.\
Stable stepsizes for the algorithm can be calculated using [@pock2011diagonal]. We recommend using diagonal matrices $\sigma$ and $\tau$ instead of scalar values for better convergence speed. Thus, the respective parameter choices refer to each element.
- $nSales$ The number of scales in the coarse-to-fine pyramid is dynamically chosen, such that the image size on the coarsest scale is 10 pixels minimum in each direction.
- $\eta$ The downsampling factor between subsequent levels is set to 0.8.
- $nWarps$ We do three warps per level of the pyramid.
- $sizeMed$ The median filtering is done in a 5x5 neighborhood.
- $\epsilon_v, \epsilon_u$ The stopping tolerance for both, the $u$ and the $\boldsymbol{v}$ problem is set to $10^{-6}$.
- $\epsilon_{main}$ The stopping criterion for the main algorithm is set to $10^{-5}$.
- $nRes$ The residual is only updated every 100 iterations, which is sufficient since the algorithm typically needs more that 1000 iterations and the calculation is computationally expensive.
- $\alpha$ Typically, we have for the regularization weight in the image sequence $\alpha\in[0.005,0.1]$ depending on the noise in the input data. We usually do an initial calculation with $\alpha=0.02$.
- $\beta$ Since the flow part incorporates $\beta$ and $\gamma$ we propose to only adjust the ratio $\frac{\beta}{\gamma}\in [0.01,0.1]$ and to start with a regularization weight for both flow components of $\frac{\beta}{\gamma} = 0.02$.
- $\gamma$ The choice for $\gamma$ regulates how strong the images are coupled by the flow component. In our experiments, $\gamma=1$ is the usual choice.
- $\sigma,\tau$ We calculate for the $\boldsymbol{v}$-problem $\sigma^v_1=\sigma^v_2=\frac{1}{2}, \sigma^v_3 = \frac{1}{\|\nabla\tilde{u}\|_1}, \tau^v_1 = \frac{1}{4 + |\tilde{u}_x|}$ and $\tau^v_2 = \frac{1}{4 + |\tilde{u}_y|}$. Analogously, we get for the $u$-problem $\sigma^u_1 = \frac{1}{\|A\|_1}, \sigma^u_2=\frac{1}{2}, \sigma^u_3=\frac{1}{\|\mathcal{W}\|_1}$ and $\tau^u = \frac{1}{\|A^T\|_1 + 4 + \|\mathcal{W}\|_1}$, where $\|\cdot\|_1$ for the matrix refers to the corresponding row-vector.
- $iterMainMax$ The maximum number of iterations for the main algorithm is set to 10 throughout our experiments.
Algorithm
=========
The algorithm can be separated into three problems:
- A $\boldsymbol{v}$-part that calculates the flow field between every two subsequent images.
- A $u$-part that solves the reconstruction problem for the image sequence using the flow information.
- A main algorithm that creates initial values and alternatingly calls the other two functions.
The main algorithm uses a stopping criterion based on the difference between consecutive values for $u$ and $\boldsymbol{v}$ or alternatively stops after a defined number of iteration $iterMainMax$.\
The flow part requires solving a series of minimization problems and, consequently, is the time consuming part of the main algorithm. The structure of the primal-dual algorithm allows parallelization of algorithm and on CPU or GPU to improve the runtime.
FlexBox
-------
An easier approach to minimize the arising variational problems for $u$ and $\boldsymbol{v}$ is given by **FlexBox** [@dirks2016flexible], a flexible primal-dual toolbox. The basic framework for this optimization toolbox is written in MATLAB, but the software comes with an optional C$^{++}$ and CUDA module to massively enhance the runtime. **FlexBox** allows to formulate the primal variational problem term-by-term via MATLAB with only a few lines of code and calculates the solution in the fastest possible way. Creation of dual variables, calculation of step-sizes and transfer to the C$^{++}$ or CUDA solver are fully automated. The full sourcecode can be downloaded from http://www.flexbox.im.\
Init $\boldsymbol{y}_1,\boldsymbol{y}_2,y_3,\boldsymbol{v}$ on coarsest scale Upsample $\boldsymbol{y}_1,\boldsymbol{y}_2,y_3,\boldsymbol{v}$ to current scale using bicubic interpolation [$\boldsymbol{\tilde{v}}$ $\gets$ $\boldsymbol{v}$]{} Calculate $\tilde{u}_2(x+\tilde{\boldsymbol{v}})$ and $\nabla\tilde{u}_2(x+\tilde{\boldsymbol{v}})$ w.r.t. $s$ [$\boldsymbol{y}_1^{k+1}$ $\gets$ $\frac{\boldsymbol{y}_1^k + \sigma \nabla \bar{v}^k_1}{\max(1,\frac{\|\boldsymbol{y}_1^k + \sigma \nabla \bar{v}^k_1\|_2}{\alpha})}$]{} [$\boldsymbol{y}_2^{k+1}$ $\gets$ $\frac{\boldsymbol{y}_2^k + \sigma \nabla \bar{v}^k_2}{\max(1,\frac{\|\boldsymbol{y}_2^k + \sigma \nabla \bar{v}^k_2\|_2}{\alpha})}$]{} [$y_3^{k+1}$ $\gets$ $\max(-1,\min(1,y_3^k + \sigma (\nabla \tilde{u} \cdot \bar{\boldsymbol{v}}^k + \tilde{u}_t)))$]{} [$v_1^{k+1}$ $\gets$ $v_1^k - \tau(\nabla^T\cdot\boldsymbol{y}_1^{k+1} + \tilde{u}_xy_3^{k+1})$]{} [$v_2^{k+1}$ $\gets$ $v_2^k - \tau(\nabla^T\cdot\boldsymbol{y}_2^{k+1} + \tilde{u}_yy_3^{k+1})$]{} [$\bar{\boldsymbol{v}}^{k+1}$ $\gets$ $2\boldsymbol{v}^{k+1} - \boldsymbol{v}^{k}$]{} Update $r^{pd}_v$ every $nRes$ iterations Apply median filter to $\boldsymbol{v}$
Generate operators $\mathcal{A},\bar{\nabla},\mathcal{W}$ [$y_1^{k+1}$ $\gets$ $\frac{1}{\sigma + 1} (y_1^k + \sigma (A\bar{u}^k - f))$]{} [$\boldsymbol{y}_2^{k+1}$ $\gets$ $\frac{\boldsymbol{y}_2^k + \sigma \bar{\nabla} \bar{u}^k}{\max(1,\frac{\|\boldsymbol{y}_2^k + \sigma \bar{\nabla} \bar{u}^k\|_2}{\alpha})}$]{} [$y_3^{k+1}$ $\gets$ $\max(-\gamma,\min(\gamma,y_3^k + \sigma \mathcal{W} \bar{u}^k ))$]{} [$u^{k+1}$ $\gets$ $u^k - \tau(A^Ty_1^{k+1} + \bar{\nabla}^T\cdot\boldsymbol{y}_2^{k+1} + \mathcal{W}^Ty_3^{k+1})$]{} [$\bar{u}^{k+1}$ $\gets$ $2u^{k+1} - u^{k}$]{} Update $r^{pd}_u$ every $nRes$ iterations
[$iteration$ $\gets$ 0]{} [$\boldsymbol{v}$ $\gets$ 0]{} [$u$ $\gets$ problemU($f,\boldsymbol{v},\alpha,0$)]{} [$u_{old}$ $\gets$ $u$]{} [$\boldsymbol{v}_{old}$ $\gets$ $\boldsymbol{v}$]{} [$v_i$ $\gets$ problemV($u_i,u_{i+1},\beta,\gamma$)]{} [$u$ $\gets$ problemU($f,\boldsymbol{v},\alpha,\gamma$)]{} [$r_{main}$ $\gets$ $\|u-u_{old}\| + \|\boldsymbol{v}-\boldsymbol{v}_{old}\|$]{} [$iteration$ $\gets$ $iteration + 1$]{}
Time-Continuous Model
---------------------
We want to make clear that the given algorithm can be used to minimize the original time-continuous model from [@dirks] with only a few modifications. By leaving out the coarse-to-fine pyramid and the intermediate warping steps in the flow problem $\boldsymbol{v}$ (i.e. setting $\tilde{\boldsymbol{v}}=0$ in the evaluation of $\tilde{u}_2$ and $\nabla\tilde{u}_2$) we solve a classical optical flow problem. The sub-problem for the image sequence $u$ can be adjusted by exchanging the warping operator $\mathcal{W}$ with a discretization of the equation $\nabla u_2 \cdot \boldsymbol{v} + u_2-u_1 = K (u_1,u_2)^t$ with $K = [-Id,diag(v_1)\nabla_x + diag(v_2)\nabla_y + Id]$ does directly apply the calculated small-scale flows and has not to be changed at all.\
The limitations from [@dirks] of course apply and the model can only be used from sequences with motion of very small magnitude.
Experimental Results {#section:experiments}
====================
We evaluate the algorithm on grayscale datasets from the Middlebury optical flow database [@baker2011database], which contain at least five consecutive frames and have available ground truth flow fields. Each image is disturbed with Gaussian noise ($\mu=0,\sigma^2=0.01$) to test both, the image denoising and the flow estimation capacities. We use static parameters presented in Section \[sec:parameters\] throughout the evaluation.\
To measure the image reconstruction quality, we calculate L$^2$-error, PSNR and SSIM (default MATLAB implementation) and average them with respect to the number of frames. To evaluate the calculated flow fields, we calculate endpoint error (EPE) [@otte1994optical] and angular error (AE) [@fleet1990computation; @barron1994performance] between our result and the given ground-truth flow field. In the Middlebury sequence, the ground truth field is only given for the flow between the fourth and fifth frame.\
The resulting errors are listed in Table \[tab:jointModelResultsOverview\]. Figure \[fig:results\] gives a visual overview of the denoising capabilities of our algorithm. Both, calculated errors and visual impression indicate a robust and superior reconstruction capability for dynamic image sequences. Moreover, the zoom in Figure \[fig:results\] shows nicely recovered edges, well reconstructed structures and less staircasing effects than comparable total variation regularization techniques. The error in the flow field is relatively high, compared to similar motion estimation techniques. However, one has to take into account that usual benchmarks are calculated on clean input images.
\
\
\
**Gove2** **Grove3** **Hydrangea** **RubberWhale** **Urban2** **Urban3**
------------- -------------- -------------- --------------- ----------------- -------------- --------------
SSIM [0.6920]{} [0.6036]{} [0.7412]{} [0.8350]{} [0.7103]{} [0.6961]{}
L$^2$-Error [3.70e-05]{} [5.62e-05]{} [3.85e-05]{} [2.10e-05]{} [2.91e-05]{} [3.11e-05]{}
PSNR [72.9669]{} [69.2809]{} [74.0326]{} [79.2496]{} [75.1465]{} [74.4912]{}
EPE [1.0290]{} [1.8428]{} [1.3262]{} [0.6566]{} [2.0765]{} [2.2277]{}
AE [0.1713]{} [0.2115]{} [0.1396]{} [0.2514]{} [0.1588]{} [0.2395]{}
: Image reconstruction and flow estimation errors for sequences (manually disturbed with Gaussian noise) from the Middlebury database [@baker2011database].[]{data-label="tab:jointModelResultsOverview"}
Conclusions
===========
In this work, we propose a detailed implementation of the work presented by Burger, Dirks and Schönlieb [@burger2016variational] and extend their time-continuous model for small-scale flows to arbitrary large-scale motion between consecutive frames. The arising variational problem is minimized using a primal-dual method and the operator-discretization is explained in detail. We present concrete parameter choices and give a pseudo-code of the implemented algorithm. Finally, a short numerical evaluation based on data from the Middlebury optical flow database is presented.\
An obvious possible extension would be to deal with color-images. The color-channels can be coupled here to improve the results. Moreover, an implementation to higher spatial dimensions is possible, which causes an increased runtime, and also may require a different motion model (e.g. mass-preservation) on the other hand.
[^1]: Institute for Computational and Applied Mathematics and Cells in Motion Cluster of Excellence, University of M[ü]{}nster, Orléans-Ring 10, 48149 M[ü]{}nster, Germany, Email: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
The implementation, deployment and testing of secure services for Internet of Things devices is nowadays still at an early stage. Several frameworks have recently emerged to help developers realize such services, abstracting the complexity of the many types of underlying hardware platforms and software libraries. Assessing the performance and usability of a given framework remains challenging, as they are largely influenced by the application and workload considered, as well as the target hardware. Since 15 years, [<span style="font-variant:small-caps;">Arm</span>]{}processors are providing support for [<span style="font-variant:small-caps;">TrustZone</span>]{}, a set of security instructions that realize a trusted execution environment inside the processor. [<span style="font-variant:small-caps;">Op-Tee</span>]{}is a free-software framework to implement trusted applications and services for [<span style="font-variant:small-caps;">TrustZone</span>]{}. In this short paper we show how one can leverage [<span style="font-variant:small-caps;">Op-Tee</span>]{}for implementing a secure service (, a key-value store). We deploy and evaluate the performance of this trusted service on common Raspberry Pi hardware platforms.
We report our experimental results with the data store and also compare it against [<span style="font-variant:small-caps;">Op-Tee</span>]{}’s built-in secure storage.
author:
- 'Christian Göttel, Pascal Felber,'
- Valerio Schiavoni
bibliography:
- 'biblio.bib'
title: |
Developing Secure Services for IoT with OP-TEE:\
A First Look at Performance and Usability[^1]
---
Introduction {#sec:intro}
============
Despite the availability of security-oriented instruction sets in consumer-grade processors, high-level frameworks that can help developers use such extensions are still at an early stage. Moreover, little has been said regarding the performance and usability of these frameworks. This is unfortunate given that the large majority of devices featuring [<span style="font-variant:small-caps;">Arm</span>]{}processors (mobile and not) feature the [<span style="font-variant:small-caps;">TrustZone</span>]{}extensions, introduced since 15 years [@armeverywhere], and are constantly being improved with new processor revisions. For instance, [<span style="font-variant:small-caps;">Arm</span>]{}recently [@armsvirt] updated its ARMv8.4 architecture of application processors enabling virtualization in the secure world. The introduction of virtualization in the secure world better improves the isolation of components and resources, and it is expected to boost the trusted applications (TA) ecosystem in developing and using common standards and APIs.
It is only very recently that the first open-source tools aiming to exploit these capabilities have emerged. Notable examples include Linaro ARM Trusted Firmware [@linaro-trusted-firmware], ARM GNU Toolchain [@arm-toolchain], Android’s Trusty [@android:trusty], Trustonic’s Kinibi [@trustonic:kinibi], NVIDIA’s TLK [@nvidia:tlk] and finally Linaro’s [<span style="font-variant:small-caps;">Op-Tee</span>]{} [@optee].
A major challenge for developers of trusted applications resides in the complexity of the secure platforms themselves. Despite the existence of standards and APIs, trusted applications remain OS-specific because of the custom libraries provided by the different vendors. Theses libraries are specialized for the various processors and are required to access secure storage and processing elements. They rely on drivers shipped with the hardware by the silicon manufacturer. Furthermore, dispatching trusted OSs requires trusted OS-specific code in the firmware, which adds up to the issue. This greatly hinders the portability of trusted applications across different trusted OSs and drives TA developers toward implementing and supporting several versions of trusted OS-specific TAs.
In this paper, we focus on a specific framework, [<span style="font-variant:small-caps;">Op-Tee</span>]{} [@optee], which has gained much attraction recently and is arguably the most mature open-source framework for developing trusted application with [<span style="font-variant:small-caps;">Arm</span>]{}’s [<span style="font-variant:small-caps;">TrustZone</span>]{}extensions. We describe its architecture and features, and we evaluate its usability and performance by developing a simple key-value store. We also execute [<span style="font-variant:small-caps;">Op-Tee</span>]{}’s secure storage benchmark and report our results. This preliminary study bring insights into the benefits of such framework, able to hide the complexity of the underlying vendor-specific libraries and processor, as well as their performance and overhead.
![Organization of components within [<span style="font-variant:small-caps;">TrustZone</span>]{}and interaction with [<span style="font-variant:small-caps;">Op-Tee</span>]{}[]{data-label="fig:tz"}](figures/tz-optee.pdf)
Background {#sec:background}
==========
TrustZone in a Nutshell {#subsec:tz}
-----------------------
The [<span style="font-variant:small-caps;">TrustZone</span>]{}technology is available in [<span style="font-variant:small-caps;">Arm</span>]{}processors since 2003 [@armsvirt]. It is a hardware-enforced mechanism isolating a *secure world* (trusted) from a *normal world* (untrusted), which includes all components within the SoC as well as peripherals. Thus, [<span style="font-variant:small-caps;">TrustZone</span>]{}provides secure endpoints and enables device root-of-trust. Software running in the normal world is unable to directly access secure components and resources. When booting up a [<span style="font-variant:small-caps;">TrustZone</span>]{}-enabled SoC, secure firmware is the first software component executed at *exception level 3* (EL3). The secure firmware code is responsible for initializing the platform, installing the trusted *operating system* (OS) and routing secure monitor calls. The trusted OS consists of a small and secure kernel to execute *trusted applications* (TA). Once the secure world is set up, the normal world OS is booted in parallel to the trusted OS running in the secure world. Worlds can be switched via a software-based *secure monitor* (ARMv8-A) or in hardware (ARMv8-M) [@arm:tzv8m]. The secure monitor acts as a gateway and runs at the highest privilege level EL3 [@armv8].
The GlobalPlatform Specifications for TEEs {#subsec:gp}
------------------------------------------
The main specifications for secure digital services and devices are published by industry associations [@gp:home; @tcg:home]. In our study, we focus on the [<span style="font-variant:small-caps;">GlobalPlatform</span>]{}specifications for [<span style="font-variant:small-caps;">TrustZone</span>]{}. A *rich execution environment* (REE) is an execution environment that involves at least one device and all its components or an OS, excluding any trusted or secure component. In contrast, a *trusted execution environment* (TEE) provides a level of security to protect against attacks and secures data access. The TEE executes alongside the REE, but is shielded from it. A trusted application executes inside a TEE and exposes secure services to applications in the REE. *Trusted storage* is a hardware or cryptographically-protected device capable of storing data [@gp:sysarch]. Data can be exchanged between an application in the REE and a TA by three types of shared memory: *whole* (an entire memory region and is allocated by the TEE), *partial* (only a a subset of the *whole* with a specified offset), and *temporarily*, for which a memory buffer region allocate by the application in the REE temporarily shared with the TA for the duration of the API call [@gp:client].
[>[-]{}lll<[-]{}]{} & &\
CPU & Intel Xeon E3-1270 v6 & Broadcom BCM2837\
CPU Frequency & &\
Memory & DDR4 & LPDDR2\
Memory Frequency & &\
& Samsung & Transcend micro SDHC\
& MZ7KM480HMHQ0D3 & UHI-I Premium\
Disk Size & &\
Disk Read Speed & &\
The OP-TEE Framework {#subsec:optee}
--------------------
[<span style="font-variant:small-caps;">Op-Tee</span>]{} [@optee] is a TEE implementation of [<span style="font-variant:small-caps;">GlobalPlatform</span>]{}specifications on top of [<span style="font-variant:small-caps;">TrustZone</span>]{}. It can be used alongside a Linux-based distribution running in the REE. TAs are single-threaded executables stored inside the REE. Users develop TAs without having to recompile the entire framework. [<span style="font-variant:small-caps;">Op-Tee</span>]{}does not provide mechanisms to verify the integrity of a TA, and consequently it exposes TAs to the untrusted REE. Upon modification, this can compromise the integrity or protection of the TEE. Alternatively, TAs can be directly integrated into [<span style="font-variant:small-caps;">Op-Tee</span>]{}as *pseudo TAs*. Pseudo TAs run inside the [<span style="font-variant:small-caps;">Op-Tee</span>]{}OS as secure privileged-level services without access to [<span style="font-variant:small-caps;">GlobalPlatform</span>]{}’s Internal Core API. Thus, pseudo TAs can only use [<span style="font-variant:small-caps;">Op-Tee</span>]{}’s core Internal API. Secure storage allows applications to offload data from a TA to either the REE file system or a *replay protected memory block* (RPMB) partition of an *embedded multi-media controller* (eMMC) device using the Internal Core API. By default, the [<span style="font-variant:small-caps;">Op-Tee</span>]{}OS is configured to use the RPMB [@optee:secstor] if available. The secure storage is accessible and visible only to the TA that created it.
Usability {#sec:usability}
=========
The communication between an application in the normal world and a TA evolves around functions handling the context, session, command and shared memory as shown in . This facilitates interoperability between different [<span style="font-variant:small-caps;">GlobalPlatform</span>]{}API compatible TEE implementations and allows REE applications to set up multiple contexts. A context is initialized by referencing the device file connecting to the TEE driver . TAs are identified by a *universally unique identifier* (UUID), which is referred to when setting up a session to a TA . To set up a session, [<span style="font-variant:small-caps;">Op-Tee</span>]{}will load the TA from the normal world to the secure world with the help of `tee-supplicant` . The `tee-supplicant` is a daemon running in the normal world used by [<span style="font-variant:small-caps;">Op-Tee</span>]{}to request services from the REE. These steps are skipped when a session to a pseudo TA is established. A TA can initialize and set up its environment upon TA creation and session establishment ( & ). From this point on, the REE application can request services from the TA by invoking commands. These commands can pass up to four parameters, which are either values or references to shared memory regions. Values are pairs of unsigned integers. Shared memory regions are allocated, registered and released through [<span style="font-variant:small-caps;">GlobalPlatform</span>]{}API calls in `libteec`. Without the availability of `libteec`, developers would have to communicate directly with the kernel driver through `ioctl` system calls.
In [<span style="font-variant:small-caps;">Op-Tee</span>]{}, TAs can use services accessible through [<span style="font-variant:small-caps;">GlobalPlatform</span>]{}Internal Core API implemented in `libutee`. TAs are statically linked against `libutee`, which wrapps the API functions around assembler macros to [<span style="font-variant:small-caps;">Op-Tee</span>]{}OS system calls. The library provides interfaces to secure storage , time, arithmetic and cryptographic operations . The secure storage API encrypts data objects by the use of a secure storage service. The encryption process involves three keys: *secure storage key* (SSK), *trusted application storage key* (TSK) and *file encryption key*. The SSK is generated from the *hardware unique key* and is used to derive TSKs. Each TA has a TSK that is generated from the SSK and the TA’s UUID. Both SSK and TSK are generated using HMAC SHA256 algorithm [@optee:secstor]. Finally, every created file a FEK is generated from the pseudo random number generator. The encrypted data objects are then transferred to the `tee-supplicant` by a series of remote procedure calls and stored in a special file. [<span style="font-variant:small-caps;">Op-Tee</span>]{}further provides TAs with libraries for TLS and SSL protocols (`libmbedtls` [@mbedtls]), arithmetic (`libmpa`) and a subset of ISO C functions (`libutils`). Once the REE application has no further service requests, the session is terminated and the context is destroyed.
[0.48]{}
[0.48]{}
\
1em
[0.48]{}
[0.48]{}
Performance Evaluation {#sec:evaluation}
======================
Setup
-----
The key-value store and [<span style="font-variant:small-caps;">Op-Tee</span>]{}’s Sanity Testsuite v3.2.0 [@optee:xtest] were deployed on the two platforms: Dell PowerEdge R330 Server and Raspberry Pi 3B v1.2. The Dell PowerEdge R330 is running Ubuntu 18.04.1 LTS with the 4.15.0-43-generic Linux kernel and is used to emulate the Raspberry Pi 3B platform with QEMU v2.12.0 [@qemu]. A comparison of the two platforms can be found in . [<span style="font-variant:small-caps;">Op-Tee</span>]{}provides a build environment which, by default, deploys and emulates its OS on an ARM Virtual Machine `virt` using a Cortex-A57 with no more than two cores. The deployment was changed to match the specification of the Raspberry Pi 3B platform as close as possible.
Shared Memory
-------------
We have ported a simple key-value store to a TA, in order to evaluate the overhead and performance of different types of shared memory. As basis, we used a modified version of the hash table implementation of `kazlib` v1.20 [@kazlib], removing support for contexts and dynamic tables. The hash table is static, uses separate chaining to resolve collisions, applies a modular hashing and has chains. We time every `DEL` (delete), `GET` and `PUT` operation for each benchmark by referring to `CLOCK_MONOTONIC` in the REE. Operations are uniformly distributed and issued times at a rate of operations per second.
When using whole or partially shared memory, the REE application requests a shared memory region of and fills it with random data from `/dev/urandom`. Similarly, the REE application allocates and initializes a buffer used as temporarily shared memory. Before every invocation of a key-value operation, a random offset into the shared memory region is computed, which is also used as key. A chunk size of beginning at the random offset is used as data object. The `PUT` benchmark starts with an empty hash table. The `DEL` and `GET` benchmarks start with a pre-populated hash table of data objects. Finally, the mixed benchmark (ratio of `GET` and `PUT` operations) begins with a pre-populated hash table relative to the percentage of `GET` operations.
shows throughput and latency for the different shared memory types and for running the key-value store entirely in the REE. On the QEMU platform, the operations do not separate as well as on the Raspberry platform; we assume due to reaching an I/O bound. The operations on the Raspberry platform separate as expected according to their throughput (lowest to highest): `PUT`, `MIX50`, `MIX20`, `GET`, and `DEL`. The overhead of the `PUT` operation is due to memory allocation, memory copy and object insertion. The `GET` operation looks up a data object and copies it to shared memory, resulting in a lower overhead. The higher the portion of `PUT` operations in the `MIX` benchmarks is, the slower the average operation speed becomes. Thus, `MIX50` ( `PUT` operations) has a lower average throughput than `MIX20`. The `DEL` operation looks up a data object and frees its memory, avoiding time consuming memory operations. Comparing TEE throughput against REE throughput yields a overhead on the QEMU platform and a overhead on the Raspberry platform. A similar experiment was conducted in [@scitepress], where they compared the time spend in normal and secure world when invoking a noop operation.
Secure Storage
--------------
The secure storage benchmark is part of the [<span style="font-variant:small-caps;">Op-Tee</span>]{}sanity test suite adhering to the *Trusted Storage API for Data and Keys* described in [@gp:core]. Neither of the platforms is equipped with an eMMC, for which reason the secure storage has to be offloaded to the REE file system. The benchmark executes three commands `WRITE`, `READ`, and `REWRITE`, for data sizes in the range of to , that are accessed in chunks of at most . The `REWRITE` command first reads data from an object, resets the cursor and writes the data back to the same object. The data to be stored in the secure storage is allocated and filled with scrambled data within the TEE.
shows the overhead of accessing data in chunks of in the secure storage. In general, the overhead becomes more significant with increasing data sizes, more precisely once the data size exceeds the chunk size. Maximum speed is achieved when the data size equals the chunk size. Overall, the `REWRITE` command has the highest overhead, because it basically executes the `READ` and `WRITE` commands in one batch.
Concluding Remarks {#sec:conclusion}
==================
Development of secure services benefits from well established APIs and standards. [<span style="font-variant:small-caps;">Op-Tee</span>]{}has implemented several of [<span style="font-variant:small-caps;">GlobalPlatform</span>]{}’s specifications and APIs and provides common interfaces for secure services. We have ported a simple key-value store to a TA and we have studied the performance and usability of secure storage and shared memory. The results of our benchmarks have shown that requesting services from TAs in [<span style="font-variant:small-caps;">TrustZone</span>]{}on ARMv8-A using [<span style="font-variant:small-caps;">Op-Tee</span>]{}incurs a significant overhead compared to service execution in the normal world. Limiting the space available to a TA is sensible, in order to minimize the trusted computing base. However, the default memory limit of for TAs in [<span style="font-variant:small-caps;">Op-Tee</span>]{}becomes a major inconvenience with respect to secure storage and shared memory.
Generating the SSK in [<span style="font-variant:small-caps;">Op-Tee</span>]{}requires the HUK. However, most platforms lack of documentation to access or obtain the HUK. [<span style="font-variant:small-caps;">Op-Tee</span>]{}avoids this issue by considering a static string value instead of the HUK. This alternative can potentially weaken the cryptographic protection of the objects stored in the REE file system of the secure storage. TEEs would greatly benefit from unrestricted access to HUKs and could so improve the protection of trusted storage.
We expect the trusted application ecosystem to improve portability of TAs among TEEs. Furthermore, we hope that our evaluation of usability and performance of TAs provides deeper insight into future development of trusted services.
Acknowledgments {#acknowledgments .unnumbered}
===============
The research leading to these results has received funding from the European Union’s Horizon 2020 research and innovation programme under the LEGaTO Project ([legato-project.eu](https://legato-project.eu/)), grant agreement No 780681.
[^1]: This is a pre-print of an article published in “Distributed Applications and Interoperable Systems” (DAIS) 2019. The final authenticated version is available online at: <https://doi.org/10.1007/978-3-030-22496-7_11>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Over the last years many technological advances were introduced in Internet television to meet user needs and expectations. However due to an overwhelming bandwidth requirements traditional IP-based television service based on simple client-server approach remains restricted to small group of clients. In such situation the use of the peer-to-peer overlay paradigm to deliver live television on the Internet is gaining increasing attention. Unfortunately the current Internet infrastructure provides only best effort services for this kind of applications and do not offer quality of service.
This paper is a research proposition which presents potential solutions for efficient IPTV streaming over P2P networks. We assume that the solutions will not directly modify existing P2P IPTV protocols but rather will be dedicated for a network engineer or an Internet service provider which will be able to introduce and configure the proposed mechanisms in network routers.
author:
-
bibliography:
- 'zotero.bib'
title: Methods of QoS improvement for P2P IPTV based on traffic modelling
---
Network Architecture and Design; Distributed applications; Wide-area networks; Emerging technologies;
Introduction
============
Television is one of the most dominant and pervasive mass media; it is watched across all age groups and by almost all countries in the world. Over the last years, many technological advances were produced by trying to meet user needs and expectations in such a widespread media. Traditional Internet TV (IPTV) service based on simple client-server approach restricted small group of clients, the overwhelming bandwidth requirement makes it impossible when the number of user grows to thousands or millions because servers have limited available resources (CPU, bandwidth) that will decrease proportionally with the number of users. By multiplying the servers and creating content distribution network (CDN), the solution will scale only to a larger audience with regards to the number of deployed servers which may be limited by the infrastructure costs. Finally, the lack of deployment of IP-Multicast limits the availability and scope of this approach for a TV service on the Internet scale. Therefore the use of the peer-to-peer overlay paradigm (P2P) to deliver live television on the Internet (P2P IPTV) is gaining increasing attention, and has become a promising alternative [@liu_opportunities_2008].
From the technical point of view, in P2P IPTV system, a local user (or peer) act both as receiver and supplier of the IPTV data. Connected to “upstream” pears, the user receives IPTV data chunks and forwards them to other “downstream” receivers, watching the program which, in turn, can forward it further down in a hierarchy of peers – fig. \[fig:camera\]. Consequently, such an approach has the potential to scale with group size, as greater demand also generates more resources. Moreover, by using the existing Internet infrastructure as a medium and by exploiting user participation for the creation of the content distribution network, P2P IPTV technologies have innovative potential by making any TV channel from any country globally available and allowing Internet users to broadcasting their own TV with low costs.
![P2P IPTV system[]{data-label="fig:camera"}](figs/camera.eps){width="3.1in"}
P2P research challenges
=======================
P2P applications are posing serious challenges to Internet infrastructures and there is continuous battle between service providers and P2P applications for traffic management. One of the major problems is QoS provisioning. Unlike file sharing, the live media need to be delivered almost synchronously to large number of users, with minimum delay in playback compared to the playback at the source. Due to the large volume of data in the media stream, it is of paramount interest to avoid redundant transmission of the stream. Constructing efficient paths for streaming is especially hard because the nodes participating in the overlay have very minimal information regarding the topology of the baseline network.
Real time services such as IPTV are inelastic, as the transmission bandwidth, transmission time and QoS requirements need to be kept within strict limits and hence are not flexible. Current Internet infrastructure provides best effort services that do not offer quality of service. These issues result into lower throughput (bandwidth management), packet losses, high transfer delay, delay variation (jitter) and out-of-order delivery. These parameters are unpredictable and never acceptable for real-time applications. Thus, we need to design solutions for efficient video streaming over P2P networks that can address the above-mentioned issues. We assume that the solutions will not directly modify existing P2P IPTV protocols but rather will be dedicated for network engineer or Internet service provider (ISP) which will be able to introduce and configure the proposed mechanisms in network routers.
In the prosed research we will try to answer the following questions: in which case proper ISP assistance is needed to help P2P IPTV system and what are the most proper services ISP should provide to support P2P IPTV. Taking into account the above assumptions we will focus on QoS provisioning for P2P IPTV services by improving the overall received video throughput with low packet drop ratio, transmission delay and jitter – treating them simultaneously as our QoS metrics. We concentrate on provisioning to P2P IPTV some of the methods influencing QoS based on differentiated services and including traffic control which regulates data flows by classifying, scheduling, shaping traffic and admission control determining which applications and users are entitled to network resources.
There are also other mechanisms which lead to improvement of QoS. One of them is redesign overlay network topology which should results in improving proposed QoS metrics. Early P2P streaming systems were designed as alternatives to IP multicast. Therefore they all attempt to build application level multicast trees. However, a tree structure is unsuited in a P2P environment. A tree is vulnerable to node failures and streaming rate of a peer cannot exceed that of its parent. To address the problem of tree based architectures, current systems adopts rather a multi-parent mesh approach for P2P streaming – fig. \[fig:classyfication1\]. The idea is to allow each peer to stream media data from multiple neighbour peers. To coordinate the streaming from multiple sources in P2P IPTV system, usually a pull-based approach is used where peer collects data availability from its neighbours, and request different data blocks from different neighbours. While the multi-parent, receiver-driven approach offers great flexibility in dealing with peer and network dynamics, a remaining question is how should peers select their neighbours and what does the resulting overlay look like? P2P applications usually do not implement any algorithms to select peers with the best connectivity (e.g. considering delays, throughput, etc.). There exists solutions improving the connection choice based on hybrids mixing a tree with mesh structure [@zhang_peer-to-peer_2005] or peer ranking [@xie_p4p:_2008][@aggarwal_isp-aided_2008] but in most of the P2P streaming platforms the overlay neighbours are chosen randomly [@zhang_coolstreaming/donet:data-driven_2005] [@wu_magellan:_2007].
![Classification of [P2P]{} streaming systems[]{data-label="fig:classyfication1"}](figs/classyfication1.eps){width="3.1in"}
QoS improvement propositions
============================
Taking into account the goals stated in previous section we proposed a few solutions which should improve the overlay topology between peers and should help deliver video content efficiently. Proposing the solution we assume that ISPs and P2P systems collaborate so that both benefit: 1) the ISP keeps a significant portion of their network traffic localised within their internal network, and hence gain cost advantages by reducing costs for traffic that leaves their network boundary 2) better management of traffic flow may provide better service to customers and ensure fairness for other applications like VoIP or web traffic.
Caching
-------
P2P caching is similar in principle to the content caching long used by ISPs to accelerate Web (HTTP) content. P2P caching temporarily stores popular content that is flowing into an ISP’s network. If the content requested by a subscriber is available from a cache, the cache satisfies the request from its temporary storage, eliminating data transfer through expensive transit links and reducing network congestion. However we must note that in case of IPTV the expiration time of the cache content will be quite short in contrary to other P2P services like file sharing. Content placement deals with how many replicas of each object has and where in the network to place them. Intuitively, cache servers should be placed in such a manner that they are closer to the clients, thereby reducing latency and bandwidth consumption – fig. \[fig:p2pCache1\]. Also, content replicas should be placed to even the load of the replica servers in P2P network, that is, trying to balance the load among cache servers. In a our case, TV content from source server is delivered to P2P network which is responsible for streaming the content between its peers. So one specific question is: how to select from the available peers a streaming server within the given ISP domain? Another question is: how to place cache servers among all possible locations to cut down cost and improve performance of the whole ISP network? After selecting the location of servers, we need to decide what amount of media content the server hold. The problem of server placement may be modelled as center placement problem: for the placement of a given number of centers, minimise the maximum distance between a node and the nearest center. There are several class of algorithms which are used for solving this problem: greedy, hot-spot and tree-based [@weigmann_center_1997]. In the particular case a cache server be may installed at the border between the local user base of the ISP and the Internet cloud. Based on destination port number for each TCP connection all P2P IPTV traffic can be redirected to this server. Thus, the server is able to intercept all downloads performed by local users [@wierzbicki_cache_2004]. The aforementioned idea is presented in literature in context of CDN and P2P networks. In [@shen_hptp:_2007] authors proposed to use caching to relieve the tension between ISPs and P2P systems. A network layer packet-level caching for reducing the volume of emerging P2P traffic is proposed in [@nakao_remedy_2008]. In [@chen_server_2008] authors formulate server placement problem and in P2P streaming system and propose solution schemes for the sub-problem of server selection and rate assignments.
![Cache server placement problem[]{data-label="fig:p2pCache1"}](figs/p2pCache1.eps){width="3.1in"}
Prioritisation and smoothing
----------------------------
One method to reduce information loss (due to buffer overflow or end-to-end delay) experienced by multimedia traffic is through the use of multiple priority traffic classes within the network. Assigning higher priority to multimedia traffic throughout the network will prevent latency tolerant data traffic from delaying time critical multimedia streams. Traffic priority classes will also improve performance in systems whose peak data rate exceeds network capacity.
When there are large bursts of data traffic, there will be severe congestion on the network. The end-to-end delay experienced by the multimedia streams will increase, resulting in poor performance. Traffic flows are aggregated in the network, so that core routers only need to distinguish a comparably small number of aggregated flows, even if those flows contain thousands or millions of individual flows. In order to support real time traffic, we need a mechanism to prioritise data. This is done by classifying traffic into service classes based on expected traffic patterns. Each service class has a data priority level and associated guarantees. Scheduling algorithms determine which packet to send next and are used primarily to manage the allocation of bandwidth between flows. We propose this approach for P2P IPTV traffic which needs real time guarantees – fig. \[fig:scheduling\]. Applying different scheduling algorithms we intend to observe the dynamic and interaction of overlay with baseline network. For example the simplest scheduling is always transmitting high priority traffic first. If the multimedia traffic is given priority, it will not be affected by these large bursts of data. This guarantees the IPTV traffic will always have good performance. However, it is possible that lower priority traffic will never be serviced. This will be true as long as the average bandwidth requirements do not exceed the capacity of the network. Once the network capacity becomes insufficient to handle the average requirements, the only solutions are to increase capacity, change scheduling algorithm or impose restrictions on the transmitting stations. Taking into account the last case we will attempt to use queue management algorithm to prevent network overloading.
![P2P IPTV packets scheduling[]{data-label="fig:scheduling"}](figs/scheduling.eps){width="3.1in"}
Compressed digital video is inherently variable-rate since its complexity and motion content affect the encoding bit rate required to maintain picture quality which leads to uncontrolled burstiness. Traffic burstiness is one of the reasons of inefficient use of network resources by occasionally requiring excessively high processing, storage (buffering), and transmission capacity from the network. By transmitting smoother traffic, the network can improve its utilisation. However the price for smoothing is either reduction of bandwidth the streams consumes or packet drops or delays in stream transmission. We propose delaying the playback of the P2P IPTV traffic to permit the source a transmission over a larger interval (window) of frames, based on the buffer space available at the client site. However even if we delay the stream transmission, if such smoother stream traverses multiple hops its queuing delay is reduced on each hop resulting in considerable decrease in whole end-to-end delay bound which overweights the initial delay [@paola_improving_1996]. We propose to implement traffic shapers in cache servers mentioned in previous proposition or in border router. The traffic shapers may be switched on and off depending on the hop distance between the cache server and a receiver – fig. \[fig:smoothing\].
![P2P IPTV traffic smoothing[]{data-label="fig:smoothing"}](figs/smoothing.eps){width="3.1in"}
Congestion avoidance
--------------------
As it was stated in previous section in most P2P streaming platforms, the overlay neighbours are chosen randomly which may lead to the situation when one link may be heavily congested while other links in the network remain lightly loaded or the P2P IPTV packets may traverse a long path with high propagation delay when a low-latency path is available. The responsibility of selecting the path, which a packet follows through the baseline network, falls to the routing protocols implemented by the individual routers in the network. Rather than using hard-wired tables to forward the packets, the routers exchange control messages with each other to compute the paths through the network in a distributed fashion. The distributed approach allows a collection of routers to adapt automatically to changes in the network topology. This makes IP networks robust in the presence of link and router failures, and easily accommodates the deployment of new equipment as the network grows. However, the routing protocols deployed in most IP networks do not incorporate information about network load and performance into the selection of the paths. Left to their own devices, the routers continue to forward packets over heavily loaded link. A near-optimal intra-domain load balancing can be achieved by altering the link weights of the network’s routing protocol (OSPF or IS-IS) [@rexford_route_2005]. The selection of weights depends on having an estimate of the offered load on the network, in terms of the volume of traffic between each pair of routers or each pair of edge links. This kind of information is necessary for designing the network and planning the outlay of new capacity. In some cases, the operator may have an estimate of the traffic demand based on past experience or customer subscription information. In other cases, the traffic demands can be obtained by measuring the traffic in the operational network. Computing estimates of the offered load requires combining measurement data from multiple locations in the network to compute the traffic demands. This allows the use of optimisation techniques for identifying parameter settings that satisfy the network’s performance goals – fig. \[fig:routing\].
![ISP routing improvement in presence of P2P IPTV traffic[]{data-label="fig:routing"}](figs/routing.eps){width="3.1in"}
Implementation draft
====================
Proposed traffic engineering tasks including caching, scheduling, smoothing and load balancing require an effective way to predict the flow of traffic through the network. In such circumstances traffic matrices (TM) are helpful which reflect the volume of traffic that flows between all possible pairs of sources and destinations in a network. Constructing a network-wide view of the traffic demands requires relatively sophisticated techniques for the collection and analysis of measurement data. Traffic statistics may be available directly from SNMP, can be computed by combining packet-level or flow-level measurements at the network edge with the information available in routing tables, may be inferred based on observations or sampling of the aggregate load on links inside the network in conjunction with routing data [@fortz_traffic_2002].
The above methods of TM estimation would require access to ISP network equipment. Additionally the measurement would present a snapshot of ISP traffic in certain period of time with no access to parametrisation like 1) dynamics of the P2P system – nodes availability; 2) nodes links capacity; 3) TV channel popularity; 4) overlay and baseline networks topology; 5) exchange protocol – e.g. peer selection, chunks scheduling; 6) video encoder used. Instead we propose to use physical modelling for P2P network traffic which tries to explicate the physical causes of certain traffic behaviour in the system based on network mechanisms and empirical established properties of the system. In order to build a comprehensive P2P network traffic model we must take into account behaviour of P2P system in three main categories: individual peer behaviour, shared contents (TV) and multiple peer characteristics. As a result of the model, we obtain internal and external ISP link utilisation in time domain – fig. \[fig:trafficModel\].
![Inputs and output for the proposed traffic model[]{data-label="fig:trafficModel"}](figs/trafficModel.eps){width="3.1in"}
Model parameters will be obtained from analysis and measurement of real open-source P2P IPTV application GoalBit [@bertinat_goalbit:first_2009]. GoalBit is capable of distributing high-bandwidth live-content using a Bittorrent-like approach where the stream is decomposed into several flows sent by different peers to each client. The system has also built-in mechanism of perceived quality measurement. P2P traffic can be broadly classified into two categories: signalling and data transfer. The signalling traffic includes TCP connection set-up, search queries and query replies. Its volume heavily depends on the type of P2P network (structured or unstructured) and protocol used. The leading content shared in the P2P IPTV systems tend to be larger in size compared to the signalling traffic although the signalling traffic packets are sent more frequently [@silverston_traffic_2009]. To obtain comprehensive view of network behaviour we will take into account both types of traffic.
The traffic model will be important element of our project and will be applied in all proposed solutions: 1) traffic caching – due to the spatio-temporal traffic model we may identify the optimal placement of the cache server; 2) traffic scheduling – the temporal traffic model will be used for selection of optimal parameters for scheduling algorithm [@lazar_modeling_1994]; 3) traffic smoothing – temporal traffic model will allow to choose optimal parameters for the smoothing mechanisms; 4) weights manipulation – the spatial traffic model may be used instead of the costly and time intensive measurements of traffic intensity in ISP network.
The next issue is the implementation and comparison of the traffic models and proposed solutions. We take into account analytical models, simulations and experiments on the real systems. Large-scale distributed systems are complex and accurately modelling them analytically is not an easy task. In many cases the first iteration of an analytical model is not tractable and the model needs to be successively simplified to produce useful insights. This leads to models that describe the system at a very coarse-grained level and with many uniformity assumptions. Despite this, we plan to use analytical models for early feasibility assessment of a P2P solutions. However, for a more complete and accurate evaluation under a wider range of conditions we incline towards a simulation and system evaluation on the real networks.
Usually the simulators are purpose-built for specific applications or classes of applications. Few simulators are designed as more general tools for system building and evaluation. What is more, even though simulation is a widely used evaluation technique there is almost no simulator code sharing among the researchers and little standardisation of the common practises [@naicken_survey_2006-1]. There are a number of dedicated P2P simulators although no all of them have the functionality that we would expect. Given these issues with current P2P simulators and the importance of reproducing the results, in this project we plan use open-source OMNeT++. In order to obtain spatio-temporal traffic model the simulation of overlay network will be based on underlying network including most important physical network elements like workstations, routers, links etc.
Validation of a model of complex P2P system based on performing experiments with the actual system requires significant resources which could be very costly in hardware and administration, and is vulnerable to node failures. There may also be factors external to the experiment that can not be controlled, yet influence experiment results, such as cross-traffic and changes in the properties of the baseline network. Thus we plan to use where possible a virtual test-bed – Planetlab [@chun_planetlab:overlay_2003] or P2P-Next [@jimenez_p2p-next:_2009] which are useful tool for performing large scale experiments on overlay protocols and for validating some results obtained from simulation results. For creation our P2P IPTV application we consider to use a peer-to-peer systems prototyping toolkit ProtoPeer that allows for switching between the event-driven simulation and live network deployment without changing any of the application code [@galuba_protopeer:simulation_2008].
For the most of the proposed solutions the P2P IPTV traffic flows must first be identified. We assume that popular traffic identification techniques are available based for example on ports numbers or signature matching. We also consider applying more advanced method of statistical traffic identification based on the developed traffic model and guidance presented in [@kim_internet_2008].
Conclusions
===========
In the paper we proposed a few solutions which should improve QoS in P2P IPTV network. We assume that the solutions will not directly modify existing P2P IPTV protocols but rather will be dedicated for network engineer or Internet service provider which will be able to introduce and configure the proposed mechanisms in network routers. The proposed methods require an effective way to predict the flow of traffic through the network. We are going to use physical modelling for P2P IPTV network traffic which tries to explicate the physical causes of certain traffic behaviour in the system based on network mechanisms and empirical established properties of the system. The challenge lies in combining relevant feature from workload modelling, network architecture (topology, protocols), users behaviour, and analytical modelling into consistent description of P2P system. Because we are to model a complex distributed system we incline towards a simulation and system evaluation on the real networks. We expect to find answer for the question: which of the proposed solutions is most suitable for QoS improvements in P2P IPTV networks taking into account its efficiency, easiness of implementation and potential side effect on other kinds of network traffic.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We present a principal component analysis method which tracks and compensates for short-timescale variability in pulsar profiles, with a goal of improving pulsar timing precision. We couple this with a fast likelihood technique for determining pulse time of arrival, marginalizing over the principal component amplitudes. This allows accurate estimation of timing errors in the presence of pulsar variability.
We apply the algorithm to the slow pulsar PSR J2139+0040 using an archived set of untargeted raster-scan observations at arbitrary epochs across four years, obtaining an improved timing solution. The method permits accurate pulsar timing in data sets with short contiguous on-source observations, opening opportunities for commensality between pulsar timing and mapping surveys.
author:
- |
Hsiu-Hsien Lin,$^{1}$[^1] Kiyoshi Masui,$^{2}$ Ue-Li Pen,$^{3,4,5,6}$ Jeffrey B. Peterson$^{1}$\
$^{1}$McWilliams Center for Cosmology, Carnegie Mellon University, Department of Physics, 5000 Forbes Ave, Pittsburgh, PA, 15213,\
USA\
$^{2}$Department of Physics and Astronomy, University of British Columbia, 6224 Agricultural Rd, Vancouver, BC, V6T 1Z1, Canada\
$^{3}$Canadian Institute of Theoretical Astrophysics, 60 St George St, Toronto, ON M5S 3H8, Canada\
$^{4}$Canadian Institute for Advanced Research, CIFAR Program in Cosmology and Gravity, Toronto, ON, M5G 1Z8\
$^{5}$Dunlap Institute for Astronomy & Astrophysics, University of Toronto, 50 St George St, Toronto, ON, M5S 3H4, Canada\
$^{6}$Perimeter Institute for Theoretical Physics, Waterloo, Ontario N2L 2Y5, Canada
bibliography:
- 'references.bib'
title: Improved Pulsar Timing via Principle Component Mode Tracking
---
\[firstpage\]
pulsars: general – pulsars: individual: J2139+0040 – methods: data analysis
Introduction
============
While there are thought to be 100,000 pulsars in the Milky Way only about 2600 have been cataloged [@2005AJ....129.1993M]. A new generation of drift-scan interferometric telescopes, including CHIME [@2014SPIE.9145E..22B], HIRAX [@2016SPIE.9906E..5XN] and HERA [@2017PASP..129d5001D], will soon begin recording data with a primary goal of creating cosmological 21-cm intensity maps. In addition, there are plans to use MeerKAT [@2017MNRAS.466.2780F] to create intensity maps using a raster scan mode. Such instruments could be used commensally to search for pulsars, and these instruments will have the collecting area, multi-beam capability, and on-sky integration time to substantially increase the pulsar discovery rate.
The pulsar search will likely have to work within the observational parameters of the hydrogen surveys. For the drift-scan telescopes, the duration of a contiguous observation will be limited to the time sources take to drift through a beam, which for CHIME will be a few minutes. MeerKAT, with its raster scan observing mode will make shorter duration passes but will pass near a pulsar’s coordinates more often. Integration time can be accumulated by combining multiple passes. Such a coherent pulsar search in time-disjointed data has been demonstrated in @Anderson:1993cna [Chapter 4].
To further develop and test time-disjoint pulsar search algorithms, we used an existing hydrogen-mapping data set, the Green Bank Hydrogen Intensity Mapping (GBTIM) survey [@2010Natur.466..463C; @2013ApJ...763L..20M; @2013MNRAS.434L..46S]. This is the data set in which fast radio burst FRB 110523 was discovered [@FRB110523]. The set consist of raster scans, so it contains a time-disjointed set of near-passes of pulsar coordinates.
Our long-term goal is to search for new pulsars, but first we need to test our algorithms and code on known examples, so we obtained a timing solution for PSR J2139+0040. The pulsar was discovered by @1996ApJ...469..819C at right ascension 21:39:42(16), declination +00:36(5), period 0.312470(3)s, and dispersion measure of 36(7)pc/cm$^3$. The pulsar is bright—we observe an average flux at 800MHz of approximately 50mJy—such that single pulses are readily detectable with the Green Bank Telescope. Our team began studying this pulsar because it lies within two degrees of FRB 110523 and could be used to constrain the Galactic component of the scintillation properties for this region of the sky.
Despite the sporadic scan-mode observations in the data set, we show that a timing solution for PSR J2139+0040 can be obtained, and we present improved timing parameters. We find timing residuals are reduced by use of a novel principal component analysis (PCA) technique to fit the time-variable pulse waveforms. The PCA technique automatically compensates for the rapid mode switching of the pulse shape, allowing a more precise timing solution. We describe this technique in detail since it may be more widely useful in pulsar timing.
Data and Processing
===================
Here we describe our data and observations as well as the processing of the data into folded pulse profiles.
Observations
------------
The data set we used came from an 21-cm intensity mapping survey, in which we raster-scanned four different Wiggle-Z (WZ) fields (1hr, 11hr, 15hr, and 22hr)[@2007ASPC..379...72G] recording spectra from 700-900 MHz with 4096 spectral frequency channels, using integration intervals of 1.024ms. The full survey is comprised of 680 hours of observations between 2011 and 2015. PSR J2139+0040 is located in the 22hr field, which was observed as part of GBT projects 10B-336 and 14B-339. Scans of this field have the beam crossing close to the pulsar often, but for short periods of time, with the beam typically moving across the sky at a rate of several degrees per minute. The set consists of sporadically-spaced observing sessions of several hours. This provides a sample of pulsar data with a wide range of crossing angles and with baseline drifts in the data due to variable ground spill and sky brightness structure. The cadence of the observing epochs is variable. It is therefore challenging to extract a precise pulsar timing solution from the set, providing a test of our processing techniques.
Our WiggleZ data are stored in blocks of 2048 time samples ([PSRFITS]{} records [@2004PASA...21..302H]). From our bulk data, we select all such blocks where at the block midpoint the telescope boresight is within 0.15$^{\circ}$ from the published pulsar position. This corresponds roughly to a half width at half maximum of the telescope beam at 800MHz. The resulting data set includes 1975 seconds of integration time, which we term the WZ data set.
In addition to the raster scan data described above, we obtained a single pointed observation of PSR J2139+0040 with a duration of one hour on MJD 57178. The frequency range of the pointed data is 720-920 MHz with 2048 spectral frequency channels, and an integration interval 8.192e-5s, grouped into records of 4096 samples. We dub this the pointed data set.
Using the [PRESTO]{}[^2] software package on the pointed data set yields DM = 31.7262 pc cm$^{-3}$, which we use for all subsequent analysis except Section \[subsection: Dispersion Measure\] where we perform a full fit for the DM.
Preprocessing {#subsection: preprocessing}
-------------
Data pre-processing produces an initial calibration of the data, removes flux from a noise injection source, mitigates radio-frequency interference (RFI), long-time-scale noise, and other sky signals such as point sources moving through the beam. To remove the system bandpass response function, we divided the recorded intensity by its time mean, converting to units of the system temperature. To mitigate RFI and long-time-scale signals we apply a stack of filters to the data consisting of: 1. deleting frequencies where the time variance is anomalously high, 2. subtracting the time mean from the data, 3. subtracting the time-linear component from the data, 4. deleting time samples where the frequency mean is outlying, 5. subtracting the frequency mean from the data, and finally repeating step 1. We use $5\sigma$ for all thresholds.
The above filters result in a zero-mean artifact in the folded profiles where the pulse profile is pushed negative just outside the main pulse. This is visible in the middle panel of Figure \[figure: folding\_dedisperse\_fit\]. However, this artifact is a result of linear operations on the data and as such does not bias our subsequent timing analysis.
Barycentric Time and Folding
----------------------------
We choose to time the pulsar in the Barycentric coordinate system because we are working toward searching such data sets for weaker pulsars. This step is not critical to the mode-tracking technique presented in Section \[subsection: Fitting phase bin number by SVD analysis\], but we describe it for completeness. We converted time stamps from Universal Time to Barycentric Time, $\tau(t, \alpha, \delta)$, defined as the time a signal from a far-off source at right ascension $\alpha$ and declination $\delta$ arriving at the observatory (here GBT) at time $t$, would arrive at the solar system barycenter.
We perform this conversion using the [TEMPO]{}[^3] software package as invoked by the [bary]{} command in [PRESTO]{}. For the right ascension and declination we initially use the previously published location of the pulsar, which we later refine. As described in Section \[subsection: Fitting a timing solution\], refining this position requires the partial derivatives of the Barycentric times with respect to right ascension and declination, which we calculate by finite difference.
To form pulse profiles, we group our sporadically spaced WZ data into groups spanning five minutes or less. These are typically very sparsely sampled with most groups containing only a few seconds of on-target data. The five-minute duration was selected to reduce the set to a manageable size of fixed cadence, with reasonable signal-to-noise for each folded waveform, and to ensure little worry of phase drift within a group. The pointed data, in contrast, is contiguously sampled and we divide it into 640 groups of duration 5.369s, which we use to characterize the time variability, and later stack into 40 pulse profiles of duration 86s used for timing.
We fold the time stream on the pulsar period 0.312470s [@1996ApJ...469..819C]. We use 200 phase bins for the WZ data, and 800 phase bins for pointed data.
We dedisperse the folded data, and then average it over frequency yielding pulse profiles as shown in Figure \[figure: folding\_dedisperse\_fit\]. This yields 289 such profiles although a number of these are noise dominated and show no evidence of the pulsar. We discard these, resulting in 232 pulse profiles in the WZ data set. Combined with the 40 profiles from the pointed data set, there are a total of 272 pulse profiles used for the timing analysis. Figure \[figure: J2139\_all\] below shows how these profiles are distributed in time.
![[**Typical Pulse profile for PSR J2139+0040.**]{} Upper panel: Dedispersed dynamic spectra. Middle panel: Blue dots are the pulse profile averaged across frequency, and red line is the best-fit model. Barycentric Arrival Times are used in these plots. Bottom panel: Fit residuals.[]{data-label="figure: folding_dedisperse_fit"}](images/fit_dedisperse_pulse.png){width="\columnwidth"}
Analysis and Results
====================
Here we describe the analysis of the folded pulse profiles. Our analysis accounts for the pulsar’s short-time-scale variability, estimates the pulsar phase, and fits a timing solution.
Principal component analysis of pulse shapes {#subsection: Fitting phase bin number by SVD analysis}
--------------------------------------------
Examining the set of folded pulse profiles, we find substantial variation in shape. In particular, some profiles have a stronger first peak, while others have a stronger second peak as shown in Figure \[figure: two\_pulses\]. This is the dominant mode of time variability for the profiles, with the preference for one peak over the other appearing to be correlated over timescales of several minutes. This is thus a form of mode switching, which is common for a substantial fraction of pulsars on a range of time scales [[@2004hpa..book.....L]]{}. To find the best-fit arrival time for each folded waveform, one could simply fit each waveform to an average profile, but waveform shape variations would then cause timing errors, since the fit would favor earlier arrival times when the first peak is stronger. To address this, we use a principal component analysis of the pulse profiles (PCA), which automatically tracks the pulsar mode. This PCA Mode Tracking compensates for mode-switching waveform-shape variation.
![[**Pulse shape variations.**]{} The two waveforms show folded data from the WZ set, focusing on the phase range -0.2 to 0.2. The pulsar shows mode-switching behavior: The ratio of the fluence between the first and second peak varies substantially. If these two waveforms were naively fit to an average profile, there would be a substantial timing error because of this profile variation.[]{data-label="figure: two_pulses"}](images/two_pulses_raw.png){width="\columnwidth"}
We use the 640 pulse profiles from the pointed data (each containing 5.369s of integration) to construct the PCA. Using Fourier techniques to achieve sub-bin alignment, we align these profiles according to a preliminary, linear timing solution valid for the pointed data. We create a matrix of 640 pulse profiles called $d_{E i}$ where $E$ is the epoch, an index of the mid-point time of each folding interval, and $i$ is the 800-point phase bin within the waveform. We carry out a Singular Value Decomposition: $$\label{equation: svd}
\centering
d_{Ei} =\sum_m {U_{Em}}{S_{mm}}{V_{mi}},$$ where for each mode $m$: $U_{Em}$ is the eigenfunction in epoch $E$, $V_{mi}$ the eigenfunction in phase $i$, and $S_{mm}$ is the singular value. We show $V_{mi}$ for the first ten $m$ modes in Figure \[figure: v\_mode\].
![The 1st to the 10th modes of ${\bm{\mathrm{V}}}$ are in the sequence of red, green, blue, yellow, cyan, black and dark to light greys, respectively. We set the offset of the 1st to the 10th modes by steps of -0.3.[]{data-label="figure: v_mode"}](images/phase_1hr_5sec_V_raw.png){width="\columnwidth"}
Inspecting the phase eigenfunctions, the first mode $V_{0 i}$ appears similar to the average waveform. We tested this equivalence by averaging the 640 pulse profiles into single template. We compared this average template with the mode $V_{0 i}$ of our PCA technique, and found the two consistent within the noise.
The second mode $V_{1i}$ allows the two peaks in the waveform to depart in amplitude from the average. Essentially, $U_{E1}$ tracks the variation with epoch of the waveform mode, changing sign when the mode switches.
Since we intend to use the $V$ modes as templates for fitting pulsar phases, a concern is that noise in these modes will bias the phase measurements. To reduce this noise for the modes above the first, we set them to zero outside of the main pulse (all phase bins except the central 160 of 800) since we see no evidence for pulsar flux outside this region. For the primary mode, the filtering artifact mentioned in Section \[subsection: preprocessing\] results in a smooth structure outside this region and so we spline the profile outside the central 160 phase bins.
Pulsar time of arrival estimation from direct integration of the likelihood {#subsection: Pulsar time of arrival estimation from direct integration of the likelihood}
---------------------------------------------------------------------------
The natural next step would be to simultaneously fit mode amplitudes and a pulse phase to each of our individual pulse profiles, using $V_{mi}$ as templates. However, we find that for this high-dimensional parameter space, the likelihood often has multiple maxima, and so the traditional least-squares fitting method fails. The multi-modal nature of the parameter space is due to the fact that for some profiles, there are multiple combinations of the $V_{mi}$ modes that, with different phases, may adequately describe the profiles such as those shown in Figure \[figure: two\_pulses\]. To deal with this, we employ a new pulse profile fitting technique that fully samples the parameter space.
We denote an individual measured folded pulsar profile $d_i$ where the index $i$ runs over the phase bins (we suppress the index $E$ for epoch in this section, since the analysis is performed independently for each profile). We model the measured profile as $$d_i = \sum_m A_m V_{mi}(\phi) + n_i,$$ where $V_{mi }(\phi)$ are the profile template modes (with $m$ running over mode number), $A_m$ is the mode amplitude, $\phi$ is the finely adjusted pulsar phase, and $n_i$ is the noise contribution. The dependence of $V_{mi }$ on $\phi$ describes the rotation of the templates to match the data. For notational brevity, we now switch to vector notation, where the above equation becomes: $${\bm{\mathrm{d}}} = {\bm{\mathrm{A}}}^T {\bm{\mathrm{V}}}(\phi) + {\bm{\mathrm{n}}}.$$ We assume Gaussian noise that is uniform and uncorrelated: $\langle n_i n_j \rangle = \delta_{ij} \sigma^2$. The template modes have already been measured so the parameters of the model are $\phi$ and the amplitudes, ${\bm{\mathrm{A}}}$.
Of these parameters we are chiefly interested in $\phi$. All the information about this parameter is in $p(\phi|{\bm{\mathrm{d}}})$, the posterior probability distribution of $\phi$ given the measurements ${\bm{\mathrm{d}}}$. For instance, the measurement mean and variance of the phase are the first and second moments of this distribution. This can be calculated from the posterior of all the parameters marginalized over the mode amplitudes: $$\label{e:margin_post}
p(\phi|{\bm{\mathrm{d}}}) = \int d^N{{\bm{\mathrm{A}}}}\, p(\phi,{{\bm{\mathrm{A}}}}|{\bm{\mathrm{d}}}),$$ where $N$ is the length of the vector ${\bm{\mathrm{A}}}$.
For flat priors on the initial parameters, Bayes’ Theorem states that the posterior distribution for the parameters is proportional to the likelihood function, $p(\phi,{{\bm{\mathrm{A}}}}|{\bm{\mathrm{d}}}) \propto p({\bm{\mathrm{d}}} | \phi,{{\bm{\mathrm{A}}}})$. Since the data is Gaussian-distributed, the latter is given by: $$\begin{aligned}
p({\bm{\mathrm{d}}} | \phi,{{\bm{\mathrm{A}}}}) &\propto
\exp\left[
{-\frac{1}{2}\chi^2({\bm{\mathrm{d}}}, \phi,{{\bm{\mathrm{A}}}})}
\right]
\\
\chi^2({\bm{\mathrm{d}}}, \phi,{{\bm{\mathrm{A}}}}) &=
\frac{1}{\sigma^2}\sum_i \left[
d_i - \sum_m A_m V_{mi}(\phi)
\right]^2.\end{aligned}$$
Combining the above, we have: $$p(\phi|{\bm{\mathrm{d}}}) \propto \int d^N{{\bm{\mathrm{A}}}}\,\exp\left\{-\frac{1}{2\sigma^2} \sum_i \left[ d_i - \sum_m A_m V_{mi}(\phi) \right]^2 \right\}$$ This is a multidimensional integral over the vector space of ${\bm{\mathrm{A}}}$, which would be prohibitively expensive to do numerically. A key insight is that since the model is linear in ${\bm{\mathrm{A}}}$, the likelihood is Gaussian not only in the data, but in ${\bm{\mathrm{A}}}$ as well (but not in $\phi$). This permits the integral to be done analytically. This same insight was used in @2013PhDT..Pennucci [Chapter 2] where a similar integral appears over the frequency dependant template amplitudes in wide-band pulsar timing. However, while in wide-band timing this is a computational convenience, here being able to perform this integral is essential due to the multi-modal nature of the likelihood space. Similarly, the integral appears in @2015MNRAS.447.2159L when marginalizing over epoch dependant profile amplitudes in profile domain timing analysis.
When performing the integral, there is a factor that depends on the expression $$\sum_i V_{mi}(\phi) V_{ni}(\phi).$$ This expression is independent of $\phi$ because the dependence of $V_{mi}$ on $\phi$ is just a shift of the template modes, which disappears after summing over phase bins. Thus, there is no need to carry out this part of the calculation. Note that this is only true if the noise is uniform. The final expression is $$\begin{aligned}
p(\phi|{\bm{\mathrm{d}}})
&\propto
\exp\left[
-\frac{1}{2\sigma^2}
\sum_m \sum_i V_{mi}(\phi) d_i \sum_j V_{mj}(\phi) d_j,
\right]\end{aligned}$$ or equivalently $$\begin{aligned}
p(\phi|{\bm{\mathrm{d}}})
&\propto
\exp\left[
-\frac{1}{2\sigma^2}
({\bm{\mathrm{V}}} {\bm{\mathrm{d}}})^T
{\bm{\mathrm{V}}} {\bm{\mathrm{d}}}
\right].\end{aligned}$$
which is proportional to $$\begin{aligned}
p(\phi|{\bm{\mathrm{d}}})
&\propto
\exp\left\{
{-\frac{1}{2}\chi^2\left[{\bm{\mathrm{d}}}, \phi,\hat{{\bm{\mathrm{A}}}}({\bm{\mathrm{d}}}, \phi)\right]}
\right\},\end{aligned}$$ where $\hat{{\bm{\mathrm{A}}}}({\bm{\mathrm{d}}}, \phi)$ is the linear-best-fit value for the template amplitudes at fixed $\phi$. Linear regression gives $$\begin{aligned}
\hat{{\bm{\mathrm{A}}}}({\bm{\mathrm{d}}}, \phi) = \left[{\bm{\mathrm{V}}}(\phi) {\bm{\mathrm{V}}}^T(\phi)\right]^{-1}{\bm{\mathrm{V}}}(\phi) {\bm{\mathrm{d}}}.\end{aligned}$$
![[**Pulse profile fit residuals in the frequency domain.**]{} Real and imaginary parts of the Fourier components are shown. The red curve and blue dots are the model and the measured flux, respectively. Green squares in lower panel are residuals. Fitting for phase in the frequency domain allows for simple interpolation with time precision finer than the 1.024 ms cadence of the dynamic spectra. []{data-label="figure: fft_Fourier"}](images/wz_9_fft.png){width="\columnwidth"}
So succinctly, the procedure for estimating the time of arrival is:
1. At fixed $\phi$ perform a *linear* fit for the mode amplitudes ${\bm{\mathrm{A}}}$, which we found in Section \[subsection: Fitting phase bin number by SVD analysis\].
2. Evaluate $\chi^2$ for these parameters and take $e^{-\chi^2/2}$ as the likelihood. \[item: evaluate the likelihood\]
3. Repeat for a range of $\phi$ values, covering the region of low $\chi^2$ that dominates the likelihood. (We repeat this step for a range of -20 to +20 phase bins from the fixed $\phi$, with steps of 0.02 phase bins.)
4. Take the zeroth, first, and second moments to get the normalization, phase estimate, and variance, respectively. In Section \[subsection: Fitting a timing solution\], we use the phase estimate and variance to find the timing solution.
While our procedure has been described in the profile space, our actual analysis is performed in the Fourier domain, as in @1992RSPTA.341..117T. We use only the first through 90^th^ harmonics to limit contamination from noise in the template modes and observed extraneous noise near 300Hz. We use $N=6$ template modes in our fits. An example of a profile fit is shown in Figure \[figure: fft\_Fourier\].
Finally, properly estimating the error on $\phi$ requires an estimate of the noise power $\sigma^2$. We use the value of $\sigma^2$ that results in a reduced chi-squared of unity ($\chi^2/{\rm \nu}$, where $\nu$ is the number of degrees of freedom) for the best fit (maximum likelihood) parameters.
Timing solution {#subsection: Fitting a timing solution}
---------------
We proceed to adjust the parameters of the pulsar timing model: $$\phi_E =
{\phi_0}
+ \frac{\Delta P}{P^2}(\tau_E - \tau_0)
+ \frac{\dot P}{2 P^2}(\tau_{E} - \tau_0)^{2}
+ {\frac{1}{P}}\left[-\frac{\partial \tau_E}{\partial \alpha} \Delta{\alpha}
- \frac{\partial \tau_E}{\partial \delta} \Delta{\delta}\right]
\label{equation: timing_model}$$ where, $\phi_E$ is the barycentric phase of the pulse profile indexed by $E$, $\tau_{E}$ is its barycentric time, $\tau_{0}$ is the reference epoch, $P$ is the period of the pulsar used for folding, ${\partial \tau_E}/{\partial \alpha} $ is the derivative of the barycentric time with respect to source right ascension, and ${\partial \tau_E}/{\partial \delta}$ is derivative of the barycentric time with respect to source declination. There are also five free parameters in the timing model, including the period derivative $\dot P$, period correction $\Delta P$, initial phase offset $\phi_0$, right ascension correction $\Delta{\alpha}$, and declination correction $\Delta{\delta}$. We proceed to adjust these parameters such that the above equation fits our phase measurements using standard weighted least squares and extract best fit parameters with uncertainties.
One challenge to obtaining a timing solution with this data set is the large gap in our data. We have phase measurements for epochs over several months in 2011 and 2015 but none in between. Compared to more uniformly sampled data, this distribution of measurements weakly constrains the total number of pulsar rotations in the gap period. Indeed we find multiple $\chi^2$ minima corresponding to changing the parameter $\Delta P$ by integer multiples of $\sim 8\times 10^{-10}$s, which changes the number of rotations in the gap. However, the lowest of these minima is $\delta \chi^2=44.5$ smaller than the next lowest. Thus this minimum is strongly preferred over the others, and our timing parameters are well constrained.
Inspecting the timing residuals, we find one phase measurement that is an extreme, $9\sigma$, outlier. Inspecting the full likelihood curve for the profile phase fit (as described in Section \[subsection: Pulsar time of arrival estimation from direct integration of the likelihood\]), we find the likelihood to be very non-Guassian, making such an outlier roughly $0.1\%$ likely. This is not terribly improbable given that we have 272 such data points, and is far more likely than a $9\sigma$ outlier for Gaussian data. Since our least-squares fit to the timing solution implicitly assumes Gaussian-distributed phase errors, we exclude this outlying data point in the timing solution fit.
The timing residuals are shown with the fitting results in Figure \[figure: J2139\_all\]. The fit has a reduced chi-squared of $\chi^2/{\nu}=1.26$ for 266 degrees of freedom, which is reasonable given the non-Gaussian nature our phase measurements which we have represented with 1-$\sigma$ error bars. We inflate the uncertainties on the timing parameters by this factor.
{width="\textwidth"}
The parameters for the final timing solution for PSR J2139+0040 are given in Table \[table: J2139+00\]. We find new right ascension and declination, period, period derivative, reference time of arrival (derived from $\phi_0$), and also provide the epoch used for the timing model. We use the period and period derivative to estimate the magnetic field strength ($B$), the characteristic age ($t_c$), and the spin-down luminosity ($\dot E$)[@2004hpa..book.....L].
------------------------------------------------------------------ --------------------------------
Right ascension, $\alpha$ $+21:39:25.20(2)$
Declination, $\delta$ $+00^{\circ}40^{'}21.6(3)^{"}$
Period, $P$ (s) 0.3124695464326(2)
Period derivative, $\dot P$ (s s$^{-1}$) $7.64(13)\times 10^{-18}$
Dispersion measure, DM (pccm$^{-3}$) 31.585(8)
Period epoch, $\tau_0$ (MJD) 56436.5
Reference arrival time (Barycentric MJD, 900MHz) 56436.506782091(13)
Magnetic field strength, $\log_{10}[ B/{\rm G}]$ 10.7
Characteristic age, $\log_{10}[ t_c/{\rm yr}]$ 8.8
Spin-down luminosity, $\log_{10}[ \dot E / ({\rm erg}/{\rm s})]$ 31.0
Dispersion-derived distance, $d$ (kpc) 1.7
------------------------------------------------------------------ --------------------------------
To check for evidence of proper motion of the pulsar’s sky position, we fix all parameters other than the position, then separately fit the 2011 and 2015 data. We find the two output positions are consistent with each other and the solution using the full data set, and as such we find no evidence for proper motion.
We used the [TEMPO]{} pulsar timing software to verify our timing solution. For this we converted our pairs of epoch-phase measurements to barycentric times of arrival then inverted [PRESTO]{}’s [bary]{} command to convert to topocentric times of arrival. Feeding these and our timing solution parameters into [TEMPO]{} we find agreement in the value of $\chi^2$. Letting [TEMPO]{} refit our timing-model parameters we find no significant shifts.
Dispersion Measure {#subsection: Dispersion Measure}
------------------
We used the pointed data to determine DM. We align the 40 folded profiles in phase based on our timing solution, then stacked them into a single profile. We fit this profile for the dispersion using the first $V$ mode as a template according to: $$\begin{aligned}
d_{if} &= A\left({\frac{f}{800\,{\rm MHz}}}\right)^{\beta}V_{0i}(\phi_f)\label{equation: dm_solution}\\
\phi_f &\equiv \phi_{900} + \frac{\rm DM}{P}
\frac{4148.808\,{\rm s\,MHz^2}}{\rm pc / cm^3}
\left[\frac{1}{f^{2}}-\frac{1}{(900 {\rm MHz})^{2}}\right]
\label{equation: dm_phase}\end{aligned}$$ Here, $d_{if}$ is the frequency-dependant profile data, which are fit with parameters for the amplitude $A$, power-law slope $\beta$, and frequency dependant phase $\phi_{f}$ for each frequency $f$. The frequency dependant phase $\phi_{f}$ contains an offset $\phi_{900}$ and the dispersion phase delay.
We subsequently calculate the dispersion-measure-based distance using Cordes-Lazio NE2001 Galactic Free Electron Density Model [@2002astro.ph..7156C] website[^4] using the new right ascension, declination, and DM. The results of this analysis are given in Table \[table: J2139+00\]
Timing improvement available from PCA Mode Tracking {#subsection: timing impovement}
---------------------------------------------------
To assess the impact of PCA mode tracking we compared the timing residuals using long and short integrations. We carried out these parallel analyses on sets of 16 individual 5.4s profiles from the pointed data. First, we stacked these profiles into a single pulse profile with a total integration time of 86s and fit this profile with the technique described in Section \[subsection: Pulsar time of arrival estimation from direct integration of the likelihood\]. Second, we fit the 16 profiles individually, then combined their resulting posterior distributions. We did this for the 40 sets of 16 profiles covering the full hour of pointed data. The first technique is similar to the conventional technique of taking long averages to smooth out pulse-to-pulse variability. The second technique uses fine grained data allowing the PCA algorithm to track and compensate for the variability.
We find that using PCA mode tracking and fitting fine-grained 5.4 second averages results in a $\sim20\%$ smaller uncertainty on average in the pulse time of arrival (calculated from the second moment of the posterior as described in Section \[subsection: Pulsar time of arrival estimation from direct integration of the likelihood\]). While the improvement is substantial for this pulsar, it may not be generic. Pulsars are highly individual: they vary substantially in the degree and time scales mode switching, so we anticipate the improvement available from PCA mode tracking will also vary strongly from case to case.
Discussion and Conclusions
==========================
Pulsar timing observations often make use of profiles averaged for minutes. These long averages smooth out the effect of rapid mode switching. Since it is in general sub-optimal to average over a time-variable signal, we suggest that PCA Mode Tracking could be applied broadly, by intentionally using short averages that display rapid mode switching—which the PCA then follows and compensates for. In particular, it may be useful to test this technique using data from millisecond pulsars, since improvement of timing precision of these pulsars allows more precise tests of general relativity, along with tighter constraints on long-wavelength gravitational wave backgrounds.
We have obtained a substantially improved timing solution for PSR J2139+0040, using untargeted observations in conjunction with a $\sim$1 hour duration targeted observation. Our data are well-described by a simple five-parameter model, including its sky location, reference phase, period, and period derivative. The measured slow-down rate of $\dot P = (7.64\pm0.13)\times10^{-18}$(ss$^{-1}$) is well below average, with $\sim 99\%$ of slow pulsars having a higher rate (@2017MNRAS.467.3493J) as shown in Figure \[figure: psr\_p0\_p1\]. This indicates that PSR J2139+0040 has an abnormally weak magnetic field.
![Periods and Period derivatives of known pulsars[@2005AJ....129.1993M]. Pulsar J2139+0040 lies at low $\dot P$ compared to other pulsars of similar period, but is not an extreme outlier.[]{data-label="figure: psr_p0_p1"}](images/psr_p0_p1.png){width="\columnwidth"}
For many of our pulse profiles, we find a highly non-Gaussian likelihood with multiple maxima for the phase. However, by analytically integrating over the template amplitudes, we are able to fully sample the marginalized posterior for phase and calculate the distribution’s mean and standard deviation. The subsequent fit to a timing model achieves a reduced chi-squared of 1.26, which is reasonable given the non-Gaussian nature of our phase errors. A more complete treatment would be to use the phase measurement posteriors directly when fitting a timing solution, rather than reducing them to their first and second moments. This would optimally extract the timing information from the phase measurements but would substantially complicate fitting a timing solution, since least-squares would no longer be applicable.
Our PCA mode tracking method addresses issues similar to those addressed via “profile-domain pulsar timing” [@2015MNRAS.447.2159L; @2015MNRAS.454.1058L; @2017MNRAS.466.3706L], particularly, what these authors refer to as “low-frequency stochasticity” and “phase-correlated stochasticity”. The profile-domain techniques uses pre-defined uncorrelated shapelet components to contribute to the individual profiles. In contrast, the PCA mode tracking method is non-parametric. There is no need to guess the shaplets, the PCA finds them automatically. Compared to the profile-domain strategy we anticipate that fewer components will be needed, yielding less degeneracy with the pulse phase and likely reducing timing error. The PCA mode tracking technique derives its mode templates from the data, and so for weak pulsars these waveforms may have noise that is absent when using the pre-defined templates of the profile domain technique.
A PCA technique was employed to characterize pulsar variability in @2011MNRAS.418.1258O, where, rather than fitting for the time of arrival and component amplitudes in high cadence data, the authors used the empirical correlation between the timing residuals and component amplitudes to retroactively correct for systematic timing error.
Our results demonstrate the feasibility of timing pulsars commensally with mapping surveys. In particular, upcoming hydrogen intensity mapping surveys using MeerKAT [@2017MNRAS.466.2780F] and Phase 1 of the Square Kilometre Array [@2015aska.confE..19S] will map large fractions of the southern sky with thousands of hours of telescope time. The telescopes will operate as a collection of single dishes and will thus need to scan rapidly to overcome $1/f$ noise. There is the potential to obtain pulsar timing solutions for free (in terms of telescope time) using the data from these surveys in much the same way as we have done here. The challenge is in storing the data at the rapid cadence required for pulsar studies, but, in single dish mode, data volumes are modest compared to interferometric mode. Mapping surveys spend a small fraction of their time pointing at known pulsars, however the mapping survey gets this data at many epochs, for every pulsar in its survey field, providing a large volume of timing data.
One of our goals in this work is to enable pulsar *searches* commensal with intensity mapping experiments. Having obtained a precise timing solution for a known pulsar in the GBTIM data, we have demonstrated that accumulation of pulsar data over five years in few-pulse snippets can be accomplished, although this may be more difficult for weaker or more erratic pulsars.
This analysis was possible because the GBTIM data used short 1 ms integrations. The intensity mapping data from CHIME [@2014SPIE.9145E..22B] and HIRAX [@2016SPIE.9906E..5XN] will use longer integrations for their intensity mapping data, but these instrument will have additional transient-search backend hardware allowing high cadence analysis of the data streams. New search algorithms have recently been proposed by @2016arXiv161006831S to reduce the cost of such long-term pulsar waveform assembly by several orders of magnitude. If these techniques could be employed at upcoming transit survey instruments such as CHIME and HIRAX the pay off could be a substantially increased rate of pulsar discovery.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank Kendrick Smith, Ingrid Stairs, Maura McLaughlin and Alexander Roman for valuable discussions. K. W. M. is supported by the Canadian Institute for Theoretical Astrophysics National Fellows program. U.-L. P. acknowledges support from the Natural Sciences and Engineering Research Council of Canada. Research at Perimeter Institute is supported by the Government of Canada through the Department of Innovation, Science and Economic Development Canada and by the Province of Ontario through the Ministry of Research, Innovation and Science. J. B. P. acknowledges support from NSF Award 1211777. Computations were performed on the GPC supercomputer at the SciNet HPC Consortium.
\[lastpage\]
[^1]: E-mail: [email protected]
[^2]: <http://www.cv.nrao.edu/~sransom/presto/>
[^3]: <http://tempo.sourceforge.net/>
[^4]: <https://www.nrl.navy.mil/rsd/RORF/ne2001>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a Molecular Dynamics study of large Lennard-Jones clusters evolving on a crystalline surface. The static and the dynamic properties of the cluster are described. We find that large clusters can diffuse rapidly, as experimentally observed. The role of the mismatch between the lattice parameters of the cluster and the substrate is emphasized to explain the diffusion of the cluster. This diffusion can be described as a Brownian motion induced by the vibrationnal coupling to the substrate, a mechanism that has not been previously considered for cluster diffusion.'
address: ' Département de Physique des Matériaux, Université Claude Bernard Lyon-1, CNRS UMR 5586, 69622 Villeurbanne Cédex, France'
author:
- 'Pierre Deltour, Jean-Louis Barrat and Pablo Jensen'
title: 'Fast diffusion of a Lennard-Jones cluster on a crystalline surface'
---
Understanding the interaction of particles of nanometer size with surfaces is important, both from a fundamental point of view and for controlled growth of thin films and nanostructures. Recently, the diffusion of large clusters containing hundreds of atoms was investigated both experimentally [@diffexp] and theoretically [@diffth]. These studies have focused on clusters epitaxially oriented on the surface, and have found relatively low diffusion coefficients, of order $10^{-17} cm^2 s^{-1}$. In contrast, Bardotti et al. [@prl] have shown experimentally that large (one hundred to a few thousand atoms) gold or antimony clusters, non-epitaxially oriented, have a surprisingly high diffusion coefficient (of the order $10^{-8} cm^2 s^{-1}$ at room temperature) on a graphite substrate. This observation is reminiscent of early work by Kern et al. [@Kern], in which a noticeable diffusion phase preceding epitaxial locking was observed for small gold crystallites on a $NaCl$ surface. The essential conclusion that can be drawn is that “large” (a few nanometers) objects can have large surface diffusion coefficients at room temperature, a result that might seem rather counterintuitive to many surface scientists. Indeed, most diffusion mechanisms that have been considered for surface diffusion of composite objects such as clusters involve a combination of single atom diffusion events (e.g. evaporation condensation [@diffth]), and yield diffusion constants much smaller than the above mentioned result. Although these mechanisms have been shown to describe correctly the diffusion of clusters of adatoms in epitaxy on the substrate, they do not seem to be relevant to explain the fast diffusion that is observed in references [@prl; @Kern].
In this Letter, we present the results of a molecular dynamics (MD) study of the cluster diffusion problem, with particular emphasis on the consequences of epitaxial or non-epitaxial cluster-substrate configurations. The non-epitaxial case is important for growth of non-epitaxial thin films or for films prepared by [*pre-formed*]{} cluster deposition [@review]. A simple model, aimed at clarifying the generic aspects of the question rather than modeling a particular case, was used. Both the cluster and the substrate are made up of Lennard-Jones atoms, interacting through potentials of the form $
V(r)=4 \epsilon \left (\left(\frac{\sigma}{r} \right)^{12}
- \left(\frac{\sigma}{r} \right)^{6} \right )
$. Empirical potentials of this type, originally developed for the description of inert gases, are now commonly used to model generic properties of condensed systems. The substrate is modeled by a single layer of atoms on a triangular lattice, attached to their equilibrium sites by weak harmonic springs that preserve surface cohesion. The Lennard-Jones parameters for cluster atoms and substrate atoms are respectively $\left (\epsilon_{cc},\sigma_{cc} \right) \mbox{ and }
\left (\epsilon_{ss},\sigma_{ss} \right)$. The parameters $\left (\epsilon_{sc}, \sigma_{sc} \right)$ for the substrate-cluster interaction are constructed by following the standard combination rule : $ \epsilon_{ss}
\sim \sigma_{ss}^{6} $ and $ \sigma_{sc} = \frac {1}{2} \left ( \sigma_{cc} + \sigma_{ss} \right )$. Finally, the unit of time is defined as $\tau = (M \sigma_{cc}^2/\epsilon_{cc})^{1/2}$, where $M$ is the mass of the atoms which is identical for cluster and substrate atoms.
Our simulation uses a standard molecular dynamics technique with thermostatting of the [*surface*]{} temperature . The equations of motion are then integrated using the Leap-Frog Algorithm, which is a simple, time reversible and very stable integration scheme [@Frenkel] The run is started from a configuration in which the cluster atoms occupy sites of a cubic lattice near the substrate. The system is then equilibrated for 500$\tau$, after which the trajectories are recorded. Diffusion constants $D$ are obtained from the mean-squared displacement of the cluster center of mass. Typically, the length of the runs used to compute $D$ was 2500 $\tau$, with some longer runs of 12500 $\tau$. This means that diffusion constants smaller than $5\ 10^{-4}\sigma_{cc}^2\tau^{-1}$, for which the cluster does not visit an area of more than $25\sigma_{cc}^2$ over the length of the simulation, are not accessible.
We first briefly describe the static properties of our supported clusters. At the temperature at which most runs were carried out, $T=0.3\epsilon_{cc}$, the clusters are in a crystalline FCC or HCP configuration. They take the spherical cap shape of a solid droplet (figure 1) partially wetting the substrate. The contact angle $\theta$, that can be defined following reference [@angle], is roughly independent of the cluster particle number $N$, for $50 < N < 500$. This angle can be changed by tuning the cluster-substrate interaction. For large enough $\epsilon_{sc}$, total wetting is observed, and the cluster dissociates. As noted in [@angle], this shows the relevance of macroscopic concepts such as contact angle and wetting even for nanometer-sized particles.
We now turn to the dynamical behavior of the supported cluster. Most of the runs were carried out at a reduced temperature of 0.3 so that the cluster is clearly solid. In that case, visual observation of atomic trajectories during the run indicate that the diffusion inside the cluster, or on its surface, is completely negligible. Except for the vibrational motion of the atoms, the cluster behaves as a rigid object. This is clearly visible in figure 1, where the left and right halves of the cluster, colored grey and white at the beginning of the run, clearly retain their identity after the cluster [*center of mass*]{} has moved over 2 lattice parameters. Hence the motion of the cluster appears to be controlled by collective motions of the cluster as a whole rather than by single atomic jumps. This collective diffusion mechanism will depend essentially on three parameters: the lattice parameter of the substrate, the temperature and the cluster size. We consider in turn the influence of each parameter.
We first investigate the effect of the ratio of the cluster lattice parameter to the substrate lattice parameter for $T=0.3 \epsilon_{cc}$ and $N=100$, by varying the interaction diameter of the substrate atoms in the range of $0.5\sigma_{cc}$ to $1.5\sigma_{cc}$ The lattice constant $d$ of the substrate atoms is also correspondingly scaled in such a way that $\sigma_{ss}/d$ remains constant, equal to its value for a Lennard-Jones solid at zero pressure. The results for the diffusion coefficient are shown in figure 2. When the substrate and cluster are commensurate ($\sigma_{ss}=\sigma_{cc}$), the cluster can lock into a low energy epitaxial configuration. A global translation of the cluster would imply overcoming an energy barrier scaling as $N^{2/3}$, the contact area between the cluster and the substrate. In that case, single atom mechanisms as described in [@diffth] will dominate, and the diffusion will be very slow. Indeed on the time scale of the MD simulations, diffusion was vanishingly small in that case. The points in figure 2 at $\sigma_{ss} \sim \sigma_{cc}$ do not in fact represent true diffusion constants, since in these cases the center of mass did not move by more than a few lattice constants over the whole duration of the run. However, for small deviations from this commensurate case, the diffusion becomes measurable on the time scale of the MD runs. This can be understood from the fact that the effective potential in which the center of mass moves is much weaker, as the cluster atoms, constrained to their lattice sites inside the rigid solid cluster, are unable to adjust to the substrate potential. The effect is rather spectacular. A fifteen percent change on the lattice parameter induces a two order of magnitude change on the diffusion coefficient. For the same reason, the diffusion coefficient decreases slightly when $\sigma_{cc}$ is getting close to $2 \sigma_{ss}$. Finally, we note that the effect is not exactly symmetric on both sides of the commensurate situation. Clusters with atoms smaller than the substrate atoms tend to diffuse more slowly, since the potential wells for these atoms will be deeper than for big adatoms.
Next, we study the influence of temperature for a fixed values of $\sigma_{ss}$ and $\epsilon_{sc}$, $\sigma_{ss}=0.9$ and $\epsilon_{sc}=0.4$. The diffusion constant can be fitted by an Arrhenius law (figure 3) , with an activation energy E of $0.66 \epsilon_{cc}$ and a prefactor $D_0$ of $0.02 D_\star$ where $D_\star$ is equal to $\sigma_{cc} \left( \epsilon_{cc} / M \right )^{1/2}$. However the mechanisms involved here are not simple, single atom activated processes. Not too much meaning should then be given to the activation energy and prefactor.
In order to establish a connection with experiments, we can identify $\epsilon_{cc}/k_B$ to a typical melting temperature (1000K for Au). Using $\sigma_{ss} \sim 5$ Å and $M\sim 10^{-25} kg$, we find that our diffusion coefficients would fall in the range of $10^{-5} cm^2 s^{-1}$ for a 100 atoms cluster. This is in reasonable agreement with the values obtained in [@prl] , but differs strikingly from the values obtained with single atom mechanisms.
Finally, for the same values of $\epsilon_{sc}$ and $T$, the effect of cluster size on the diffusion constant is considered for different lattice parameter values. As the number $N$ of atoms in the cluster is varied between $N=10$ and $N=500$, the diffusion constant decreases, roughly following a power law $D\sim N^{\alpha}$ (figure 4). This power law exponent $\alpha$ depends significantly on the mismatch between the cluster and the substrate lattice parameters. For high mismatches ($\sigma_{ss}=0.7,0.8$), $\alpha$ is close to $-0.66$. As the diffusion constant is inversely proportional to the cluster-substrate friction coefficient, this result is in agreement with a simple “surface of contact” argument yielding $D\sim N^{-2/3}$. On the other hand, when the lattice mismatch is equal to $0.9$, $\alpha$ one obtains $\alpha \approx -1.4$, although the shape of the cluster, characterized by the contact angle, does not appreciably change. Moreover, the trajectory followed by the the cluster center of mass changes qualitatively. In the runs with a large mismatch, this trajectory is brownian like, with no apparent influence of the substrate. In contrast, when the mismatch is small, the center of mass of the cluster follows a “hopping-like” trajectory, jumping from site to site on the honeycomb lattice defined by the substrate . When $\sigma_{ss}=\sqrt{3}/2$, there seems to be a transition between the two regimes around $N=200$.
In an attempt to disentangle the contribution of the cluster internal vibrations to the diffusion process from that of the substrate vibrations in the small mismatch case ($\sigma_{ss}=0.9$), we now consider idealized cases in which one of the two subsystems is artificially “frozen”, so that its contribution vanishes. First, we consider the case of cluster supported by a “frozen” substrate, with atoms constrained to their equilibrium positions. The external potential experienced by the cluster atoms is thus purely static. The cluster is first equilibrated using constant temperature MD at $T=0.3\epsilon_{cc}$, and this equilibration period is followed by a constant energy simulation. The results for the diffusion constant are very close to those obtained with a thermalized substrate. This shows that such clusters have enough internal degrees of freedom to play the role of being their own thermostat, and that their internal vibration modes can be an efficient motor for the diffusion.
Next, we consider the other extreme case of a “frozen” cluster deposited on a thermalized substrate. The cluster is first equilibrated, either on a perfectly flat substrate exerting an average potential equivalent to that of the triangular lattice of Lennard-Jones atoms, or in free space. In the first case, it adopts the usual spherical cap shape, while in the second case a quasi-spherical, faceted shape is observed. This last shape would be reasonable for a cluster made of a highly cohesive material. After this equilibration phase, the cluster is “frozen” and deposited on the thermalized substrate as a solid body. The center of mass trajectory is integrated using the quaternion algorithm .
Not surprisingly, the diffusion is dependent on the way the system has been equilibrated. For a spherical rigid cluster, the system rotates until a facet comes into contact with the substrate, then diffuses without rotating for the rest of the MD run. Hence despite the more or less spherical shape of the cluster, rotation does not seem to give an important contribution to the diffusion. In that case, the diffusion depends on the shape and size of the facet the cluster rests on, so that the results in this case are not well reproducible. For a cluster equilibrated on a flat surface, the diffusion constant follows the same power law as in the free cluster case. This surprising result suggests that the diffusion mechanism in our case cannot be just simply explained in terms of dislocation migration within the cluster as proposed to explain the diffusion of 2D islands in [@Hamilton]. Rather, the motor for diffusion is here the vibrational motion of the substrate, and its efficiency appears to be comparable to that of the internal cluster modes.
In summary, the surprisingly high diffusivity of large clusters can be understood with two main ingredients. When the cluster is not commensurate with the substrate, the modulation of the potential felt by the center of mass is small. In this case the cluster is not locked by the substrate and can vibrate relatively freely. The vibrations (phonons) of the substrate and the internal vibrations of the cluster both create a “random” force on the cluster center of mass, which executes a brownian motion in this weak external potential. The two components of this random force appear to have similar intensities, and, at the temperature we consider, are sufficiently strong to overcome the small energy barriers, resulting in a rapid diffusive motion. Our calculation demonstrates the efficiency of this brownian-like mechanism, which was not considered in earlier studies of cluster diffusion, for explaining the diffusion of rather large objects. Further analysis of the vibrational coupling between the substrate and the cluster will be necessary to fully understand the mass dependence of the diffusion constant.
We thank the Centre Pour le Développement du Calcul Scientifique Parallèle at Université Claude Bernard Lyon-I for the allocation of computer time that made this work possible.
[99]{}
J.M. Wen et al., Phys. Rev. Lett. [**73**]{}, 2591 (1994); K. Morgenstern et al., Phys. Rev. Lett. [**74**]{}, 2058 (1995); G.L. Kellogg, Phys. Rev. Lett. [**73**]{}, 1833 (1994) and references therein.
A. F. Voter, Phys. Rev. B [**34**]{}, 6819 (1986); J. M. Soler, Phys. Rev. B [**50**]{}, 5578 (1994); S. V. Khare et al. Phys. Rev. Lett. [**75**]{}, 2148 (1995); D. S. Sholl and R. T. Skodje Phys. Rev. Lett. [**75**]{}, 3158 (1995); C. deW. Van Siclen Phys. Rev. Lett. [**75**]{}, 1574 (1995); J. M. Soler, Phys. Rev. B [**53**]{}, 16 R10540 (1996).
L. Bardotti et al., Phys. Rev. Lett. [**74**]{}, 4694 (1995); Surf. Sci. [**367**]{}, 267 (1996)
R. Kern, G. Le Laye and J.J. Metois, Current Topics in Materials Science, vol 3 chap 3, North Holland (1979)
P. Melinon et al., Int. J. of Mod. Phys. B [**9**]{}, 339-397 (1995)
M. P. Allen and T. E. Tidesley, Computer simulation of Liquids, Oxford University Press (1987)
D. Frenkel and B. Smit, Understanding Molecular Simulation, Academic Press (1996)
J. Hautman et al., Phys. Rev. Lett. [**67**]{}, 1763 (1991)
J.C. Hamilton, Phys. Rev. Lett. [**77**]{}, 885 (1996)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Temporally localizing actions in a video is a fundamental challenge in video understanding. Most existing approaches have often drawn inspiration from image object detection and extended the advances, e.g., SSD and Faster R-CNN, to produce temporal locations of an action in a 1D sequence. Nevertheless, the results can suffer from robustness problem due to the design of predetermined temporal scales, which overlooks the temporal structure of an action and limits the utility on detecting actions with complex variations. In this paper, we propose to address the problem by introducing Gaussian kernels to dynamically optimize temporal scale of each action proposal. Specifically, we present Gaussian Temporal Awareness Networks (GTAN) — a new architecture that novelly integrates the exploitation of temporal structure into an one-stage action localization framework. Technically, GTAN models the temporal structure through learning a set of Gaussian kernels, each for a cell in the feature maps. Each Gaussian kernel corresponds to a particular interval of an action proposal and a mixture of Gaussian kernels could further characterize action proposals with various length. Moreover, the values in each Gaussian curve reflect the contextual contributions to the localization of an action proposal. Extensive experiments are conducted on both THUMOS14 and ActivityNet v1.3 datasets, and superior results are reported when comparing to state-of-the-art approaches. More remarkably, GTAN achieves 1.9% and 1.1% improvements in mAP on testing set of the two datasets.'
author:
- |
Fuchen Long$^{\dag}$, Ting Yao$^{\ddag}$, Zhaofan Qiu$^{\dag}$, Xinmei Tian$^{\dag}$, Jiebo Luo$^{\S}$ and Tao Mei$^{\ddag}$\
$^{\dag}$University of Science and Technology of China, Hefei, China\
$^{\ddag}$JD AI Research, Beijing, China\
$^{\S}$University of Rochester, Rochester, NY USA\
[{longfc.ustc, tingyao.ustc, zhaofanqiu}@gmail.com; [email protected];]{}\
[ [email protected]; [email protected]]{}\
bibliography:
- 'egbib.bib'
title: 'Gaussian Temporal Awareness Networks for Action Localization[^1]'
---
Introduction
============
With the tremendous increase of online and personal media archives, people are generating, storing and consuming a large collection of videos. The trend encourages the development of effective and efficient algorithms to intelligently parse video data. One fundamental challenge that underlies the success of these advances is action detection in videos from both temporal aspect [@Gaidon:PAMI13; @Geest:ECCV16; @Lea:CVPR17; @Shou:CVPR16; @yao2017msr; @Xiong:ICCV17] and spatio-temporal aspect [@Gkioxari:CVPR15; @Dong:ECCV18]. In this work, the main focus is temporal action detection/localization, which is to locate the exact time stamps of the starting and the ending of an action, and recognize the action with a set of categories.
![The intuition of a typical one-stage action localization (upper) and our GTAN (lower). The typical method fixes temporal scale in each feature map and seldom explores temporal structure of an action. In contrast, temporal structure is taken into account in our GTAN through learning a set of Gaussian kernels.[]{data-label="fig1:1"}](intro_fig_camera){width="42.00000%"}
One natural way of temporal action localization is to extend image object detection frameworks, e.g., SSD [@Liu:ECCV16] or Faster R-CNN [@Ren:NIPS15], for producing spatial bounding boxes in a 2D image to temporal localization of an action in a 1D sequence [@Chao:CVPR18; @Lin:MM17]. The upper part of Figure \[fig1:1\] conceptualizes a typical process of one-stage action localization. In general, the frame-level or clip-level features in the video sequence are first aggregated into one feature map, and then multiple 1D temporal convolutional layers are devised to increase the size of temporal receptive fields and predict action proposals. However, the temporal scale corresponding to the cell in each feature map is fixed, making such method unable to capture the inherent temporal structure of an action. As such, one ground-truth action proposal in the green box is detected as three ones in this case. Instead, we propose to alleviate the problem by exploring the temporal structure of an action through learning a Gaussian kernel for each cell, which dynamically indicates a particular interval of an action proposal. A mixture of Gaussian kernels could even be grouped to describe an action, which is more flexible to localize action proposals with various length as illustrated in the bottom part of Figure \[fig1:1\]. More importantly, the contextual information is naturally involved with the feature pooling based on the weights in Gaussian curve.
By delving into temporal structure of an action, we present a novel Gaussian Temporal Awareness Networks (GTAN) architecture for one-stage action localization. Given a video, a 3D ConvNet is utilized as the backbone to extract clip-level features, which are sequentially concatenated into a feature map. A couple of convolutional layers plus max-pooling layer are firstly employed to shorten the feature map and increase the temporal size of receptive fields. Then, a cascaded of 1D temporal convolutional layers (anchor layers) continuously shorten the feature map and output anchor feature map, which consists of features of each cell (anchor). On the top of each anchor layer, a Gaussian kernel is learnt for each cell to dynamically predict a particular interval of an action proposal corresponding to that cell. Multiple Gaussian kernels could even be mixed to capture action proposals with arbitrary length. Through Gaussian pooling, the features of each cell is upgraded by aggregating the features of contextual cells weighted by the values in the Gaussian curve for final action proposal prediction. The whole architecture is end-to-end optimized by minimizing one classification loss plus two regression losses, i.e., localization loss and overlap loss.
The main contribution of this work is the design of an one-stage architecture GTAN for addressing the issue of temporal action localization in videos. The solution also leads to the elegant view of how temporal structure of an action should be leveraged for detecting actions with various length and how contextual information should be utilized for boosting temporal localization, which are problems not yet fully understood in the literature.
{width="88.00000%"}
Related Work
============
We briefly group the related works into two categories: temporal action proposal and temporal action detection. The former focuses on investigating how to precisely localize video segments which contain actions, while the latter further classifies these actions into known classes.
We summarize the approaches on temporal action proposal mainly into two directions: content-independent proposal and content-dependent proposal. The main stream of content-independent proposal algorithms is uniformly or sliding window-ly sampling in a video [@Oneata:ICCV13; @Tang:ICCV13; @Yuan:CVPR16], which leads to huge computations for further classification. In contrast, content-dependent proposal methods, e.g., [@Buch:CVPR17; @Escorcia:ECCV16; @Gao:ECCV18; @Gao:ICCV17; @Lin:ECCV18], utilize the label of action proposals during training. For instance, Escorcia *et al.* [@Escorcia:ECCV16] leverage Long Short-Term Memory cells to learn an appropriate encoding of a video sequence as a set of discriminative states to indicate proposal scores. Though the method avoids running sliding windows of multiple scales, there is still the need of executing an overlapping sliding window that is inapplicable when the video duration is long. To address this problem, Single Stream Temporal proposal (SST) [@Buch:CVPR17] generates proposals with only one single pass by utilizing a recurrent GRU-based model, and Temporal Unit Regression Network (TURN) [@Gao:ICCV17] builds video units in a pyramid manner to avoid window overlapping. Different from the above methods which generate proposals in a fixed multi-scale manner, Boundary Sensitive Network (BSN) [@Lin:ECCV18] localizes the action boundaries based on three actionness curves in a more flexible way. Nevertheless, such actionness-based methods may fail in locating dense and short actions because of the difficulty to discriminate between very close starting and ending peaks in the curve.
Once the localization of action proposals completes, the natural way for temporal action detection is to further classify the proposals into known action classes, making the process in two-stage manner [@Chao:CVPR18; @Heilbron:CVPR17; @Shou:CVPR17; @Shou:CVPR16; @Xu:ICCV17; @Xiong:ICCV17]. However, the separate of proposal generation and classification may result in sub-optimal solutions. To further facilitate temporal action detection, there have been several one-stage techniques [@Buch:BMVC17; @Lin:MM17; @Yeung:CVPR16] being proposed recently. For example, Single Stream Temporal Action Detection (SS-TAD) [@Buch:BMVC17] utilizes the Recurrent Neural Network (RNN) based architecture to jointly learn action proposal and classification. Inspired by SSD [@Liu:ECCV16], Lin *et al.* [@Lin:MM17] devise 1D temporal convolution to generate multiple temporal action anchors for action proposal and detection. Moreover, with the development of reinforcement learning, Yeung *et al.* [@Yeung:CVPR16] explore RNN to learn a glimpse policy for predicting the starting and ending points of actions in an end-to-end manner. Nevertheless, most of one-stage methods are still facing the challenge in localizing all the action proposals due to the predetermined temporal scales.
In short, our approach belongs to one-stage temporal action detection techniques. Different from the aforementioned one-stage methods which often predetermine temporal scales of action proposals, our GTAN in this paper contributes by studying not only learning temporal structure through Gaussian kernels, but also how the contextual information can be better leveraged for action localization.
Gaussian Temporal Awareness Networks
====================================
In this section we present the proposed Gaussian Temporal Awareness Networks (GTAN) in detail. Figure \[fig2:1\] illustrates an overview of our architecture for action localization. It consists of two main components: a base feature network and a cascaded of 1D temporal convolutional layers with Gaussian kernels. The base feature network is to extract feature map from sequential video clips, which will be fed into cascaded 1D convolutional layers to generate multiple feature maps in different temporal resolution. For each cell in one feature map, a Gaussian kernel is learnt to control temporal scale of an action proposal corresponding to that cell as training proceeds. Furthermore, a Gaussian Kernel Grouping algorithm is devised to merge multiple Gaussian kernels with high overlap to a larger one for capturing long actions with arbitrary length. Specifically, each action proposal is generated by aggregating the features of contextual cells weighted by the values in the Gaussian curve. The whole network is jointly optimized with action classification loss plus two regression losses, i.e., localization loss and overlap loss, which are utilized to learn action category label, default temporal boundary adjustment and overlap confidence score for each action proposal, respectively.
Base Feature Network
--------------------
The ultimate target of action localization is to detect action instances in temporal dimension. Given an input video, we first extract clip-level features from continuous clips via a 3D ConvNet which could capture both appearance and motion information of the video. Specifically, a sequence of features are extracted from 3D ConvNet, where $T$ is the temporal length. We concatenate all the features into one feature map and then feed the map into two 1D convolutional layers (“conv1” and “conv2” with temporal kernel size 3, stride 1) plus one max-pooling layer (“pool1” with temporal kernel size 3, stride 2) to increase the temporal size of receptive fields. The base feature network is composed of 3D ConvNet, two 1D convolutional layers and max-pooling layer. The outputs of the base feature network are further exploited for action proposal generation.
Gaussian Kernel Learning {#sec:3.2}
------------------------
Given the feature map output from the base feature network, a natural way for one-stage action localization is to stack 1D temporal convolutional layers (anchor layers) to generate proposals (anchors) for classification and boundary regression. This kind of structure with predetermined temporal scale in each anchor layer can capture action proposals whose temporal intervals are well aligned with the size of receptive fields, however, posts difficulty to the detection of proposals with various length. The design limits the utility on localizing actions with complex variations.
To address this issue, we introduce temporal Gaussian kernel to dynamically control the temporal scales of proposals in each feature map. In the literature, there has been evidences on the use of Gaussian kernels for event detection in videos [@Piergiovanni:AAAI17; @Piergiovanni:CVPR18]. In particular, as shown in Figure \[fig2:1\], eight 1D temporal convolutional layers (anchor layers) are first cascaded for action proposal generation in different temporal resolution. For each cell in the feature map of the anchor layer, a Gaussian kernel is learnt to predict a particular interval of an action proposal corresponding to that cell. Formally, we denote the feature map of $j$-th convolutional layer as , $1\leq j\leq 8$, where $T^j$ and $D^j$ are the temporal length and feature dimension of the feature map. For a proposal $P^j_{t}$ whose center location is $t$, we leverage its temporal scale by a Gaussian kernel $G^j_{t}$. The standard deviation $\sigma^j_{t}$ of $G^j_{t}$ is learnt via a 1D convolutional layer on a $3 \times D^j$ feature map cell, and the value is constrained within the range $(0,1)$ through a sigmoid operation. The weights of the Gaussian kernel $G^j_{t}$ are defined as $$\label{Eq2:1}\small
\begin{split}
&W^j_t[i] = \frac{1}{Z}\exp(-\frac{(p_i-\mu_t)^2}{2{\sigma^j_{t}}^2})~, \\
&s.t.~~~p_i = \frac{i}{T^j},~ \mu_t = \frac{t}{T^j}~,\\
&~~~~~~~~i \in \{0,1,...,T^j-1\},~~~~ t \in \{0,1,...,T^j-1\},
\end{split}$$ where $Z$ is the normalizing constant. Taking the spirit from the theory that the $\sigma^j_{t}$ could be considered as a measure of width (Root Mean Square width, RMS) in Gaussian kernel $G^j_{t}$, we utilize $\sigma^j_{t}$ as the interval measure of action proposal $P^j_{t}$. Specifically, the $\sigma^j_{t}$ can be multiplied with a certain ratio to represent the default temporal boundary: $$\label{Eq2:3}\small
a_c = (t+0.5)/T^{j},~~~~a_w = r_d \cdot 2\sigma^j_{t}/T^{j},$$ where $a_c$ and $a_w$ are the center location and width of default temporal boundary and $r_d$ represents temporal scale ratio. The $W^j_t$ is also utilized for feature aggregation with a pooling mechanism to generate action proposals, which will be elaborated in Section \[sec:3.4\].
![Visualization of Gaussian Kernel Grouping.[]{data-label="fig2:2"}](grouping_camera){width="34.00000%"}
Compared to the conventional 1D convolutional anchor layer which fixes the temporal scale as $1/T^j$ in $j$-th layer, ours employs the dynamic temporal scales by leveraging the learned Gaussian kernel of each proposal to explore the action instances with complex variations.
Gaussian Kernel Grouping
------------------------
\
Original Gaussian kernel set $\mathbb{S} = \{G(t_i, \sigma_i)\}^{T-1}_{i=0}$;\
Intersection over Union (IoU) threshold $\varepsilon$;\
\
Mixed Gaussian kernel set $\mathbb{G}$;\
Choose the beginning grouping position $p = 0$; Initialize mixed Gaussian kernel set $\mathbb{G} = \varnothing$; Initialize base Gaussian kernel $G_{bs} = G(t_p, \sigma_p)$, the ending grouping position $z = p+1$; Compute IoU value $O$ between kernel $G_{bs}$ and $G(t_z, \sigma_z)$; Group $G_{bs}$ and $G(t_z, \sigma_z)$ to $G'$ according to Eq.(\[Eq2:4\]), replace $G_{bs}$ with the new mixed kernel $G'$; Add kernel $G_{bs}$ to mixed kernel set $\mathbb{G}$; $p = z$, $G_{bs} = G(t_p, \sigma_p)$; $z = z + 1$; $\mathbb{G}$
Through learning temporal Gaussian kernels, the temporal scales of most action instances can be characterized with the predicted standard deviation. However, if the learned Gaussian kernels span and overlap with each other, that may implicitly indicate a long action centered at a flexible position among these Gaussian kernels. In other words, utilizing the center locations of these original Gaussian kernels to represent this long proposal may not be appropriate. To alleviate this issue, we attempt to generate a set of new Gaussian kernels to predict center location and temporal scales of proposals for long action. Inspired by the idea of temporal actionness grouping in [@Xiong:ICCV17], we propose a novel Gaussian Kernel Grouping algorithm for this target.
Figure \[fig2:2\] illustrates the process of temporal Gaussian Kernel Grouping. Given two adjacent Gaussian kernels $G(t_1, \sigma_1)$ and $G(t_2, \sigma_2)$ whose center location and standard deviation are $t$ and $\sigma$, we compute the temporal intersection and union between two kernels by using the width $a_w$ of the default temporal boundary defined in Section \[sec:3.2\]. In upper part of Figure \[fig2:2\], the length of temporal intersection between two kernels is $H$, while the length of union is $L$. If the Intersection over Union (IoU) between the two kernels $H/L$ exceeds a certain threshold $\varepsilon$, we merge them into one Gaussian kernel (bottom part of Figure \[fig2:2\]). The new mixed Gaussian kernel is formulated as follows $$\label{Eq2:4}\small
\begin{split}
&W[i] = \frac{1}{Z}\exp(-\frac{(p_i-\mu')^2}{2{\sigma'}^2})~, \\
&s.t.~~~ p_i = \frac{i}{T}, ~~~ \mu' = \frac{t_1+t_2}{2 \cdot T}~, ~~~ \sigma' = \frac{L}{2}~,\\
&~~~~~~~~i \in \{0,1,...,T-1\}.
\end{split}$$
In each feature map, Algorithm \[GKG\] details the grouping steps to generate merged kernels.
Gaussian Pooling {#sec:3.4}
----------------
![Comparisons of manual extension plus average-pooling strategy (left) and Gaussian pooling strategy (right) for involving temporal contextual information of action proposals.[]{data-label="fig2:3"}](context){width="44.00000%"}
With the learned and mixed Gaussian kernels, we calculate the weighted sum of the feature map based on the values in Gaussian curve and obtain the aggregated feature $F$. Specifically, given the weighting coefficients $W^j_{t}$ of Gaussian kernel $G^j_{t}$ at center location $t$ in $j$-th layer, the aggregated feature for proposal $P^j_{t}$ is formulated as $$\label{Eq2:2}\small
\begin{split}
&F^j_t = \frac{1}{T^j}\sum\nolimits_{i=0}^{T^j-1} W^j_t[i] \cdot f_i, \\
\end{split}$$ where the representation $F^j_{t}$ is further exploited for the action classification and temporal boundary regression.
The above Gaussian pooling mechanism inherently takes the contextual contributions around each action proposal into account. In contrast to the manual extension plus average-pooling strategy to capture video context information (left part of Figure \[fig2:3\]), ours provides an elegant alternative to adaptively learn the weighted representation (right part of Figure \[fig2:3\]) based on the importance.
Network Optimization
--------------------
Given the representation of each proposal from Gaussian pooling, three 1D convolutional layers are utilized in parallel to predict action classification scores, localization parameters and overlap parameter, respectively. Action classification scores $\mathbf{y^a}=[y^a_0, y^a_1, ..., y^a_{C}]$ indicate the probabilities belonging to $C$ action classes plus one “background” class. Localization parameters $(\Delta c, \Delta w)$ denote temporal offsets relative to default center location $a_c$ and width $a_w$, which are leveraged to adjust the temporal coordinate $$\label{Eq2:6}\small
\begin{split}
\varphi_c = a_c + \alpha_1 a_w \Delta c ~~~{\rm{and}} ~~~\varphi_w = a_w \exp{(\alpha_2 \Delta w)}~,
\end{split}$$ where , are refined center location and width of the proposal. The $\alpha _1$, $\alpha _2$ are utilized to control the impact of temporal offsets. In particular, we define an overlap parameter $y_{ov}$ to represent the precise IoU prediction of the proposal, which benefits the proposal re-ranking in prediction.
In the training stage, we accumulate all the proposals from Gaussian pooling and produce the action instances through prediction layer. The overall training objective in our GTAN is formulated as a multi-task loss by integrating action classification loss ($L_{cls}$) and two regression losses, i.e., localization loss ($L_{loc}$) and overlap loss ($L_{ov}$): $$\label{Eq2:9}
\small
{L} = {L}_{cls} + \beta {L}_{loc} + \gamma {L}_{ov},$$ where and are the trade-off parameters. Specifically, we measure the classification loss $L_{cls}$ via the softmax loss: $$\label{Eq2:5}\small
L_{cls} = -\sum\limits_{n=0}^{C}I_{n=c}\log(y^a_{n}),$$ where indicator function $I_{n=c}=1$ if $n$ equals to ground truth action label $c$, otherwise $I_{n=c}=0$. We denote $g_{iou}$ as the IoU between default temporal boundary of this proposal and its corresponding closest ground truth. If the of this proposal is larger than $0.8$, we set it as a foreground sample. If is lower than $0.3$, it will be set as background sample. The ratio between foreground and background samples is set as 1.0 during training. The localization loss is devised as Smooth L1 loss [@Girshick:ICCV15] () between the predicted foreground proposal and the closest ground truth instance of the proposal, which is computed by $$\label{Eq2:7}
\small
{L}_{loc} = S_{L1}(\varphi_{c}-g_{c})+S_{L1}(\varphi_{w}-g_{w}),$$ where and represents the center location and width of the proposal’s closest ground truth instance, respectively. For overlap loss, we adopt the mean square error (MSE) loss to optimize it as follows: $$\label{Eq2:8}
\small
L_{ov} = (y_{ov} - g_{iou})^2.$$
Eventually, the whole network is trained in an end-to-end manner by penalizing the three losses.
Prediction and Post-processing
------------------------------
During prediction of action localization, the final ranking score $y_f$ of each candidate action proposal depends on both action classification scores $\mathbf{y^a}$ and overlap parameter $y_{ov}$: $$\label{Eq2:10}
\small
y_f = \max(\mathbf{y^a}) \cdot y_{ov}.$$ Given the predicted action instance $\phi=\{\varphi_{c},\varphi_{w},C_a,y_f\}$ with refined boundary ($\varphi_{c},\varphi_{w}$), predicted action label $C_a$, and ranking score $y_f$, we employ the soft non-maximum suppression (soft-NMS) [@Bodla:ICCV17] for post-processing. In each iteration of soft-NMS, we represent the action instance with the maximum ranking score $y_{f_m}$ as $\phi_m$. The ranking score $y_{f_k}$ of other instance $\phi_k$ will be decreased or not, according to the IoU computed with $\phi_m$: $$\label{Eq2:11}
\small
y'_{f_k} =
\begin{cases}
~~~y_{f_k}~~~~~~~~~~~~~~~~~~~~~~~~~~~~~,~~ \text{\footnotesize if~~$iou(\phi_k,\phi_m)< \rho$} \\
~~~y_{f_k} \cdot e^{-\frac{iou(\phi_k,\phi_m)^2}{\xi}}~~,~~ \text{\footnotesize if~~$iou(\phi_k,\phi_m) \geq \rho$}
\end{cases},$$ where $\xi$ is the decay parameter and $\rho$ is the NMS threshold.
Experiments {#sec:EX}
===========
We empirically verify the merit of our GTAN by conducting the experiments of temporal action localization on two popular video recognition benchmarks, i.e., ActivityNet v1.3 [@ActivityNet] and THUMOS14 [@Thumos].
Datasets
--------
The **ActivityNet v1.3** dataset contains 19,994 videos in 200 classes collected from YouTube. The dataset is divided into three disjoint subsets: training, validation and testing, by 2:1:1. All the videos in the dataset have temporal annotations. The labels of testing set are not publicly available and the performances of action localization on ActivityNet dataset are reported on validation set. The **THUMOS14** dataset has 1,010 videos for validation and 1,574 videos for testing from 20 classes. Among all the videos, there are 220 and 212 videos with temporal annotations in validation and testing set, respectively. Following [@Xiong:ICCV17], we train the model on validation set and perform evaluation on testing set.
Experimental Settings
---------------------
**Implementations.** We utilize Pseudo-3D [@Qiu:ICCV17] network as our 3D backbone. The network input is a 16-frame clip and the sample rate of frames is set as $8$. The 2,048-way outputs from pool5 layer are extracted as clip-level features. Table \[table3:1\] summarizes the structures of 1D anchor layers. Moreover, we choose three temporal scale ratios $\{r_d\}_{d=1}^{3}=[2^0, 2^{1/3}, 2^{2/3}]$ derived from [@LinYi:ICCV17]. The IoU threshold $\varepsilon$ in Gaussian grouping is set as $0.7$ by cross validation. The balancing parameters $\beta$ and $\gamma$ are also determined on a validation set and set as $2.0$ and $75$. $\xi$ and $\rho$ are set as $0.8$ and $0.75$ in soft-NMS. The parameter $\alpha_1$ and $\alpha_2$ are all set as $1.0$ by cross validation. We implement GTAN on Caffe [@Yang:caffe] platform. In all the experiments, our networks are trained by utilizing stochastic gradient descent (SGD) with $0.9$ momentum. The initial learning rate is set as $0.001$, and decreased by $10\%$ after every $2.5k$ iterations on THUMOS14 and $10k$ iterations on ActivityNet. The mini-batch size is $16$ and the weight decay parameter is $0.0001$.
\[0.83\]
[[c|c|c|c|c|c]{}]{} & & & & & \
& & & & & 11\
& & & & & 19\
& & & & & 35\
& & & & & 67\
& & & & & 131\
& & & & & 259\
& & & & & 515\
& & & & & 1027\
\[table3:1\]
**Evaluation Metrics.** We follow the official evaluation metrics in each dataset for action detection task. On ActivityNet v1.3, the mean average precision (mAP) values with IoU thresholds between $0.5$ and $0.95$ (inclusive) with a step size $0.05$ are exploited for comparison. On THUMOS14, the mAP with IoU threshold $0.5$ is measured. We evaluate performances on top-100 and top-200 returned proposals in ActivityNet v1.3 and THUMOS14, respectively.
Evaluation on Temporal Action Proposal
--------------------------------------
We first examine the performances on temporal action proposal task, which is to only assess the boundary quality of action proposals, regardless of action classes. We compare the following advanced approaches: (1) Structure Segment Network (SSN) [@Xiong:ICCV17] generates action proposals by temporal actionness grouping. (2) Single Shot Action Detection (SSAD) [@Lin:MM17] is the 1D variant version of Single Shot Detection [@Liu:ECCV16], which generates action proposals by multiple temporal anchor layers. (3) Convolution-De-Convolution Network (CDC) [@Shou:CVPR17] builds a 3D Conv-Deconv network to precisely localize the boundary of action instances at frame level. (4) Boundary Sensitive Network (BSN) [@Lin:ECCV18] locates temporal boundaries with three actionness curves and reranks proposals with neural networks. (5) Single Stream Temporal action proposal (SST) [@Buch:CVPR17] builds a RNN-based action proposal network, which could be implemented in a single stream over long video sequences to produce action proposals. (6) Complementary Temporal Action Proposal (CTAP) [@Gao:ECCV18] balances the advantages and disadvantages between sliding window and actionness grouping approaches for final action proposal.
\[0.80\]
[[|c|c|c@[ ]{}c|c|]{}]{} & & &\
& AR & AR & AUC & AUC\
& 37.9 & - & -& -\
& 50.1 & 73.2 & 65.7 & -\
& 53.2 & 74.2 & 66.2 & 66.3\
& **54.3** & **74.8** & **67.1** & **67.4**\
\[table4:1\]
{width="91.00000%"}
We adopt the standard metric of Average Recall in different IoU (AR) for action proposal on both datasets. Moreover, following the official evaluations in ActivityNet, we plot both Recall-IoU curve and Average Recall vs. Average Number of proposals per video (AR-AN) curve in Figure \[fig4:1\]. In addition to AR metric, the area under AR-AN curve (AUC) is also reported in Table \[table4:1\] as AUC is the measure on test server of ActivityNet. Overall, the performances across different metrics and two datasets consistently indicate that our GTAN leads to performance boost against baselines. In particular, AR of GTAN achieves 54.3% and 74.8% on THUMOS14 and ActivityNet respectively, making the absolute improvement over the best competitor BSN by 1.1% and 0.6%. GTAN surpasses BSN by 1.1% in AUC when evaluating on online ActivityNet test server. The results demonstrate the advantages of exploiting temporal structure for localizing actions. Furthermore, as shown in Figure \[fig4:1\], the improvements are constantly attained across different IoU. In terms of AR-AN curve, GTAN also exhibits better performance on different number of top returned proposals. Even in the case when only less than 10 proposals are returned, GTAN still shows apparent improvements, indicating that GTAN is benefited from the mechanism of dynamically optimizing temporal scale of each proposal and the correct proposals are ranked at the top.
Evaluation on Gaussian Kernel and Grouping
------------------------------------------
\[0.76\]
[[|c|c|c| c|c | c|c|]{}]{} & &\
& & & & & &\
& & & & & &\
& & & & & &\
& 33.5 & 37.1 & 38.2 & 29.8 & 31.6 & 34.3\
\[table4:2\]
\[0.84\]
[[|c|c|c|c|c|]{}]{} & &\
& $\geq 128$ & & $\geq 2048$ & \
& 22.1 & 37.1 & 49.4 & 31.6\
& 25.9 & 38.2 & 54.2 & 34.3\
\[table4:3\]
Next, we study how each design in GTAN influences the overall performance on temporal action localization task. Fixed Scale simply employs a fixed temporal interval for each cell or anchor in an anchor layer and such way is adopted in SSAD. Gaussian Kernel leverages the idea of learning one Gaussian kernel for each anchor to model temporal structure of an action and dynamically predict temporal scale of each action proposal. Gaussian Grouping further mixes multiple Gaussian kernels to characterize action proposals with various length. In the latter two cases, Gaussian pooling is utilized to augment the features of each anchor with contextual information.
Table \[table4:2\] details the mAP performances by considering one more factor in GTAN on both datasets. Gaussian Kernel successfully boosts up the mAP performance from 33.5% to 37.1% and from 29.8% to 31.6% on THUMOS14 and ActivityNet v1.3, respectively. This somewhat reveals the weakness of Fixed Scale, where the temporal scale of each anchor is independent of temporal property of the action proposal. Gaussian Kernel, in comparison, models temporal structure and predicts a particular interval for each anchor on the fly. As such, the temporal localization or boundary of each action proposal is more accurate. Moreover, the features of each action proposal are simultaneously enhanced by contextual aggregation through Gaussian pooling and lead to better action classification. Gaussian grouping further contributes a mAP increase of 1.1% and 2.7%, respectively. The results verify the effectiveness and flexibility of mixing multiple Gaussian kernels to capture action proposals with arbitrary length. To better validate the impact of Gaussian grouping, we additionally evaluate GTAN on long action proposals. Here, we consider actions longer than 128 frames in THUMOS14 and 2048 frames in ActivityNet v1.3 as long actions, since the average duration of action instances in THUMOS14 is $\sim$4 seconds which is much smaller than that ($\sim$50 seconds) of ActivityNet. Table \[table4:3\] shows the mAP comparisons between GTAN and its variant GTAN$^-$ which excludes Gaussian grouping. As expected, larger degree of improvement is attained on long action proposals by involving Gaussian grouping.
Evaluation on the Number of Anchor Layers
-----------------------------------------
In existing one-stage methods, e.g., SSAD, temporal scale is fixed in each anchor layer and the expansion of multiple temporal scales is implemented through increasing the number of anchor layers. Instead, our GTAN learns one Gaussian kernel for each anchor in every anchor layer and dynamically predicts temporal scale of the action proposal corresponding to each anchor. The grouping of multiple Gaussian kernels makes the temporal scale more flexible. Even with a small number of anchor layers, our GTAN should be more responsible to localize action proposals with various length in theory. Figure \[fig4:2\] empirically compares the performances between SSAD and our GTAN on ActivityNet v1.3 when capitalizing on different number of anchor layers. As indicated by the results, GTAN consistently outperforms SSAD across different depths of anchor layers from 4 to 8 on both temporal action proposal and localization tasks. In general, more anchor layers provide better AUC and mAP performances. It is expected that the performance of SSAD decreases more sharply than that of GTAN when reducing the number of anchor layers. In the extreme case of 4 layers, GTAN still achieves 26.77% in average mAP while SSAD only reaches 5.12%, which again confirms the advantage of exploring temporal structure and predicting temporal scale of action proposals.
Comparisons with State-of-the-Art
---------------------------------
\[0.76\]
[[c|c@[ ]{}c@[ ]{}c@[ ]{}c@[ ]{}c]{}]{}\
& 0.1 & 0.2 & 0.3 & 0.4 & 0.5\
\
& 18.2 & 17.0 & 14.0 & 11.7 & 8.3\
& - & - & - & - & 13.5\
& - & - & - & - & 13.9\
& 36.6 & 33.6 & 27.0 & 20.8 & 14.4\
& 51.4 & 42.6 & 33.6 & 26.1 & 18.8\
& 47.7 & 43.5 & 36.3 & 28.7 & 19.0\
& - & - & 37.8 & - & 23.0\
& - & - & 40.1 & 29.4 & 23.3\
& 54.0 & 50.9 & 44.1 & 34.9 & 25.6\
& 54.5 & 51.5 & 44.8 & 35.6 & 28.9\
& 66.0 & 59.4 & 51.9 & 41.0 & 29.8\
& - & - & - & - & 29.9\
& - & - & 53.5 & 45.0 & 36.9\
\
& 39.7 & 35.7 & 30.0 & 23.2 & 15.2\
& 48.9 & 44.0 & 36.0 & 26.4 & 17.1\
& 51.0 & 45.2 & 36.5 & 27.8 & 17.8\
& 50.1 & 47.8 & 43.0 & 35.0 & 24.6\
& - & - & 45.7 & - & 29.2\
& 67.2 & 61.1 & 56.9 & 46.5 & 37.9\
& **69.1** &**63.7** & **57.8** & **47.2** & **38.8**\
\[table4:4\]
We compare with several state-of-the-art techniques on THUMOS14 and ActivityNet v1.3 datasets. Table \[table4:4\] lists the mAP performances with different IoU thresholds on THUMOS14. For fair comparison, we additionally implement GTAN using C3D [@Tran:ICCV15] as 3D ConvNet backbone. The results across different IoU values consistently indicate that GTAN exhibits better performance than others. In particular, the [email protected] of GTAN achieve 37.9% with C3D backbone, making the improvements over one-stage approaches SSAD and SS-TAD by 13.3% and 8.7%, which also employ C3D. Compared to the most advanced two-stage method BSN, our GTAN leads to 1.0% and 1.9% performance gains with C3D and P3D backbone, respectively. The superior results of GTAN demonstrate the advantages of modeling temporal structure of actions through Gaussian kernel.
On ActivityNet v1.3, we summarize the performance comparisons on both validation and testing set in Table \[table4:5\]. For testing set, we submitted the results of GTAN to online ActivityNet test server and evaluated the performance on the localization task. Similarly, GTAN surpasses the best competitor BSN by 0.6% and 1.1% on validation and testing set, respectively. Moreover, our one-stage GTAN is potentially simpler and faster than two-stage solutions, and tends to be more applicable to action localization in videos.
Figure \[fig4:3\] showcases temporal localization results of one video from ActivityNet. The Gaussian kernels and grouping learnt on the outputs of “” layer are also visualized. As shown in the Figure, Gaussian kernels nicely capture the temporal structure of each action proposal and predict accurate default boxes for the final regression and classification.
\[0.76\]
[[c|c@[ ]{}c@[ ]{}c@[ ]{}c@[ ]{}|c]{}]{}\
& &\
& 0.5 & 0.75 & 0.95 & Average & Average\
& 45.11 & 4.11 & 0.05 & 16.41 & 14.62\
& 26.01 & 15.22 & 2.61 & 14.62 & 17.68\
& 22.71 & 10.82 & 0.33 & 11.31 & 17.83\
& 45.30 & 26.00 & 0.20 & 23.80 & 22.90\
& 39.12 & 23.48 & 5.49 & 23.98 & 26.05\
& - & - & - & - & 28.28\
& 48.99 & 32.91 & 7.87 & 32.26 & 33.40\
& 52.50 & 33.53 & 8.85 & 33.72 & 34.42\
& **52.61** & **34.14** & **8.91** & **34.31** & **35.54**\
\[table4:5\]
Conclusions
===========
We have presented Gaussian Temporal Awareness Networks (GTAN) which aim to explore temporal structure of actions for temporal action localization. Particularly, we study the problem of modeling temporal structure through learning a set of Gaussian kernels to dynamically predict temporal scale of each action proposal. To verify our claim, we have devised an one-stage action localization framework which measures one Gaussian kernel for each cell in every anchor layer. Multiple Gaussian kernels could be even mixed for the purpose of representing action proposals with various length. Another advantage of using Gaussian kernel is to enhance features of action proposals by leveraging contextual information through Gaussian pooling, which benefits the final regression and classification. Experiments conducted on two video datasets, i.e., THUMOS14 and ActivityNet v1.3, validate our proposal and analysis. Performance improvements are also observed when comparing to both one-stage and two-stage advanced techniques.
**Acknowledgments** This work was supported in part by the National Key R&D Program of China under contract No. 2017YFB1002203 and NSFC No. 61872329.
[^1]: [This work was performed at JD AI Research.]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We study the problem of the emergence of cooperation in dynamic signed networks where agent strategies coevolve with relational signs and network topology. Running simulations based on an agent-based model, we compare results obtained in a regular lattice initialization with those obtained on a comparable random network initialization. We show that the increased degree heterogeneity at the outset enlarges the parametric conditions in which cooperation survives in the long run. Furthermore, we show how the presence of sign-dependent emotional strategies catalyze the evolution of cooperation with both network topology initializations.
*Keywords*: Evolution of cooperation, signed graphs, network dynamics, negative ties, agent-based models, degree heterogeneity
author:
- 'Simone Righi [^1]'
- 'Károly Takács[^2]'
bibliography:
- 'RighiTakacsECMS.bib'
title: Degree Variance and Emotional Strategies Catalyze Cooperation in Dynamic Signed Networks
---
Introduction and Related Literature
===================================
Cooperation among individuals is a key element for the survival and functioning of human and many animal societies. While cooperation is socially optimal, it is difficult to explain its existence in a population of selfish individuals. The Prisoner’s Dilemma (PD) is frequently used to study this puzzle as it describes the situation in which the self interest of the individual is opposed to the emergence of cooperation. Two players are given two alternative strategies: to cooperate or to defect. Defection guarantees a higher payoff regardless of what the partner does and is thus the dominant strategy. However, cooperation - if played mutually - provides higher payoffs than mutual defection.
A natural framework in which to study the emergence of cooperative behaviour is evolutionary game theory. This literature burgeoned following the seminar papers of [@maynardsmith1973the; @maynardsmith1982evolution] and [@axelrod1981evolution], with a large number of contributions being dedicated to the puzzle of cooperation (see [@rowthorn2009theories] for a recent survey of this subject). One strand of this literature looks into the effects of the structure of interactions on the outcome of the evolutionary process. In the context of the single shot PD, they find that cooperation in an unstructured population of randomly interacting individuals is not viable. Natural selection favors selfish defection, thus leading to groups composed entirely agents playing this strategy ([@taylor1978evolutionary; @hofbauer1998evolutionary]). Introducing a more stringent structure for the social contacts and thus allowing only agents that are interconnected on a network to interact ([@nakamaru1997evolution; @nowak1992evolutionary]) seems to provide a solution. Indeed, structuring the interactions increases both realism of models and the realism of conclusions allowing the survival of cooperation in the population. When considering structured interactions, the impact of network topology on the diffusion of cooperation needs can be addressed ([@Virtuallabs; @santos2005scale; @johnson2003social; @ohtsuki2006simple]) and the realism of model can be improved allowing the interaction structure to co-evolve with agent strategies. In this case, chances for cooperation are enhanced ([@santos2006cooperation; @yamagishi1996selective; @yamagishi1994prisoner]). More specifically, among the mechanisms that improve the conditions of cooperation in dynamic networks are the possibility of parter selection, exclusion of defecting agents, and exit from relationships ([@schuessler1989exit; @vanberg1992rationality; @yamagishi1996selective]).
A recent series of our ([@righi2014emotional; @righi2014parallel]) and other authors’ papers ([@szolnoki2013evolution]), extended the analysis of the emergence of cooperation to signed networks. We introduced the possibility of network ties to turn positive or negative, or to be deleted and relinked as a consequence of previous interactions. We showed that the presence of emotional strategies - that use the [*emotional*]{} content implied by the relational signs in social interactions when considering the strategy to play - is pivotal for the survival and diffusion of cooperation. Indeed, in some cases, this strategy acts as a catalyst for unconditional cooperation rather than gaining dominance itself. We characterized the conditions in terms of the speed of evolution and selection pressure that allow the emergence of cooperation. In line with the literature, we found that relatively low rates of strategy adoptions and high rates of rewiring of stressed links are required in order to sustain cooperation.
In this paper, we further extend the study of the emergence of cooperation in signed networks studying the impact of variance (or heterogeneity) in the number of connections of agents at the outset. In particular, we compare the results obtained on a regular lattice with those obtained on a comparable random network. We show that the increased degree variance at the outset extends the parameters’ range under which cooperation survives in the long run. In this sense, our results confirm and extend those of [@santos2005scale], and show that networks with high heterogeneity in degrees improve the conditions for the emergence of cooperation. Moreover, we show that the benefits in terms of increased space for cooperation by introducing the emotional strategy extend to both random networks and regular lattices.
In the remaining of this paper, we proceed as follows. First we discuss the characteristics of the agent-based model, then we report our results, and conclude with a brief discussion.
Model
=====
We consider a population of size $N$, connected by an undirected and non-weighted signed network. We restrict our interest to networks that are single components. Each agent $i\in\{1,2, ..., n\}$ plays the single-shot Prisoner’s Dilemma (PD) with each of his current neighbors, i.e. with a subset of the whole population $\mathcal{F}_i^t \subset N$. The cardinality $k_i^t$ of $\mathcal{F}_i^t$ is the degree (or number of network contacts) of the agent $i$, at time $t$. The network is signed and each tie is labelled either *negative* or *positive*.
We assume that the social network constrains the possible interactions so that only currently connected agents can play the game together. The payoff structure of the PD is reported in Table \[PDPayff\]. When two agents cooperate with each other, each gets a reward (R). When they both defect, they are both punished (P). When one agent defects and the other cooperates, the first gets a temptation payoff (T), while his partner obtains the sucker payoff (S). The PD is defined with payoffs $T>R>P>S$. A typical additional assumption, that we adopt here, is $T+S>R+P$ ([@axelrod2006evolution]).
$\begin{array}{|p{0.7cm}|c|c|}
\hline
& \textrm{C} & \textrm{D} \\
\hline \textrm{C} & (R=3,R=3) & (S=0,T=5) \\
\hline \textrm{D} & (T=5,S=0) & (P=1,P=1) \\
\hline
\end{array}$
Agents play cooperation or defection in the PD according to their type. We consider three possible strategy types:
- Unconditional Cooperation (UC) that always cooperates, without taking into account the sign of the tie he shares with his interacting partner.
- Unconditional Defection (UD) that always defects.
- Conditional Action (COND) that cooperates with agents he shares a positive tie with and defects with those he has a negative tie with[^3]. We label this strategy as *emotional*, because it is a trigger response to the valence of the relation.
We let our agent based model to run in time steps. Steps are iterated until an equilibrium is reached. The conditions for considering one configuration as an equilibrium are stringent. It is required that: (1) a transitory period of 150 steps has passed from the beginning of the simulation (2) in five randomly chosen periods of time since (each time has a probability $0.1$ to be selected) the configuration of both relational sign, network topology and agent types needs to be precisely the same. [^4] Each time step (say $t$), a set of actions are performed by each agent, with the updates being done in parallel. Agents interact with peers they were connected with at the previous time step ($t-1$) and eventual updates in signs or network topology are observed by partners only in the following step $t+1$. Following a typical implementation of the literature, we assume that each agent plays the PD with all agents in his first order social neighborhood (i.e. with each $j\in\mathcal{F}_i^{t-1}$) and the average payoff is used when updating the agent strategy. The interested reader can find in [@righi2014parallel] a discussion of the effects of using an alternative, sequential, updating protocol.
The dynamics of our model allows for the co-evolution of network signs, agent strategies, and network structure. At each time step, network signs and agents behavior influence each other and the latter also affects the evolution of network topology. More in detail, after each dyadic interaction, stressed network signs are updated (with probabilities $P_{neg}$ and $P_{pos}$) or deleted and substituted with a new one with a certain probability ($P_{rew}$). At the end of each time step, when all payoffs are calculated, agents update their strategy to one that has been more successful in their neighborhood, with a certain probability $P_{adopt}$ (see Algorithm \[parallelalgo\]).
Let’s discuss each of the elements described in Algorithm \[parallelalgo\] more in detail.
[**Update of the relational sign between $i$ and $j$.**]{} After each dyadic PD game, agents might update their relational sign with each other. Given the nature of the PD, there are three possible situations:
- [*Both players cooperate*]{}. In this case an existing positive connection remains positive, while a negative one turns positive.
- [*Both players defect*]{}. Similarly, an existing positive relation is turned negative and a negative one remains so.
- [*One agent cooperates and the other defects*]{}. In this case, the emotional content of the relationship is subject to stress. We assume that if the link is positive, then the cooperator is [*frustrated*]{} to have a positive relation to a defector. Therefore, we assume that the valence of the tie can turn negative with probability $P_{neg}$. If the link is negative, the defector might be interested in turning it into a positive tie. We assume that it happens with probability $P_{pos}$. There are two possible justifications for such behaviour. The first is that the defector feels remorse or moral guilt (as suggested by [@gaudou2014moral]). The second is instead purely selfish. The defecting partner is content to remain friends with the cooperator. This type of relationship provides him with a strictly higher payoff, in case he is paired with a COND player, whose action is sensitive to the sign of their relationship. It is logical to assume, however, that the frustration from the cooperator is larger, therefore we impose $P_{neg}>>P_{pos}$. [^5]
[**Delete the link between $i$ and $j$ and create a new tie.**]{} An agent, frustrated by the current behavior of the partner, may decide to delete the social connection completely with probability $P_{rew}$. In this sense, our network topology co-evolves with agents’ strategies endogenously (similarly to [@santos2006cooperation]). $P_{rew}$, called rewiring probability, is assumed to be equal for the whole population and it non-strategic. When rewiring takes place, once the old link is erased, a new one is created with another agent. In line with the sociological literature ([@granovetter1973strength]), we assume that there is a tendency towards transitive closure.[^6] New connections are created to friends of friends (excluding the possibility of connections to friends of enemies, to enemies of friends, or to enemies of enemies). In order to introduce some social noise, with a probability $P_{rand}$, rewiring takes place to a randomly selected agent in the population. [^7] The network structure evolves dynamically through rewiring. This implies that, while the initial topology is either a regular lattice or a random network, it does not necessarily remain of this type - and in general, it does not.
[**Adopt a better strategy.**]{} Agents observe their average payoffs as well as the ones of the agents in their social neighborhood, and are thus able to measure the relative local efficiency of their strategy. If a subset of agents in $\mathcal{F}_i^{t-1}$ has a payoff at time $t$ higher than his own, then agent $i$ will adopt the strategy played in $t$ by one of them, selected uniformly at random. Evolutionary update happens, for each agent, with probability $P_{adopt}$ which is assumed to be equal for all agents.
[**Simulations Calibration.**]{} Concerning the initial structure of the social network, we provide results for two cases. In a first set of simulations we assume that agents are laid on a [*regular lattice*]{} in which every agent has precisely 16 connections (the degree distribution is therefore degenerate as shown in the Left Panel of Figure \[distros\]). Then we introduce heterogeneity in degree distribution and we study networks initialized as Erdős-Rényi ([@erdHos1959random]) [*random graph*]{} (an example of the resulting degree distribution is provided in the Right Panel of Figure \[distros\]). In order to make the results comparable we impose that each pair of nodes is connected with an independent probability $P_{link}=0.16$ so that the degree distribution is centered around 16 with a standard deviation of about 4. Moreover, agents are assigned with one of the three strategies randomly in equal proportions. In the absence of conditional players, the proportion of UDs and of UCs are $1/2$. When CONDs are added, then the starting proportion of each type of agent is $1/3$. Finally, the relational signs are randomly distributed and initialized so that each link has a 50
![Degree distributions in typical networks used for initialize our simulations. In both the regular lattice case (Left Panel) and the random network case (Right Panel) the average number of connections per agent is 16. $N=200$.[]{data-label="distros"}](degdistroRL.eps "fig:"){width="48.00000%"} ![Degree distributions in typical networks used for initialize our simulations. In both the regular lattice case (Left Panel) and the random network case (Right Panel) the average number of connections per agent is 16. $N=200$.[]{data-label="distros"}](degdistroRand.eps "fig:"){width="48.00000%"}
Results
=======
Our aim with this study is to characterize the parameter configurations that favor the evolution of cooperation in dynamic signed networks. We focus on the effect of conditional (or emotional) strategy in two different network initializations: in a regular lattice and in an Erdős-Rényi random graph. The two main dynamic forces that operate in our model are the evolutionary pressure ($P_{adopt}$) and the network update dynamics ($P_{rew}$). Our strategy is to analyze their impact, changing their relative strength progressively. For each possible combination of the two probabilities (each studied for values between $0$ and $1$ with a granularity of $0.05$) we show results concerning the average proportion (calculated in 50 simulations) of the agents and network ties surviving at the steady state. [^8]
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
Figures \[FixedPFlipRL\] and \[FixedPFlipRand\] report results for two alternative cases each. In Figure \[FixedPFlipRL\], the social network is initialized so that every agent has initially the same amount of connections, which defines a regular lattice. In Figure \[FixedPFlipRand\]), results are shown for a setup where degree variance is introduced and the network is initialized as an Erdős-Rényi random network. In the Top Panels of the Figures, the population is initialized as equally divided between unconditional cooperators and unconditional defectors. This simulation is compared with the case (Lower Panels), in which the population is initialized as divided equally among UCs, UDs and CONDs.
One can observe several similarities in the results from the two kinds of starting configurations. As noted in our previous contributions ([@righi2014emotional; @righi2014parallel]), and coherently to what observed by [@santos2006cooperation], the relative speed (i.e. the probability) of the network topology update and of strategy adoption have two opposite effects on the viability of cooperative strategies. Increasing the speed of adoption of better strategies favors defection, as this is the strategy that maximizes payoffs in dyadic terms. At the opposite, a relatively high degree of network updating leads to higher proportions of cooperation, as it helps the formation of clusters of cooperators.
From Figures \[FixedPFlipRL\] and \[FixedPFlipRand\], we can observe that defectors suddenly lose dominance when a certain ratio between the two dynamic forces is reached. In the case of the regular lattice initialization without emotional strategies, the cooperation survives if the approximate relation $P_{rew} > 2 P_{adopt}$ holds. The chances of cooperation are increased for ER networs compared to a regular lattice initialization for any combination of $P_{rew}$ and $P_{adopt}$. In this case, the condition for cooperation to survive is $P_{rew} > 5/3 P_{adopt} (\sim 1.6 P_{adopt})$.
Let’s speculate about the reason for this improvement. In the absence of CONDs, the only force preventing UCs from being eliminated from the network is the rewiring of stressed ties. As we discussed, this process tends to create clusters of cooperators that can then survive. When all agents have the same connectivity, they all require similar amounts of time to rewire their connections to UDs, which can then spread locally and dominate the final population. When the network is initialized as random, some of the agents have less connections than the average, and they become isolated from the defector faster; substituting negative stressed connections with positive ones via transitive closure. The new connections are more likely to be with CONDs (when present) or UCs (which cooperate at least when given the opportunity and thus tend to develop positive ties) than with defectors, given the positive relations involved. These agents constitute therefore the nucleus of cooperative clusters around which more connected cooperative peers can survive.
In summary, degree heterogeneity provides [*time*]{} for clusters to form, even when the ratio between $P_{rew}$ and $P_{adopt}$ is less favorable. The positive effect of the increased heterogeneity for cooperation is stronger in more dynamic networks (higher $P_{rew}$) since agents with few connections extricate more efficiently their leverage effect on the formation of cooperative clusters. Introducing a variability in the degrees of the agents, thus increases the range of parameters in which cooperation survives and diffuses in the population. This result confirms the one obtained by [@santos2005scale]. We consider, however, a dynamic environment in which agent strategies co-evolve with relational signs and network topology. Moreover, from the purely topological point of view, we show that cooperation can be increased through heterogeneity also without recurring alterations of the randomness of the network (such as preferential attachment or network growth).
In both types of network initializations, when the conditions for cooperation to survive are met and CONDs are absent, the results for different parameter combinations are rather similar. They indicate that about 25% of the signs in the network are negative and the population turns out to be equally split between UCs and UDs. Regardless of the starting network, we observe that when cooperation survives, it does not diffuses. In both cases, the proportion of cooperators remains similar (or just below) its initial setup value. We can understand this result observing that the sole driving force allowing the survival of cooperation (in the absence of emotional strategies) is the rewiring of stressed links. In this sense, $P_{rew}$ has a purely positive effect on cooperation and $P_{adopt}$ has a purely negative one. When the first dominates, cooperation survives, when the second dominates cooperation disappears; hence there is the sharp phase transition between the two states.
As noted in our previous work ([@righi2014emotional]), the introduction of the COND strategy relaxes the parametric conditions for cooperation to survive. In the context of this paper, we note that this happens both for the regular lattice initialization, where the approximate condition for the survival of cooperation becomes $P_{rew} > 20/13 P_{adopt} (\sim1.53 P_{adopt})$, and for the random network where it becomes $P_{rew} > 4/3 P_{adopt} (\sim 1.3 P_{adopt})$. The relative effect of introducing CONDs in the population is thus similar in terms of the proportion of parameters in which cooperation becomes viable and thus the two initializations can be discussed together.
Understanding this result requires a closer look at the final proportions of agents and relational signs in the area that allows the survival of cooperation. Here, the final proportion of UC agents ranges from 25% to 75% of the population, with this probability decreasing monotonically as the adoption and rewiring probability increases. Confronting the results regarding UCs with those regarding UDs, one can notice that the decrease in the proportion of UCs benefits UDs little (their proportion passes from a minimum of about 5% to a maximum of 17%, but much more the conditional players (see Figure \[conds\]).
![Effect of the competing dynamics of strategy adoption (vertical axis) and of rewiring of stressed links (horizontal axis) on the final proportion conditional (COND) agents. Left Panel: results for the regular lattice initialization. Right Panel: results for the random network initialization[]{data-label="conds"}](RL_cond_with.eps "fig:"){width="48.00000%"} ![Effect of the competing dynamics of strategy adoption (vertical axis) and of rewiring of stressed links (horizontal axis) on the final proportion conditional (COND) agents. Left Panel: results for the regular lattice initialization. Right Panel: results for the random network initialization[]{data-label="conds"}](Rand_cond_with.eps "fig:"){width="48.00000%"}
In the same area, also the proportion of negative ties progressively increases from about 4% to about 25%, but never exceeds this value. We can thus conclude that, when cooperation is viable, clusters are formed in which CONDs and UCs are intertwined by positive links and therefore are functionally indistinguishable. Moreover, while in presence of conditional agents cooperative behaviour spread in the population, the dominant type of cooperation (conditional or unconditional) depends on the relative strength of our two main dynamic variables.
For environments with relatively [*low frequency*]{} of network and evolutionary updates, the COND strategy acts as a shield for UCs. When in contact with both pure cooperators and pure defectors, agents playing this strategy tend to enjoy higher payoffs than those who always defect (COND gets payoffs from cooperation when interacting with the UC type, while avoiding the [*sucker*]{} position when interacting with UDs), and therefore tends to replace them due to the evolutive pressure. As the proportion of UDs decreases and segregation increases, pure cooperation becomes the optimal strategy, as it avoids “errors” due to the mis-interpretation of network signs. Thus, UCs tend to diffuse at the expense of CONDs, and the final proportion of unconditional cooperators tends to be high. At the opposite, when the two dynamic updates happen relatively fast, this dynamics reverse in favor of COND. When adoption of strategies with higher payoffs is faster, the cooperation is in general more difficult to sustain and pure defection diffuses more. Under these conditions, emotional agents, being able to discern among cooperators (with whom they tend to form positive ties) and defectors (with whom they tend to form negative ties) suffer less [*sucker*]{} payoffs from pure defection than pure cooperators; which therefore tend to disappear faster. In this more dynamic setup, the number of cooperators reduces too fast to regain dominance later, and conditional cooperation turns out as dominant.
Conclusions
===========
The problem of evolution of cooperation has been widely studied in social sciences. Unlike most of the previous literature that considered only positive relations, we introduced negative ties as a force that is able to influence agents behavior. We presented results from an agent based model, where we studied the evolution of cooperation in dynamic signed networks in which agent strategies co-evolved with relational signs and network topology. Agents played the Prisoner’s Dilemma with their current neighbors and the result of dyadic interactions drove the evolution of relational signs and network relations. The average performance of a strategy across all interactions of one individual was defined as the fitness value that determined the evolutionary process.
In this paper, we performed an extensive simulation analysis of our model focusing on the effects on the survival of cooperative strategies as (1) network topology was varied, considering a regular lattice and a random network initialization; and as (2) a sign dependent strategy that considers the network signs when deciding whether to cooperate or to defect was introduced. We provided results for all possible combinations of the two main dynamic forces of this model by progressively changing both the probability of adopting more fitting strategies and the probability of rewiring tense connections.
In all cases, and in line with the literature, we showed that higher strategy evolution rates reduced the combinations of parameters in which cooperation survived (favoring the dyadic dominant strategy: to defect), while increasing the rewiring probability helped isolating cooperators from defectors thus favoring the survival of cooperation.
Random networks provided more place for the emergence of cooperative behavior for a larger set of parameters than regular lattices. This result is similar to the one of [@santos2005scale], however our outcome follows from a different mechanism. In [@santos2005scale], cooperation diffusion followed from the the presence of very connected hubs. In our setup, there were no such hubs (degrees have a bell shaped distribution around a characteristic degree and connections are purely random). The process allowing the diffusion of cooperation is the presence of individuals which are less connected than the average. By segregating early on in the simulations from defectors, these created the nuclei around which cooperative clusters could emerge.
Extending the analysis of [@righi2014emotional] to regular lattices, we studied the effects of the introduction of a conditional strategy that considers the relational signs to the partner to decide whether to cooperate or defect. The conditional strategy enlarged the parametric space in which cooperation evolved. Despite the advantage of being able to use more information, however, and regardless of the network topology adopted, the conditional strategy gained dominance itself only in a few, rapidly changing environment (where both adoption and rewiring happened relatively frequently). In these situations, the better performance of the COND strategy against pure defectors made the spread in the population possible. When the network and strategies were more stable, the conditional strategy acted instead as catalyst for the diffusion of unconditionally cooperative behavior.
The work presented in this paper is a first step in understanding the role of network topology in the diffusion of cooperation in dynamic signed networks. While a preliminary analysis has been introduced in [@righi2013signed], further studies are required to address the issue of the evolution of cooperation in non-random signed networks systematically. In particular, more realistic network initializations, such as scale-free and small-world networks, could be analyzed.
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors wish to thank the “Lendület” program of the Hungarian Academy of Sciences and the Centre for Social Sciences for their financial and organizational support and three anonymous referees for their useful comments.
Authors Biographies {#authors-biographies .unnumbered}
===================
[**SIMONE RIGHI**]{} is currently a Research Fellow at “Lendület” Research Center for Educational and Network Studies (RECENS) of the Hungarian Academy of Sciences. He studied mathematics and economics at the University of Namur (Belgium) where he obtained a Ph.D. in Economics in 2012 with a thesis on “Information aggregation and Political Economics”. His main research interests are: opinion dynamics, evolutionary game theory, network theory and industrial organization. His e-mail address is: [email protected] and up to date information and research papers can be found at his web-page: <http://perso.fundp.ac.be/~srighi/>.
[**KÁROLY TAKÁCS**]{} is Director of the “Lendület” Research Center for Educational and Network Studies (RECENS) of the Hungarian Academy of Sciences and associate professor of Sociology at the Corvinus University of Budapest. He obtained his Ph.D. in Sociology from the University of Groningen (Netherlands) with a thesis on “Social Networks and Intergroup Conflict”. During and after his graduate studies he has been visiting scholar at Cornell University, University of Brescia and at the Netherlands Institute for Advanced Study in Humanities and Social Sciences. His main research interests are: social networks, intergroup conflict, discrimination, evolution of altruism and of cooperation, agent-based simulations and experiments. His e-mail address is: [email protected]. Up to date information and research papers can be found at his web-page: <http://web.uni-corvinus.hu/~tkaroly/>.
[^1]: MTA TK “Lendület” Research Center for Educational and Network Studies (RECENS), Hungarian Academy of Sciences. Mailing address: Országház utca 30, 1014 Budapest, Hungary. Email: [email protected]
[^2]: MTA TK “Lendület” Research Center for Educational and Network Studies (RECENS), Hungarian Academy of Sciences. Mailing address: Országház utca 30, 1014 Budapest, Hungary. Email: [email protected]
[^3]: The opposite strategy, that of defecting with cooperators and cooperating with defectors is not considered as deemed to be unrealistic.
[^4]: Robustness checks with alternative parametrizations have been performed and they do not influence the results.
[^5]: For the runs reported here, we fixed $P_{neg}=0.2$ and $P_{pos}=0.1$. These values are assumed equal for all agents. We run a sensitivity analysis of this parameter in [@righi2014emotional]
[^6]: The assumption of existence of transitive closure makes the model more realistic and increase cooperation. As shown in [@righi2014emotional] however, our results are qualitatively robust when we relax this assumption and consider totally random rewiring.
[^7]: This parameter is assumed to be small but positive. Its value is fixed to $P_{rand}=0.01$ in our simulations.
[^8]: Standard deviations are not reported here and are available upon request. The variability of the results is quite small except in the area of the phase transition between the configurations in which cooperation survives and those where it disappears completely. Only statistically significant phenomena are studied and discussed in the following.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Tensor Minkowski Functionals (TMFs) are tensorial generalizations of the usual Minkowski Functionals which are scalar quantities. We introduce them here for use in cosmological analysis, in particular to analyze the Cosmic Microwave Background (CMB) radiation. They encapsulate information about the shapes of structures and the orientation of distributions of structures. We focus on one of the TMFs, namely $W_2^{1,1}$, which is the $(1,1)$ rank tensor generalization of the genus. The ratio of the eigenvalues of the average of $W_2^{1,1}$ over all structures, $\alpha$, encodes the net orientation of the structures; and the average of the ratios of the eigenvalues of $W_2^{1,1}$ for each structure, $\beta$, encodes the net intrinsic anisotropy of the structures. We have developed a code that computes $W_2^{1,1}$, and from it $\alpha$ and $\beta$, for a set of structures on the 2-dimensional Euclidean plane. We use it to compute $\alpha$ and $\beta$ as functions of chosen threshold levels for simulated Gaussian and isotropic CMB temperature and $E$ mode fields. We obtain the value of $\alpha$ to be one for both temperature and $E$ mode, which means that we recover the statistical isotropy of density fluctuations that we input in the simulations. We find that the standard inflationary $\Lambda$CDM predicts that the level of intrinsic anisotropy of hotspot and coldspot structures in the CMB fields is quantified by $\beta\sim 0.62$. Then we compute $\alpha$ and $\beta$ for temperature and $E$ mode data from the PLANCK mission. We find that the temperature field agrees with the standard $\Lambda$CDM prediction of no net orientation within $3-\sigma$. However, we find that the structures in $E$ mode data have a net orientation that deviates from the theoretical expectation at $14-\sigma$. The possible origin of this deviation may be due to instrumental effects or other sources and needs to be investigated further. For the net intrinsic anisotropy of structures we obtain values of $\beta$ for both temperature and $E$ mode that are consistent with the expectations from the standard $\Lambda$CDM simulations. Accurate measurements of $\alpha$ and $\beta$ can be used to test the standard model of cosmology and to search for deviations from it.'
address: |
$^1$ Indian Institute of Astrophysics, Koramangala II Block, Bangalore 560 034, India\
$^2$Indian Institute of Science, C V Raman Ave, Bangalore 560 012, India
author:
- 'Vidhya Ganesan$^{1,2}$'
- 'Pravabati Chingangbam$^{1}$'
title: 'Tensor Minkowski Functionals: first application to the CMB'
---
Introduction
============
From the time of its first detection by Penzias and Wilson in 1965 [@penzias:1965], the cosmic microwave background (CMB) radiation has provided us with a wealth of information about the Universe. The information is encoded in the statistical properties of temperature and polarization anisotropies [@Bond:1987ub]. Fluctuations in the CMB temperature were first detected by COBE [@smoot:1992]. Less than 10$\%$ of CMB photons are linearly polarized due to the presence of quadrupole anisotropies in the plasma just before the matter-radiation decoupling phase [@Coulson:1994]. This polarization was first detected by DASI [@Kovac:2002fg]. Polarization observations measure the Stokes parameters $Q$ and $U$. They are however not invariant under rotations about the lines of sight. The complex quantities $Q\pm iU$ transform as spin $\mp 2$ objects, respectively. These are usually expressed in terms of gradient component and curl component known as $E$ mode and $B$ mode respectively. As $E$ mode and $B$ mode transform as scalars under these rotation transformations, these fields are widely used for the polarization analysis. The PLANCK team has recently released high resolution map of $E$ mode [@adam:2015]. The additional information contained in the polarization data has significantly improved the constraints on cosmological parameters. The main contribution of $E$ mode polarization comes from the primordial density fluctuations [@Bardeen:1983] while the $B$ mode polarization can be generated by primordial gravitational waves [@Starobinsky:1979] or secondary effects such as gravitational lensing of CMB photons by large scale structure. $B$ mode polarization of primordial origin has not been detected yet, while the component that is sourced by lensing has been detected by SPTpol [@hanson:2013].
CMB fluctuations have been analyzed so far using various tools such as power spectrum, bispectrum, trispectrum, scalar Minkowski Functionals (SMFs) and so on. SMFs are used to study the geometric and topological properties of cosmological fields [@Tomita:1986; @gott:1990; @mecke:1994; @sch:1997; @winit:1998; @mat:2003; @hik:2006; @mat:2010; @Schmalzing:1998]. It has been applied to study primordial non-Gaussianity using CMB temperature [@wmap:2008; @planck:2013] and $E$ mode fields [@planck:2015; @Ganesan:2014lqa]. To name a few other applications, SMFs have also been used to study the effect of foreground contamination [@fore:2013; @bmode:2015], cold spot anomaly [@cold:2014] and modified gravity theories through lensing [@lens:2016] in CMB. Vector extensions of SMFs have been introduced into cosmology and applied to study the morphology and dynamical evolution of galaxy clusters [@beis:2000; @beis:2001]. Some of the other methods that have been used to study the topological properties are through clustering strength [@ross:2009; @ross:2010; @ross:2011], number of hot and cold spots [@coles:1987; @prava:2012; @park:2013; @Ganesan:2014lqa] and extrema counts [@pogo:2009].
The observed CMB data is consistent with the standard model of cosmology. However, there are a few anomalies seen in the CMB temperature anisotropy that contradict the expectations from the standard model. Some of these anomalies are low variance, hemispherical asymmetry, point parity asymmetry, mirror parity asymmetry and cold spot [@ade:2015]. These have been detected by both WMAP and PLANCK experiments, implying that their origin is not due to the systematics in these experiments. The PLANCK team’s final data release which will include the analysis of large angular scale polarization will further reveal the statistical significance of these anomalies. In order to get a better understanding and resolution of these anomalies it is desirable to develop efficient methods which can analyze the data.
Tensor Minkowski Functionals (TMFs) are $(a,b)$ rank tensorial extensions of SMFs [@Alesker:1999; @Hug:2008] that are constructed by taking tensor product of $a$ copies of the position vector and $b$ copies of the unit normal at each point of the boundary of a given structure (see section 2 for detailed mathematical definition). By virtue of being tensor quantities and having more degrees of freedom, they carry additional information in comparison to SMFs about the morphology of a single structure or a set of many structures. In this paper we introduce TMFs as tools to analyze cosmological fields, in particular for analyzing the CMB fields which are defined on two dimensions. In this work we focus attention on one of the TMFs, denoted by $W_2^{1,1}$, which is a generalization of the genus (one of the three SMFs in two dimensions), and which is a $(1,1)$ rank tensor. For a given set of many structures there are two ways of obtaining the eigenvalues of $W_2^{1,1}$. The first is to average the elements of $W_2^{1,1}$ over all the structures and then calculate the eigenvalues. The ratio of the two eigenvalues gives a measure of the degree of alignment or relative orientation of the structures. The second way is to calculate the eigenvalues for each structure, take their ratio, and then obtain the average ratio over all the structures. The average of the ratio of eigenvalues calculated in the second way gives a measure of the average intrinsic anisotropy of the structures. These new information about net orientation and net intrinsic anisotropy of a set of structures can be useful in resolving the anomalies in the CMB data, test the standard model of cosmology and search for deviations from it, and to discriminate different early Universe physics models. They can also be used to study characteristic signatures of instrumental and foreground signals. Moreover, TMFs are quite promising for analyzing data of the large scale structure of the Universe and 21 cm emissions from the epoch of reionization and extracting cosmological information from them.
G. E. Schroder-Turk [*et. al.*]{} [@schroder2D:2009] have provided with an explicit formulae to calculate TMFs for structures on two dimensions. We implement the formulae and develop a code, that we refer to as TMFCode, to calculate TMFs, in particular $W_2^{1,1}$, and from it the two kinds of ratio of eigenvalues described above. We carry out a detailed analysis of the numerical inaccuracies in these quantities due to pixelization of the plane. For application to the CMB fields we use the stereographic projection to project the CMB data which is given on the spherical sky, onto a Euclidean plane to calculate the TMFs. We first calculate $W_2^{1,1}$ and the two eigenvalue ratios for simulated CMB fields in order to calculate the net orientation and net anisotropy of structures predicted by the standard $\Lambda$CDM cosmology. Then we apply to temperature and $E$ mode polarization data from PLANCK. We find no net orientation in the temperature data in agreement with the standard $\Lambda$CDM prediction of no net orientation within $3-\sigma$. However, we find $14-\sigma$ evidence for a net orientation in $E$ mode. This deviation may be due to instrumental effects or other sources and requires further investigation for a proper understanding. We will pursue it further after the full PLANCK data release. For the net intrinsic anisotropy of structures we obtain values for both temperature and $E$ mode that are consistent with the expectations from the standard model.
TMFs have been used in various research areas to study a wide range of phenomena ranging from the shape of neuronal cells in the brain [@Beisbart:2006], to the shape of ice crystals in Antarctica [@Durand:2004], to the shape of galaxies and clusters of galaxies in our Universe [@Beisbart:2002].
This paper is organized as follows. In section \[sec2\] we give the definition of tensor Minkowski Functionals, describe its numerical calculation and the corresponding numerical inaccuracies due to the pixelization of the data. In section \[sec3\] we apply the TMF $(W_2^{1,1})$ to simulated CMB fields and we calculate the variation of net orientation $(\alpha)$ and net anisotropy $(\beta)$ with the threshold value, and we also give an estimate of $\alpha$ and $\beta$ that we expect for a Gaussian, isotropic CMB field. In section \[sec4\] we apply the TMF $(W_2^{1,1})$ to PLANCK data, we compare the values of $\alpha$ and $\beta$ with that of the simulated CMB fields and estimate the level of consistency of PLANCK data with the standard model of cosmology. In appendix A, we describe how to quantify the level of orientation for a set of structures on the sphere. Finally, we summarize the results along with a discussion of their implications and future directions in section \[sec5\].
Tensor Minkowski Functionals {#sec2}
============================
In this section we give the mathematical definition of tensor Minkowski Functionals and describe their numerical calculation for 2-dimensional random fields on Euclidean space.
Definition
----------
Let us denote a subset of 2-dimensional Euclidean space with smooth closed boundary (or boundaries) by $K$, and let the boundary contour (or contours) be denoted by $\partial K$. $K$ may consist of one or more simply and/or multiply connected regions. A simply connected region will have no hole and a multiply connected region may have one or a countable number of holes in them. To connect with the terminology used in geometrical and topological analyses of CMB fields, which is our main interest here, we refer to $K$ as an [*excursion set*]{}. A connected region is referred to as a [*hotspot*]{} and a hole as a [*coldspot*]{}. We will use the word [*structure*]{} to mean either a connected region or a hole. The contour that encloses a hotspot is assigned anticlockwise direction while a contour that encloses a hole is assigned clockwise direction.
The morphological properties of $K$ can be characterized by three scalar Minkowski Functionals (SMFs) which are defined as $$W_0(K)= \int_K d^2r, \quad W_{i}(K)=\frac{1}{2} \int_{\partial K} G_i dr,
\label{eqn:smf}$$ where $i = 1,2$, and the functions $G_i$ are $G_1=1$ and $G_2=\kappa$ [@gott:1990; @Schmalzing:1998]. The symbol $\kappa$ denotes the local curvature at each point on $\partial K$. Physically $W_0, W_{1}$ are the area and the length of the boundary of connected region of $K$ and $W_2$ is the number of connected regions minus the number of holes in $K$. Note that these definitions have normalization constants that are different from the usual SMFs used in the CMB analysis. However, their physical interpretation remains the same. Being scalar quantities they are invariant under coordinate transformations and insensitive to the orientation and anisotropy of the structures (connected regions and holes).
Tensor Minkowski Functionals are defined by generalizing the SMFs [@Alesker:1999; @Hug:2008], as follows. For simplicity let us focus on a single simply or multiply connected region $K_s \subset K $. The tensor Minkowski Functionals of rank $a+b$, with $a+b\ge 0$, are constructed by taking $a$ number of copies of the position vector $\vec{r}$ and $b$ number of copies of the normal vector $\vec{n}$ at each point on the contour $\partial K_s$ and taking their tensor product. The tensor Minkowski Functionals of $K_s$ with rank $a+b$ are then defined as $$\begin{array}{l}
\displaystyle W_0^{a,0}(K_s) = \int_{K_{s}} \vec{r}^{\,a} d^2 r\\
\displaystyle W_{\nu}^{a,b}(K_s) = \frac{1}{2} \int_{\partial K_{s}} \vec{r}^{\,a} \otimes \vec{n}^{\,b} G_i dr,
\end{array}
\label{eqn:tmf}$$ for $i = 1,2$. The functions $G_i$ are the same as in Eq. (\[eqn:smf\]). The tensor product between two vectors $\vec{A}$ and $\vec{B}$ is given by $(\vec{A} \otimes \vec{B})_{ij}=(A_iB_j+A_jB_i)/2$. For $a+b=0$, Eq. (\[eqn:tmf\]) reduces to the three SMFs. For $a=1$ and $b=0$ we get three vectorial Minkowski Functionals, and for $a+b=2$ we get seven tensor Minkowski Functionals of rank 2.
In general, the tensor Minkowski Functionals, unlike their scalar counterparts, are not invariant under coordinate transformations. They behave like tensors of corresponding rank. Specifically, tensors $W_1^{1,1}, W_1^{0,2}, W_2^{1,1}$ and $W_2^{0,2}$ are invariant under translation operation while others vary. In this paper we focus only on translation invariant tensors as translation covariant tensors are sensitive to position or the choice of origin. Specifically, we choose the tensor $W_{2}^{1,1}$ for the study as other translation invariant tensors are related to either $W_2^{1,1}$ or $W_{\nu}$.
In practice real data is pixelized and hence in this case the excursion set $K$ will be a finite set consisting of pixels that are included in the structures. $K_s \subset K$ will be a finite set of pixels which form a single structure. Therefore, the boundary of a single simply connected region will be a polygon, a doubly connected region (having one hole) will consist of two polygons, with the one corresponding to the hole being located inside the external one, and so on. The formulae for calculating TMFs of each polygon of $K_s$ is given in [@schroder2D:2009]. In particular, the tensor $W_2^{1,1}$ is given by $$\begin{array}{l}
\displaystyle W_2^{1,1}(K_s) = \sum_{(i,j)} \frac{1}{2} |e_{ij}|^{-1} (\vec{e}_{\,ij} \otimes \vec{e}_{\,ij}),\\
\end{array}
\label{eq:for}$$ where the pair $(i,j)$ labels the edge of the polygon between $i$ and $j$ vertices. $\vec{e}_{\,ij}$ is the edge of the polygon with length $|\vec{e}_{\,ij}|$. As mentioned above, the direction of $\vec{e}_{\,ij}$ is such that the boundary of a hotspot is anticlockwise while the boundary of a coldspot is clockwise.
Measure of anisotropy and orientation of structures
---------------------------------------------------
For each structure (hotspot or coldspot) of $K$, $W_2^{1,1}$ being a $2\times 2$ matrix has two real eigenvalues, which we denote by $\lambda_1$ and $\lambda_2$ such that $\lambda_1 \le \lambda_2$. For the entire $K$ let $\langle \dots \rangle$ denote averaging over all structures (hotspots or coldspots). Then let $\Lambda_1$ and $\Lambda_2$, such that $\Lambda_1\le \Lambda_2$, denote the eigenvalues of $\left\langle W_2^{1,1} \right\rangle$, where the averaging is done for each element of $W_2^{1,1}$. Then, we define the ratios $\alpha$ and $\beta$ as $$\alpha\equiv \frac{\Lambda_1}{\Lambda_2}, \quad
\beta\equiv\left\langle \frac{\lambda_1}{\lambda_2}\right\rangle.
\label{eq:avg}$$ As observed from the above formula, $\alpha$ and $\beta$ are both ratios of eigenvalues but they differ in the averaging process. $\alpha$ is obtained by first averaging $W_2^{1,1}$ over all structures, calculating the eigenvalues of $\left\langle W_2^{1,1} \right\rangle$ and then calculating their ratio. On the other hand $\beta$ is obtained by first calculating the ratio of eigenvalues of $W_2^{1,1}$ for each structure and then averaging the ratio over all the structures.
If the excursion set consists of only one structure then $\alpha=\beta$. For a given structure $\beta$ gives a measure of [*the intrinsic anisotropy*]{} or the deviation from rotational symmetry in the shape of the structure. For the case of many structures it gives the net anisotropy of all the structures. It is a straightforward exercise to show that $\beta=1$ for isotropic shapes such an equilateral triangle, square and circle. For a general shape the value of $\beta$ lies between 0 and 1, with $\beta<1$ implying that the shape of the structure is anisotropic with the value indicating the extent of anisotropy in the shape of the structure.
The arrangement of many anisotropic structures can be associated with an [*orientation*]{}. $\alpha$ gives a measure of the orientation or the deviation from rotational symmetry in the distribution of structures. If the structures are oriented randomly with no preferred direction then $\alpha=1$, else $\alpha$ lies between 0 and 1. If the structures with certain fixed anisotopic shape are arranged such that they are all orientated along the same direction then $\alpha$ will be equal to $\beta$. We define the quantity $\mathscr{O}$ which is a measure of the degree of alignment of the structures as $$\mathscr{O}\equiv\frac{1-\alpha}{1-\beta}.
\label{eqn:O}$$ $\mathscr{O}=1$ means that the structures are completely aligned, while $\mathscr{O}=0$ means no net orientation. Any value between zero and one will indicate the extent to which the structures are orientated. As this is a normalized form of net orientation $\alpha$, this will be a useful quantity for the purpose of comparison of the extent of orientation in the structures between different fields.
Numerical calculation of TMFs
-----------------------------
We have developed a code that computes TMFs of individual structures of an excursion set on a plane and then estimates the ratio of eigenvalues $\alpha$, $\beta$ of these structures. We refer to this code as TMFCode. When an excursion set is input to the TMFCode it first estimates the line segments which forms the edges of the structures at each vertex (point where four neighbouring pixels meet) on the pixelized plane based on the configurations of the surrounding pixels (see Fig. 7 of [@schroder2D:2009]). The details of each structure such as $\vec{e}_{\,ij}$ corresponding to each of its edges are stored. Then the code tracks all of the individual structures in the excursion set and assigns labels to each of these structures which allows us to retrieve the details of an individual structure whenever necessary. Then $W_2^{1,1}$ can be calculated for each of these structures using Eq. \[eq:for\], and from it $\alpha$ and $\beta$ are calculated.
### Anisotropy of a single elliptical disk on a plane and estimation of pixelization error:
The approximation of a continuous contour by a polygon on a pixelized space will inherently lead to numerical inaccuracy in the calculation of $W_2^{1,1}$. We quantify this for an excursion set consisting of a single elliptic disk (see Fig. \[fig:ellipse\]) for which $W_2^{1,1}$ is given by the following analytic expression, $$\begin{array}{l}
\displaystyle W_2^{1,1}= \begin{bmatrix} f_2^{1,1}(p,q) & 0 \\ 0 & f_2^{1,1}(q,p), \end{bmatrix} \\
\vspace{0.2cm} \\
\displaystyle f_2^{1,1}(p,q)=\frac{1}{2} p^2 q^2 \int_0^{2\pi} d \varphi \frac{cos^2 \varphi}{\left[ p^2-(p^2-q^2)cos^2 \varphi \right]^{3/2}},
\end{array}
\label{eq.tensorformula}$$ where $p$ and $q$ are the semi-major and semi-minor axes, respectively.
-------- -------------- ---------- ------------ ---------- -------
$q/p$ $\beta$ from $\%$ error
analytical
formula
$1000^2$ $2000^2$ $3000^2$
pixels pixels pixels
1.0000 1.0000 1.0000 1.0000 1.0000 0.0
0.8000 0.7154 0.7642 0.7641 0.7641 6.8
0.6000 0.4638 0.5418 0.5417 0.5418 16.8
0.5000 0.3518 0.4371 0.4370 0.4370 24.2
0.3000 0.1602 0.2432 0.2433 0.2432 51.8
0.1000 0.0274 0.0739 0.0741 0.0741 170.4
-------- -------------- ---------- ------------ ---------- -------
: Values of $\beta$ for a single elliptical disk, with different values of the aspect ratio $q/p$ (column 1), obtained from the analytical formula Eq. \[eq.tensorformula\] (column 2) and from the numerical calculation using TMFCode (column 3). Column 4 shows the percentage numerical error in $\beta$ corresponding to $3000^2$ pixels. []{data-label="table:plane"}
Table \[table:plane\] summarizes the values of $\beta$ for a continuous ellipse obtained using Eq. \[eq.tensorformula\] and the corresponding values for a polygon obtained from the TMFCode that uses Eq. \[eq:for\], for different resolutions corresponding to $1000^2,\ 2000^2$ and $3000^2$ pixels. Note that $\beta=\alpha$ in this case since there is only one structure. For a fixed aspect ratio, we find that the value of $\beta$ does not vary significantly with the resolution. However for a fixed resolution, we find that the numerical inaccuracy in $\beta$ increases as the aspect ratio $q/p$ decreases. This is because the polygon edges fails to capture the true curvature at every point of the contour of the continuous ellipse. Hence the cause of this numerical inaccuracy is the pixelization of the continuous ellipse. The parts of the contour having higher curvature will lead to larger error. We can notice that the polygons systematically over-estimate the value of $\beta$ in comparison to the analytically expected value of $\beta$ for the continuous ellipse. We will use this fact to make approximate corrections due to the pixel effect when we apply to the CMB fields.
### Orientation of two elliptical disks on a plane and estimation of pixelization error:
In order to quantify the numerical error in the calculation of $\alpha$ due to the pixelization we compare $\alpha$ obtained for the excursion set consisting of two continuous ellipses and the corresponding value obtained with its approximation as two polygons on a pixelized space. We fix the aspect ratio $q/p$ of both the ellipses to be 0.5. Table. \[tle:orient\] shows $\alpha$ for two continuous ellipses obtained using Eq. \[eq.tensorformula\], and the corresponding values for two polygons obtained from the TMFCode which uses Eq. \[eq:for\], for various relative orientations of the major axes of the two ellipses. $0^{\circ}$ corresponds to the case where the ellipses are completely aligned with each other, while $90^{\circ}$ corresponds to the case where the major axes are perpendicular to each other, and hence are completely unoriented with each other. For a fixed relative orientation, we find that the value of $\alpha$ does not vary significantly with the resolutions. However, for a fixed resolution, we find that the numerical inaccuracy decreases as the two ellipses become more and more unoriented with each other. When both the ellipses are completely aligned then $\alpha = \beta$. Hence we find that the numerical inaccuracy in $\alpha$ for the case where the ellipses are completely oriented is the same as the corresponding numerical inaccuracy in $\beta$. But as the ellipses become more and more unoriented the value of $\alpha$ approaches one irrespective of the value of $\beta$, hence the numerical inaccuracy also decreases. Therefore, for the application to real data with $\alpha$ close to one, the corrections to $\alpha$ due to the pixel effect can be neglected as the numerical inaccuracy in $\alpha$ is not significant when $\alpha$ is close to one. For the case when the real data has $\alpha$ close to $\beta$, then the corrections due to the pixel effect has to be taken into account. We will use this result when we apply to the CMB fields.
----------------- --------------- ---------- ------------ ---------- ------
Angle between $\alpha$ from $\%$ error
major axis analytical
of the ellipses formula
$1000^2$ $2000^2$ $3000^2$
pixels pixels pixels
$0^{\circ}$ $0.3518$ $0.4369$ $0.4371$ $0.4369$ 24.2
$20^{\circ}$ $0.3787$ $0.4668$ $0.4668$ $0.4674$ 23.4
$45^{\circ}$ $0.4936$ $0.5665$ $0.5660$ $0.5661$ 14.7
$60^{\circ}$ $0.6132$ $0.6720$ $0.6724$ $0.6727$ 9.7
$90^{\circ}$ $1.0000$ $1.0000$ $1.0000$ $1.0000$ 0.0
----------------- --------------- ---------- ------------ ---------- ------
: Values of $\alpha$ for two elliptical disk, with various relative orientation angle between the major axis of the ellipses (column 1), obtained from the analytical formula Eq. \[eq.tensorformula\] and using Eq. \[eq:avg\] (column 2), and from the numerical calculation using TMFCode (column 3). Column 4 shows the percentage numerical error in $\alpha$ corresponding to $3000^2$ pixels.[]{data-label="tle:orient"}
$W_2^{1,1}$ for excursion sets on the sphere: application to simulations of CMB data {#sec3}
====================================================================================
A CMB field value is associated with each point on a 2-dimensional spherical space. Therefore, the TMFCode is not directly applicable to such data. A straightforward way to proceed in this case would be to project the excursion set onto an Euclidean plane and then calculate the TMFs.
Stereographic projection of excursion sets
------------------------------------------
We choose to use the stereographic projection which is a conformal map and therefore preserves angles and shapes of structures. However, it does not preserve the size. Structures located further away from the north pole will have projected images on the plane that have the same shape but the size will be scaled up. This can be seen in Fig. \[fig:deform\] which shows the projection of an elliptic disk with its location on the sphere shifted further and further away from the north pole. Roughly speaking, if $a$ denotes the size of the scaling factor of a structure, then $W_2^{1,1}$ scales as $a$ because $\vec r$ scales as $a$, $\kappa$ as $a^{-1}$, $dl$ as $a$ and the unit normal is invariant. Therefore, two structures that have the same shape but different sizes will have the same value of $\beta$. Hence, stereographic projection is well suited for calculating $\beta$.
For a collection of structures with a particular distribution of shapes, there are many possible ways in which these structures can be arranged with different relative orientations. In each of these possibilities, the value of $\alpha$ quantifies the extent of orientation in the corresponding collection of structures. This statement remains true for structures on the stereographic projection also, as the shape of the structure remains unaffected as described in the previous paragraph. But due to the stereographic projection, scaling of the structures further away from the north pole leads to errors in the average $\left\langle W_2^{1,1} \right\rangle$ and hence also in the estimation of $\alpha$. However, as the scaling of the structures in the stereographic projection becomes significant as the structures get close to the equator, these errors can be reduced by removing the structures close to the equator.
$W_2^{1,1}$ for simulated CMB data
----------------------------------
For a generic random field with zero mean, the set of all pixels having values greater than or equal to a chosen value, usually referred to as [*threshold*]{}, forms an excursion set. Let the threshold value be denoted by $\nu$. The boundary contours that enclose hotspots and coldspots are iso-threshold contours. The excursion set changes systematically as $\nu$ is varied. Consequently, we can expect the TMFs to follow a corresponding systematic behaviour.
In this section we calculate $W_2^{1,1}$, and from it $\alpha$ and $\beta$, for simulated CMB temperature and $E$ mode data, and study the intrinsic shapes and orientations of the structures of the excursion set. We work with the field rescaled by the corresponding rms value, and so the typical threshold value is of order one. We choose the threshold range $-6<\nu<6$ with 20 bins for our calculations. We focus on Gaussian fields with isotropic distribution of fluctuations. In the absence of analytic expressions for $\alpha$ and $\beta$ our results here will serve the purpose of providing an estimate of their values that we should expect for Gaussian, isotropic random fields.
For our analysis we simulate Gaussian and isotropic CMB fields with input $\Lambda$CDM cosmological parameter values given by $\Omega_c h^2$=0.1198, $\Omega_b h^2$=0.02225, $H_0$=67.27, $n_s$=0.9645, $ln(10^{10}A_s)$=3.094 and $\tau$=0.079, from the 2015 Planck data release [@planck:cosmopara2015]. We study the imprint of only the scalar primordial perturbations and hence we simulate $T$ and $E$ fields only. The angular power spectrum for each of the fields is first calculated using the publicly available CAMB package [@Lewis:2000ah]. We take the maximum multipole to be $\ell_{\rm max}=2200$. The $T$ and $E$ maps are then simulated using the HEALPIX [@Gorski:2005] package[^1]. We choose the HEALPIX resolution parameter value $N_{\rm side}=1024$. We have chosen a Gaussian smoothing of FWHM=$20'$ for $T$, and FWHM=$50'$ for $E$ mode.
To calculate $W_2^{1,1}$ for each threshold value we apply the procedure outlined in the beginning of this section. The number of pixels on the projected plane is $(3\times N_{\rm side})^2$. To minimize numerical inaccuracies arising from structures near the boundary of the stereographic projected disk we remove the structures that fall in the range $\theta$ = $70^{\circ}$ to $110^{\circ}$. To do this we create a map which marks these pixels. Then this is projected on to a plane using stereographic projection. The list of pixels on the projected map which corresponds to the initially marked pixels on the sphere are stored, so that it can be used repeatedly. For each $\nu$ structures that overlap with the above pixel list are then removed.
Our results for $\alpha$ and $\beta$ for the temperature field (denoted by superscript $T$) are shown in Fig. \[fig:tnu\]. The top left and top right panels show $\alpha^T$ and $\beta^T$ calculated using all hotspots and cold spots. All plots are average over 100 maps. The error bars are the sample variances at each $\nu$ obtained using the 100 realizations. We find that $\alpha^T$ is roughly one for $|\nu| < 1$ with small error bars. The conclusion we draw from this is that the simulated CMB temperature maps has [*no net orientation*]{} of the structures. This result is consistent with what we expect from a statistically isotropic distribution of a large number of structures, which was our assumption during the map simulation. Note that our result here is obtained for one particular choice of the stereographic projection plane. For a realistic search for net orientation in observed data we will have to work with several choices of projection planes. The shape of $\alpha^T$ as a function of $\nu$ appears symmetric about $\nu=0$. $\alpha^T$ becomes less than one for larger values of $|\nu|$ with larger error bars. This can be explained by the fact that for larger $|\nu|$ the number of structures gradually becomes fewer. The reduced sample size of structures leads to a net orientation with the degree of alignment of the structures growing with $|\nu|$.
From the flat part of the plot of $\beta^T$ we conclude that, the structures are on average [*intrinsically anisotropic*]{}, and the field has a net anisotropy of roughly $\beta^T \simeq 0.68$. Error bars are small in the flat part of the plot since the number of structures are large, as in the case of $\alpha^T$. The shape of $\beta^T$ as a function of $\nu$ appears to be symmetric about $\nu=0$. At larger values of $|\nu|$, large number of sample maps do not contain any structures, hence when the $\beta$ from all the sample maps are averaged it results in smaller value of $\beta_{T}$ and larger error bars. The drop from the flat to the decreasing parts are much more sharp in comparison to $\alpha^T$.
We have further calculated $\alpha$ and $\beta$ separately for only hotspots and only coldspots for the temperature field. The results for hotspots (denoted by subscript $h$) are shown in the middle of the left and right panels of Fig. \[fig:tnu\]. The corresponding results for coldspots (denoted by subscript $c$) are shown in the bottom part of the left and right panels of the same figure. The ‘peaks’ (central part of the flat region) of the plots for hotspots is located at roughly $\nu=1$, while for coldspots they are located at roughly $\nu=-1$. These shifts are due to the fact that we find more hotspots for positive values of $\nu$ and more coldspots for negative values. Apart from this shift of the ‘peaks’, the salient features that we observe are similar to what we described above for the combined result of hotspots and coldspots and hence the overall conclusions remain the same. Hotspots and coldspots independently do not exhibit any net orientation and their net intrinsic anisotropy are the same.
The results for $\alpha$ and $\beta$ for the $E$ mode field (denoted by superscript $E$) are shown in Fig. \[fig:enu\]. The behaviour of all the corresponding plots are the same as that was seen earlier for the temperature field. Hence the overall conclusions about the net orientation and intrinsic anisotropy of structures are the same for E mode field.
Correction of the numerical inaccuracy due to pixelization for $\alpha$ and $\beta$
------------------------------------------------------------------------------------
The true iso-threshold contours of the excursion sets of the CMB fields are continuous curves, at least up to quantum scales. The results for $\alpha$ and $\beta$ quoted in the previous subsection have been obtained using polygons on pixelised space. As we have shown in Section 2.3.1 the resulting numerical inaccuracy due to the effect of pixelization systematically increases with the rising ellipticity. In order to correct for this inaccuracy we make use of two points. The first is that, upon inspection of a typical CMB temperature or $E$ mode map one can see that contours with very high curvature are rare at any threshold. Secondly, any curved part of a typical contour can be approximated by a portion of an ellipse. Using these two observations we conclude that a reasonable correction of $\beta$ can be made by subtracting the numerical error corresponding to $\beta=0.68$ for a elliptic polygon by interpolating the values in Table \[table:plane\]. After doing this correction we conclude that the standard $\Lambda$CDM cosmology predicts that the structures of the temperature and $E$ mode CMB fields must have an intrinsic anisotropy given by $\beta=0.62$. This anisotropy encodes the anisotropy of the primordial fluctuations as well as tidal gravitational effects (first order perturbations) during the recombination epoch.
As shown in section 2.3.2, the calculation of $\alpha$ has negligible numerical error when the structures are completely unoriented. The value $\alpha=1$ that we obtain here for the simulated CMB fields show that the structures do not have any net orientation and hence no correction of numerical error is needed.
$W_2^{1,1}$ for masked CMB data
-------------------------------
Since the TMFCode is implemented directly on the pixel space it is straightforward to use it for masked CMB fields also. The CMB field should not be masked before the projection process as it may result in errors in the estimation of TMFs. If there are structures at the boundary of the mask then masking the CMB field directly without projecting it on a plane, will change the shape of these structures resulting in errors. In order to avoid these errors CMB field and the mask should be separately projected onto the plane. Then the structures in the projected CMB map which overlap with projected masked region can be removed completely without resulting in any errors. The above procedure applies for the case when no smoothing is involved. In the case of the smoothed CMB field, during the smoothing process contaminated pixels at the boundary of mask results in the contamination of the surrounding pixels and the extent to which the contamination spreads depends on the value of the smoothing angle. In order to remove these contaminated pixels, first the mask has to be smoothed with the chosen Gaussian smoothing angle. Then we choose a threshold level whose value represents how far away we stay from the boundary of the original mask map so as to reduce the inaccuracies due to the inclusion of the pixels that is contaminated due to smoothing. Then we change the value of all the pixels which have value above the threshold level into 1 and all the pixels which have value below the threshold level into 0. This will give the final mask for a given smoothing angle. Now the previously described procedure for the case of unsmoothed CMB field can be followed to get the masked projected CMB map for this case also.
Application to CMB data from PLANCK {#sec4}
===================================
For the analysis in this section we use CMB data from component-separation algorithm SMICA which combines the observed frequency maps efficiently to minimize the foreground contamination and also the corresponding masks included in the 2015 PLANCK data [@PLA].
In this section we estimate the net orientation and net anisotropy of Planck CMB fields, choosing $\nu=1$ for hotspots and $\nu=-1$ for coldspots. These values of $\nu$ were chosen since the number of structures are the largest around these threshold values. We choose a Gaussian smoothing such that the number of structures are about the order of $10^3$ at the chosen threshold level. This corresponds to a smoothing of FWHM = $20'$ for temperature field and $50'$ for $E$ mode field. The $1-\sigma$ error bar for the net orientation and net anisotropy can be obtained by averaging over a sample of simulated CMB fields which gives the cosmic variance. The simulated CMB maps that we have used for the present analysis does not include any instrumental effects. Therefore, we expect it to have no net orientation and can serve as a reference with which we can compare the value obtained for PLANCK map.
Net anisotropy of structures
----------------------------
Table \[ma\] shows $\beta$ separately for hotspots and coldspots for PLANCK temperature and $E$ mode data (column 2). For comparison, we show $\overline{\beta}$, which is the average over 100 simulation maps (described in Section 3), and the corresponding $1-\sigma$ error bars, in column 4. We find $\beta$ for both temperature and $E$ mode fields of PLANCK to be roughly $0.68$. By applying the correction for pixelization error as was done for the value of $\beta$ calculated from simulated maps in Section 3.2, to temperature and $E$ mode fields for both PLANCK and theoretical expectations from simulations in Table \[ma\], we obtain $\beta=0.62$.
In order to quantify the level of deviations of the measurements for PLANCK data from the theoretical expectations, we define the quantity $\mathscr{D}$ as $$\mathscr{D} = \left| \frac{X - \overline{X}}{\sigma_X} \right|,
\label{eqn:D}$$ where $X$ can be either $\alpha$ or $\beta$ and $\sigma_{X}$ is the corresponding one sigma errorbar.The values of $\mathscr{D}$ for $\beta$ are shown in table \[asig\]. The calculations shown in tables \[ma\] and \[asig\] where obtained for a particular choice of stereographic projection planes. In order to remove any dependence of the results on the choice of the projection plane, we have also repeated the calculations for various choices of projection planes. We find that the value of $\beta$ agrees with $\overline{\beta}$ obtained by averaging over 100 simulation maps to within $3-\sigma$ for all of these various choices of projection planes. In conclusion the net anisotropy of structures of the CMB fields of PLANCK agrees with the theoretical expectations of standard model of cosmology to within $3 \sigma$.
Net orientation of structures
-----------------------------
------------- ---------- ----------------------------- ---------- ------------------------------ ------------------------------ ----------
Field $\mathscr{O}$ using $\beta$ $\mathscr{O}$ using $\beta$
and corrected for corrected for
structure pixelization pixelization
error error
$\alpha$ $\beta$ $\overline{\alpha}$ $\overline{\beta}$
Temperature $0.9889$ $0.6795$ $0.0290$ $0.9911^{+0.0034}_{-0.0054}$ $0.6754^{+0.0026}_{-0.0030}$ $0.0231$
hotspot
Temperature $0.9936$ $0.6791$ $0.0167$ $0.9910^{+0.0038}_{-0.0052}$ $0.6754^{+0.0030}_{-0.0026}$ $0.0233$
coldspot
E mode $0.9673$ $0.6820$ $0.0861$ $0.9930^{+0.0034}_{-0.0034}$ $0.6858^{+0.0022}_{-0.0028}$ $0.0186$
hotspot
E mode $0.9593$ $0.6812$ $0.1069$ $0.9928^{+0.0028}_{-0.0038}$ $0.6854^{+0.0032}_{-0.0030}$ $0.0191$
coldspot
------------- ---------- ----------------------------- ---------- ------------------------------ ------------------------------ ----------
: [*Column 2:*]{} $\alpha, \beta$ at a chosen threshold, for temperature and $E$ mode data from PLANCK calculated separately for hotspots and coldspots. The threshold value chosen was $\nu=1$ for hotspots and $\nu=-1$ for coldspots. [*Column 3 :*]{} $\mathscr{O}$ defined in Eq. \[eqn:O\] calculated using the values of $\alpha$, and $\beta$ corrected for pixelization error. [*Column 4:*]{} The corresponding average values (denoted by over bar) calculated from 100 simulated temperature and $E$ mode maps. The corresponding $1 \sigma$ error bars calculated using 100 simulations. [*Column 5:*]{} The corresponding value of $\mathscr{O}$ calculated using average values $\overline{\alpha}$ and $\overline{\beta}$ corrected for pixelization error.[]{data-label="ma"}
Table \[ma\] shows $\alpha$ separately for hotspots and coldspots for PLANCK temperature and $E$ mode data (column 2). For comparison, in column 4 we show $\overline\alpha$, which is the average over 100 simulation maps (described in Section 3), and the corresponding $1\sigma$ error bars. Comparison of the level of orientation of the structures in various fields can be conveniently done using $\mathscr{O}$ defined in Eq. \[eqn:O\]. $\mathscr{O}$ calculated using $\alpha$ and with $\beta$ corrected for pixelization effect, from PLANCK data are shown in column 3 of Table. \[ma\]. The calculations of $\mathscr{O}$ corresponding to $\overline{\alpha}$ and with $\overline{\beta}$ corrected for pixelization effect, which are the averages over $100$ simulation maps are shown in column 5. We find that the extent of orientation in the structures of the $E$ mode field, $\mathscr{O}$, from PLANCK data is large compared to what is obtained from the average over many realizations while the corresponding values obtained for temperature field are comparable. Table \[asig\] shows $\mathscr{D}$ calculated for hotspots and coldspots for temperature and $E$ mode fields. It can be observed that for temperature field, all values of $\mathscr{D}$ are less than one. However, for $E$ mode the values are much larger than one. For these calculations we had chosen the Galactic plane as the stereographic projection plane.
The results described above were obtained for a particular choice of projection plane. We then ask whether these results are a fluke arising due to the choice of stereographic projection plane (described in \[A1\]). So in order to further analyze the net orientation for temperature and $E$ mode fields, we have repeated the calculations for several other choices of projection planes. We still find that the deviation of $\alpha$ from $\overline{\alpha}$ for the temperature field is within $3-\sigma$, but for $E$ mode the deviation is significant and of about $14-\sigma$ from the theoretical expectation of no net orientation. Hence, we conclude that the structures in the $E$ mode data from PLANCK do exhibit a net orientation. The reason behind this behaviour may be due to the instrumental effects that has not been included in the simulated CMB maps or this may be due to the contamination present in the PLANCK $E$ mode map. Or this may be a signature of the existence of some net orientation in the structures of the $E$ mode field. The exact reason can only be revealed by further investigation.
--------------------- ------------------------------------------------------------------------------------ ---------------------------------------------------------------------------------
Field and structure $\mathscr{D} = \left| \dfrac{\alpha - \overline{\alpha}}{\sigma_{\alpha}} \right|$ $\mathscr{D} = \left| \dfrac{\beta - \overline{\beta}}{\sigma_{\beta}} \right|$
Temperature $0.5000$ $1.4643$
hotspots
Temperature $0.5778$ $1.3214$
coldspots
$E$ mode $7.5588$ $1.5200$
hotspots
$E$ mode $10.1515$ $1.3548$
coldspots
--------------------- ------------------------------------------------------------------------------------ ---------------------------------------------------------------------------------
: The table shows the quantity $\mathscr{D}$, defined in Eq. \[eqn:D\], for $\alpha$ (column 2) and $\beta$ (column 3).[]{data-label="asig"}
Conclusion {#sec5}
==========
We have introduced tensor Minskowski Functionals as a new statistic to analyze CMB data. The new information about the morphology of structures, namely, the net orientation and net anisotropy of structures, opens up the possibility for a wide range of applications in cosmology. The applications include new tests for the standard cosmological model and searches for deviations from it, resolving the anomalies in the CMB data, and to constrain the physics of the early Universe. They can also be used to understand characteristic signatures of instrumental and foreground signals. They are also promising for the analysis of data of the large scale structure and 21 cm emissions from the epoch of reionization.
We have developed a code for calculating TMFs, focussing on the $(1,1)$ rank tensor $W_2^{1,1}$ which is a generalization of the genus, for excursion sets of two dimensional random fields. The code then uses $W_2^{1,1}$ to measure the net orientation, $\alpha$, and net intrinsic anisotropy, $\beta$, of structures. Our intended application is to random fields on continuous space and the pixelization of the data introduces numerical errors. We have done a careful estimation of this numerical error by using the known formula for $W_2^{1,1}$ for ellipses. We find that the error in the measure of the intrinsic anisotropy increases as the curvature of the boundaries of the structure increases. For the net orientation we find that for a distribution of structures that are completely unaligned with each other, the error is negligible. However, as the structures become more and more aligned, the error approaches the value of the error for the intrinsic anisotropy.
CMB data is associated with each point on the spherical sky. So in order to apply the TMFCode, we use stereographic projection of the CMB field onto a plane. We calculate $W_2^{1,1}$, and then compute $\alpha$ and $\beta$ as functions of different threshold levels for simulated Gaussian and isotropic CMB temperature and $E$ mode fields. We find that the standard $\Lambda$CDM predicts that the level of intrinsic anisotropy of hotspots and coldspots in both the CMB fields to be $\beta = 0.62$, where correction due to pixelization has been taken into account. Further, we find the value of $\alpha$ to be one for botih temperature and $E$ mode fields, which implies that there is no net orientation in the structures of these fields. This is a recovery of the statistical isotropy of density fluctuations that we have input into the CMB simulations. Then, we use TMFCode to compute $\alpha$ and $\beta$ for temperature and $E$ mode data from PLANCK mission. We find that $\beta$ for both temperature and $E$ mode data are consistent with the expectations from standard $\Lambda$CDM simulations within $3-\sigma$. Further, we find that the temperature field agrees with the standard $\Lambda$CDM prediction of no net orientation within $3-\sigma$. However, we find $14-\sigma$ evidence for a net orientation in $E$ mode data. The reason behind this may be instrumental effects that we have not included in the simulated CMB maps, or this may be due to the contamination present in the PLANCK $E$ mode data. This may also be a signature of the existence of some net orientation in the structures of the $E$ mode field. The exact reason can only be revealed by further investigation and we plan to come back to this issue after the PLANCK team releases the full polarization data.
This work is the initiation of several lines of investigation. We list here some of our ongoing and planned work. We have ignored tensor perturbations, and consequently $B$ mode field in this work. We are currently studying the net orientation and net anisotropy for $B$ mode sourced by primordial tensor perturbation as well as sourced by lensing due to the large scale structure of the Universe. Further, non-Gaussianity of primordial fluctuations generated during the very early Universe has an imprint on the CMB fields. Understanding of the effect of primoridial non-Gaussianity on the TMFs of CMB fields may eventually lead to an alternative tool to constrain $f_{NL}$ which can be used as a consistency check. And with further development it may be used to distinguish different origins of non-Gaussianity in CMB fields. We are currently pursuing this issue.
The effects which are of primordial origin should be dominant on angular scales larger than roughly $1^{\circ}$. We plan to use this fact to study whether TMFs of CMB fields are capable of distinguishing the primordial origin of fluctuations from the late time effects. Further, our study here was carried out for one set of cosmological parameters. It would be important to understand the effect of various cosmological parameters on the TMFs of the CMB fields. This will lead to a better understanding of the power of using TMFs as a tool to constrain cosmology. Since a precise measurement of $\alpha$ and $\beta$ is necessary for model testing and constraining, it is very important to get a better handle on the numerical inaccuracies induced due to the pixelization and we are searching for ways to minimize them.
Since the TMFs are quite general and applicable to any random field on dimensions higher than or equal to two, they can be applied to other areas of cosmology as well. The two-dimensional case developed here can be used to study 21cm emissions from the epoch of reionization on two-dimensional redshift slices in a straightforward manner. This can be used to probe cosmological parameters and also different models of reionization. Further, we plan to develop codes which can calculate TMFs for 3-dimensional data [@schroder3D:2013] with the aim of constraining cosmology using $21$cm data from future radio interferometers and large scale structure data.
Orientation measure of the structures of excursion set on a sphere {#A1}
==================================================================
The orientation measure $\alpha$ can detect the existence of any alignment in structures of the excursion set on a plane (described in Section 2.3.2). In this section we extend this study to the excursion set on a sphere. We consider an excursion set on the sphere with many ellipses whose aspect ratio are fixed to be 0.7. We consider all of these ellipses to be aligned towards the pole of the sphere and hence has net orientation. In order to reduce the numerical inaccuracies due to the stereographic projection we remove the ellipses which fall in the range $\theta = 70^{\circ}$ to $110^{\circ}$.
Table \[table:ellipse\_orient\] shows the net orientation $\alpha$ and the normalized form of net orientation, $\mathscr{O}$ defined in Eq. \[eqn:O\], for various choices of stereographic projection plane. In each of these choices the equator of the unit sphere with the north pole specified in column 1 of Table \[table:ellipse\_orient\] is the projection plane. We find that $\alpha$ and $\mathscr{O}$ does not vary significantly with $\varphi$ but varies with $\theta$. At $\theta = 0^{\circ}$, the value of $\alpha$ and $\mathscr{O}$ shows that the structures on the corresponding projection plane has no net orientation. While for the cases with $\theta=90^{\circ}$, the values show that the structures on the corresponding projection plane has a high level of orientation. The stereographic projection image corresponding to these two cases are shown in Fig. \[fig:orient\]. We find that these images also show the same results, that the structures corresponding to $\theta=0^{\circ}$ are not oriented with each other while for the case of $\theta=90^{\circ}$ shows a net orientation. Hence we can come to the conclusion that even though the structures of the excursion set on the sphere has net orientation, its stereographic projection onto a particular choice of projection plane may or may not show a net orientation. While for the case where the structures of the excursion set are randomly oriented then for any choice of the projection plane, the structures on these plane will also be randomly oriented and hence will have no net orientation. So in order to detect the orientation in the structures of the excursion set on a sphere one has to analyze the orientation for various choices of stereographic projection planes. Further we can infer the extent of orientation in the structures of the excursion set on the sphere by estimating $\alpha$ corresponding to the projection plane which shows maximum net orientation. In the present case of excursion set on a sphere with many ellipses, this corresponds to the choice of projection plane with $\theta = 90^{\circ}$ and, hence $\alpha = 0.74, \mathscr{O} = 0.93$. The value of $\mathscr{O}$ is close to one which implies that the structures in the present case of the excursion set are highly oriented with each other which is the result that we expected. Finally in conclusion, the orientation in the structures of the excursion set on the sphere can be detected and its extent of orientation can be quatified by estimating $\alpha$ for various choices of stereographic projection planes.
\
------------------------------ ---------- ---------------
North pole $\alpha$ $\mathscr{O}$
$(\theta,\varphi)$
($0^{\circ}$,$0^{\circ}$) $0.9963$ $0.0107$
($45^{\circ}$,$0^{\circ}$) $0.9322$ $0.2204$
($45^{\circ}$,$90^{\circ}$) $0.9212$ $0.2529$
($45^{\circ}$,$180^{\circ}$) $0.9322$ $0.2204$
($45^{\circ}$,$270^{\circ}$) $0.9212$ $0.2529$
($90^{\circ}$,$0^{\circ}$) $0.7259$ $0.9259$
($90^{\circ}$,$45^{\circ}$) $0.7368$ $0.9236$
($90^{\circ}$,$90^{\circ}$) $0.7208$ $0.9310$
($90^{\circ}$,$135^{\circ}$) $0.7368$ $0.9236$
------------------------------ ---------- ---------------
: The table showing the net orientation $\alpha$ (column 2) and $\mathscr{O}$ (column 3) defined in Eq. \[eqn:O\], for various choices of projection plane, for the excursion set on a sphere with many ellipses. In each of these various choices the equator of the unit sphere with the north pole specified in column 1 is the chosen projection plane.[]{data-label="table:ellipse_orient"}
References {#references .unnumbered}
==========
[599]{}
A. A. Penzias and R. W. Wilson, [*A Measurement of Excess Antenna Temperature at 4080 Mc/s,*]{} [[*Astrophys. J.* ]{} [**142**]{} p.419-421 (1965)](https://dx.doi.org/10.1086/148307).
J. R. Bond and G. Efstathiou, [*The statistics of cosmic background radiation fluctuations,*]{} [[*Mon. Not. Roy. Astron. Soc.*]{} [**226**]{} 655 (1987)](http://dx.doi.org/10.1093/mnras/226.3.655).
G. F. Smoot et al., [*Structure in the COBE differential microwave radiometer first-year maps,*]{} [[*Astrophys. J.*]{} [**396**]{} p.L1-L5 (1992)](http://adsabs.harvard.edu/abs/1992ApJ...396L...1S).
D. Coulson, R. G. Crittenden and N. G. Turok, [*Polarization and anisotropy of the microwave sky,*]{} [[*Phys. Rev. Lett.*]{} [**73**]{} 2390 (1994)](http://dx.doi.org/10.1103/PhysRevLett.73.2390) \[[arXiv:astro-ph/9406046](http://arxiv.org/abs/astro-ph/9406046)\] \[[[IN]{}SPIRE](https://inspirehep.net/search?p=find+eprint+ASTRO-PH/9406046)\].
J. Kovac, E. M. Leitch, C. Pryke, J. E. Carlstrom, N. W. Hatverson and W. L. Holzapfel, [*Detection of polarization in the cosmic microwave background using DASI,*]{} [[*Nature.*]{} [**420**]{} 772 (2002)](http://www.nature.com/nature/journal/v420/n6917/full/nature01269.html) \[[arXiv:astro-ph/0209478](http://arxiv.org/abs/astro-ph/0209478)\] \[[[IN]{}SPIRE](https://inspirehep.net/search?p=find+eprint+ASTRO-PH/0209478)\].
R. Adam et al. \[Planck Collaboration\], [*Planck 2015 results. IX. Diffuse component separation: CMB maps,*]{} [arXiv:1502.05956 \[astro-ph.CO\]](http://xxx.lanl.gov/abs/1502.05956) \[[[IN]{}SPIRE](http://inspirehep.net/search?p=find+eprint+1502.05956)\].
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**]{} 679 (1983)](http://dx.doi.org/10.1103/PhysRevD.28.679).
A. A. Starobinsky, [*Spectrum of relict gravitational radiation and the early state of the universe,*]{} [[*JETP Lett.*]{} [**30**]{} 682 (1979)](http://adsabs.harvard.edu/abs/1979JETPL..30..682S).
D. Hanson et al., [*Detection of B-mode Polarization in the Cosmic Microwave Background with Data from the South Pole Telescope,*]{} [[*Phys. Rev. Lett.*]{} [**111**]{} 141301 (2013)](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.111.141301) \[[arXiv:1307.5830 \[astro-ph.CO\]](https://arxiv.org/abs/1307.5830)\] \[[[IN]{}SPIRE](https://inspirehep.net/search?p=find+eprint+1307.5830)\].
H. Tomita, [*Curvature invariants of random interface generated by Gaussian fields,*]{} [[*Prog. Theor. Phys.*]{} [**76**]{} 952 (1986)](http://ptp.oxfordjournals.org/content/76/4/952.abstract).
J.R. Gott, C. Park, R. Juzkiewicz, W. E. Bies, D. P. Bennett, F. R. Bouchet and A. Stebbins, [*Topology of microwave background fluctuations-theory,*]{} [[*Astrophys. J.*]{} [**352**]{} 1 (1990)](http://dx.doi.org/10.1086/168511).
K. R. Mecke, T. Buchert, and H. Wagner, [*Robust Morphological Measures for Large-Scale Structure in the Universe,*]{} [[*Astron. Astrophys.*]{} [**288**]{} 697 (1994)](http://adsabs.harvard.edu/abs/1994A%26A...288..697M) \[[arXiv:astro-ph/9312028](http://arxiv.org/abs/astro-ph/9312028)\] \[[[IN]{}SPIRE](https://inspirehep.net/search?p=find+eprint+ASTRO-PH/9312028)\].
J. Schmalzing and T. Buchert, [*Beyond genus statistics: a unifying approach to the morphology of cosmic structure,*]{} [[*Astrophys. J.*]{} [**482**]{} L1-L4 (1997)](http://arxiv.org/ct?url=http%3A%2F%2Fdx.doi.org%2F10%252E1086%2F310680&v=7c2ef39c) \[[arXiv:astro-ph/9702130](http://arxiv.org/abs/astro-ph/9702130)\] \[[[IN]{}SPIRE](https://inspirehep.net/search?p=find+eprint+ASTRO-PH/9702130)\].
S. Winitzki and A. Kosowsky, [*Minkowski functional description of microwave background Gaussianity,*]{} [[*New Astron.*]{} [**3**]{} 75 (1998)](http://dx.doi.org/10.1016/S1384-1076(97)00046-8) \[[arXiv:astro-ph/9710164](http://arxiv.org/abs/astro-ph/9710164)\] \[[[IN]{}SPIRE](https://inspirehep.net/search?p=find+eprint+ASTRO-PH/9710164)\].
T. Matsubara, [*Statistics of smoothed cosmic fields in perturbation theory. I. Formulation and useful formulae in second-order perturbation theory,*]{} [[*Astrophys. J.*]{} [**584**]{} 1 (2003)](http://dx.doi.org/10.1086/345521) \[[arXiv:astro-ph/0006269](http://arxiv.org/abs/astro-ph/0006269)\] \[[[IN]{}SPIRE](https://inspirehep.net/search?p=find+eprint+ASTRO-PH/0006269)\].
C. Hikage, E. Komatsu and T. Matsubara, [*Primordial non-Gaussianity and analytical formula for Minkowski functionals of the cosmic microwave background and large-scale structure,*]{} [[*Astrophy. J.*]{} [**653**]{} 11 (2006)](http://dx.doi.org/10.1086/508653) \[[arXiv:astro-ph/0607284](http://arxiv.org/abs/astro-ph/0607284)\] \[[[IN]{}SPIRE](https://inspirehep.net/search?p=find+eprint+ASTRO-PH/0607284)\].
T. Matsubara, [*Analytic Minkowski functionals of the cosmic microwave background: second-order non-Gaussianity with bispectrum and trispectrum,*]{} [[*Phys. Rev. D*]{} [**81**]{} 083505 (2010)](http://dx.doi.org/10.1103/PhysRevD.81.083505) \[[arXiv:1001.2321\[astro-ph.CO\]](https://arxiv.org/abs/1001.2321)\].
J. Schmalzing and K. M. Gorski, [*Minkowski functionals used in the morphological analysis of cosmic microwave background anistropy maps,*]{} [[*Mon. Not. Roy. Astron. Soc.*]{} [**297**]{} 355 (1998)](http://dx.doi.org/10.1046/j.1365-8711.1998.01467.x) \[[arXiv:astro-ph/9710185](http://arxiv.org/abs/astro-ph/9710185)\] \[[[IN]{}SPIRE](https://inspirehep.net/search?p=find+eprint+ASTRO-PH/9710185)\].
C. Hikage, T. Matsubara, P. Coles, M. Liguori, F. K. Hansen and S. Matarrese., [*Limits on primordial non-Gaussianity from Minkowski Functionals of the WMAP temperature anisotropies,*]{} [[*Mon. Not. Roy. Astron. Soc.*]{} [**389**]{} 1439 (2008)](https://dx.doi.org/10.1111/j.1365-2966.2008.13674.x) \[[arXiv:0802.3677 \[astro-ph\]](http://arxiv.org/abs/0802.3677)\] \[[[IN]{}SPIRE](https://inspirehep.net/search?p=find+eprint+0802.3677)\].
P. A. R. Ade et. al. \[Planck Collaboration\], [*Planck 2013 results. XXIV. Constraints on primordial non-Gaussianity,*]{} [[*Astronom. and Astrophys.*]{} [**571**]{} A24 (2014)](https://dx.doi.org/10.1051/0004-6361/201321554) \[[arXiv:1303.5084 \[astro-ph.CO\]](https://arxiv.org/abs/1303.5084)\] \[[[IN]{}SPIRE](https://inspirehep.net/search?p=find+eprint+1303.5084)\].
P. A. R. Ade et al. \[Planck Collaboration\], [*Planck 2015 results. XVII. Constraints on primordial non-Gaussianity*]{}, [arXiv:1502.01592 \[astro-ph.CO\]](http://adsabs.harvard.edu/cgi-bin/bib_query?arXiv:1502.01592) \[[[IN]{}SPIRE](https://inspirehep.net/search?p=find+eprint+1502.01592)\].
V. Ganesan, P. Chingangbam, K.P. Yogendran and C. Park, [*Primordial non-Gaussian signatures in CMB polarization,*]{} [[*JCAP*]{} [**02**]{} 028 (2015)](http://iopscience.iop.org/1475-7516/2015/02/028) \[[arXiv:1411.5256 \[astro-ph.CO\]](http://arxiv.org/abs/1411.5256)\] \[[[IN]{}SPIRE](https://inspirehep.net/search?p=find+eprint+1411.5256)\].
P. Chingangbam and C. Park, [*Residual foreground contamination in the WMAP data and bias in non-Gaussianity estimation,*]{} [[*JCAP*]{} [**02**]{} 031 (2013)](http://dx.doi.org/10.1088/1475-7516/2013/02/031) \[[arXiv:1210.2250 \[astro-ph.CO\]](http://arxiv.org/abs/1210.2250)\].
W. Zhao and L. Santos, [*Probing the statistical properties of CMB B-mode polarization through Minkowski functionals,*]{} [[*JCAP* ]{} [**07**]{} 029 (2016)](http://iopscience.iop.org/article/10.1088/1475-7516/2016/07/029/meta) \[[arXiv:1510.07779 \[astro-ph.CO\]](http://arxiv.org/abs/1510.07779)\] \[[[IN]{}SPIRE](https://inspirehep.net/search?p=find+eprint+1510.07779)\].
Wen Zhao, [*Probing CMB cold spot through local Minkowski functionals,*]{} [[*Res. Astron. Astrophys.*]{} [**14**]{} 625 (2014)](http://iopscience.iop.org/article/10.1088/1674-4527/14/6/002/meta) \[[arXiv:1209.4021 \[astro-ph.CO\]](http://arxiv.org/abs/1209.4021)\].
D. Munshi, B. Hu, T. Matsubara, P. Coles and A. Heavens, [*Lensing-induced morphology changes in CMB temperature maps in modified gravity theories,*]{} [[*JCAP*]{} [**04**]{} 056 (2016)](http://iopscience.iop.org/article/10.1088/1475-7516/2016/04/056/meta) \[[arXiv:1602.00965v1 \[astro-ph.CO\]](http://arxiv.org/abs/1602.00965)\] \[[[IN]{}SPIRE](https://inspirehep.net/search?p=find+eprint+1602.00965)\].
C. Beisbart, T. Buchert and H. Wagner, [*Morphometry of Spatial Patterns,*]{} [[*Physica A*]{} [**293**]{} (2001) 592-604](http://arxiv.org/ct?url=http%3A%2F%2Fdx.doi.org%2F10%252E1016%2FS0378-4371%252800%252900612-9&v=593ea017) \[[arXiv:astro-ph/0007459](https://arxiv.org/abs/astro-ph/0007459)\] \[[[IN]{}SPIRE](https://inspirehep.net/search?p=find+eprint+ASTRO-PH/0007459)\].
C. Beisbart, R. Valdarnini and T. Buchert, [*The morphological and dynamical evolution of simulated galaxy clusters,*]{} [[*Astronom and Astrophys*]{} [**379**]{} 412-425 (2001)](http://arxiv.org/ct?url=http%3A%2F%2Fdx.doi.org%2F10%252E1051%2F0004-6361%253A20011319&v=eabb3f37) \[[arXiv:astro-ph/0109459](http://arxiv.org/abs/astro-ph/0109459)\] \[[[IN]{}SPIRE](https://inspirehep.net/search?p=find+eprint+ASTRO-PH/0109459)\].
G. Rossi, R. K. Sheth, C. Park, C. Hernandez-Monteagudo, [*Non-Gaussian distribution and clustering of hot and cold pixels in the five-year WMAP sky,*]{} [[*Mon. Not. Roy. Astron. Soc.*]{} [**399**]{} 304 (2009)](http://dx.doi.org/10.1111/j.1365-2966.2009.15278.x) \[[arXiv:0906.2190 \[astro-ph.CO\]](http://arxiv.org/abs/0906.2190)\].
G. Rossi, P. Chingangbam and C. Park, [*Excursion set statistics with primordial non-Gaussianity,*]{} [[*J. Korean Phys. Soc.*]{} [**57**]{} 563 (2010)](http://old.kps.or.kr/jkps/abstract_view.asp?articleuid=B8006122-7F4B-4163-B85D-76D5BE8E712E&globalmenu=3&localmenu=10).
G. Rossi, P. Chingangbam and C. Park, [*Statistics of the excursion sets in models with local primordial non-Gaussianity,*]{} [[*Mon. Not. Roy. Astron. Soc.*]{} [**411**]{} 1880 (2011)](http://dx.doi.org/10.1111/j.1365-2966.2010.17825.x) \[[arXiv:1003.0272 \[astro-ph.CO\]](http://arxiv.org/abs/1003.0272)\].
P. Coles and J. D. Barrow, [*Non-Gaussian statistics and the microwave background radiation,*]{} [[*Mon. Not. Rot. Astron. Soc.*]{} [**228**]{} 407 (1987)](http://dx.doi.org/10.1093/mnras/228.2.407).
P. Chingangbam, C. Park, K. P. Yogendran and R. van de Weygaert, [*Hot and cold spot counts as probes of non-Gaussianity in the cosmic microwave background,*]{} [[*Astrophys. J.*]{} [**755**]{} 122 (2012)](http://dx.doi.org/10.1088/0004-637X/755/2/122) \[[arXiv:1206.0436\[astro-ph.CO\]](https://arxiv.org/abs/1206.0436)\].
C. Park, P. Pranav, P. Chingangbam, R. van de Weygaert, B. Jones, G. Vegter, I. Kim, J. Hidding and W. A. Hellwing, [*Betti numbers of gaussian fields,*]{} [[*JKAS*]{} [**46**]{} 125 (2013)](http://dx.doi.org/10.5303/JKAS.2013.46.3.125) \[[arXiv:1307.2384\[astro-ph.CO\]](https://arxiv.org/abs/1307.2384)\] \[[[IN]{}SPIRE](https://inspirehep.net/search?p=find+eprint+1307.2384)\].
D. Pogosyan, C. Gay and C. Pichon, [*Invariant joint distribution of a stationary random field and its derivatives: Euler characteristic and critical point counts in 2 and 3D,*]{} [[*Phys. Rev. D*]{} [**80**]{} 081301 (2009)](http://dx.doi.org/10.1103/PhysRevD.80.081301) \[[arXiv:0907.1437\[astro-ph.CO\]](http://arxiv.org/abs/0907.1437)\].
P. A. R. Ade et al. \[Planck Collaboration\], [*Planck 2015 results. XVI. Isotropy and statistics of the CMB,*]{} \[[arXiv:1506.07135 \[astro-ph.CO\]](http://arxiv.org/abs/1506.07135)\] \[[[IN]{}SPIRE](https://inspirehep.net/search?p=find+eprint+1506.07135)\].
S. Alesker, [*Description of continuous isometry covariant valuations on convex sets,*]{} [[*Geom. Dedicata*]{} [**74**]{} 241-248 (1999)](https://dx.doi.org/10.1023/A:1005035232264).
D. Hug, R. Schneider and R. Schuster, [*The space of isometry covariant tensor valuations,*]{} [[*Math. J.*]{} [**19**]{} 137-158 (2008)](https://dx.doi.org/10.1090/S1061-0022-07-00990-9).
G.E. Schroder-Turk, S. Kapfer, B. Breidenbach, C. Beisbart, and K. Meche, [*Tensorial Minkowski functionals and anisotropy measures for planar patterns,*]{} [[*J. Microsc.*]{} [**238**]{} 57 (2010)](http://onlinelibrary.wiley.com/doi/10.1111/j.1365-2818.2009.03331.x/full).
C. Beisbart, M. S. Barbosa, H. Wagner and L. da F. Costa, [*Extended morphometric analysis of neuronal cells with Minkowski valuations,*]{} [[*Eur. Phys. J. B*]{} [**52**]{} 531-546](https://dx.doi.org/10.1140/epjb/e2006-00328-1) \[[arXiv:cond-mat/0507648 \[cond-mat.dis-nn\]](http://arxiv.org/abs/cond-mat/0507648)\].
G. Durand, F. Graner and J. Weiss, [*Deformation of grain boundaries in polar ice,*]{} [[*Europhy. Lett.*]{} [**67**]{} 1038-1044](https://dx.doi.org/10.1209/epl/i2004-10139-0) \[[arXiv:cond-mat/0309081 \[cond-mat.mtrl-sci\]](http://arxiv.org/abs/cond-mat/0309081)\].
C. Beisbart, R. Dahlke, K. Mecke and H. Wagner, [*Vector- and tensor- valued descriptors for spacial patterns, Morphology of Condensed Matter - Physics and Geometry of Spatially Complex Systems (ed. by K. Meche and D. Stoyan) Vol. 600 of Lecture Notes in Physics pp. 249-271*]{} \[[arXiv:physics/0203072 \[physics.data-an\]](http://arxiv.org/abs/physics/0203072)\].
P. A. R. Ade et al. \[Planck Collaboration\], [*Planck 2015 results. XIII. Cosmological parameters,*]{} [arXiv:1502.01589 \[astro-ph.CO\]](https://arxiv.org/abs/1502.01589) \[[[IN]{}SPIRE](https://inspirehep.net/search?p=find+eprint+1502.01589)\].
A. Lewis A. Challinor and A. Lasenby, [*Efficient Computation of Cosmic Microwave Background Anisotropies in Closed Friedmann-Robertson-Walker Models,*]{} [[*Astrophys. J D*]{} [**538**]{}, 473 (2000)](http://iopscience.iop.org/article/10.1086/309179/meta) \[[arXiv:astro-ph/9911177](https://arxiv.org/abs/astro-ph/9911177)\] \[[[IN]{}SPIRE](https://inspirehep.net/search?p=find+eprint+ASTRO-PH/9911177)\].
CAMB, <http://camb.info/>.
K. M. Gorski, E. Hivon, A. J. Banday, B. B. Wandelt, F. K. Hansen, M. Reinecke and M. Bartelmann, [*HEALPix: A Framework for High-Resolution Discretization, and Fast Analysis of Data Distributed on the Sphere,*]{} [[*Astrophys. J. *]{} [**622**]{}, 759 (2005)](https://arxiv.org/ct?url=http%3A%2F%2Fdx.doi.org%2F10%252E1086%2F427976&v=fc60ff6a) \[[arXiv:astro-ph/0409513](https://arxiv.org/abs/astro-ph/0409513)\] \[[[IN]{}SPIRE](https://inspirehep.net/search?p=find+eprint+ASTRO-PH/0409513)\].
HEALPIX, <http://healpix.sourceforge.net>.
Planck Legacy Arxive, [http://pla.esac.esa.int/pla/$\#$maps](http://pla.esac.esa.int/pla/#maps).
G.E. Schroder-Turk, W. Mickel, S.C. Kapfer, F.M. Schaller, B. Breidenbach, D. Hyg, and K. Meche, [*Minkowski Tensors of Anisotropic Spatial Structure,*]{} [[*New J. Phys.*]{} [**15**]{} 083028 (2013)](http://dx.doi.org/10.1088/1367-2630/15/8/083028) \[[arXiv:1009.2340 \[cond-mat.soft\]](http://arxiv.org/abs/1009.2340)\].
[^1]: http://healpix.sourceforge.net
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The problem of recovering a matrix of low rank from an incomplete and possibly noisy set of linear measurements arises in a number of areas such as quantum state tomography, machine learning and the PhaseLift approach to phaseless reconstruction problems. In order to derive rigorous recovery results, the measurement map is usually modeled probabilistically and convex optimization approaches including nuclear norm minimization are often used as recovery method. In this article, we derive sufficient conditions on the minimal amount of measurements that ensure recovery via convex optimization. We establish our results via certain properties of the null space of the measurement map. In the setting where the measurements are realized as Frobenius inner products with independent standard Gaussian random matrices we show that $m > 10 r (n_1 + n_2)$ measurements are enough to uniformly and stably recover an $n_1 \times n_2$ matrix of rank at most $r$. Stability is meant both with respect to passing from exactly rank-$r$ matrices to approximately rank-$r$ matrices and with respect to adding noise on the measurements. We then significantly generalize this result by only requiring independent mean-zero, variance one entries with four finite moments at the cost of replacing $10$ by some universal constant. We also study the particular case of recovering Hermitian rank-$r$ matrices from measurement matrices proportional to rank-one projectors. For $r=1$, such a problem reduces to the PhaseLift approach to phaseless recovery, while the case of higher rank is relevant for quantum state tomography. For $m \geq C r n$ rank-one projective measurements onto independent standard Gaussian vectors, we show that nuclear norm minimization uniformly and stably reconstructs Hermitian rank-$r$ matrices with high probability. Subsequently, we partially de-randomize this result by establishing an analogous statement for projectors onto independent elements of a complex projective 4-designs at the cost of a slightly higher sampling rate $m \geq C rn \log n$. Complex projective $t$-designs are discrete sets of vectors whose uniform distribution reproduces the first $t$ moments of the uniform distribution on the sphere. Moreover, if the Hermitian matrix to be recovered is known to be positive semidefinite, then we show that the nuclear norm minimization approach may be replaced by the simpler optimization program of minimizing the $\ell_2$-norm of the residual subject to the positive semidefinite constraint. This has the additional advantage that no estimate of the noise level is required a priori. We discuss applications of such a result in quantum physics and the phase retrieval problem. Apart from the case of independent Gaussian measurements, the analysis exploits Mendelson’s small ball method.'
author:
- 'Maryia Kabanava^1^, Richard Kueng^2,3,4^, Holger Rauhut^1^, Ulrich Terstiege^1^'
title: 'Stable low-rank matrix recovery via null space properties'
---
^1^ Lehrstuhl C für Mathematik (Analysis), RWTH Aachen University, Germany\
^2^ School of Physics, The University of Sydney, Australia\
^3^ Institute for Physics & FDM, University of Freiburg, Germany\
^4^ Institute for Theoretical Physics, University of Cologne, Germany
[**Keywords.**]{} low rank matrix recovery, quantum state tomography, phase retrieval, convex optimization, nuclear norm minimization, positive semidefinite least squares problem, complex projective designs, random measurements
[**MSC 2010.**]{} 94A20, 94A12, 60B20, 90C25, 81P50
Introduction
============
In recent years, the recovery of objects (signals, images, matrices, quantum states etc.) from incomplete linear measurements has gained significant interest. While standard compressive sensing considers the reconstruction of (approximately) sparse vectors [@FoucartRauhut], we study extensions to the recovery of (approximately) low rank matrices from a small number of random measurements. This problem arises in a number of areas such as quantum tomography [@gross_quantum_2010; @flammia_quantum_2012; @gross_focus_2013], signal processing [@ahro15], recommender systems [@care09; @cata10] and phaseless recovery [@candes_phase_2013; @castvo13; @gross_partial_2014; @grkrku14]. On the one hand, we consider both random measurement maps generated by independent random matrices with independent entries and on the other hand, measurements with respect to independent rank one measurements. We derive bounds for the number of required measurements in terms of the matrix dimensions and the rank of the matrix that guarantee successful recovery via nuclear norm minimization. Our results are uniform and stable with respect to noise on the measurements and with respect to passing to approximately rank-$r$ matrices. For rank-one measurements the latter stability result is new.
Let us formally describe our setup. We consider measurements of an (approximately) low-rank matrix $X\in{\mathbb{C}}^{n_1\times n_2}$ of the form $b = {\mathcal{A}}(X)$, where the linear measurement map ${\mathcal{A}}$ is given as $$\label{eq:MeasurementProcess}
{\mathcal{A}}:{\mathbb{C}}^{n_1\times n_2}\to{\mathbb{C}}^m,\quad Z\mapsto\sum_{j=1}^m\operatorname{tr}(ZA_j^*)e_j.$$ Here, $e_1,\ldots,e_m$ denote the standard basis vectors in ${\mathbb{C}}^m$ and $A_1,\ldots,A_m\in{\mathbb{C}}^{n_1\times n_2}$ are called measurement matrices. A prominent approach [@Fazel; @RechtFazelParrilo] for recovering the matrix $X$ from $b = {\mathcal{A}}(X)$ consists in computing the minimizer of the convex optimization problem $$\label{eqnnMinimization}
\underset{Z\in{\mathbb{C}}^{n_1\times n_2}}\min{\Vert Z \Vert}_*\quad\mbox{subject to } {\mathcal{A}}(Z) = b,$$ where ${\Vert Z \Vert}_* = {\Vert Z \Vert}_1 = \sum_{j=1}^n \sigma_j(Z)$ denotes the nuclear norm with $\sigma_j(Z)$ being the singular values of $Z \in {\mathbb{C}}^{n_1 \times n_2}$ and $n = \min\{n_1,n_2\}$. Efficient optimization methods exist for this problem [@bopa14; @bova04]. In practice the measurements are often perturbed by noise, i.e., $$\label{eq:measurements}
b={\mathcal{A}}(X)+w,$$ where $w\in{\mathbb{C}}^m$ is a vector of perturbations. In this case, we replace (\[eqnnMinimization\]) by the noise constrained nuclear norm minimization problem $$\label{eqNNMinimization}
\underset{Z\in{\mathbb{C}}^{n_1\times n_2}}\min{\Vert Z \Vert}_*\quad\mbox{subject to}\;{\Vert {\mathcal{A}}(Z)-b \Vert}_{\ell_2}\leq\eta,$$ where $\eta$ corresponds to a known estimate of the noise level, i.e., $\|w\|_{\ell_2} \leq \eta$ with $\|x\|_{\ell_p} = (\sum_{j} |x_j|^p)^{1/p}$ being the usual $\ell_p$-norm. In some cases it is known a priori that the matrix $X$ of interest is both Hermitian and positive semidefinite ($X \succcurlyeq 0$). Then one may replace by the optimization problem $$\label{eq:postracemin}
\underset{Z \succcurlyeq 0 }\min \operatorname{tr}(Z) \quad\mbox{subject to}\;{\Vert {\mathcal{A}}(Z)-b \Vert}_{\ell_2}\leq\eta.$$ However, as we will see, the simpler least squares problem $$\label{eq:posleastsquares}
\underset{Z \succcurlyeq 0 }\min {\Vert {\mathcal{A}}(Z)-b \Vert}_{\ell_2}$$ works equally well or even better in terms of recovery under certain natural conditions. Apart from simplicity and computational efficiency it has the additional advantage that no estimate $\eta$ of the noise level is required. We note that other efficient recovery methods exist as well [@brle10; @forawa11; @tawe13], but we will not go into details here.
A question of central interest concerns the minimal number $m$ of required measurements that guarantees exact (in the noiseless case) or approximate recovery. While it is very hard to study this question for deterministic measurement maps ${\mathcal{A}}$, several results are available for certain models of random maps. We will study several scenarios which all have in common that the matrices $A_1,\hdots,A_m \in {\mathbb{R}}^{n_1 \times n_2}$ in are independent draws of a random matrix $\Phi = (X_{ij})_{ij}$. We first consider the real-valued case, where all entries $X_{ij}$ are independent and then move to a complex-valued scenario where $\Phi = a a^* \in {\mathbb{C}}^{n \times n}$ is a rank one matrix generated by a random vector $a \in {\mathbb{C}}^n$. For the latter scenario we consider $a$ being a complex Gaussian random vector, or $a$ being randomly drawn from a so-called (approximate) $t$-design. This last setup has implications for quantum tomography and this part of the article can be seen as a continuation of the investigations in [@krt14]. Next, we describe the present state of the art of of the various setups and present our results.
Robust recovery from measurement matrices with independent entries {#sub:rank_one}
------------------------------------------------------------------
We call ${\mathcal{A}}$ a Gaussian measurement map if the matrices $A_1,\hdots,A_m \in {\mathbb{R}}^{n_1 \times n_2}$ in are independent realizations of Gaussian random matrices, i.e., all entries of the $A_j$ are independent standard Gaussian random variables. More generally, ${\mathcal{A}}$ is called subgaussian, if the entries of all the $A_j$ are independent, mean zero, variance one, subgaussian random variables, where we recall that a random variable $\xi$ is called subgaussian if ${\mathbb{P}}(|\xi| \geq t) \leq 2 e^{-c t^2}$ for some constant $c > 0$. If $$\label{bound:m:Gaussian}
m \geq C r (n_1 + n_2)$$ for some universal constant $C>0$, then with probability at least $1- e^{-cm}$ any rank $r$ matrix $X \in {\mathbb{C}}^{n_1 \times n_2}$ is reconstructed exactly from subgaussian measurements $b = {\mathcal{A}}(X)$ via nuclear norm minimization [@RechtFazelParrilo; @CandesPlan]. Moreover, if noisy measurements $b = {\mathcal{A}}(X)+ w$ with $\|w \|_2 \leq \eta$ of an arbitrary matrix $X \in {\mathbb{C}}^{n_1 \times n_2}$ are taken, then the minimizer $X^\sharp$ of satisfies, again with probability at least $1-e^{-cm}$, $$\label{error:estimate}
\|X - X^\sharp\|_F \leq \frac{C'}{\sqrt{r}} \inf_{Z:\operatorname{rank}(Z) \leq r} \|X-Z\|_* + \frac{C'' \eta}{\sqrt{m}},$$ where $\|A\|_F = \sqrt{\operatorname{tr}(A^*A)}$ denotes the Frobenius norm, $\operatorname{tr}$ being the trace. Note that $$\inf_{Z:\operatorname{rank}(Z) \leq r} \|X-Z\|_* = \sum_{j=r+1}^n \sigma_j(X) = \|X_c\|_*,$$ where the singular values $\sigma_j(X)$ are arranged in decreasing order and for $X$ with singular value decomposition $\sum_{j=1}^n \sigma_j(X) u_j v_j^*$ the matrix $X_c = \sum_{j=r+1}^n \sigma_j(X) u_j v_j^*$. The error estimate means that reconstruction is robust with respect to noise on the measurements and stable with respect to passing to only approximately low rank matrices. These statements are uniform in the sense that they hold for all matrices $X$ simultaneously once the matrix $A$ has been drawn. They have been established in [@CandesPlan; @MohanFazel; @RechtFazelParrilo] via the rank restricted isometry property (rank-RIP), see e.g. [@FoucartRauhut] for the standard RIP and its implications.
While the RIP is a standard tool by now, recovery of low rank matrices via nuclear norm minimization is characterized by the so-called null space property [@MohanFazel2; @RechtXuHassibiArticle; @RechtXuHassibiProceedings; @FoucartRauhut; @forawa11], see below for details. By using this concept, we are able to significantly relax from subgaussian distributions of the entries to distributions with only four finite moments.
\[th:indepNSP\] Let ${\mathcal{A}}:{\mathbb{R}}^{n_1\times n_2}\to{\mathbb{R}}^m$, ${\mathcal{A}}(X) = \sum_{j=1}^n \operatorname{tr}(XA_j) e_j$, where the $A_j$ are independent copies of a random matrix $\Phi=(X_{ij})_{i,j}$ with independent mean zero entries obeying $\mathbb E X_{ij}^2 = 1$ and $$\mathbb E X_{ij}^4\leq C_4 \quad \mbox{ for all } i,j \mbox{ and some constant } C_4.$$ Fix $1\leq r\leq\min\{n_1,n_2\}$ and $0<\rho<1$ and set $$m\geq c_1\rho^{-2}r(n_1 + n_2).$$ Then with probability at least $1-e^{-c_2m}$, for any $X\in{\mathbb{R}}^{n_1\times n_2}$ the solution $X^{\sharp}$ of (\[eqNNMinimization\]) with $b={\mathcal{A}}(X)+w$, ${\Vert w \Vert}_{\ell_2}\leq\eta$, approximates $X$ with error $$\label{eq:ErrorEstimate}
{\Vert X-X^{\sharp} \Vert}_F\leq \frac{2(1+\rho)^2}{(1-\rho)\sqrt r}
\|X_c\|_* +\frac{(3+\rho)}{(1-\rho)c_3}\cdot\frac{\eta}{\sqrt m}.$$ Here $c_1,c_2,c_3$ are positive constants that only depend on $C_4$.
In the special case, when $\Phi$ has independent standard Gaussian entries, we apply Gordon’s escape through a mesh theorem [@Gordon] in order to obtain an explicit constant in the estimate for the number of measurements, see Theorem \[th:GaussianMeas\]. Roughly speaking, with high probability, any $n_1\times n_2$ matrix of rank $r$ is stably recovered from $m>10r(n_1 + n_2)$ Gaussian measurements. We remark that the explicit bound $m > 3 r (n_1 + n_2)$ has been derived in [@chparewi10], (see also [@mctr14] and [@amlomctr13 Section 4.4] for a phase transition result in this context), but this bound considers nonuiform recovery, i.e. recovery of a fixed low rank matrix with a random draw of a Gaussian measurement matrix with high probability. Moreover, no stability under passing to approximately low rank matrices has been considered there. Our recovery result is therefore stronger than the one in [@chparewi10], but requires more measurements.
Robust recovery of Hermitian matrices from rank-one projective measurements {#sub:rank_one_measurements}
---------------------------------------------------------------------------
Let us now focus on the particular case of recovering complex Hermitian $n \times n$ matrices from noisy measurements of the form , where the measurement matrices are proportional to rank-one projectors, i.e., $$\begin{aligned}
A_j = a_j a_j^* \in \mathcal{H}_n \label{eq:rank_one_measurements}\end{aligned}$$ where $a_j \in \mathbb{C}^n$. Here, $\mathcal{H}_n$ denotes the space of complex Hermitian $n \times n$ matrices, which has real dimension $n^2$. Measurements of that type occur naturally in convex relaxations of the phase retrieval problem [@candes_phase_2013; @castvo13; @gross_partial_2014; @grkrku14]. In fact, suppose phaseless measurements of the form $b_j = |\langle x, a_j\rangle|^2$ of a vector $x \in \mathbb{C}^n$ are given. Then we can rewrite $b_j = \operatorname{tr}(x x^*a_j a_j^* ) = \operatorname{tr}(X A_j)$ as linear measurements of the rank one matrix $X = x x^*$. We will expand on this aspect below in Section \[sub:phase\_retrieval\]. Rank one measurements of low rank matrices feature prominently in quantum state tomography as well, see also below.
The prior information that the desired matrix is Hermitian limits the search space in the convex optimization problem and it simplifies to $$\label{eqNNMinimizationHerm}
\underset{Z\in\mathcal H_n}\min{\Vert Z \Vert}_*\quad\mbox{subject to}\;{\Vert {\mathcal{A}}(Z)-b \Vert}_{\ell_2}\leq\eta.$$
Arguably, the most generic measurement matrices of the form result from choosing each $a_j$ to be an independent complex standard Gaussian vector. For the particular case of phase retrieval — i.e., where the matrix of interest $X = x x^*$ is itself proportional to a rank-one projector — uniform recovery guarantees by means of have been established for $m = Cn$ independent measurements in [@candes_solving_2012]. Recently, this result has been generalized to recovery of any Hermitian rank $r$-matrix by means of $m = Crn$ such measurements in [@krt14]. Our refined analysis of the null space property enables us to further strengthen this result by additionally guaranteeing stability under passing to approximately low rank matrices:
\[mainTh1\] Consider the measurement process described in (\[eq:MeasurementProcess\]) with $m$ measurement matrices of the form ,where each $a_i$ is an independent complex standard Gaussian vector. Fix $r\leq n$, $0<\rho<1$ and suppose that $$m\geq C_1\rho^{-2}nr.$$ Then with probability at least $1 - \mathrm{e}^{-C_2 m}$ it holds that for any $X \in \mathcal H_n$, any solution $X^\sharp$ to the convex optimization problem (\[eqNNMinimizationHerm\]) with noisy measurements $b = \mathcal A (X)+\epsilon$, where $\| \epsilon \|_{\ell_2} \leq \eta$, obeys $$\label{err:bound1}
\| X - X^\sharp \|_F \leq \frac{2(1+\rho)^2}{(1-\rho)\sqrt r}{\Vert X_c \Vert}_*+\frac{(3+\rho)C_3}{(1-\rho)}\cdot\frac{\eta}{\sqrt m}.$$ Here, $C_1,C_2$ and $C_3$ denote positive universal constants. (In particular, for $\eta=0$ and $X$ of rank at most $r$ one has exact reconstruction.)
In addition to the Gaussian measurement setting, we also consider measurement matrices that arise from taking the outer product of elements chosen independently from an approximate complex projective 4-design. Complex projective $t$-designs are finite sets of unit vectors in $\mathbb{C}^n$ that exhibit a very particular structure. Roughly speaking, sampling independently from a complex projective $t$-design, reproduces the first $t$ moments of sampling uniformly from the complex unit sphere. Likewise, approximate complex projective $t$-designs obey such a structural requirement approximately — for a precise introduction, we refer to Definition \[def:approx\_design\] below. As a consequence, they serve as a general purpose tool for partially de-randomizing results that initially required Gaussian random vectors [@kueng_spherical_2015; @gross_partial_2014]. This is also the case here and employing complex projective $4$-designs allows for partially de-randomizing Theorem \[mainTh1\] at the cost of a slightly larger sampling rate. Here, we content ourselves with presenting and shortened version of this result and refer the reader to Theorem \[Th2\] where precise requirements on the approximate design are stated.
\[mainTh2\] Let $r,\rho$ be as in Theorem \[mainTh1\] and suppose that each measurement matrix $A_j$ is of the form , where $a_j$, $j=1,\hdots,m$, are chosen independently from a (sufficiently accurate approximate) complex projective 4-design. If $$m \geq C_4 \rho^{-2} nr\log n,$$ then the assertions of Theorem \[mainTh1\] remains valid, possibly with different universal constants.
Note that Theorems \[th:indepNSP\], \[mainTh1\], \[mainTh2\] resp. Theorem \[th:GaussianMeas\] below and their proofs are presented in condensed versions in the conference papers [@KRTSampta1] resp. [@KRTSampta2].
Recovery of positive semidefinite matrices reduces to a feasibility problem
---------------------------------------------------------------------------
Imposing additional structure on the matrices to be recovered can further strengthen low rank recovery guarantees. Positive semidefiniteness is one such structural prerequisite that, for instance, occurs naturally in the phase retrieval problem, quantum mechanics and kernel-based learning methods [@scholkopf_learning_2002]. Motivated by the former, Demanet and Hand [@demanet_stable_2014] pointed out that minimizing the nuclear norm — in the sense of algorithm — can be superfluous for recovering positive semidefinite matrices of rank one. Instead, they propose to reduce the recovery algorithm to a mere feasibility problem and proved that such a reduction works w.h.p. for rank one projective measurements onto Gaussian vectors (the measurement scenario considered in Theorem \[mainTh1\]). Subsequently, this recovery guarantee was strengthened by Candès and Li [@candes_solving_2012]. Here, we go one step further and generalize these results to cover uniform and stable recovery of positive semidefinite matrices of arbitrary rank. Relying on ideas presented in [@kalev_informationally_2015], we establish the following statement. (We refer to Section \[Notation\] for the definition of the Schatten $p$-norm ${\Vert \cdot \Vert}_p$ used in .)
\[mainTh3\] Fix $r \leq n$ and consider the measurement processes introduced in Theorem \[mainTh1\] (Gaussian vectors), or Theorem \[mainTh2\] (complex projective 4-designs), respectively. Assume that $m\geq C_1 n r $ (in the Gaussian case) resp. $m\geq C_2 s nr \log n $ (in the design case), where $s\geq 1$ is arbitrary. Then, for $1 \leq p \leq 2$ and any two positive semidefinite matrices $X,Z \in \mathcal{H}_n$, $$\left\| Z - X \right\|_p \leq \frac{C_3}{r^{1-1/p}} {\Vert X_c \Vert}_1 + \frac{C_4 r^{1/p-1/2}}{\sqrt{m}} \left\| \mathcal{A}(Z) - \mathcal{A}(X) \right\|_{\ell_2}
\label{eq:mainTh3}$$ holds universally with probability exceeding $ 1 - \mathrm{e}^{-C_5 m}$ for the Gaussian case and $1- \mathrm{e}^{- s r}$ in the design case. Here, $C_1,\hdots, C_5$ denote suitable positive universal constants.
This statement renders nuclear norm minimization in the sense of redundant and allows for a regularization-free estimation. Moreover, knowledge of a noise bound $\| w \|_{\ell_2} \leq \eta$ for the measurement process is no longer required, since we can estimate any $X \succcurlyeq 0$ by solving a least squares problem of the form , i.e., $$\min_{Z \in \mathcal{H}_n} \left\| \mathcal{A} (Z) - b \right\|_{\ell_2} \quad \textrm{subject to} \quad Z \succcurlyeq 0. \label{eq:least_squares}$$ Theorem \[mainTh3\] then in particular assures that the minimizer $Z^\sharp$ of this optimization program obeys $$\| Z^\sharp - X \|_F \leq
\frac{C_3}{\sqrt{r}} \| X_c \|_1 + \frac{C_4}{\sqrt{m}} \left\| \mathcal{A}(Z^\sharp) - \mathcal{A}(X) \right\|_{\ell_2}
\leq \frac{C_3}{\sqrt{r}} \| X_c \|_1 + \frac{2C_4}{\sqrt{m}} \| w \|_{\ell_2},$$ where $w \in \mathbb{R}^m$ represents additive noise in the measurement process. It is worthwhile to mention that if a matrix $X$ of interest has rank at most $r$ and no noise is present in the sampling process , Theorem \[mainTh3\] assures $$\left\{ Z: \; Z \succcurlyeq 0, \; \mathcal{A}(Z) = \mathcal{A}(X) \right\} = \left\{ X \right\} \label{eq:feasibility_problem}$$ with high probability. Hence, recovering $X$ from noiseless measurements indeed reduces to a feasibility problem.
We emphasize that Theorem \[mainTh3\] is only established for rank one projective measurements. For the other measurement ensembles considered here — matrices with independent entries — one cannot expect such a statement to hold. This pessimistic prediction is due to negative results recently established in [@slawski_regularization_2015 Proposition 2]. Focusing on real matrices, the authors show that if the measurement matrices $A_j$ are chosen independently from a Gaussian orthogonal ensemble, then estimating any symmetric, positive semidefinite matrix $X$ via becomes ill-posed, unless the number of measurements obeys $$m \geq \frac{1}{4}n(n+1) = \mathcal{O}(n^2).$$
Finally, we want to point out that the fruitfulness of plain least squares regression for recovering positive semidefinite matrices was already pointed out and explored by Slawski, Li and Hein [@slawski_regularization_2015]. However, there is a crucial difference in the mindset of [@slawski_regularization_2015] and the results presented here. The main result [@slawski_regularization_2015 Theorem 2] of Slawski et al. assumes a fixed signal $X \succcurlyeq 0$ of interest and provides bounds for the reconstruction error in terms of geometric properties of both $X$ and the measurement ensemble. Conversely, Theorem \[mainTh3\] assumes fixed measurements (e.g. $m = C rn$ projectors onto Gaussian random vectors) and w.h.p. assures robust recovery of all matrices $X \succcurlyeq 0$ having approximately rank-$r$ simultaneously.
Notation
--------
[\[Notation\]]{} The Schatten $p$-norm of $Z\in{\mathbb{C}}^{n_1\times n_2}$ is given by $${\Vert Z \Vert}_p = {\left(\sum_{j=1}^n\sigma_j(Z)^p\right)}^{1/p},\quad p\geq 1,$$ where $\sigma_j(Z)$, $j=1,\ldots,n$, denote the singular values of $Z$. It reduces to the nuclear norm ${\Vert \cdot \Vert}_{*}$ for $p=1$ and the Frobenius norm ${\Vert \cdot \Vert}_F$ for $p=2$. It is a common convention that the singular values of $Z$ are non-increasingly ordered. We write $Z=Z_r+Z_c$, where $Z_r$ is the best rank-$r$ approximation of $Z$ with respect to any Schatten $p$-norm of $Z$.
Applications
============
Phase retrieval {#sub:phase_retrieval}
---------------
The problem of retrieving a complex signal $x \in \mathbb{C}^n$ from measurements that are ignorant towards phase information has long been abundant in many areas of science. Measurements of that type correspond to $$b_i = \left| \langle a_i, x \rangle \right|^2 + w_i\quad i=1,\ldots,m, \label{eq:phaseless_measurements}$$ where $a_1,\ldots,a_m \in \mathbb{C}^n$ are measurement vectors and $w_i$ denotes additive noise. Recently, the problem’s mathematical structure has received considerable attention in its own right. It is clearly ill-posed, since all phase information is lost in the measurement process and, moreover, the measurements are of a non-linear nature. This second obstacle can be overcome by a trick [@balan_painless_2009] well known in conic programming: the quadratic expressions are linear in the outer products $x x^*$ and $a_i a_i^*$: $$b_i = \left| \langle a_i, x \rangle \right|^2 + w_i = \operatorname{tr}\left( \left(a_i a_i \right)^* \left( x x^* \right) \right) + w_i. \label{eq:lifting}$$ Note that such a “lift” allows for reinterpreting the phase-less sampling process as $\mathcal{A}(x x^*) = b + w$. Also, the new object of interest $X := x x^*$ is an Hermitian, positive semidefinite matrix of rank one. In turn, the measurement matrices $A_i = a_i a_i^*$ are constrained to be proportional to rank-one projectors. Consequently, such a “lift” turns the phase retrieval problem into a very particular instance of low rank matrix recovery — a fact that was first observed by Candès, Eldar, Strohmer and Voroninski [@candes_phase_2013; @castvo13]. Subsequently, uniform recovery guarantees for $m = C n $ complex standard Gaussian measurement vectors $a_i$ have been established which are stable towards additive noise. The main result in [@candes_solving_2012] establishes with high probability that for any $X = x x^*$, solving the convex optimization problem (PhaseLift) $$\begin{aligned}
\underset{Z \in {\mathcal{H}}_n}{\operatorname{min}}\; \| \mathcal{A}(Z) - b \|_{\ell_1} \quad \textrm{subject to} \quad Z \succcurlyeq 0 \label{eq:PhaseLift}\end{aligned}$$ yields an estimator $Z^\sharp$ obeying $\| Z^\sharp - x x^* \|_2 \leq C \| w \|_1 / m$. If a bound $\| w \|_{\ell_2} \leq \eta$ on the noise in the sampling process is available, an extension of [@krt14 Theorem 2] (see section 2.3.2 in loc. cit) establishes a comparable recovery guarantee via solving $$\begin{aligned}
\underset{Z \in {\mathcal{H}}_n}{\operatorname{min}}\; \operatorname{tr}(Z) \quad \textrm{subject to} \quad \| \mathcal{A}(Z) - b \|_{\ell_2} \leq \eta, \; Z \succcurlyeq 0 \label{eq:krt_phaselift}\end{aligned}$$ instead of PhaseLift. Our findings allow for establishing novel recovery guarantees for retrieving phases. Indeed, since assures that any signal of interest is positive semidefinite and has precisely rank one, Theorem \[mainTh3\] is applicable and yields the following corollary.
\[cor:phaselift\] Consider $m \geq C n$ phaseless measurements of the form , where each $a_i$ is a complex standard Gaussian vector. Then with probability at least $1 - \mathrm{e}^{-C' m}$ these measurements allow for estimating any signal $x \in \mathbb{C}^n$ via solving $$\underset{Z \in {\mathcal{H}}_n}{\operatorname{min}}\; \| \mathcal{A}(Z) - b \|_{\ell_2} \quad \textrm{subject to} \quad Z \succcurlyeq 0. \label{eq:our_phaselift}$$ The resulting minimizer $Z^\sharp$ of obeys $$\| Z^\sharp - x x^* \|_{\ell_2}
\leq \frac{ C \| w \|_{\ell_2}}{\sqrt{m}},$$ where $C$ denotes a positive constant and $w \in \mathbb{R}^m$ represents additive noise in the sampling process .
An analogous statement is true — with a weaker probability of success $1 - \mathrm{e}^{-s}$ for $s\geq 1$ — for $m \geq C' s n \log (n)$ rank one projective measurements onto independent elements of an approximate 4-design.
This recovery procedure is in spirit very similar to , but it utilizes an $\ell_2$-regression instead of an $\ell_1$-norm minimization. Numerical studies indicate that algorithm outperforms as well as . These studies were motivated and accompany actual quantum mechanical experiments and will be published elsewhere [@kueng_optics_2015].
Finally, we want to relate Corollary \[cor:phaselift\] to a non-convex phaseless recovery procedure devised by Candès, Li and Soltanolkotabi [@candes_wirtinger_2015]. There, the authors refrain from applying the aforementioned “lifting” trick to render the phase retrieval problem linear. Instead, they use a careful initialization step, followed by a gradient descent scheme (based on Wirtinger derivatives) to minimize the problem’s least squares loss function directly over complex vectors $z \in \mathbb{C}^n$. Mathematically, such an optimization is equivalent to solving $$\underset{Z \in {\mathcal{H}}_n}{\textrm{min}} \; \| \mathcal{A} (Z) - b \|_{\ell_2}
\quad \textrm{subject to} \quad Z \succcurlyeq 0, \; \mathrm{rank}(Z) = 1 \label{eq:wirtinger}$$ and the rank-constraint manifests the problem’s non-convex nature. Hence, the convex optimization problem can be viewed as a convex relaxation of , obtained by omitting the non-convex rank constraint.
Quantum information {#sub:quantum}
-------------------
In this section we describe implications and possible applications of our findings to problems in quantum information science. For the sake of being self-contained, we have included a brief introduction to crucial notions of quantum mechanics in the appendix. Quantum mechanics postulates that a finite $n$-dimensional quantum system is described by an Hermitian, positive semidefinite matrix $X$ with unit trace, called a *density operator*. This “quantum shape constraint” assures that all density operators meet the requirements of Theorem \[mainTh3\]. Furthermore, the rank-one projective measurements assumed in that theorem can be recast as valid quantum mechanical measurements — see [@krt14 Section 3] for possible implementations and further discussion on this topic. Note, however, that such a reinterpretation is in general not possible for the measurement matrices with independent entries considered in Theorem \[th:indepNSP\], because these matrices fail to be Hermitian. With Theorem \[mainTh3\] at hand, we underline its implications for two prominent issues in (finite dimensional) quantum mechanics.
### Quantum state tomography
Inferring a quantum mechanical description of a physical system is equivalent to assigning it a *density operator* (or quantum state) — a process referred to as *quantum state tomography* [@gross_focus_2013; @ferrie_have_2015]. Tomography is now a routine task for designing, testing and tuning qubits in the quest of building quantum information processing devices. Since the size of controllable quantum mechanical systems is ever increasing[^1] it is very desirable to exploit additional structure — if present — when performing such a task. One such structural property — often encountered in actual experiments — is *approximate purity*, i.e., the density operator $X$ is well approximated by a low rank matrix. Performing quantum state tomography under such a prior assumption therefore constitutes a particular instance of low rank matrix recovery [@gross_quantum_2010; @flammia_quantum_2012].
The results presented in this paper provide recovery guarantees for tomography protocols that stably tolerate noisy measurements and moreover are *robust* towards the prior assumption of approximate purity. In the context of tomography, results of this type so far have already been established for $m = Cnr \log^6n$ random (generalized) Pauli measurements [@liu_universal_2011 Proposition 2.3] via proving a rank-RIP for such measurement matrices and then resorting to [@CandesPlan Lemma 3.2]. However, this auxiliary result manifestly requires additive Gaussian noise and using a type of Dantzig, or Lasso selector to recover the best rank-$r$ approximation of a given density operator. This is not the case for the result established here, where performing a plain least squares regression of the form is sufficient.
\[cor:tomography\] Fix $r \leq n$ and suppose that the measurement operator $\mathcal{A}: \; \mathcal{H}_n \to \mathbb{R}^m$ is of the form $$\mathcal{A} (X) = \sum_{i=1}^m \sqrt{\frac{(n+1)n}{m}} \langle a_i, X a_i \rangle e_i + w \in \mathbb{R}^m \quad \textrm{with} \quad m \geq C_1 r n \log n,$$ where each $a_i \in \mathbb{C}^n$ is chosen independently from an approximate 4-design and $w \in \mathbb{R}^m$ denotes additive noise. Then, the best rank-$r$ approximation of any density operator $X$ can be obtained from such measurements via solving $$\min_{Z \in \mathcal{H}_n} \left\| \mathcal{A} (Z) - \mathcal{A} (X) \right\|_{\ell_2}
\quad \textrm{subject to} \quad Z \succcurlyeq 0, \; \operatorname{tr}\left( Z \right) = 1.
\label{eq:least_squares2}$$ With probability at least $1 - \mathrm{e}^{-C_2 m}$, the minimizer $Z^\sharp$ of this optimization obeys $$\| X - Z^\sharp \|_1
\leq C_3\| X_c \|_1 + C_4 \sqrt{r} \| w \|_{\ell_2}, \label{eq:tomography_bound}$$ where $C_1,C_2,C_3$ and $C_4$ denote positive constants.
This statement is a direct consequence of Theorem \[mainTh3\]. For the sake of clarity, we have re-scaled each projective measurement with $\sqrt{\frac{(n+1)n}{m}}$. This simplifies the resulting expression and moreover facilitates[^2] direct comparison with the main result in [@liu_universal_2011], as it closely mimics the scaling employed there.
Corollary \[cor:tomography\] is valid for any type of additive noise and no a priori knowledge of its magnitude is required. This includes the particularly relevant case of a Bernoulli error model — see e.g. [@carpentier_uncertainty_2015 Section 2.2.2] and also [@flammia_quantum_2012] — which is particularly relevant for tomography experiments. Also, note that the recovery error is bounded in nuclear norm, instead of Frobenius norm. Such a bound is very meaningful for tomography, since quantum mechanics is a probabilistic theory and the nuclear norm encapsulates total variational distance. Moreover, Helstrom’s theorem [@helstrom_quantum_1969] provides an operational interpretation of the nuclear norm distance bounded in : it is proportional to the maximal bias achievable in the task of distinguishing the two quantum states $X$ and $Z^\sharp$, provided that any physical measurement can be implemented.
Finally, note that the bound on the probability of failure in Corollary \[cor:tomography\] is much stronger than the one provided in Theorem \[mainTh3\]. Such a strengthening is possible, because the trace of any density operator equals one. We comment on this in Remark \[rem:quantum\_improvement\] below.
### Distinguishing quantum states
One crucial prerequisite in the task of inferring density operators from measurement data, is the ability to faithfully distinguish any two density operators via quantum mechanical measurements. The most general notion of a quantum measurement is a *positive operator valued measure* (POVM) $\mathcal{M} = \left\{ E_m: E_m \succcurlyeq 0, \sum_{m} E_m = \operatorname{id}\right\}$ [@nielsen_quantum_2010 Chapter 2.2]. A POVM $\mathcal{M}$ is called *informationally complete* (IC) [@scott_tight_2006] if for any two density operators $X \neq Z \in \mathcal{H}_n$ there exists $E_m \in \mathcal{M} \subseteq \mathcal{H}_n$ such that $$\operatorname{tr}\left( E_m X \right) \neq \operatorname{tr}\left( E_m Z \right). \label{eq:IC}$$ This assures the possibility of discriminating any two quantum states via such a measurement in the absence of noise. Without additional restrictions, such an IC POVM must contain at least $n^2$ elements. However, such a lower bound can be too pessimistic, if the density operators of interest have additional structure. Approximate purity introduced in the previous subsection can serve as such an additional structural restriction:
For $ r \leq n$, we call a POVM $\mathcal{M}=\left\{ E_m \right\}_{m \in I}$ *rank-$r$ restricted informationally complete* (rank-$r$ IC), if holds for any two density operators of rank at most $r$.
Bounds for the number $m$ of POVM elements required to assure rank-$r$-IC have been established in [@heinosaari_quantum_2013; @kech_role_2015; @kech_quantum_2015]. These approaches exploit topological obstructions of embeddings for establishing lower bounds and explicit POVM constructions for upper bounds. For instance, in [@heinosaari_quantum_2013] a particular rank-$r$-IC POVM containing $m = 4r(n-r)-1$ elements is constructed.
Focusing less on establishing tight bounds and more on identifying entire families of rank-$r$ IC measurements, Kalev et al. [@kalev_informationally_2015] observed that each measurement ensemble fulfilling the rank-RIP for some $r \leq n$ is also rank-$r$ IC. This in particular applies with high probability to $m = C \log^6 n \; n r$ random (generalized) Pauli measurements [@liu_universal_2011]. Theorem \[mainTh3\], and likewise Corollary \[cor:tomography\], allow us to draw similar conclusions without having to rely on any rank-RIP. Indeed, in the absence of noise, these results guarantee for any rank-$r$ density operator $X$ $$\left\{ Z: \; Z \succcurlyeq 0, \; \mathcal{A}(Z) = \mathcal{A} (X) \right\} = \left\{ X \right\} \label{eq:injectivity}$$ with high probability. If this is the case, the measurement operator $\mathcal{A}$ allows for uniquely identifying any rank-$r$ density operator $X$. This in turn implies that $\mathcal{A}$ is rank-$r$ IC and the following corollary is immediate:
Fix $r \leq n$ arbitrary and let $C,C'$ be absolute constants of sufficient size. Then
1. Any POVM containing $m = C nr$ projectors onto Haar[^3] random vectors is rank-$r$ IC with probability at least $1-\mathrm{e}^{C_2 m}$.
2. Any POVM containing $m = C' nr\log n$ projectors onto random elements of a (sufficiently accurate approximate) $4$-design is rank-$r$ IC with probability at least $1- \mathrm{e}^{-\tilde{C}_2 m}$.
This statement is reminiscent of a conclusion drawn in [@ambainis_quantum_2007; @matthews_distinguishability_2009]: In the task of distinguishing quantum states, a POVM containing a 4-design essentially performs as good as as the uniform POVM (the union of all rank-one projectors).
In the process of finishing this article we became aware of recent work by Kech and Wolf [@kewo15], who showed that the elements of a generic Parseval frame generate a rank-$r$ IC map $\mathcal{A}$ if $m \geq 4r(n-r)$. In fact, Xu showed in [@xu15] that $m \geq 4r(n-r)$ is both a sufficient and necessary condition for identifiability of complex rank $r$ matrices in ${\mathbb{C}}^{n \times n}$. We emphasize, however, that these results are only concerned with pure identifiability and do not come with a practical and stable recovery algorithm.
The null space property for low-rank matrix recovery {#sec:NSP}
====================================================
Let $X\in{\mathbb{C}}^{n_1\times n_2}$. If $X$ is only approximately of low-rank, then we would like to find a condition on the measurement map ${\mathcal{A}}:{\mathbb{C}}^{n_1\times n_2}\to{\mathbb{C}}^m$ that provides the control of the recovery error by the error of its best approximation by low rank matrices. Moreover, it should also take into account that the measurements might be noisy.
We say that ${\mathcal{A}}:{\mathbb{C}}^{n_1\times n_2}\to{\mathbb{C}}^m$ satisfies the Frobenius robust rank null space property of order $r$ with constants $0<\rho<1$ and $\tau>0$ if for all $M\in{\mathbb{C}}^{n_1\times n_2}$, the singular values of $M$ satisfy $${\Vert M_r \Vert}_2\leq\frac{\rho}{\sqrt r}{\Vert M_c \Vert}_1+\tau{\Vert {\mathcal{A}}(M) \Vert}_{\ell_2}.$$
The stability and robustness of (\[eqNNMinimization\]) are established by the following theorem.
\[th:FrobeniusNSP\] Let ${\mathcal{A}}:{\mathbb{C}}^{n_1\times n_2}\to{\mathbb{C}}^m$ satisfy the Frobenius robust rank null space property of order $r$ with constants $0<\rho<1$ and $\tau>0$. Let $n=\min\{n_1,n_2\}$. Then for any $X\in{\mathbb{C}}^{n_1\times n_2}$ any solution $X^{\sharp}$ of (\[eqNNMinimization\]) with $b={\mathcal{A}}(X)+w$, ${\Vert w \Vert}_{\ell_2}\leq\eta$, approximates $X$ with error $${\Vert X-X^{\sharp} \Vert}_2\leq \frac{2(1+\rho)^2}{(1-\rho)\sqrt r}{\Vert X_c \Vert}_1+\frac{2\tau(3+\rho)}{1-\rho}\eta.$$
Theorem \[th:FrobeniusNSP\] can be deduced from the following stronger result.
\[th:DifferenceBetweenSignalAndFeasibleElements\] Let $1\leq p\leq 2$ and $n=\min\{n_1,n_2\}$. Suppose that ${\mathcal{A}}:{\mathbb{C}}^{n_1\times n_2}\to{\mathbb{C}}^m$ satisfies the Frobenius robust rank null space property of order $r$ with constants $0<\rho<1$ and $\tau>0$. Then for any $X,Z\in{\mathbb{C}}^{n_1\times n_2}$, $$\label{eq:DifferenceBetweenSignalAndFeasibleElements}
{\Vert Z-X \Vert}_p\leq\frac{(1+\rho)^2}{(1-\rho) r^{1-1/p}}{\left({\Vert Z \Vert}_1-{\Vert X \Vert}_1+2{\Vert X_c \Vert}_1\right)}+\frac{\tau(3+\rho)}{1-\rho}r^{1/p-1/2}{\Vert {\mathcal{A}}(Z-X) \Vert}_{\ell_2}.$$
The proof requires some auxiliary lemmas. We start with a matrix version of Stechkin’s bound.
\[lm:2NormNuclearNorm\] Let $M\in{\mathbb{C}}^{n_1\times n_2}$ and $r\leq\min\{n_1,n_2\}$. Then, for $p>0$, $${\Vert M_c \Vert}_p\leq \frac{{\Vert M \Vert}_1}{r^{1-1/p}}.$$
This follows immediately from [@FoucartRauhut Proposition 2.3], but for convenience we give the proof. Since the singular values of $M$ are non-increasingly ordered, it holds $$\begin{aligned}
{\Vert M_c \Vert}_p^p=\sum_{j=r+1}^n(\sigma_j(M))^p&\leq(\sigma_r(M))^{p-1}\sum_{j=r+1}^n\sigma_j(M)\leq{\left[\frac{1}{r}\sum_{j=1}^r\sigma_j(M)\right]}^{p-1}\sum_{j=r+1}^n\sigma_j(M)\\
&\leq\frac{1}{r^{p-1}}{\Vert M \Vert}_1^{p-1}{\Vert M \Vert}_1=\frac{{\Vert M \Vert}_1^p}{r^{p-1}}. \qquad \qquad
\end{aligned}$$
The next result shows that under the Frobenius robust rank null space property the distance between two matrices is controlled by the difference between their norms and the $\ell_2$-norm of the difference between their measurements.
\[lm:BestRApproxOfDifference\] Suppose that ${\mathcal{A}}:{\mathbb{C}}^{n_1\times n_2}\to{\mathbb{C}}^m$ satisfies the Frobenius robust rank null space property of order $r$ with constants $0<\rho<1$ and $\tau>0$. Let $X,Z\in{\mathbb{C}}^{n_1\times n_2}$ and $n=\min\{n_1,n_2\}$. Then $${\Vert X-Z \Vert}_1\leq \frac{1+\rho}{1-\rho}{\left({\Vert Z \Vert}_1-{\Vert X \Vert}_1+2{\Vert X_c \Vert}_1\right)}+\frac{2\tau\sqrt r}{1-\rho}{\Vert {\mathcal{A}}(X-Z) \Vert}_{\ell_2}.$$
Theorem 7.4.9.1 in [@HornJohnson] states that for matrices $A,B$ of the same size over ${\mathbb{C}}$ $${\Vert A-B \Vert}\geq {\Vert \Sigma(A)-\Sigma(B) \Vert},$$ where ${\Vert \cdot \Vert}$ is any unitarily invariant norm and $\Sigma(\cdot)$ denotes the diagonal matrix of singular values of its argument. Hence, $$\begin{aligned}
{\Vert Z \Vert}_1&={\Vert X-(X-Z) \Vert}_1\geq\sum_{j=1}^n{\left| \sigma_j(X)-\sigma_j(X-Z) \right|}\\
&=\sum_{j=1}^r{\left| \sigma_j(X)-\sigma_j(X-Z) \right|}+\sum_{j=r+1}^n{\left| \sigma_j(X)-\sigma_j(X-Z) \right|}\\
&\geq \sum_{j=1}^r{\left(\sigma_j(X)-\sigma_j(X-Z)\right)}+\sum_{j=r+1}^n{\left(\sigma_j(X-Z)-\sigma_j(X)\right)}.
\end{aligned}$$ Hence, $$\begin{aligned}
{\Vert (X-Z)_c \Vert}_1=\sum_{j=r+1}^n\sigma_j(X-Z)&\leq{\Vert Z \Vert}_1- \sum_{j=1}^r\sigma_j(X)+ \sum_{j=1}^r\sigma_j(X-Z)+ {\Vert X_c \Vert}_1\\
&\leq{\Vert Z \Vert}_1-{\Vert X \Vert}_1+ \sqrt r{\Vert (X-Z)_r \Vert}_2+2{\Vert X_c \Vert}_1.
\end{aligned}$$ Applying the Frobenius robust null space property of ${\mathcal{A}}$ we obtain $${\Vert (X-Z)_c \Vert}_1\leq{\Vert Z \Vert}_1-{\Vert X \Vert}_1+\rho{\Vert (X-Z)_c \Vert}_1+\tau\sqrt r{\Vert {\mathcal{A}}(X-Z) \Vert}_{\ell_2}+2{\Vert X_c \Vert}_1.$$ By rearranging the terms in the above inequality we obtain $${\Vert (X-Z)_c \Vert}_1\leq\frac{1}{1-\rho}{\left({\Vert Z \Vert}_1-{\Vert X \Vert}_1+\tau\sqrt r{\Vert {\mathcal{A}}(X-Z) \Vert}_{\ell_2}+2{\Vert X_c \Vert}_1\right)}.$$ In order to bound ${\Vert X-Z \Vert}_1$ we use Hölder’s inequality, the Frobenius robust rank null space property of ${\mathcal{A}}$ and the inequality above, $$\begin{aligned}
{\Vert X-Z \Vert}_1&={\Vert (X-Z)_r \Vert}_1+{\Vert (X-Z)_c \Vert}_1\leq\sqrt r{\Vert (X-Z)_r \Vert}_2+{\Vert (X-Z)_c \Vert}_1\\
&\leq(1+\rho){\Vert (X-Z)_c \Vert}_1+\tau\sqrt r{\Vert {\mathcal{A}}(Z-X) \Vert}_{\ell_2}\\
&\leq \frac{1+\rho}{1-\rho}{\left({\Vert Z \Vert}_1-{\Vert X \Vert}_1+\tau\sqrt r{\Vert {\mathcal{A}}(X-Z) \Vert}_{\ell_2}+2{\Vert X_c \Vert}_1\right)}+\tau\sqrt r{\Vert {\mathcal{A}}(X-Z) \Vert}_{\ell_2}\\
&= \frac{1+\rho}{1-\rho}{\left({\Vert Z \Vert}_1-{\Vert X \Vert}_1+2{\Vert X_c \Vert}_1\right)}+\frac{2\tau\sqrt r}{1-\rho}{\Vert {\mathcal{A}}(X-Z) \Vert}_{\ell_2}.
\end{aligned}$$ This concludes the proof.
Now we return to the proof of the theorem.
By Hölder’s inequality, Lemma \[lm:2NormNuclearNorm\] and the Frobenius robust rank null space property of ${\mathcal{A}}$ $$\begin{aligned}
{\Vert Z-X \Vert}_p&\leq{\Vert (X-Z)_r \Vert}_p+{\Vert (X-Z)_c \Vert}_p\leq r^{1/p-1/2}{\Vert (X-Z)_r \Vert}_2+{\Vert (X-Z)_c \Vert}_p\notag\\
&\leq\frac{\rho}{r^{1-1/p}}{\Vert (X-Z)_c \Vert}_1+\tau r^{1/p-1/2}{\Vert {\mathcal{A}}(X-Z) \Vert}_{\ell_2}+\frac{1}{r^{1-1/p}}{\Vert X-Z \Vert}_1\notag\\
&\leq \frac{1+\rho}{r^{1-1/p}}{\Vert X-Z \Vert}_1+\tau r^{1/p-1/2}{\Vert {\mathcal{A}}(X-Z) \Vert}_{\ell_2}\label{eq:PreliminaryBound}.\end{aligned}$$ Substituting the result of Lemma \[lm:BestRApproxOfDifference\] into (\[eq:PreliminaryBound\]) yields the desired inequality.
As a corollary of Theorem \[th:DifferenceBetweenSignalAndFeasibleElements\] we obtain that if $X\in{\mathbb{C}}^{n_1\times n_2}$ is a matrix of rank at most $r$ and the measurements are noiseless ($\eta=0$), then the Frobenius robust rank null space property implies that $X$ is the unique solution of $$\label{eq:NoNoiseNNMinimization}
\underset{Z\in{\mathbb{C}}^{n_1\times n_2}}\min{\Vert Z \Vert}_1\quad\mbox{subject to}\;{\mathcal{A}}(Z)=b.$$ It was first stated in [@RechtXuHassibiProceedings] that a slightly weaker property is actually equivalent to the successful recovery of $X$ via (\[eq:NoNoiseNNMinimization\]).
Given ${\mathcal{A}}:{\mathbb{C}}^{n_1\times n_2}\to{\mathbb{C}}^m$, every $X\in{\mathbb{C}}^{n_1\times n_2}$ of rank at most $r$ is the unique solution of (\[eq:NoNoiseNNMinimization\]) with $b={\mathcal{A}}(X)$ if and only if, for all $M\in\ker{\mathcal{A}}\setminus\{0\}$, it holds $$\label{eq:NSP}
{\Vert M_r \Vert}_1<{\Vert M_c \Vert}_1.$$
For the proof we refer to [@RechtXuHassibiProceedings] and [@FoucartRauhut Chapter 4.6]. According to Lemma \[lm:BestRApproxOfDifference\], another implication of the Frobenius robust rank null space property consists in the following error estimate in ${\Vert \cdot \Vert}_1$ for the case of noiseless measurements, $${\Vert X-X^{\sharp} \Vert}_1\leq\frac{2(1+\rho)}{1-\rho}{\Vert X_c \Vert}_1.$$ The above estimate remains true, if we require that for all $M\in\ker{\mathcal{A}}$, the singular values of $M$ satisfy $${\Vert M_r \Vert}_1\leq\rho{\Vert M_c \Vert}_1,\quad 0<\rho<1.$$ This property is known as the stable rank null space property of order $r$ with constant $\rho$. It is clear that if ${\mathcal{A}}:{\mathbb{C}}^{n_1\times n_2}\to {\mathbb{C}}^m$ satisfies the Frobenius robust rank null space property, then it satisfies the stable rank null space property. The approach used in [@OymakMohanFazelHassibi] to verify that the stable null space property accounts for stable recovery of matrices which are not exactly of low rank, exploits the similarity between the sparse vector recovery and the low-rank matrix recovery. It shows that if some condition is sufficient for stable and robust recovery of any sparse vector with at most $r$ non-zero entries, then the extension of this condition to the matrix case is sufficient for the stable and robust recovery of any matrix up to rank $r$.
In order to check whether the measurement map ${\mathcal{A}}:{\mathbb{C}}^{n_1\times n_2}\to{\mathbb{C}}^m$ satisfies the Frobenius robust rank null space property, we introduce the set $$T_{\rho,r}:=\left\{M\in{\mathbb{C}}^{n_1\times n_2}:{\Vert M \Vert}_2=1,{\Vert M_r \Vert}_2>\frac{\rho}{\sqrt r}{\Vert M_c \Vert}_1\right\}.$$
\[nsplemma\] If $$\inf\{{\Vert {\mathcal{A}}(M) \Vert}_{\ell_2}:M\in T_{\rho,r}\}>\frac{1}{\tau},$$ then ${\mathcal{A}}$ satisfies the Frobenius robust rank null space property of order $r$ with constants $\rho$ and $\tau$.
Suppose that $$\label{eq:InfL2NormMeasurements}
\inf\{{\Vert {\mathcal{A}}(M) \Vert}_{\ell_2}:M\in T_{\rho,r}\}>\frac{1}{\tau}.$$ It follows that for any $M\in{\mathbb{C}}^{n_1\times n_2}$ such that ${\Vert {\mathcal{A}}(M) \Vert}_{\ell_2}\leq\frac{{\Vert M \Vert}_2}{\tau}$ it holds $$\label{eq:MatrixL2StableNSP}
{\Vert M_r \Vert}_2\leq\frac{\rho}{\sqrt r}{\Vert M_c \Vert}_1.$$ For the remaining $M\in{\mathbb{C}}^{n_1\times n_2}$ with ${\Vert {\mathcal{A}}(M) \Vert}_{\ell_2}>\frac{{\Vert M \Vert}_2}{\tau}$ we have $${\Vert M_r \Vert}_2\leq{\Vert M \Vert}_2<\tau{\Vert {\mathcal{A}}(M) \Vert}_{\ell_2}.$$ Together with (\[eq:MatrixL2StableNSP\]) this leads to $${\Vert M_r \Vert}_2\leq\frac{\rho}{\sqrt r}{\Vert M_c \Vert}_1+\tau{\Vert {\mathcal{A}}(M) \Vert}_{\ell_2}.$$ for any $M\in{\mathbb{C}}^{n_1\times n_2}$.
It is natural to expect that the recovery error gets smaller as the number of measurements increases. This can be taken into account by establishing the null space property for $\tau=\frac{\kappa}{\sqrt m}$. Then the error bound reads as follows $${\Vert X-X^{\sharp} \Vert}_2\leq \frac{2(1+\rho)^2}{(1-\rho)\sqrt r}{\Vert X_c \Vert}_1+\frac{2\kappa(3+\rho)}{\sqrt m(1-\rho)}\eta.$$
An important property of the set $T_{\rho,r}$ is that it is imbedded in a set with a simple structure. The next lemma relies on the ideas presented in [@RudelsonVershynin] for the compressed sensing setting.
\[lm:SetD\] Let $D$ be the set defined by $$\label{eqDefinitionOfD}
D:=\operatorname{conv}{\left\{M\in{\mathbb{C}}^{n_1\times n_2}:{\Vert M \Vert}_2=1,\operatorname{rank}M\leq r\right\}},$$ where $\operatorname{conv}$ stands for the convex hull.
(a) \[itUnitBall\] Then $D$ is the unit ball with respect to the norm $${\Vert M \Vert}_D:=\sum_{j=1}^L{\left[\sum_{i\in I_j}{\left(\sigma_i(M)\right)}^2\right]}^{1/2},$$ where $L=\lceil\frac{n}{r}\rceil$, $$I_j=\left\{\begin{array}{ll}
{\left\{r(j-1)+1,\ldots,rj\right\}}, & j=1,\ldots, L-1,\\
{\left\{r(L-1)+1,\ldots,n\right\}}, & j=L.
\end{array}\right.$$
(b) \[itInclusion\] It holds $$\label{eqInclusionInUniversalSet}
T_{\rho,r}\subset \sqrt{1+(1+\rho^{-1})^2}D.$$
Let us argue briefly why ${\Vert \cdot \Vert}_D$ is a norm. Define $g:{\mathbb{C}}^n\to[0,\infty)$ by $$g(x):=\sum_{j=1}^L{\left(\sum_{i\in I_j}{\left(x_i^*\right)}^2\right)}^{1/2},$$ where $L$ and $I_j$ are defined in the same way as in item \[itUnitBall\] of Lemma \[lm:SetD\]. Then $g$ is a symmetric gauge function and ${\Vert M \Vert}_D=g(\sigma(M))$ for any $M\in{\mathbb{C}}^{n_1\times n_2}$. The norm property follows from [@HornJohnson Theorem 7.4.7.2].
\[itUnitBall\] Any $M\in D$ can be written as $$M=\sum_i\alpha_iX_i$$ with $$\operatorname{rank}X_i\leq r,\;{\Vert X_i \Vert}_2=1,\; \alpha_i\geq 0,\;\sum_i\alpha_i=1.$$ Thus $${\Vert M \Vert}_D\leq\sum_i\alpha_i{\Vert X_i \Vert}_D=\sum_i\alpha_i{\Vert X_i \Vert}_2=\sum_i\alpha_i=1.$$ Conversely, suppose that ${\Vert M \Vert}_D\leq 1$, and let $M$ have a singular value decomposition $M=U\Sigma V^*=\sum\limits_{j=1}^L\sum\limits_{i\in I_j}\sigma_i(M)u_iv_i^*$, where $u_i\in {\mathbb{C}}^{n_1}$ and $v_i\in{\mathbb{C}}^{n_2}$ are column vectors of $U$ and $V$ respectively. Set $M_j:=\sum\limits_{i\in I_j}\sigma_i(M)u_iv_i^*$ and $\alpha_j:={\Vert M_j \Vert}_2$, $j=1,\ldots, L$. Then each $M_j$ is a sum of $r$ rank-one matrices, so that $\operatorname{rank}M_j\leq r$, and we can write $M$ as $$M=\sum_{j:\alpha_j\neq 0}\alpha_j{\left(\frac{1}{\alpha_j}M_j\right)}$$ with $$\sum_{j:\alpha_j\neq 0}\alpha_j=\sum_j{\Vert M_j \Vert}_2={\Vert M \Vert}_D\leq 1\quad\text{and}\quad{\Vert \frac{1}{\alpha_j}M_j \Vert}_2=\frac{1}{\alpha_j}{\Vert M_j \Vert}_2=1.$$ Hence $M\in D$.
\[itInclusion\] To prove the embedding of $T_{\rho,r}$ into a scaled version of $D$, we estimate the norm of an arbitrary element $M$ of $T_{\rho,r}$. According to the definition of the ${\Vert \cdot \Vert}_D$-norm $$\begin{aligned}
{\Vert M \Vert}_D&=\sum_{\ell=1}^L{\left[\sum_{i\in I_{\ell}}{\left(\sigma_i(M)\right)}^2\right]}^{\frac{1}{2}}={\Vert M_r \Vert}_2+{\left[\sum_{i=r+1}^{2r}{\left(\sigma_i(M)\right)}^2\right]}^{\frac{1}{2}}+\sum_{\ell\geq 3}^L{\left[\sum_{i\in I_{\ell}}{\left(\sigma_i(M)\right)}^2\right]}^{\frac{1}{2}}\label{eq:EstimateForDNormInBlocks}.\end{aligned}$$ To bound the last term in the inequality above, we first note that for each $i\in I_{\ell}$, $\ell\geq 3$, $$\sigma_i(M)\leq\frac{1}{r}\sum_{j\in I_{\ell-1}}\sigma_j(M)$$ and hence $$\quad {\left[\sum_{i\in I_{\ell}}(\sigma_i(M))^2\right]}^{1/2}\leq\frac{1}{\sqrt r}\sum_{j\in I_{\ell-1}}\sigma_j(M).$$ Summing up over $\ell\geq 3$ yields $$\sum_{\ell\geq 3}^L{\left[\sum_{i\in I_{\ell}}{\left(\sigma_i(M)\right)}^2\right]}^{\frac{1}{2}}\leq\frac{1}{\sqrt r}\sum_{l\geq 2}\sum_{j\in I_{\ell}}\sigma_j(M)=\frac{1}{\sqrt r}\sum_{j=r+1}^n\sigma_j(M)=\frac{1}{\sqrt r}{\Vert M_c \Vert}_1.$$ and taking into account the inequality for the singular values of $M\in T_{\rho,r}$ $$\sum_{\ell\geq 3}^L{\left[\sum_{i\in I_{\ell}}{\left(\sigma_i(M)\right)}^2\right]}^{\frac{1}{2}}\leq\rho^{-1}{\Vert M_r \Vert}_2.$$ Applying the last estimate to (\[eq:EstimateForDNormInBlocks\]) we derive that $$\begin{aligned}
&{\Vert M \Vert}_D\leq(1+\rho^{-1}){\Vert M_r \Vert}_2\!\!+{\left[\sum_{i=r+1}^{2r}{\left(\sigma_i(M)\right)}^2\right]}^{\frac{1}{2}} \leq(1+\rho^{-1}){\Vert M_r \Vert}_2+{\left(1-{\Vert M_r \Vert}_2^2\right)}^{\frac{1}{2}}.
\end{aligned}$$ Set $a={\Vert M_r \Vert}_2$. The maximum of the function $$f(a):=(1+\rho^{-1})a+\sqrt{1-a^2}, \quad 0\leq a\leq 1,$$ is attained at the point $$a = \frac{1+\rho^{-1}}{\sqrt{1+(1+\rho^{-1})^2}}$$ and is equal to $\sqrt{1+(1+\rho^{-1})^2}$. Thus for any $M\in T_{\rho,r}$ it holds $${\Vert M \Vert}_D\leq\sqrt{1+(1+\rho^{-1})^2},$$ which proves (\[eqInclusionInUniversalSet\]).
The previous results hold true in the real-valued case and in the case of Hermitian matrices, when the nuclear norm minimization problem is solved over the set of matrices of that special type. As a set $D$ we then take the convex hull of corresponding matrices of rank $r$ and unit Frobenius norm. The only difference in the proof of Lemma \[lm:SetD\] occurs at the point, where we have to show that any $M$ with ${\Vert M \Vert}_D\leq 1$ belongs to $D$. Say, $M\in{\mathbb{C}}^{n_\times n}$ is Hermitian and ${\Vert M \Vert}_D\leq 1$. Then $M=U\Lambda U^*=\sum\limits_{j=1}^L\sum\limits_{i\in I_j}\sigma_i(M)u_iu_i^*$, where $u_i\in{\mathbb{C}}^n$, and $M_j:=\sum\limits_{i\in I_j}\sigma_i(M)u_iu_i^*$ is Hermitian. The rest of the proof remains unchained.
Employing the matrix representation of the measurement map ${\mathcal{A}}$, the problem of estimating the probability of the event (\[eq:InfL2NormMeasurements\]) is reduced to the problem of giving a lower bound for the quantities of the form $\underset{x\in T}\inf{\Vert Ax \Vert}_2$. This is not an easy task for deterministic matrices, but the situation significantly changes for matrices chosen at random.
Gaussian measurements
=====================
Our main result for Gaussian measurements reads as follows.
\[th:GaussianMeas\] Let ${\mathcal{A}}:{\mathbb{R}}^{n_1\times n_2}\to{\mathbb{R}}^m$ be the linear map generated by a sequence $A_1, \hdots, A_m$ of independent standard Gaussian matrices, let $0<\rho<1$, $\kappa >1$ and $0<{\varepsilon}<1$. If $$\label{eq:NumberOfMeasurementsForL2RobustRecovery}
\frac{m^2}{m+1}\geq \frac{r(1+(1+\rho^{-1})^2)\kappa^2}{(\kappa-1)^2}\left[\sqrt{n_1}+\sqrt{n_2}+\sqrt\frac{2\ln({\varepsilon}^{-1})}{r(1+(1+\rho^{-1})^2)}\right]^2,$$ then with probability at least $1-{\varepsilon}$, for every $X\in{\mathbb{R}}^{n_1\times n_2}$, a solution $X^{\sharp}$ of (\[eqNNMinimization\]) with $b={\mathcal{A}}(X)+w$, ${\Vert w \Vert}_{\ell_2}\leq\eta$, approximates $X$ with error $${\Vert X-X^{\sharp} \Vert}_2\leq \frac{2(1+\rho)^2}{(1-\rho)\sqrt r}{\Vert X_c \Vert}_1+\frac{2\kappa\sqrt 2(3+\rho)}{\sqrt m(1-\rho)}\eta.$$
In order to prove Theorem \[th:GaussianMeas\] we employ Gordon’s escape through a mesh theorem that provides an estimate of the probability of the event (\[eq:InfL2NormMeasurements\]). First we recall some definitions. Let $g\in{\mathbb{R}}^m$ be a standard Gaussian random vector, that is, a vector of independent mean zero, variance one normal distributed random variables. Then for $$E_m:=\operatorname{\mathbb{E}}{\Vert g \Vert}_2=\sqrt{2}\;\frac{\Gamma{\left((m+1)/2\right)}}{\Gamma{\left(m/2\right)}}$$ we have $$\frac{m}{\sqrt{m+1}}\leq E_m\leq\sqrt{m},$$ see [@Gordon; @FoucartRauhut]. For a set $T\subset{\mathbb{R}}^n$ we define its Gaussian width by $$\ell(T):=\operatorname{\mathbb{E}}\underset{x\in T}\sup{\langlex,g\rangle},$$ where $g\in{\mathbb{R}}^n$ is a standard Gaussian random vector.
\[thGordonsEscapeThroughTheMesh\] Let $A\in{\mathbb{R}}^{m\times n}$ be a Gaussian random matrix and $T$ be a subset of the unit sphere ${\mathbb{S}}^{n-1}$. Then, for $t>0$, $$\label{eqGordonsEscapeThroughTheMesh}
{\mathbb{P}}{\left(\underset{x\in T}\inf{\Vert Ax \Vert}_2> E_m-\ell(T)-t\right)}\geq 1-e^{-\frac{t^2}{2}}.$$
In order to apply this result to our measurement process (\[eq:MeasurementProcess\]) we unravel the columns of $A_j$, $j=1,\ldots,m$, into a single row and collect all of these in a $m\times n_1n_2$-matrix $A$, so that $n = n_1 n_2$ when applying . In order to give a bound on the number of Gaussian measurements, Theorem \[thGordonsEscapeThroughTheMesh\] requires to estimate the Gaussian width of the set $T_{\rho,r}$ from above. As it was pointed out in the previous section, $T_{\rho,r}$ is a subset of a scaled version of $D$, which has a relatively simple structure. So instead of evaluating $\ell(T_{\rho,r})$, we consider $\ell(D)$.
For the set $D$ defined by (\[eqDefinitionOfD\]) it holds $$\label{eq:EstimateOfGWOfD}
\ell(D)\leq\sqrt{r}(\sqrt{n_1}+\sqrt{n_2}).$$
Let $\Gamma\in{\mathbb{R}}^{n_1\times n_2}$ have independent standard normal distributed entries. Then $\ell(D)=\operatorname{\mathbb{E}}\underset{M\in D}\sup{\langle\Gamma, M\rangle}$. Since a convex continuous real-valued function attains its maximum value at one of the extreme points, it holds $\ell(D)=\operatorname{\mathbb{E}}\underset{\substack{{\Vert M \Vert}_2=1\\ \operatorname{rank}M\leq r}}\sup{\langle\Gamma, M\rangle}$. By Hölder’s inequality, $$\ell(D)\leq\operatorname{\mathbb{E}}\underset{\substack{{\Vert M \Vert}_2=1\\ \operatorname{rank}M\leq r}}\sup{\Vert \Gamma \Vert}_{\infty}{\Vert M \Vert}_1\leq\sqrt r\underset{\substack{{\Vert M \Vert}_2=1\\ \operatorname{rank}M\leq r}}\sup{\Vert M \Vert}_2\operatorname{\mathbb{E}}\sigma_1(\Gamma)\leq\sqrt r(\sqrt{n_1}+\sqrt{n_2}),$$ where the last inequality follows from an estimate for the expectation of the largest singular value of a Gaussian matrix, see [@FoucartRauhut Chapter 9.3].
Set $t:=\sqrt{2\ln({\varepsilon}^{-1})}$. If $m$ satisfies (\[eq:NumberOfMeasurementsForL2RobustRecovery\]), then $$E_m{\left(1-\frac{1}{\kappa}\right)}\geq\sqrt{r(1+(1+\rho^{-1})^2)}(\sqrt{n_1}+\sqrt{n_2})+t.$$ Together with (\[eqInclusionInUniversalSet\]) and (\[eq:EstimateOfGWOfD\]) this yields $$E_m-\ell(T_{\rho,r})-t\geq\frac{E_m}{\kappa}\geq\frac{1}{\kappa}\sqrt\frac{m}{2}.$$ According to Theorem \[thGordonsEscapeThroughTheMesh\] $$\begin{aligned}
{\mathbb{P}}{\left(\underset{M\in T_{\rho,r}}\inf{\Vert {\mathcal{A}}(M) \Vert}_2>\frac{\sqrt m}{\kappa\sqrt 2}\right)}\geq 1-{\varepsilon},
\end{aligned}$$ which means that with probability at least $1-{\varepsilon}$ map ${\mathcal{A}}$ satisfies the Frobenius robust rank null space property with constants $\rho$ and $\frac{\kappa\sqrt 2}{\sqrt m}$. The error estimate follows from Theorem \[th:FrobeniusNSP\].
Measurement matrices with independent entries and four finite moments
=====================================================================
In this section we prove Theorem \[th:indepNSP\], which is the generalization of Theorem \[th:GaussianMeas\] to the case when the map ${\mathcal{A}}:{\mathbb{R}}^{n_1\times n_2}\to{\mathbb{R}}^m$ is obtained from $m$ independent samples of a random matrix $\Phi=(X_{ij})_{i,j}$ with the following properties:
- The $X_{ij}$ are independent random variables of mean zero,
- $\mathbb E X_{ij}^2=1$ and $\mathbb E X_{ij}^4\leq C_4 $ for all $i,j$ and some constant $C_4$.
Note that (by Hölder’s inequality) $C_4\geq 1$.
As before the idea of the proof is to show that the event (\[eq:InfL2NormMeasurements\]) holds with high probability. In order to do so we apply Mendelson’s small ball method [@KoltchinskiiMendelson; @Mendelson; @tr14] in the manner of [@tr14].
\[KMT\] Fix $E\subset \mathbb R^d$ and let $\phi_1,\hdots,\phi_m$ be independent copies of a random vector $\phi$ in $\mathbb R^d$. For $\xi >0$ let $$Q_{\xi}(E;\phi) =\inf_{u\in E}\mathbb P\{ \vert \langle \phi, u \rangle \vert \geq \xi\}$$ and $$W_m(E;\phi)=\mathbb E \sup_{u\in E}\langle h,u\rangle,$$ where $h=\frac{1}{\sqrt m} \sum_{j=1}^{m}\varepsilon_j \phi_j$ with $(\varepsilon_j)$ being a Rademacher sequence [^4]. Then for any $\xi >0$ and any $t\geq 0$ with probability at least $1-e^{-2 t^2}$ $$\inf_{u\in E} \left( \sum_{i=1}^{m}\vert \langle \phi_i, u \rangle \vert^2 \right )^{1/2}\geq \xi \sqrt m Q_{2\xi}(E;\phi)-2W_m(E;\phi)-\xi t.$$
We start with two lemmas.
\[qabsch\] $$\inf_{\{Y, \Vert Y \Vert_2=1 \}}\mathbb P(\vert\langle\Phi, Y \rangle \vert\geq \frac{1}{\sqrt 2} )\geq \frac{1}{4C_5},$$where $C_5=\max\{3,C_4\}$.
Assume that $Y$ has Frobenius norm one. The Payley-Zygmund inequality (see e.g. [@FoucartRauhut Lemma 7.16], and also [@tr14]), implies $$\label{PZ}
\mathbb P \{\vert \langle \Phi,Y\rangle \vert^2\geq \frac{1}{2}(\mathbb E \vert \langle \Phi,Y\rangle \vert^2) \}\geq \frac{1}{4}\cdot \frac{(\mathbb E \vert \langle \Phi,Y \rangle \vert^2)^2 }{\mathbb E \vert \langle \Phi,Y \rangle \vert^4 }.$$ We compute numerator and denominator. $$\mathbb E \vert \langle \Phi,Y\rangle \vert^2=\sum_{i,j,k,l}\mathbb E (X_{ij}X_{kl})\cdot Y_{ij}Y_{kl}=\sum_{i,j}\mathbb E X_{ij}^2\cdot Y_{ij}^2=\sum_{i,j} Y_{ij}^2=1.$$ Likewise, $$\begin{aligned}
\mathbb E \vert \langle \Phi,Y\rangle \vert^4&=\sum_{i_1,\hdots,i_4, j_1,\hdots,j_4}\mathbb E (X_{i_1j_1}\cdots X_{i_4j_4})\cdot Y_{i_1j_1}\cdots Y_{i_4j_4}\\&=\sum_{i,j}\mathbb E X_{ij}^4\cdot Y_{ij}^4+3\sum_{i_1,i_2, j_1,j_2\atop (i_1,j_1)\neq (i_2,j_2)}\mathbb E (X_{i_1j_1}^2 X_{i_2j_2}^2)\cdot Y_{i_1j_1}^2 Y_{i_2j_2}^2\\
&=\sum_{i,j}\mathbb E X_{ij}^4\cdot Y_{ij}^4+3\sum_{i_1,i_2, j_1,j_2\atop (i_1,j_1)\neq (i_2,j_2)} Y_{i_1j_1}^2 Y_{i_2j_2}^2
\leq \sum_{i,j}C_4\cdot Y_{ij}^4+3\sum_{i_1,i_2, j_1,j_2\atop (i_1,j_1)\neq (i_2,j_2)} Y_{i_1j_1}^2 Y_{i_2j_2}^2\\
&\leq C_5\sum_{i_1,i_2, j_1,j_2} Y_{i_1j_1}^2 Y_{i_2j_2}^2=C_5(\sum_{i,j} Y_{ij}^2)^2=C_5.\end{aligned}$$ Combining this with $(\mathbb E \vert \langle \Phi,Y\rangle \vert^2)^2=1$ and the estimate (\[PZ\]), the claim follows.
\[opnormabsch\] Let $\Phi_1,\hdots,\Phi_m$ be independent copies of a random matrix $\Phi$ as above. Let $\varepsilon_1,\hdots,\varepsilon_m$ be independent Rademacher variables independent of everything else and let $H=\frac{1}{\sqrt m}\sum_{k=1}^{m}\varepsilon_k \Phi_k$. Then $$\mathbb E \Vert H\Vert_{\infty}\leq C_1\sqrt{n}.$$ Here $C_1$ is a constant that only depends on $C_4$.
Let $S=\sum_{k=1}^{m} \Phi_k$. We first desymmetrize the sum $H$ (see [@lt91 Lemma 6.3]) and obtain $$\mathbb E\Vert H\Vert_{\infty}\leq \frac{2}{\sqrt{m}}\mathbb E\Vert S\Vert_{\infty}.$$
Therefore, it is enough to show that $\mathbb E\Vert S\Vert_{\infty}\leq c_3\sqrt{mn}$ for a suitable constant $c_3.$ The matrix $S$ has independent mean zero entries, hence by a result Lata[ł]{}a (see [@la05]) the following estimate holds for some universal constant $C_2$, $$\mathbb E \Vert S\Vert_{\infty}\leq C_2\left(\max_{i} \sqrt{\sum_j \mathbb E S_{ij}^2}+\max_{j} \sqrt{\sum_i \mathbb E S_{ij}^2}+\sqrt[4]{\sum_{i,j} \mathbb E S_{ij}^4}\right).$$ Denoting the entries of $\Phi_k$ by $X_{k;ij}$, we have $S_{ij}=\sum_k X_{k;ij}$. Hence, using the independence of the $X_{k;ij}$, we obtain $\mathbb E S_{ij}^2=\mathbb E (\sum_k X_{k;ij})^2=\sum_k \mathbb E X_{k;ij}^2=m$. Thus, $\sqrt{\sum_j \mathbb E S_{ij}^2} \leq \sqrt{nm}$ for any $i$ and $ \sqrt{\sum_i \mathbb E S_{ij}^2}\leq \sqrt{nm}$ for any $j$. Finally to estimate $\sqrt[4]{\sum_{i,j} \mathbb E S_{ij}^4}$ we calculate $\mathbb E S_{ij}^4=\mathbb E (\sum_k{X_{k;ij}})^4$. Using again that the $X_{k;ij}$ are independent and have mean zero we obtain $$E S_{ij}^4=\sum_k \mathbb E X_{k;ij}^4 + 3\sum_{k_1\neq k_2}\mathbb E X_{k_1;ij}^2\mathbb E X_{k_2;ij}^2.$$ Using that $\mathbb E X_{k;ij}^2=1$ for all $i,j,k$, we obtain $\mathbb E S_{ij}^4\leq C_5m^2$, where $C_5=\max\{3,C_4\}$ and hence $$\sqrt[4]{\sum_{i,j} \mathbb E S_{ij}^4}\leq \sqrt[4]{C_5m^2n^2}=\sqrt[4]{C_4}\sqrt{mn}.$$ Hence, indeed $\mathbb E\Vert S\Vert_{\infty}\leq c_3\sqrt{mn}$ for a suitable constant $c_3$ that depends only on $C_4$.
Let now $T_{\rho,r}$ and $D$ be the sets defined in Section \[sec:NSP\], but restricted to the real-valued matrices. By Hölder’s inequality, for any $n_1\times n_2$ matrix $Y$ of Frobenius norm $1$ and rank at most $r$ and any $n_1\times n_2$ matrix $H$, $$\langle H, Y \rangle \leq \Vert Y \Vert_1 \Vert H \Vert_{\infty}\leq \sqrt r \Vert H \Vert_{\infty}.$$ Hence $$\label{wmabsch}
\sup_{Y\in D} \langle H, Y \rangle \leq \sqrt r \Vert H \Vert_{\infty}.$$
Let $H=\frac{1}{\sqrt m} \sum_{j=1}^{m}\varepsilon_j \Phi_j$ and let $\xi = \frac{1}{2\sqrt 2}$ and $E=T_{\rho,r}$. Then it follows from Theorem \[KMT\] that for any $t\geq 0$ with probability at least $1-e^{-2 t^2}$ $$\label{KMTAnw}
\inf_{Y\in T_{\rho,r}} \left( \sum_{i=1}^{m}\vert \langle \Phi_i, Y \rangle \vert^2 \right )^{1/2}\geq \frac{\sqrt m}{2\sqrt 2}Q_{\frac{1}{\sqrt 2}}(T_{\rho,r};\Phi)-2 W_m(T_{\rho,r},\Phi) -\frac{1}{2\sqrt 2} t.$$ Using Lemma \[qabsch\] and the fact that all elements of $T_{\rho,r}$ have Frobenius norm $1$, we obtain $$\label{Q1/2}
Q_{\frac{1}{\sqrt 2}}(T_{\rho,r};\Phi)\geq \frac{1}{4C_5}.$$ Combining now the fact that $T_{\rho,r}\subseteq \sqrt{1+(1+\rho^{-1})^2}D $ (see Lemma \[lm:SetD\]) with estimate (\[wmabsch\]) and Lemma \[opnormabsch\] leads to $$\label{WmT}
W_m(T_{\rho,r},\Phi)\leq \sqrt{1+(1+\rho^{-1})^2} \sqrt r\ \mathbb E {\Vert H \Vert}_{\infty} \leq C_1 \sqrt{1+(1+\rho^{-1})^2} \sqrt r \sqrt n.$$ Using (\[KMTAnw\]), (\[Q1/2\]) and (\[WmT\]) we see that choosing $m\geq c_1\rho^{-2}nr$ and $t=c_4m$ for suitable constants $c_1,c_4$, we obtain with probability at least $1-e^{-c_2m}$ $$\inf_{Y\in T_{\rho,r}} \left( \sum_{i=1}^{m}\vert \langle \Phi_i, Y \rangle \vert^2 \right )^{1/2}\geq c_3 \sqrt m$$ for suitable constants $c_2,c_3$. Now the claim follows from Lemma \[nsplemma\] and Theorem \[th:FrobeniusNSP\] (both of which also hold in the real valued version by the same proofs respectively).
Rank one Gaussian measurements
==============================
In this section we prove Theorem \[mainTh1\]. The proof technique is an application of Mendelson’s small ball method analogous to the proof of Theorem \[th:indepNSP\]. Let $$T^{\mathcal H}_{\rho,r}:={\left\{M\in\mathcal H_n: {\Vert M \Vert}_2=1,\ {\Vert M_r \Vert}_2>\frac{\rho}{\sqrt r}{\Vert M_c \Vert}_1\right\}}.$$ Let $T_{\rho,r}$ be defined as $T^{\mathcal H}_{\rho,r}$ but with $\mathcal H_n$ replaced by the set of all complex $n\times n$-matrices (i.e. it is defined as before with $n_1=n_2=n$). Then $T^{\mathcal H}_{\rho,r}\subseteq T_{\rho, r}$. It is enough to show that with high probabiliy $$\inf_{Y\in T_{\rho, r}^{\mathcal H}} \left( \sum_{j=1}^{m}\vert \langle a_ja_j^*, Y \rangle \vert^2 \right )^{1/2}\geq \sqrt m/C_3 \label{eq:Gauss_tau}$$ We apply Theorem \[KMT\] with $E=T_{\rho, r}^{\mathcal H}$. The next lemma estimates the small ball probability $Q_{\frac{1}{\sqrt 2}}(E;\phi) $ used in Mendelson’s method.
$Q_{\frac{1}{\sqrt 2}}(E;\phi) :=\inf_{u\in E}\mathbb P\{ \vert \langle aa^*, u \rangle \vert \geq \frac{1}{\sqrt 2}\} \geq \frac{1}{96}$.
Let now (as in [@tr14; @krt14]) $$H=\frac{1}{\sqrt m}\sum_{j=1}^m\varepsilon_ja_ja_j^*, \label{eq:H}$$ where the $\varepsilon_j$ form a Rademacher sequence. For any $M\in \mathcal H_n$ and any $n\times n$ matrix $Y$ of Frobenius norm $1$ and rank at most $r$ $$\langle M, Y \rangle \leq \Vert Y \Vert_1 \Vert M \Vert_{\infty}\leq \sqrt r \Vert M \Vert_{\infty}.$$ Since $E=T_{\rho, r}^{\mathcal H} \subseteq T_{\rho, r}\subseteq \sqrt{1+(1+\rho^{-1})^2}D$, this implies $$W_m(E,\phi) =\mathbb E \sup_{Y\in E}\langle H,Y\rangle \leq \sqrt{1+(1+\rho^{-1})^2}\sqrt r \mathbb E \Vert H\Vert_{\infty}.$$ As in [@krt14] we use now that by the arguments in [@ve12 Section 5.4.1] we have $\mathbb E \Vert H\Vert_{\infty}\leq c_2 \sqrt n$ if $m\geq c_3 n$ for suitable constants $c_2, c_3$, see also [@tr14 Section 8]. Now the claim of Theorem \[mainTh1\] follows from Theorem \[KMT\], comp. the proof of Theorem \[th:indepNSP\].
Inspecting the above proof, resp. the proofs of the cited statements in [@krt14], we see that the real valued analogue of Theorem \[mainTh1\] is also true. We even may assume for this that the $a_j$ are i.i.d. subgaussian with $k$-th moments, where $k\leq 8$, equal to the corresponding $k$-th moments of the Gaussian standard distribution. The constants then depend only on the distribution of the $a_j$. We also note that a similar statement in the real case for the recovery of positive semidefinite matrices using subgaussian measurements has been shown by Chen, Chi and Goldsmith in [@cchg13] using the rank restricted isometry property.
Rank one measurements generated by 4-designs
============================================
Recall the definition of an approximate, weighted $t$-design.
\[def:approx\_design\] We call a weighted set $\left\{p_i,w_i \right\}_{i=1}^N$ of normalized vectors an approximate $t$-design of $p$-norm accuracy $\theta_p$, if $$\left\| \sum_{i=1}^N p_i \left( w_i w_i^* \right)^{\otimes t} - \int_{\|w \|_{\ell_2} = 1} \left( w w^* \right)^{\otimes t} \mathrm{d}w \right\|_p \leq \binom{n+t-1}{t}^{-1} \theta_p.
\label{eq:approx_designs}$$
A set of unit vectors obeying $\theta_p = 0 $ for $1 \leq p \leq \infty$ is called an *exact $t$-design*, see [@scott_tight_2006] and also [@krt14; @gross_partial_2014].
\[Th2\] Let $\left\{p_i,w_i \right\}_{i=1}^N$ be a an approximate $4$-design with either $\theta_\infty \leq 1/(16r^2)$, or $\theta_1 \leq 1/4$ that furthermore obeys $
\left\| \sum_{i=1}^N p_i w_i w_i^* - \frac{1}{n} \operatorname{id}\right\|_\infty \leq \frac{1}{n}
$. Suppose that the measurement operator $\mathcal{A}$ is generated by $$m\geq C_4\rho^{-2}nr\log n$$ measurement matrices $A_j = \sqrt{n(n+1)}a_j a_j^*$, where each $a_j$ is drawn independently from $\left\{p_i, w_i \right\}_{i=1}^N$. Then, with probability at least $1 - \mathrm{e}^{-C_5 m}$, $\mathcal{A}$ obeys the Frobenius robust rank null space property of order $r$ with constants $0 < \rho < 1$ and $ \tau = C_6/\sqrt{m}$. Here, $C_4,C_5$ and $C_6$ denote positive constants depending only on the design.
Theorem \[mainTh2\] readily follows from combining this statement with Theorem \[th:DifferenceBetweenSignalAndFeasibleElements\].
We start by presenting a proof for measurements drawn from an exact 4-design. Paralleling the proof of Theorem \[mainTh1\], the statement can be deduced from Theorem \[KMT\] by utilizing results from [@krt14]. Provided that $a$ is randomly chosen from a re-scaled, weighted $4$-design (such that each element has Euclidean length $\| w_i \|_{\ell_2} = \sqrt[4]{(n+1)n}$), [@krt14 Proposition 12] implies that $$\inf_{Z \in T_{\rho,r}} \mathbb{P} \left( | \mathrm{tr} \left( a a^* Z \right)| \geq \xi \right) \geq \inf_{\| Z \|_2 =1} \mathbb P \left( | \mathrm{tr} \left( a a^* Z \right)| \geq \xi \right) \geq \frac{(1-\xi^2)^2}{24} \label{eq:designs_Q}$$ is valid for all $\xi \in [0,1]$. Now let $H = \sum_{i=1}^m \epsilon_i a_i a_i^*$ be as in Theorem \[KMT\]. Lemma \[lm:SetD\] together with the fact that $D$ is the convex hull of all matrices of rank at most $r$ and Frobenius norm 1 allows us to conclude for $m \geq 2 n \log n$, that, $$\begin{aligned}
W_m \left( T_{\rho,r}, aa^* \right)
&= \mathbb{E} \sup_{M \in T_{\rho,r}} \operatorname{tr}\left( H M \right)
\leq \sqrt{1+ (1+\rho^{-1})^2} \;\mathbb{E} \sup_{M \in D} \operatorname{tr}\left( H M \right) \\
&\leq \sqrt{1+ (1+\rho^{-1})^2} \sup_{M \in D} \| M \|_1 \mathbb{E} \| H \|_\infty
\leq \sqrt{1+ (1+\rho^{-1})^2} \sqrt{r} \; \mathbb{E} \| H \|_\infty \\
& \leq 3.1049 \sqrt{1+ (1+\rho^{-1})^2 r n \log (2n)},\end{aligned}$$ where the last bound is due to [@krt14 Proposition 13]. Fixing $0 < \xi < 1/2$ arbitrarily and inserting these two bounds into Theorem \[KMT\] completes the proof.
An analogous statement for approximate 4-designs — with slightly worse absolute constants — can be obtained by resorting to the generalized versions of [@krt14 Propositions 12 and 13] presented in Section 4.5.1 in loc. cit.which are valid for approximate 4-designs that satisfy the conditions stated in Theorem \[Th2\].
The positive semidefinite case
==============================
Finally, we focus on the case, where the matrices of interest are Hermitian and positive semidefinite and establish Theorem \[mainTh3\]. In order to arrive at such a statement, we closely follow the ideas presented in [@kalev_informationally_2015] which in turn were inspired by [@bruckstein_uniqueness_2008] containing an analogous statement for a non-negative compressed sensing scenario.
We require two further concepts from matrix analysis. For every positive semidefinite matrix $W \succcurlyeq 0$ with eigenvalue decomposition $W = \sum_{i=1}^n \lambda_i w_i w_i^*$ we define its square root to be $W^{1/2} := \sum_{i=1}^n \sqrt{\lambda_i} w_i w_i^*$. In other words, $W^{1/2}$ is the unique positive semidefinite matrix which acts on the eigenspace corresponding to the eigenvalue $\lambda_i$ of $W$ by multiplication by $\sqrt{\lambda_i}$. Note that this matrix obeys $W^{1/2}\cdot W^{1/2} = W$. Also, recall that the condition number $\kappa (W)$ of a matrix $W$ is the ratio between its largest and smallest nonzero singular value. For an invertible Hermitian matrix with inverse $W^{-1}$ this number equals $$\kappa (W) = \| W \|_\infty \| W^{-1} \|_\infty.$$
Suppose that the measurement process is such that there exists $t\in{\mathbb{R}}^m$ which assures that $W:=\sum_{j=1}^m t_jA_j$ is positive definite. We define the artificial measurement map $${\mathcal{A}}_{W^{1/2}}:{\mathcal{H}}_n\to{\mathbb{R}}^m, \quad Z\mapsto {\mathcal{A}}(W^{-1/2}Z W^{-1/2}) \label{eq:artificial_measurements}$$ and the endomorphism $$Z \mapsto \tilde{Z} := W^{1/2} Z W^{1/2} \label{eq:mapping}$$ of $\mathcal{H}_n$. Note that these definitions assure $$\label{eq:RelationBetweenAAB}
{\mathcal{A}}(Z)={\mathcal{A}}_{W^{1/2}}(\tilde Z) \quad \mbox{ for all } Z \in \mathcal{H}_n$$ and the singular values of $Z$ and $\tilde Z$ satisfy $$\label{eq:SingularValuesZandTildeZ}
\sigma_j(\tilde Z)\leq {\Vert W^{1/2} \Vert}_{\infty}^2\sigma_j(Z)={\Vert W \Vert}_{\infty}\sigma_j(Z),\quad \sigma_j(Z)\leq {\Vert W^{-1/2} \Vert}_{\infty}^2\sigma_j(\tilde Z)={\Vert W^{-1} \Vert}_{\infty}\sigma_j(\tilde Z),$$ see [@BhatiaMatrixAnalysis p. 75]. Consequently, the mapping preserves the rank of any matrix. The following result assures that the artificial measurement operator $\mathcal{A}_{W^{1/2}}$ obeys the Frobenius robust rank null space property, if the original $\mathcal{A}$ does.
\[NSPvgl\] Suppose that ${\mathcal{A}}$ satifies the Frobenius robust rank null space property of order $r$ with constants $\rho$ and $\tau$ and suppose that $W=\sum_{j=1}^m t_jA_j$ is positive definite. Then ${\mathcal{A}}_{W^{1/2}}$ also obeys the Frobenius robust rank null space property of order $r$, but with constants $\tilde \rho= \kappa (W) \rho$ and $\tilde \tau={\Vert W \Vert}_{\infty}\tau$.
Let $\tilde Z\in{\mathcal{H}}_n$. Relations (\[eq:RelationBetweenAAB\]), (\[eq:SingularValuesZandTildeZ\]) together with the Frobenius robust rank null space property of ${\mathcal{A}}$ imply that $$\begin{aligned}
{\Vert \tilde Z_r \Vert}_2&\leq{\Vert W^{1/2} \Vert}_{\infty}^2{\Vert Z_r \Vert}_2\leq{\Vert W \Vert}_{\infty}{\left(\frac{\rho}{\sqrt r}{\Vert Z_c \Vert}_1+\tau{\Vert {\mathcal{A}}(Z) \Vert}_{\ell_2}\right)}\\
&\leq \frac{{\Vert W \Vert}_{\infty}{\Vert W^{-1} \Vert}_{\infty}\rho}{\sqrt r}{\Vert \tilde Z_c \Vert}_1+{\Vert W \Vert}_{\infty}\tau{\Vert {\mathcal{A}}_{W^{1/2}}(\tilde Z) \Vert}_{\ell_2}. \qquad
\end{aligned}$$
\[Normdiff\] Suppose there is $t\in{\mathbb{R}}^m$ such that $W:=\sum_{j=1}^m t_jA_j$ is positive definite. Let $\tilde X,\tilde Z$ be positive semidefinite. Then, $${\Vert \tilde{Z} \Vert}_1-{\Vert \tilde{X} \Vert}_1\leq {\Vert t \Vert}_{\ell_2}{\Vert {\mathcal{A}}_{W^{1/2}}(\tilde Z-\tilde X) \Vert}_{\ell_2}.$$
The claim follows from positive semidefiniteness of both $\tilde{Z}$ and $\tilde{X}$ and our choice of the endomorphism (\[eq:mapping\]). Indeed, $$\begin{aligned}
{\Vert \tilde Z \Vert}_1&=\operatorname{tr}(\tilde Z- \tilde X)+{\Vert \tilde X \Vert}_1=\operatorname{tr}(W^{1/2}(Z-X)W^{1/2})+{\Vert \tilde X \Vert}_1=\operatorname{tr}(W(Z-X))+{\Vert \tilde X \Vert}_1\\
&=\sum_{j=1}^m t_j\operatorname{tr}(A_j(Z-X))+{\Vert \tilde X \Vert}_1=\langle t,{\mathcal{A}}(Z-X) \rangle +{\Vert \tilde X \Vert}_1\\
&={\langlet,{\mathcal{A}}_{W^{1/2}}(\tilde Z-\tilde X)\rangle}+{\Vert \tilde X \Vert}_1\leq{\Vert t \Vert}_{\ell_2}{\Vert {\mathcal{A}}_{W^{1/2}}(\tilde Z-\tilde X) \Vert}_{\ell_2}+{\Vert \tilde X \Vert}_1.\end{aligned}$$ Here $X$ resp. $Z$ denote the preimage of $\tilde X$ resp $\tilde Z$ under the map (\[eq:mapping\]).
This simple technical statement allows us to establish the main result of this section.
\[th:NSPforPositiveMatrices\] Suppose there exists $t\in{\mathbb{R}}^m$ such that $W:=\sum_{j=1}^m t_jA_j$ is positive definite and ${\mathcal{A}}$ satisfies the Frobenius robust rank null space property with constants $0<\rho<\frac{1}{\kappa(W)}$ and $\tau>0$. Let $1\leq p\leq 2$. Then, for any $X,Z\succcurlyeq 0$, $${\Vert Z-X \Vert}_p\leq\frac{2C\kappa(W)}{r^{1-1/p}}{\Vert X_c \Vert}_1+r^{1/p-1/2}{\Vert {\mathcal{A}}(Z)-{\mathcal{A}}(X) \Vert}_{\ell_2} {\Vert W^{-1} \Vert}_{\infty} {\left(\frac{C{\Vert t \Vert}_2}{\sqrt r}+D{\Vert W \Vert}_{\infty}\tau\right)}
\label{eq:NSPforPositiveMatrices}$$ with constants $
C=\frac{(1+\kappa(W)\rho)^2}{1-\kappa(W)\rho}
$ and $ D=\frac{3+\kappa(W)\rho}{1-\kappa(W)\rho}.
$
Let $X,Z \succcurlyeq 0 $ be arbitrary. Then $$\| Z - X \|_p = \left\| W^{-1/2} \left( \tilde{Z} - \tilde{X} \right) W^{-1/2} \right\|_p \leq \| W^{-1} \|_\infty \| \tilde{Z} - \tilde{X} \|_p$$ holds and the resulting matrices $\tilde{Z},\tilde{X}$ are again positive-semidefinite. Also, since ${\mathcal{A}}$ satisfies the Frobenius robust rank null space property with constants $0<\rho<\frac{1}{\kappa(W)}$ and $\tau>0$, Lemma \[NSPvgl\] assures that $\mathcal{A}_{W^{1/2}}$ does the same with constants $0 < \tilde{\rho} < 1$ and $\tilde{\tau} = \| W \|_\infty \tau >0$. Combining this with Theorem \[th:DifferenceBetweenSignalAndFeasibleElements\] and Lemma \[Normdiff\] implies $$\begin{aligned}
{\Vert \tilde Z-\tilde X \Vert}_p& \leq\frac{C}{r^{1-1/p}}{\left({\Vert \tilde Z \Vert}_1-{\Vert \tilde X \Vert}_1+2{\Vert \tilde X_c \Vert}_1\right)}+D{\Vert W \Vert}_{\infty}\tau r^{1/p-1/2}{\Vert {\mathcal{A}}_{W^{1/2}}(\tilde Z-\tilde X) \Vert}_{\ell_2}\\
&\leq\frac{C}{r^{1-1/p}}{\left({\Vert t \Vert}_{\ell_2}{\Vert {\mathcal{A}}_{W^{1/2}}(\tilde Z-\tilde X) \Vert}_{\ell_2}+2{\Vert \tilde X_c \Vert}_1\right)}+D{\Vert W \Vert}_{\infty}\tau r^{1/p-1/2}{\Vert {\mathcal{A}}_{W^{1/2}}(\tilde Z-\tilde X) \Vert}_{\ell_2}\\
&\leq \frac{2C}{r^{1-1/p}}{\Vert \tilde X_c \Vert}_1+r^{1/p-1/2}{\Vert {\mathcal{A}}_{W^{1/2}}(\tilde Z-\tilde X) \Vert}_{\ell_2}{\left(\frac{C{\Vert t \Vert}_{\ell_2}}{\sqrt r}+D{\Vert W \Vert}_{\infty}\tau \right)}.
\end{aligned}$$ The desired statement follows from this estimate by taking into account (\[eq:RelationBetweenAAB\]) and (\[eq:SingularValuesZandTildeZ\]).
Note that in contrast to other recovery guarantees established here, Theorem \[th:NSPforPositiveMatrices\] does not require any convex optimization procedure. However, it does require the measurement process to obey an additional criterion: the intersection of the span of measurement matrices with the cone of positive definite matrices must be non-empty. We show that this is the case for the rank-one projective measurements introduced in the previous section with high probability. Since it has already been established that sufficiently many measurements of this kind obey the Frobenius robust rank null space property with high probability (see Theorems \[mainTh1\] and \[Th2\] and their respective proofs), Theorem \[mainTh3\] can then be established by taking the union bound over the individual probabilities of failure.
\[prop:WGauss\] Suppose $m \geq 4 n$ and let $A_1,\ldots,A_m$ be matrices of the form $a_j a_j^*$, where each $a_i \in \mathbb{C}^n$ is a random complex standard Gaussian vector. Then with probability at least $1- 2\mathrm{e}^{-C_{10} m}$, $W := \frac{1}{m} \sum_{j=1}^m A_j$ is positive definite and obeys $$\max \left\{ \| W \|_\infty, \| W^{-1} \|_\infty, \kappa (W) \right\} \leq C_{11}.
\label{eq:WGauss}$$ Here, $C_9, C_{10}, C_{11}>0$ denote universal positive constants.
Note that such a construction corresponds to setting $t = \frac{1}{m} (1,\ldots,1)^T \in \mathbb{R}^m$ which obeys $\| t \|_{\ell_2} = 1/\sqrt{m}$.
For the sake of simplicity, we are going to establish the statement for real standard Gaussian vectors. Establishing the complex case can be done analogously and leads to slightly different constants. Let $e_1,\ldots,e_m$ denote the standard basis in $\mathbb{R}^m$. We define the auxiliary $m \times n$ matrix $A := \sum_{i=1}^m e_i a_i^*$ which obeys $$\frac{1}{m} A^T A = \frac{1}{m} \sum_{i=1}^m a_i e_i^* \sum_{j=1}^m e_j a_j^* = \frac{1}{m} \sum_{i=1}^m a_i a_i^* = \frac{1}{m} \sum_{i=1}^m A_i = W. \label{eq:wishart}$$ Also, by construction, $A$ is a random matrix with standard Gaussian entries. Essentially, this relation implies that $m W$ is Wishart-distributed. From and the defining properties of eigen- and singular values we infer that $$\sqrt{ \lambda_{\min} (W)} = \frac{1}{\sqrt{m}} \sqrt{ \lambda_{\min} \left( A^T A \right) } = \frac{1}{\sqrt{m}} \lambda_{\min} \left( \sqrt{A^T A} \right) = \frac{1}{\sqrt{m}} \sigma_{\min} (A)
\label{eq:eig_sing}$$ and an analogous statement is true for the largest eigenvalue $\lambda_{\max}(W)$. Since $A$ is a Gaussian $m \times n$ matrix, concentration of measure implies that for any $\tilde{\tau} > 0$ $$\sqrt{m} - \sqrt{n} -\tilde{\tau} \leq \sigma_{\min}(A) \leq \sigma_{\max}(A) \leq \sqrt{m} + \sqrt{n} + \tilde{\tau} \label{eq:wishart_aux1}$$ with probability at least $1 - 2\mathrm{e}^{-\tilde{\tau}^2/2}$ — see e.g. [@ve12 Corollary 5.35] or [@FoucartRauhut Theorem 9.26]. Combining this with , recalling the assumption $m \geq 4 n$ and defining $\tau = \tilde{\tau}/\sqrt{m}$ allows for establishing $$\frac{1}{2} - \tau \leq 1 - \sqrt{\frac{n}{m}} - \tau \leq \sqrt{\lambda_{\min}(W)} \leq \sqrt{\lambda_{\max} (W)} \leq 1 + \sqrt{\frac{n}{m}} + \tau \leq \frac{3}{2} + \tau$$ with probability at least $1 - 2\mathrm{e}^{-m\tau^2/2}$. This inequality chain remains valid, if we square the individual terms. Setting $\tau = 1/4$ thus allows us to conclude $$\max \left\{ \lambda_{\max}(W), \lambda_{\min}^{-1} (W), \frac{\lambda_{\max}(W)}{\lambda_{\min}(W)} \right\}
\leq \left( \frac{ 3/2 + \tau}{1/2 - \tau} \right)^2 = 49 = C_{11},$$ with probability at least $1 - 2\mathrm{e}^{-m/32}$.
Alternatively, we could have relied on bounds on the condition number of Gaussian random matrices presented in [@chen_condition_2005]. While these bounds would be slightly tighter, we feel that our derivation is more illustrative and it suffices for our purpose.
\[prop:W4design\] Suppose $m \geq \tilde{C}_4 nr \log n $ and let $A_1,\ldots,A_m$ be matrices of the form $a_j a_j^*$, where each $a_j \in \mathbb{C}^n$ is chosen independently from a weighted set $\left\{p_i,w_i \right\}_{i=1}^N$ of vectors obeying $\| w_i \|_{\ell_2}^2 = \sqrt{n(n+1)}$ for all $1 \leq i \leq N$ and $$\left\| \sum_{i=1}^N p_i w_i w_i^* - \sqrt{\frac{n+1}{n}} \operatorname{id}\right\|_\infty \leq \frac{1}{2}.
\label{eq:approx_tight_frame}$$ Then with probability at least $1 - \mathrm{e}^{-\gamma \tilde{C}_4 r}$, the matrix $W := \frac{1}{m}\sum_{j=1}^m A_j$ is positive definite and obeys $$\max \left\{ \| W \|_\infty,\ \| W^{-1} \|_\infty,\ \kappa (W) \right\} \leq 8.
\label{eq:W4design}$$ Here, $\tilde{C}_4 >1$ and $0 < \gamma \leq 1$ denote absolute constants of adequate size.
Note that condition (\[eq:approx\_tight\_frame\]) is slightly stronger than the corresponding condition in Theorem \[Th2\]. Also, the construction of $W$ again uses $t = \frac{1}{m} \left( 1 \ldots,1 \right)^T \in \mathbb{R}^m$.
In order to show this statement, we are going to employ the matrix Bernstein inequality[^5] [@tropp_user_2012 Theorem 6.1], see also [@ahwi02], in order to establish $$\left\| W - \sqrt{\frac{n+1}{n}} \operatorname{id}\right\|_\infty \leq \frac{3}{4}
\label{eq:W4design_aux1}$$ with high probability. Let $\lambda_1 (W),\ldots,\lambda_n (W)$ denote the eigenvalues of $W$. Then such a bound together with the definition of the operator norm assures $$\begin{aligned}
1 - \lambda_{\min}(W) &\leq \sqrt{\frac{n+1}{n}}- \lambda_{\min}(W) \leq \left| \sqrt{\frac{n+1}{n}} - \lambda_{\min}(W) \right| \leq \max_{ 1 \leq i \leq n} \left| \sqrt{\frac{n+1}{n}}- \lambda_i (W) \right| \\&= \left\| \sqrt{\frac{n+1}{n}}\operatorname{id}- W \right\|_\infty \leq 3/4, \\
\lambda_{\max}(W) - \sqrt{\frac{n+1}{n}} & \leq \left| \lambda_{\max}(W) - \sqrt{\frac{n+1}{n}} \right| \leq \max_{1 \leq i \leq n} \left| \sqrt{\frac{n+1}{n}} - \lambda_i (W) \right|\\ & = \left\| W - \sqrt{\frac{n+1}{n}}\operatorname{id}\right\|_\infty \leq 3/4.\end{aligned}$$ This in turn implies $
\lambda_{\min} (W) \geq 1/4
$ as well as $
\lambda_{\max}(W)\leq 3/4+\sqrt{\frac{n+1}{n}} \leq 2
$ for $n\geq 2$ and the desired bound readily follows.
It remains to assure the validity of with high probability. To this end, for $1 \leq k \leq m$, we define the random matrices $M_k := \frac{1}{m} \left( a_k a_k^* - \mathbb{E} \left[ a_k a_k^* \right] \right)$, where each $a_k$ is chosen independently at random from the weighted set $\left\{ p_i,w_i \right\}_{i=1}^N$. This definition assures $$\left\| W - \sqrt{\frac{n+1}{n}}\operatorname{id}\right\|_\infty
= \left\| \sum_{k=1}^m \big(M_k + \mathbb{E} \left[ a_k a_k^* \right]\big) - \sqrt{\frac{n+1}{n}} \operatorname{id}\right\|_\infty
\leq \left\| \sum_{k=1}^m M_k \right\|_\infty + \frac{1}{2}
\label{eq:W4design_aux2}$$ via the triangle inequality and assumption and along similar lines $$\left\| \mathbb{E} \left[ a_k a_k^* \right] \right\|_\infty \leq \frac{1}{2} + \sqrt{\frac{n+1}{n}} \leq 2$$ readily follows for any $1 \leq k \leq m$. The random matrices $M_k$ have mean-zero by construction and each of them obeys $$\begin{aligned}
\left\| M_k \right\|_\infty &= \frac{1}{m} \left\| a_k a_k^* - \mathbb{E} \left[ a_k a_k^* \right] \right\|_\infty
\leq \frac{1}{m} \max \left\{ \| a_k a_k^* \|_\infty, \| \mathbb{E} \left[ a_k a_k^* \right] \|_\infty \right\}
= \frac{1}{m}\| a_k \|_{\ell_2}^2 = \frac{\sqrt{(n+1)n}}{m},\end{aligned}$$ as well as $$\begin{aligned}
\left\| \mathbb{E} \left[ M_k^2 \right] \right\|_\infty
&= \frac{1}{m^2} \left\| \mathbb{E} \left[ \left( a_k a_k^* \right)^2 \right] - \mathbb{E} \left[ a_k a_k^* \right]^2 \right\|_\infty
= \frac{1}{m^2}\left\| \sqrt{(n+1)n} \mathbb{E } \left[ a_k a_k^* \right] - \mathbb{E} \left[ a_k a_k^* \right]^2 \right\|_\infty \\
&=\frac{2}{m^2} \max \left\{ \sqrt{(n+1)n} \left\| \mathbb{E} \left[ a_k a_k^* \right] \right\|_\infty, \left\| \mathbb{E} \left[ a_k a_k^* \right] \right\|_\infty^2 \right\}
\leq \frac{2 \sqrt{(n+1)n}}{m^2}. \end{aligned}$$ Hence $$\left\|\sum_{k=1}^m \mathbb{E} \left[ M_k^2 \right] \right\|_\infty\leq \frac{2 \sqrt{(n+1)n}}{m}.$$ These bounds allow us to set $R:= \frac{\sqrt{(n+1)n}}{m}$, $\sigma^2 := \frac{2 \sqrt{(n+1)n}}{m}$ and apply the matrix Bernstein inequality ([@tropp_user_2012 Theorem 6.1], [@ahwi02]) in order to establish $$\Pr \left[ \left\| \sum_{k=1}^m M_k \right\|_\infty \geq \tau \right] \leq n\; \mathrm{exp}\left(-\frac{\tau^2/2}{\sigma^2+R\tau} \right) \leq n \; \mathrm{exp} \left( - \frac{3 \tau^2 m}{16 \sqrt{(n+1)n}} \right)$$ for $0 < \tau \leq \sigma^2/R =2$. Setting $\tau = 1/4$ and inserting $m \geq \tilde{C}_4 n r \log(n) $ (where $\tilde{C}_4 $ is large enough) assures that holds with probability of failure smaller than $\mathrm{e}^{-\gamma \tilde{C}_4 r}$ via for a suitable $\gamma >0$.
Finally, we are ready to prove Theorem \[mainTh3\].
We content ourselves with establishing the design case and point out that the Gaussian case can be proved analogously (albeit with different constants). Fix $0 < \rho < 1/8$ and suppose that $$m \geq C_3 \left( 1 + \left( 1 + \rho^{-1} \right)^2 \right) n r \log n$$ measurement vectors have been chosen independently from an approximate 4-design. Theorem \[Th2\] then assures that the resulting measurement operator $\mathcal{A}$ obeys the robust Frobenius rank null space property with constants $\rho < 1/8$ and $\tau \leq \tilde{C}_6/ \sqrt{m}$ with probability at least $1 - \mathrm{e}^{-\tilde{C}_5 m}$. Likewise, Proposition \[prop:W4design\] assures that with probability at least $1 - \mathrm{e}^{-\gamma \tilde{C}_4 r}$, setting $t = \frac{1}{\sqrt{m}} (1,\ldots,1)^T \in \mathbb{R}^m$ leads to a positive definite $W = \sum_{j=1}^m t_j A_j$ obeying $\kappa (W) \leq 8$. Note that such a $t$ obeys $\| t \|_{\ell_2} = 1/\sqrt{m}$ and also $0 < \rho <1/8 \leq 1/ \kappa (W)$ holds by construction. The union bound over these two assertions failing implies that the requirements of Theorem \[th:NSPforPositiveMatrices\] are met with probability at least $$1 - \mathrm{e}^{-\tilde{C}_5 m} - \mathrm{e}^{-\gamma \tilde{C}_4 r} \geq 1 - \mathrm{e}^{- \tilde{\gamma} \tilde{C}_4 r},$$ where $\tilde{\gamma}$ denotes a sufficiently small absolute constant and $ \tilde{C}_4=m/n r \log n$. The constants $C_4$ and $s$ presented in Theorem \[mainTh3\] then amount to $s=\tilde{\gamma} \tilde{C}_4$ and $C_2\geq \tilde{C}_4$. Inserting $\| t \|_{\ell_2} = 1/\sqrt{m}$ and the bounds on $\| W \|_\infty, \| W^{-1} \|_\infty, \kappa (W)$ from Proposition \[prop:W4design\] into yields $$\begin{aligned}
{\Vert Z-X \Vert}_p & \leq\frac{2C\kappa(W)}{r^{1-1/p}}{\Vert X_c \Vert}_1+r^{1/p-1/2}{\Vert {\mathcal{A}}(Z)-{\mathcal{A}}(X) \Vert}_{\ell_2} {\Vert W^{-1} \Vert}_{\infty} {\left(\frac{C{\Vert t \Vert}_2}{\sqrt r}+D{\Vert W \Vert}_{\infty}\tau\right)} \\
&\leq \frac{16 C}{r^{1-1/p}} {\Vert X_c \Vert}_1 + 8 r^{1/p-1/2} \| \mathcal{A} (Z) - \mathcal{A}(X) \|_{\ell_2} \left( \frac{C}{\sqrt{rm}} + \frac{9D \tilde{C}_6}{\sqrt{m}} \right) \\
& \leq \frac{C_3}{r^{1-1/p}} {\Vert X_c \Vert}_1 + \frac{C_4 r^{1/p-1/2}}{\sqrt{m}} \left\| \mathcal{A} (Z) - \mathcal{A}(X) \right\|_{\ell_2}\end{aligned}$$ with constants $C_3 = 16 C$ and $C_4 = 8 C + 8 D \tilde{C}_6$ (where $C,D$ were introduced in Theorem \[th:NSPforPositiveMatrices\] and $\tilde{C}_6$ is ).
\[rem:quantum\_improvement\] In Corollary \[cor:tomography\] we focus on recovering density operators, i.e., positive semidefinite matrices $X$ with trace one. This trace constraint can be re-interpreted as an additional perfectly noiseless measurement $$b_0 = \operatorname{tr}\left( \operatorname{id}X \right) = \operatorname{tr}(X) = 1$$ corresponding to the measurement matrix $A_0 = \operatorname{id}$. Setting $t = (1,0,\ldots,0)^T \in \mathbb{R}^{m+1}$ in Theorem \[th:NSPforPositiveMatrices\] then leads to $W = \operatorname{id}$ which obeys $\| W \|_{\infty} = \| W^{-1} \|_\infty = \kappa (W) = 1$ and furthermore assures that the endomorphism is trivial, i.e. $\tilde{Z} = Z$ for all $Z \in \mathcal{H}_n$. Moreover, these properties render the estimate provided in Lemma \[Normdiff\] redundant, because any two density operators $X,Z$ obey $$\| \tilde{Z} \|_1 - \| \tilde{X} \|_1
= \| Z \|_1 - \| Z \|_1 = \operatorname{tr}\left( Z \right) - \operatorname{tr}\left( X \right) =0.$$ Such a refinement then allows for dropping the term containing $\| t \|_{\ell_2}$ in and by inserting $W = \operatorname{id}$ we arrive at the following conclusion: Any measurement operator $\mathcal{A}$ that obeys the Frobenius robust rank null space property with constants $0 < \rho <1 $ and $\tau >0$ assures for $1 \leq p \leq 2$ and any two density operators $X,Z$: $$\| Z - X\|_p \leq \frac{2 \left( 1 + \rho \right)^2}{1 - \rho} {\Vert X_c \Vert}_1 + \tau \frac{r^{1/p - 1/2}(3 + \rho)}{1 - \rho} \| \mathcal{A}(Z) - \mathcal{A} (X) \|_{\ell_2}.$$ Corollary \[cor:tomography\] then follows from combining this assertion with Theorem \[Th2\] and setting $p=1$.
Acknowledgements {#acknowledgements .unnumbered}
----------------
MK, HR and UT acknowledge funding by the European Research Council through the Starting Grant StG 258926. The work of RK is supported by the Excellence Initiative of the German Federal and State Governments (Grants ZUK 43 & 81), the ARO under contracts W911NF-14-1-0098 and W911NF-14-1-0133 (Quantum Characterization, Verification, and Validation), the Freiburg Research Innovation Fund, the DFG (GRO 4334 & SPP1798), and the State Graduate Funding Program of Baden-Württemberg.
Appendix {#appendix .unnumbered}
========
A brief review of finite-dimensional quantum mechanics {#a-brief-review-of-finite-dimensional-quantum-mechanics .unnumbered}
------------------------------------------------------
For the sake of being self-contained we briefly recapitulate crucial concepts of (finite dimensional) quantum mechanics without going too much into detail. For further reading on the topics introduced here, we defer the interested reader to [@nielsen_quantum_2010 Chapter 2.2].
An isolated quantum mechanical system is fully described by its *density operator*. For a finite $n$-dimensional quantum system, such a density operator corresponds to an Hermitian, positive semidefinite matrix $\rho$ with unit trace.
The most general notion of a measurement is that of a *positive operator-valued measure* (POVM). For an $n$-dimensional quantum system, a POVM corresponds to a collection $\mathcal{M}=\left\{ E_m \right\}_{m \in I}$ of positive semidefinite $n \times n$ matrices that sum up to identity, i.e., $$\sum_{m \in I} E_m = \operatorname{id}.$$ The indices $m\in I$ indicate the possible measurement outcomes of performing such a POVM measurement. Upon performing $\mathcal{M}$ on a system described by $\rho$, quantum mechanics then postulates that the probability of obtaining the outcome (labeled by) $m$ corresponds to $$p (m, \rho) = \operatorname{tr}\left( E_m \rho \right).$$ Repeating the same measurement (i.e., preparing $\rho$ and measuring $\mathcal{M}$) many times allows one to estimate the $n$ probabilities $p (\lambda_i, \rho )$ ever more accurately.
Note that the definitions of $\rho$ and $\mathcal{M}$ assure that ${p(m, \rho)}_{m \in I}$ is in fact a valid probability distribution. Indeed, $p(m,\rho) \geq 0$ follows from positive-semidefiniteness of both $\rho$ and $E_m$. Unit trace of $\rho$ assures proper normalization via $$\begin{aligned}
\sum_{m \in I} p (m, \rho ) = \sum_{m \in I} \operatorname{tr}\left( E_m \rho \right) = \operatorname{tr}\left( \operatorname{id}\rho \right) = \operatorname{tr}(\rho) = 1.\end{aligned}$$
[10]{}
R. [A]{}hlswede and A. [W]{}inter. trong converse for identification via quantum channels. , 48(3):569 –579, 2002.
A. [A]{}hmed and J. [R]{}omberg. ompressive multiplexing of correlated signals. , 61(1):479–498, 2015.
A. Ambainis and J. Emerson. . In [*[22nd Annual IEEE Conference on Computational Complexity, Proceedings]{}*]{}, pages [129–140]{}, [2007]{}.
D. [A]{}melunxen, M. [L]{}otz, M. B. [M]{}c[C]{}oy, and J. A. [T]{}ropp. iving on the edge: [P]{}hase transitions in convex programs with random data. , 3(3):224–294, 2014.
R. [Balan]{}, B. G. [Bodmann]{}, P. G. [Casazza]{}, and D. [Edidin]{}. , 15:488–501, 2009.
K. Banaszek, M. Cramer, and D. Gross. Focus on quantum tomography. , 15:125020, 2013.
R. Bhatia. . Springer, 1997.
S. [B]{}oyd and L. [V]{}andenberghe. ambridge [U]{}niv. [P]{}ress, 2004.
A. M. Bruckstein, M. Elad, and M. Zibulevsky. On the uniqueness of nonnegative sparse solutions to underdetermined systems of equations. , 54(11):4813–4820, 2008.
E. [C]{}and[è]{}s, T. [S]{}trohmer, and V. [V]{}oroninski. hase[L]{}ift: [E]{}xact and stable signal recovery from magnitude measurements via convex programming. , 66:1241–1274, 2013.
E. [C]{}andes and T. [T]{}ao. he power of convex relaxation: near-optimal matrix completion. , 56(5):2053–2080, 2010.
E. J. Cand[è]{}s, Y. C. Eldar, T. Strohmer, and V. Voroninski. Phase retrieval via matrix completion. , 6:199–225, 2013.
E. J. Cand[è]{}s and X. Li. Solving quadratic equations via [PhaseLift]{} when there are about as many equations as unknowns. , pages 1–10, 2013.
E. J. Cand[è]{}s, X. Li, and M. Soltanolkotabi. Phase retrieval via [W]{}irtinger flow: [T]{}heory and algorithms. , 61(4):1985–2007, 2015.
E. J. Cand[è]{}s and Y. Plan. Tight oracle inequalities for low-rank matrix recovery from a minimal number of noisy random measurements. , 57(4):2342–2359, 2011.
E. J. [C]{}and[è]{}s and B. [R]{}echt. xact matrix completion via convex optimization. , 9:717–772, 2009.
A. Carpentier, J. Eisert, D. Gross, and R. Nickl. Uncertainty quantification for matrix compressed sensing and quantum tomography problems. , 2015.
V. [C]{}handrasekaran, B. [R]{}echt, P. [P]{}arrilo, and A. [W]{}illsky. he convex geometry of linear inverse problems. , 12(6):805–849, 2012.
Y. Chen, Y. Chi, and A. Goldsmith. Exact and stable covariance estimation from quadratic sampling via convex programming. , 61(7):4034–4059, July 2015.
Z. Chen and J. J. Dongarra. Condition numbers of [G]{}aussian random matrices. , 27(3):603–620, 2005.
L. Demanet and P. Hand. Stable optimizationless recovery from phaseless linear measurements. , 20(1):199–221, 2014.
M. Fazel. Matrix rank minimization. , 2002.
C. Ferrie and R. Kueng. Have you been using the wrong estimator? [T]{}hese guys bound average fidelity using this one weird trick von [N]{}eumann didn’t want you to know. , 2015.
S. T. Flammia, D. Gross, Y.-K. Liu, and J. Eisert. Quantum tomography via compressed sensing: error bounds, sample complexity and efficient estimators. , 14:095022, 2012.
M. [F]{}ornasier, H. [R]{}auhut, and R. [W]{}ard. ow-rank matrix recovery via iteratively reweighted least squares minimization. , 21(4):1614–1640, 2011.
S. Foucart and H. Rauhut. . Applied and Numerical Harmonic Analysis. Birkhäuser/Springer, New York, 2013.
Y. Gordon. On [M]{}ilman’s inequality and random subspaces which escape through a mesh in [${\mathbb R}^n$]{}. In [*Geometric aspects of functional analysis (1986/87)*]{}, volume 1317 of [*Lecture Notes in Math.*]{}, pages 84–106. Springer, Berlin, 1988.
D. Gross, F. Krahmer, and R. Kueng. A partial derandomization of [P]{}haselift using spherical designs. , pages 1–38, 2014.
D. [G]{}ross, F. [K]{}rahmer, and R. [K]{}ueng. mproved recovery guarantees for phase retrieval from coded diffraction patterns. , 2015.
D. Gross, Y.-K. Liu, S. T. Flammia, S. Becker, and J. Eisert. State tomography via compressed sensing. , 105:150401, 2010.
T. Heinosaari, L. Mazzarella, and M. M. Wolf. Quantum tomography under prior information. , 318(2):355–374, 2013.
C. W. Helstrom. Quantum detection and estimation theory. , 1(2):231–252, 1969.
R. Horn and C. Johnson. . Cambridge University Press, Cambridge, 1991.
M. Kabanava, H. Rauhut, and U. Terstiege. Analysis of low rank matrix recovery via mendelson’s small ball method. In [*11th international conference on Sampling Theory and Applications (SampTA 2015)*]{}, Washington, USA, May 2015.
M. Kabanava, H. Rauhut, and U. Terstiege. On the minimal number of measurements in low-rank matrix recovery. In [*11th international conference on Sampling Theory and Applications (SampTA 2015)*]{}, Washington, USA, May 2015.
A. Kalev, C. Riofrio, R. Kosut, and I. Deutsch. Informationally complete measurements from compressed sensing methodology. , 60, 2015.
M. Kech, P. Vrana, and M. Wolf. The role of topology in quantum tomography. , 2015.
M. Kech and M. Wolf. From quantum tomography to phase retrieval and back. In [*11th international conference on Sampling Theory and Applications (SampTA 2015)*]{}, Washington, USA, May 2015.
M. [K]{}ech and M. M. [W]{}olf. uantum tomography of semi-algebraic sets with constrained measurements. , 2015.
V. Koltchinskii and S. Mendelson. Bounding the smallest singular value of a random matrix without concentration. , page rnv096, 2015.
R. Kueng, S. Daniel, and D. Gross. Direct characterization of linear-optical networks via [PhaseLift]{}. in preparation, 2015.
R. Kueng, D. Gross, and F. Krahmer. Spherical designs as a tool for derandomization: The case of [PhaseLift]{}. In [*11th international conference on Sampling Theory and Applications (SampTA 2015)*]{}, Washington, USA, May 2015.
R. [K]{}ueng, H. [R]{}auhut, and U. Terstiege. ow rank matrix recovery from rank one measurements. , 2014.
R. [L]{}ata[ł]{}a. Some estimates of norms of random matrices. , 133(5):1273–1282, 2005.
M. [L]{}edoux and M. [T]{}alagrand. pringer-[V]{}erlag, 1991.
K. [L]{}ee and Y. [B]{}resler. i[R]{}[A]{}: [A]{}tomic decomposition for minimum rank approximation. , 56(9):4402 – 4416, 2010.
Y.-K. Liu. Universal low-rank matrix recovery from [P]{}auli measurements. , pages 1638–1646, 2011.
W. Matthews, S. Wehner, and A. Winter. Distinguishability of quantum states under restricted families of measurements with an application to quantum data hiding. , 291(3):813–843, 2009.
M. B. [M]{}c[C]{}oy and J. A. [T]{}ropp. harp recovery bounds for convex demixing, with applications. , 14(3):503–567, 2014.
S. [M]{}endelson. earning without [C]{}oncentration. , 62(3):1–25, 2015.
K. Mohan and M. Fazel. Iterative reweighted least squares for matrix rank minimization. In [*Proceedings of the Allerton Conference*]{}, pages 653–661, 2010.
K. Mohan and M. Fazel. New restricted isometry results for noisy low-rank recovery. In [*Proc. International Symposium Information Theory*]{}, 2010.
M. A. Nielsen and I. L. Chuang. . Cambridge [U]{}niversity [P]{}ress, 2010.
S. Oymak, K. Mohan, M. Fazel, and B. Hassibi. A simplified approach to recovery conditions for low rank matrices. In [*Information Theory Proceedings (ISIT), 2011 IEEE International Symposium on*]{}, pages 2318–2322, July 2011.
N. [P]{}arikh and S. [B]{}oyd. roximal algorithms. , 1(3):123–231, 2014.
B. Recht, M. Fazel, and P. A. Parrilo. Guaranteed minimum-rank solutions of linear matrix equations via nuclear norm minimization. , 52(3):471–501, 2010.
B. Recht, W. Xu, and B. Hassibi. Necessary and sufficient conditions for success of the nuclear norm heuristic for rank minimization. In [*Proc. 47th IEEE Conference on Decision and Control*]{}, pages 3065–3070, 2008.
B. Recht, W. Xu, and B. Hassibi. Null space conditions and thresholds for rank minimization. , Ser B, 127:175–211, 2011.
M. Rudelson and R. Vershynin. On sparse reconstruction from fourier and gaussian measurements. , 61(8):1025–1045, 2008.
P. Schindler, D. Nigg, T. Monz, J. T. Barreiro, E. Martinez, S. X. Wang, S. Quint, M. F. Brandl, V. Nebendahl, C. F. Roos, et al. A quantum information processor with trapped ions. , 15(12):123012, 2013.
B. Sch[ö]{}lkopf and A. J. Smola. . MIT press, 2002.
A. [Scott]{}. , 39:13507–13530, 2006.
M. Slawski, P. Li, and M. Hein. Regularization-free estimation in trace regression with symmetric positive semidefinite matrices. , 2015.
J. [T]{}anner and K. [W]{}ei. ormalized iterative hard thresholding for matrix completion. , 59(11):7491–7508, 2013.
J. A. Tropp. User-friendly tail bounds for sums of random matrices. , 12(4):389–434, 2012.
J. A. [T]{}ropp. onvex recovery of a structured signal from independent random linear measurements. , 2014.
R. [V]{}ershynin. ntroduction to the non-asymptotic analysis of random matrices. In Y. [E]{}ldar and G. [K]{}utyniok, editors, [*[C]{}ompressed [S]{}ensing: [T]{}heory and [A]{}pplications*]{}, pages 210–268. [C]{}ambridge [U]{}niv [P]{}ress, 2012.
Z. [X]{}u. he minimal measurement number for low-rank matrices recovery. , 2015.
[^1]: Nowadays, experimentalists are able to create and control multi-partite systems of overall dimension $n = 2^8$ in their laboratories [@schindler_quantum_2013]. This results in a density operator of size $256 \times 256$ (a priori $65\;536$ parameters).
[^2]: In fact by resorting to the Frobenius norm bound in Theorem \[mainTh3\] (instead of the nuclear norm bound employed to arrive at Corollary \[cor:tomography\]), one obtains a performance guarantee that strongly resembles [@liu_universal_2011 Equation (8)] — the main recovery guarantee in that paper.
[^3]: Haar random vectors are vectors drawn uniformly from the complex unit sphere in $\mathbb{C}^n$. They can be obtained from complex standard Gaussian vectors by rescaling them to unit length. Property is invariant under such a re-scaling and Theorem \[mainTh1\] therefore assures rank-$r$ IC for both Gaussian and Haar random vectors.
[^4]: i.e., the $\varepsilon_j$ are independent and assume the values $1$ and $-1$ with probability $1/2$, respectively.
[^5]: Resorting to the matrix Chernoff inequality would allow for establishing a similar result. However, in the case of an exact tight frame, the numerical constants obtained by doing so are slightly worse.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
The circumference $c(G)$ of a graph $G$ is the length of a longest cycle. By exploiting our recent results on resistance of snarks, we construct infinite classes of cyclically $4$-, $5$- and $6$-edge-connected cubic graphs with circumference ratio $c(G)/|V(G)|$ bounded from above by $0.876$, $0.960$ and $0.990$, respectively. In contrast, the dominating cycle conjecture implies that the circumference ratio of a cyclically $4$-edge-connected cubic graph is at least $0.75$.
In addition, we construct snarks with large girth and large circumference deficit, solving Problem 1 proposed in \[J. Hägglund and K. Markström, On stable cycles and cycle double covers of graphs with large circumference, Disc. Math. 312 (2012), 2540–2544\].
author:
- |
Edita Máčajová${}^1$, Ján Mazák${}^2$\
\
[{macajova, mazak}@dcs.fmph.uniba.sk]{}\
${}^1$ Univerzita Komenského, Mlynská dolina, 842 48 Bratislava, Slovakia\
${}^2$ Trnavská univerzita, Priemyselná 4, 918 43 Trnava, Slovakia
title: Cubic graphs with large circumference deficit
---
[**Keywords:**]{} circumference, cubic graph, snark, girth
[**Classification:**]{} 05C15, 05C38
Introduction
============
A cycle is one of the most basic structures in a graph, so it comes as no surprise that cycles have been analysed from the very beginnings of graph theory. This article focuses on longest cycles in cubic graphs. The [*circumference*]{} $c(G)$ of a graph $G$ is the length of a longest cycle. The [*circumference ratio*]{} is the ratio of circumference to order. The [*circumference deficit*]{} is the difference between order and circumference.
A lot of attention was given to Hamiltonian graphs, that is, graphs with zero circumference deficit. Compared to the vast tomes written on hamiltonicity, non-Hamiltonian graphs appear rather neglected despite there is plenty of investigation to be done. The problem of determining the circumference of a given graph is NP-hard and even approximation is a very tough problem [@NP], so no simple characterisations are expected.
It transpired in many areas that the most interesting cubic graphs are those with chromatic index four. Such graphs are called [*snarks*]{}; we will additionally require snarks to have girth at least five and cyclic edge-connectivity at least four. We will encounter them in Section \[snarks\] where we prove the existence of a snark of girth at least $g$ and circumference deficit at least $g$ for every integer $g$. Section \[ratio\] is devoted to upper bounds on circumference ratio; we provide linear bounds for certain classes of cyclically $4$-, $5$-, and $6$-edge-connected cubic graphs. These bounds serve as the best presently known upper bounds on general lower bounds on circumference ratio.
Each subcubic graph can be transformed into a $3$-edge-colourable graph by removing sufficiently many edges. The least number of edges that need to be removed is the [*resistance*]{} of the graph. There are snarks with arbitrarily large resistance (see e.g. [@steffen; @oddness]). Most of our constructions are based on building blocks with large resistance; the usefulness of such blocks is demonstrated in Lemma \[lemma1\] which plays the key role in our proofs.
Circumference ratio of cubic graphs {#ratio}
===================================
Circumference ratio of cubic graphs strongly depends on connectivity. Since each vertex of a cubic graph is separated by three edges from the rest of the graph, the classical notions of vertex-connectivity and edge-connectivity are of limited use. A refined measure of connectivity, much more appropriate for our purpose, is provided by cyclic edge-connectivity. A graph is [*cyclically $k$-edge-connected*]{} if at least $k$ edges must be removed to disconnect it into components among which there are at least two containing a cycle. For cubic graphs, the notion of cyclic $k$-edge-connectivity coincides with $k$-vertex-connectivity and $k$-edge-connectivity for $k\in\{1,2,3\}$ [@atoms].
If we allow bridges in our graphs, there are infinitely many trivial cubic graphs with circumference $5$. What is more interesting, Bondy and Entringer [@be3] proved that every $2$-edge-connected cubic graph $G$ contains a cycle of length at least $4\log |V(G)|-4\log \log |V(G)|-20$. This bound is essentially best possible, as shown by Lang and Walther [@lw10]. Bondy and Simonovits [@bs5] conjectured the existence of a constant $c$ such that every $3$-connected cubic graph $G$ has circumference at least $|V(G)|^c$ and showed that $c \le \log_98\approx 0.946$. The conjecture was verified by Jackson [@jackson] for $c = \log (1+\sqrt5)-1\approx 0.694$; the constant $c$ has recently been improved to $0.753$ [@bbmy]. Bondy also conjectured the following.
There exists a constant $c > 0$ such that every cyclically $4$-edge-connected cubic graph $G$ has circumference at least $c\,|V(G)|$. \[conj:bondy\]
The dominating cycle conjecture [@fleischner] implies that $c\ge 0.75$ since a dominating cycle in a cubic graph $G$ has length at least $0.75|V(G)|$. In addition, Thomassen [@thomassen] conjectured that there exists an integer $k$ such that every cyclically $k$-edge-connected cubic graph is Hamiltonian. This would mean $c = 1$ for sufficiently connected cubic graphs.
We summarize the currently known results together with our contributions in Table \[tabulka\]. The column LB displays a lower bound which holds for all graphs; the column UB shows an upper bound on circumference for a certain infinite class of graphs with required cyclic edge-connectivity. All the bounds are asymptotic; each of them is expressed as a function of order $n$.
connectivity LB conjectured LB UB
-------------- ------------- ---------------- ------------------
2 $\log n$ $\log n$
3 $n^{0.753}$ $n^{0.946}$
4 $n^{0.753}$ $0.75n$ $0.875n$ $\star$
5 $n^{0.753}$ $0.75n$ $0.960n$ $\star$
6 $n^{0.753}$ $0.75n$ $0.990n$ $\star$
7+ $n^{0.753}$ $0.75n$ $n$ $n$
: Summary of results on circumference (our contribution are marked by $\star$).[]{data-label="tabulka"}
The crucial observation used in our construction of graphs with large circumference deficit is captured in the following lemma.
\[lemma1\] Let $H$ be a subgraph of a bridgeless cubic graph $G$. If $H$ has resistance $k$, then any cycle of $G$ not contained in $H$ misses at least $k$ vertices of $H$.
If a cycle $C$ of $G$ is not contained in $H$, then its intersection with $H$ is a union of vertex-disjoint paths. Take each of those paths in turn and alternately colour the edges along the path by colours $1$ and $2$. Colour all the remaining edges by $3$. There are two types of vertices in $H$: those that have their incident edges coloured by $1$, $2$, $3$ and those with all incident edges coloured by $3$. We call the vertices of the latter type [*bad*]{}. By removing all the bad vertices we obtain a $3$-edge-colourable graph from $H$. However, $H$ has resistance $k$, and thus there are at least $k$ bad vertices. Obviously, no bad vertex belongs to $C$, hence $C$ misses at least $k$ vertices of $H$.
In order to construct infinite classes of graphs with circumference ratio promised in Table \[tabulka\], we employ cubic construction blocks described in [@oddness]. (We do not include construction details or proofs of their properties in this article; an interested reader can find them in [@oddness].)
![The building block $N_2$ with $26$ vertices and resistance $2$.[]{data-label="fig:N2"}](bloky-1)
The cubic graph $N_2$ has order $26$, resistance $2$, and two pairs of dangling edges (see Figure \[fig:N2\] and [@oddness Section 7]). We create a graph $G$ by arranging $m\ge 2$ copies of $N_2$ along a circle and for each copy $K$, we connect one pair of dangling edges of $K$ to a pair of dangling edges of the following copy and the other pair of dangling edges to a pair in the previous copy. (The exact way of how we do it does not matter because we only need to preserve cyclic $4$-edge-connectivity.) According to Lemma \[lemma1\], a cycle of $G$ either belongs to one copy of $N_2$ or misses at least $2$ vertices in each of the $m$ copies of $N_2$, hence $G$ has circumference deficit at least $2m$ and circumference ratio at most $24m/26m = 12/13\approx 0.92$. A different idea used in Theorem \[thm7/8\] leads to a better upper bound for cyclically $4$-edge-connected graphs.
The construction described in the previous paragraph can be repeated with the cyclically $5$-edge-connected building block $Z$ with order $25$ and resistance $1$ (see [@oddness Section 8 and Figure 5] for a description of $Z$; this block was also used by Steffen under the name $T$ [@steffen Theorem 2.3]). The graph $Z$ has seven dangling edges naturally split into two triples and one single dangling edge. We repeat the circular construction with $m$ copies of $Z$ instead of $N_2$; the role of the pairs of dangling edges is now played by the triples. The single dangling edges are joined to a cycle of length $m$ (one dangling edge to each vertex of the cycle). The resulting graph has circumference ratio at most $24/25 = 0.96$ and is cyclically $5$-edge-connected.
A similar construction can also be used to construct cyclically $6$-edge-connected graphs; however, the details are more complicated. Section 9 of [@oddness] describes a cyclically $6$-edge-connected graph $M_r$ of order $99r$ with resistance at least $r$ for each even positive integer $r$, but there are no dangling edges in this graph, thus we cannot use it directly: we first have to cut a few suitable edges to obtain a block which would allow a construction of cyclically $6$-edge-connected graphs.
The graph $M_r$ is obtained by symmetrically applying superposition to a graph $L_r$ composed of $r$ circularly arranged isomorphic copies of the block $P_3$ (that is, the Petersen graph with one vertex removed). Therefore, $M_r$ also contains $r$ isomorphic blocks $A_1, A_2, \dots, A_r$ arranged along a circle. Each two consecutive blocks of $M_r$ are joined by three edges. We cut all the edges between $A_1$ and $A_2$ to form a cubic graph $M_r'$ with two triples of dangling edges. The graph $M_r'$ has order $99r$ and resistance at least $r-3$. (According to the definition of resistance, the removal of an edge can decrease resistance by at most $1$. The resistance of $M_r'$ is actually $r$, but that would require a detailed proof; we will use the obvious lower bound of $r-3$ here because it is sufficient for our purpose.)
Consequently, we can use $M_r'$ in place of $N_2$ in the above-described construction (with triples of dangling edges instead of pairs). The resulting cyclically $6$-edge-connected cubic graph $G$ has order $m\cdot 99r$ and circumference deficit at least $m(r-3)$, thus its circumference ratio is at most $1-(r-3)/99r$. By taking a sufficiently large $r$ we can make this ratio to be arbitrarily close to $98/99\approx 0.990$.
For each integer $m$, there exists a cyclically $4$-edge-connected cubic graph with order $8m$ and circumference $7m+2$. \[thm7/8\]
Let $u$ and $v$ be two adjacent vertices of the Petersen graph $P$. We remove the path $uv$, but keep the dangling edges incident to exactly one of its endvertices; the dangling edges incident to $u$ are [*input edges*]{} and the dangling edges incident to $v$ are [*output edges*]{}. We say that a path [*passes through*]{} $B$ if it starts with a vertex incident to an input edge and ends in a vertex incident to an output edge. We say that a cycle passes through $B$ if a portion of this cycle (a path) passes through $B$. The resulting graph $B$ has two properties interesting to us.
First, if a path passes through $B$, it cannot pass through all the vertices of $B$: otherwise we would be able to extend this path by $u$ and $v$ to a Hamiltonian cycle of $P$, but $P$ has no such cycle. Second, if we take two disjoint paths passing through $B$, there is at least one vertex of $B$ missed by both of these paths. Otherwise, we can extend the first path by $u$, extend the other path by $v$, and then concatenate them together by adding two edges to form a Hamiltonian cycle of $P$ which is a contradiction.
Let $G$ be the graph obtained from $m$ copies of $B$ arranged along a circle in such a way that the output edges of each copy are identified with the input edges of the following copy (see Figure \[fig:chain\]). The graph $G$ is cyclically $4$-edge-connected and has order $8m$.
![The building block $B$ and a sketch of the graph $G$.[]{data-label="fig:chain"}](obr-1 "fig:") 1.5cm
Let $C$ be a cycle in $G$. Note that $C$ passes through each copy of $B$ at most twice. According to the two properties of $B$ proved above, no matter how many times $C$ passes through $B$, at least one vertex of $B$ is missed. The only possibility for $C$ to contain all vertices of $B$ is to enter by an input edge and then leave by the other input edge (of course, it can also both enter and leave by output edges, which is essentially the same situation). Consequently, the cycle $C$ misses at least one vertex in each of at least $m-2$ copies of $B$, and thus the circumference of $G$ is at most $8m-(m-2) = 7m+2$. Since $G$ contains a cycle of length $7m+2$, the derived upper bound on its circumference is tight.
We propose the following strengthening of Conjecture \[conj:bondy\].
Every cyclically $4$-edge-connected cubic graph has circumference ratio at least $7/8$.
Large girth and large circumference deficit {#snarks}
===========================================
This section is motivated by the cycle double cover conjecture (CDCC). Huck [@huck] showed that the smallest possible counterexample to CDCC has girth at least $12$. Brinkmann et al. [@bbb; @hm] proved that if a bridgeless cubic graph $G$ has a cycle of length at least $|V(G)|-10$, then $G$ has a cycle double cover. Put together, smallest counterexamples to CDCC can only be found in the class of snarks with girth at least $12$ and with circumference deficit at least $11$. Since no such snark has been known before, the following problem is very relevant.
[**Problem**]{} (Hägglund and Markström [@hm]). For each integer $g$, construct a snark of girth at least $g$ and circumference deficit at least $g$.
We solve this problem in Theorem \[thm3\]. The construction used in the proof of Theorem \[thm3\] can be modified to produce snarks with arbitrarily large girth and linear circumference deficit.
For every integer $g$ there exists a snark with girth at least $g$ and circumference deficit at least $g$. \[thm3\]
We construct the desired snark for every integer $g\ge 5$ which is enough to prove the theorem.
Let $H_0$ be a snark of girth at least $g$; the existence of such snarks has been proved by Kochol [@kochol]. Let $v_1$ and $v_2$ be two adjacent vertices of $H_0$ and let $e_i$ and $f_i$, for $i\in\{1,2\}$, be the two edges incident to $v_i$ and not incident to $v_{3-i}$. Let $H_1$ be the cubic graph obtained from $H_0$ by removing $v_1$ and $v_2$ while keeping the dangling edges $e_1$, $e_2$, $f_1$, $f_2$. The well-known parity lemma assures that the edges $e_1$ and $f_1$ have the same colour in every $3$-edge-colouring of $H_1$ (otherwise $H_0$ would be $3$-edge-colourable, but it is a snark).
![The graph $H$.[]{data-label="fig:H_1"}](obr-3)
Take two copies of $H_1$ and join them as indicated in Fig. \[fig:H\_1\] (the edge $f_1$ of the first copy is identified with the edge $f_1$ of the second copy and the edges $e_1$ of both copies are attached to an additional vertex $v$). If $H$ was $3$-edge-colourable, then the colour of $e_1$ of the first copy of $H_1$ would be the same as the colour of $f_1$ and, in turn, the same as the colour of $e_1$ of the second copy, leading to a contradiction at $v$. Hence, $H$ is not colourable and has resistance at least $1$. Moreover, there is no cycle of length less than $g$ in $H$, and any path with endvertices incident to dangling edges of $H$ passes through at least $g-1$ vertices of $H$.
Let $G$ be a cubic graph obtained from $g$ copies of $H$ arranged along a circuit in such a way that two dangling edges of a copy of $H$ are attached to the previous copy and two of them are attached to the next. The remaining $g$ edges can be joined to a cycle of length $g$ in an arbitrary way preserving maximum degree $3$. The graph $G$ clearly has girth at least $g$, is cyclically $4$-edge-connected and is not $3$-edge-colourable. Any cycle of $G$ not contained in $H$ misses at least one vertex in each copy of $H$ thanks to Lemma \[lemma1\], and thus $G$ has circumference deficit at least $g$.
[**Acknowledgements.**]{} This work was supported from the APVV grants APVV-0223-10 and ESF-EC-0009-10 within the EUROCORES Programme EUROGIGA (project GReGAS) of the European Science Foundation.
[99]{}
M. Bilinski, B. Jackson, J. Ma, X. Yu, Circumference of 3-connected claw-free graphs and large Eulerian subgraphs of 3-edge-connected graphs, J. of Comb. Theory Ser. B 101 (2011), 214–236.
J. A. Bondy and R. C. Entringer, Longest cycles in 2-connected graphs with prescribed maximum degree, Canad. J. Math. 32 (1980), 1325–1332.
J. A. Bondy and M. Simonovits, Longest cycles in 3-connected cubic graphs, Canad. J. Math. 32 (1980), 987–992.
G. Brinkmann, J. Goedgebeur, J. Hägglund, K. Markström, Generation and properties of snarks, J. Comb. Theory Ser. B 103 (4) (2013), 468–488.
G. Chen, J. Xu, and X. Yu, Circumference of graphs with bounded degree, SIAM J. Comput. 33 (5) (2004), 1136–1170.
H. Fleischner, Some blood, sweat, but no tears in Eulerian graph theory, Congr. Numer. 63 (1988), 8–48.
J. Hägglund, K. Markström, On stable cycles and cycle double covers of graphs with large circumference, Disc. Math. 312 (2012) 2540–2544.
A. Huck, Reducible configurations for the cycle double cover conjecture, in: Proceedings of the 5th Twente Workshop on Graphs and Combinatorial Optimization (Enschede, 1997), vol. 99 (2000), 71–90.
B. Jackson, Longest cycles in 3-connected cubic graphs, J. Comb. Theory Ser. B 41 (1986), 17–26.
M. Kochol, Snarks without small cycles, J. Combin. Theory Ser. B 67 (1996), 34–47.
R. Lang and H. Walther, Über längste Kreise in regulären Graphen, in “Beitrage zur Graphenteorie, Kolloquium, Manebach 1967”, Teubner, Leipzig (1968), 91–98.
R. Lukoťka, E. Máčajová, J. Mazák, and M. Škoviera, Small snarks with large oddness, [arXiv:1212.3641]{}.
R. Nedela and M. Škoviera, Atoms of cyclic connectivity in cubic graphs, Math. Slovaca 45 (1995), 481–499.
E. Steffen, Measurements of edge-uncolorability, Disc. Math. 280 (2004), 191–214.
C. Thomassen, personal communication, a conference in Vienna (1991).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Directional data are constrained to lie on the unit sphere of ${\mathbb{R}}^q$ for some $q\geq 2$. To address the lack of a natural ordering for such data, depth functions have been defined on spheres. However, the depths available either lack flexibility or are so computationally expensive that they can only be used for very small dimensions $q$. In this work, we improve on this by introducing a class of distance-based depths for directional data. Irrespective of the distance adopted, these depths can easily be computed in high dimensions too. We derive the main structural properties of the proposed depths and study how they depend on the distance used. We discuss the asymptotic and robustness properties of the corresponding deepest points. We show the practical relevance of the proposed depths in two applications, related to (i) spherical location estimation and (ii) supervised classification. For both problems, we show through simulation studies that distance-based depths have strong advantages over their competitors.'
author:
- |
Giuseppe Pandolfo$^*$, Davy Paindaveine$^\dagger$ and Giovanni Porzio$^\ddagger$\
[University of Naples Federico II$^*$, Université libre de Bruxelles$^\dagger$]{},\
[and University of Cassino and Southern Lazio$^\ddagger$]{}
bibliography:
- 'ManuscriptRevised.bib'
title: 'Distance-based Depths for Directional Data'
---
Introduction
============
Directional data analysis is relevant when the sample space is the unit hypersphere $\mathcal{S}^{q-1}\linebreak :=\left\{x \in \mathbb{R}^{q}:x^{T}x = 1\right\}$ in $\mathbb{R}^{q}$, which occurs when observations are directions, axes, rotations, or cyclic events. Applications arise in numerous fields, including astronomy, earth sciences, biology, meteorology and political science; see [@GiHa2010] for an exemple in the latter field. Directional data analysis can also be exploited to study patterns of unit vectors in $\mathbb{R}^{q}$, such as those encountered in text mining [@Hoetal2012].
Statistically, analyzing and describing directional data requires tackling some interesting problems associated with the lack of a reference direction and with a sense of rotation not uniquely defined. Another important issue when dealing with such data is the lack of a natural ordering, which generates a special interest in depth functions on the sphere. Parallel to their role in the usual Euclidean case, directional depths are to measure the degree of centrality of a given spherical location with respect to a distribution on the sphere and to provide a center-outward ordering of spherical locations; see [@AgoRom2013a].
Depth concepts for directional data were first considered by [@Sma1987] and [@LiuSin1992]. Following the pioneering work of [@Sma1987], [@LiuSin1992] popularized the concept of *angular Tukey depth* (ATD), which is the directional analog of the celebrated *halfspace depth* [@Tuk1975]. The same paper introduced two further depths for directional data, namely the *angular simplicial depth* (ASD), which is the directional version of the *simplicial depth* from [@Liu1990], and the *arc distance depth* (ADD), which is based on the concept of arc length distance.
Unlike the ADD, the ATD and ASD have been studied and used in the literature. For instance, [@RouStr2004] investigated some of the properties of the ATD, while [@AgoRom2013a] considered some of the possible applications of the ASD and ATD. packages are also available for these depths: the package ([@Genetal2012]) allows to compute ATD values for $q = 2$ or $3$, whereas the package ([@AgoRom2013b]) implements specific functions for the evaluation of the ATD for $q=2$, and of the ASD for an arbitrary $q\geq 2$.
The main drawback of both the ASD and ATD is the computational effort they require, especially for higher dimensions $q$. The *angular Mahalanobis depth* of [@Leyetal2014], that is based on a concept of directional quantiles, is computationally much more affordable, but suffers from other disadvantages: it requires the preliminary choice of a spherical location functional and it is less flexible than the ASD/ATD in the sense that it produces rotationally symmetric depth contours, even if the underlying distribution is not rotationally symmetric.
On the one hand, depth functions for directional data are useful, yet on the other hand, they lack flexibility (and depend on some user’s choice) or are computationally too demanding. In order to improve on this, this work introduces a new class of directional depth functions that is based on spherical distances and contains the ADD as a particular case. These depth functions are computationally feasible even in high dimensions and are generally more flexible. Distance-based directional depths show several other advantages over their ASD/ATD competitors: they take positive values everywhere on $\mathcal{S}^{q-1}$ (but in the uninteresting case of a point mass distribution), whereas the ASD/ATD can take zero values (which is undesirable when performing supervised classification). Further advantages of the proposed distance-based depths is that they typically do not provide ties in the sample case (whereas ties are unavoidable for the ASD/ATD, due to their step function nature) and that they do not require any assumption on the underlying distribution (unlike the angular Mahalanobis depth that, when based on the spherical mean, is not defined for zero-mean distributions).
The paper is organized as follows. In Section \[secdef\], we introduce the proposed class of distance-based depth functions for directional data, and we consider three particular cases, namely the arc distance depth (ADD), the cosine distance depth (CDD) and the chord distance depth (ChDD). In Section \[sec:StructProperties\], we derive the main structural properties of the proposed depths and study how they depend on the distance used. In Section \[secillu\], we compare the various depths considered for several empirical distributions on the circle ($q=2$), which also allows us to illustrate the theoretical results of Section \[sec:StructProperties\]. In Section \[sec:DistrProperties\], we discuss the asymptotic and robustness properties of the proposed concepts. In Section \[secSimu\], we show the practical relevance of the distance-based depths in two applications, related to (i) spherical location estimation (Section \[secSimusub1\]) and (ii) supervised classification (Section \[secSimusub2\]). For both problems, we perform simulations that show the advantages of the proposed depths over their competitors. Final comments are provided in Section \[secfinal\]. Finally, an appendix collects technical proofs.
Distance-based depths for directional data {#secdef}
==========================================
In Definition \[defclass\] below, we introduce a class of depths on the unit sphere $\mathcal{S}^{q-1}$. A particular member of this class will be obtained by fixing a particular (bounded) distance $d(\cdot,\cdot)$ on $\mathcal{S}^{q-1}$. For such a distance, $d^{\rm sup}:=\sup \{ d(\theta, \psi): \theta,\psi\in \mathcal{S}^{q-1}\}$ will throughout denote the upper bound of the distance between any two points on $\mathcal{S}^{q-1}$.
\[defclass\] Let $d(\cdot,\cdot)$ be a bounded distance on $\mathcal{S}^{q-1}$ and $H$ be a distribution on $\mathcal{S}^{q-1}$. Then the *directional $d$-depth of $\theta(\in \mathcal{S}^{q-1})$ with respect to $H$* is $$\begin{aligned}
\label{eq:class}
D_{d}\left(\theta, H\right)
:=
d^{\rm sup}
-
E_H[d(\theta, W)]
, \end{aligned}$$ where $E_H$ is the expectation under the assumption that $W$ has distribution $H$.
While, in principle, any distance $d$ can be used in this definition, it is natural to consider distances that are *rotation-invariant* in the sense that $d(O\theta,O\psi)=d(\theta,\psi)$ for any $\theta,\psi\in\mathcal{S}^{q-1}$ and any $q\times q$ orthogonal matrix $O$. As we show for the sake of completeness in the appendix (see Proposition \[prodistinv\]), any rotation-invariant distance $d$ is of the form $$d(\theta,\psi)=d_{\delta}(\theta,\psi)=\delta(\theta'\psi)$$ for some function $\delta:[-1,1]\to{\mathbb{R}}^+$. The standard distance axioms impose that $\delta(1)=0$ but do not impose that $\delta$ is monotone non-increasing (unexpectedly, the triangle inequality may hold without this monotonicity condition). All classical choices, however, are monotone non-increasing; these include the *arc length distance* $d_{\rm arc}$ and the *cosine distance* $d_{\cos}$, that are associated with $\delta(t)=\delta_{\rm arc}(t)=\arccos t$ and $\delta(t)=\delta_{\cos}(t)=1-t$, respectively. Another rotation-invariant distance for which this monotonicity condition holds is the *chord distance* $d_{\rm chord}$ defined through $d_{\rm chord}(\theta,\psi)=\|\theta-\psi\|=\sqrt{2(1-\theta'\psi)}
\linebreak =:\delta_{\rm chord}(\theta'\psi)$. Throughout, we will denote the corresponding *arc distance depth* (ADD), *cosine distance depth* (CDD) and *chord distance depth* (ChDD) as $D_{\rm arc}$, $D_{\cos}$ and $D_{\rm chord}$, respectively.
The ADD is the arc distance depth introduced by [@LiuSin1992]. For the CDD, a direct computation yields $$\label{cosinexplic}
D_{\cos}(\theta, H)
=
2
-
E_{H}[1-\theta' W]
=
1
+
\theta' E_{H}[W]
.$$ Under the assumption that $E_{H}[W]$ is non-zero, this rewrites $D_{\cos}(\theta, H)
=
1
+
\|E_{H}[W]\|
\linebreak (\theta' \mu_H)
$, where $\mu_H:=E_{H}[W]/\|E_{H}[W]\|$ is the spherical mean of $H$. This shows that the CDD is then in a one-to-one relationship with the *angular Mahalanobis depth* of [@Leyetal2014], provided that the location functional needed in the latter is chosen as the spherical mean. We stress, however, that, unlike the angular Mahalanobis depth, the CDD does not require choosing a location functional on the sphere and is defined also in cases where $\mu_H=0$. To the best of our knowledge, the ChDD has not been considered in the literature.
Structural properties {#sec:StructProperties}
=====================
In this section, we derive the main properties of a generic directional $d$-depth. We start with the following invariance result.
[**(Rotational invariance)**]{} \[thmrotainv\] Let $d=d_\delta$ be a rotation-invariant distance and $H$ be a distribution on $\mathcal{S}^{q-1}$. Then $D_{d_{\delta}}(\theta, H)$ is a rotation-invariant depth, in the sense that $D_{d_{\delta}}(O\theta, H_O)=D_{d_{\delta}}(\theta, H)$ for any $q\times q$ orthogonal matrix $O$, where $H_O$ denotes the image of $H$ by the transformation $x\mapsto Ox$, that is, $H_O$ is the distribution of $OW$ when $W$ has distribution $H$.
A corollary is that if $H$ is rotationally symmetric about $\theta_0$ in the sense that $H_O=H$ for any $q\times q$ orthogonal matrix $O$ fixing $\theta_0$, then $d_\delta (O\theta,H)=d_\delta (\theta, H)$ for any such $O$. In particular, for any $\alpha$, the $\alpha$-depth region — that, as usual, is defined as the collection of $\theta$ values with a depth larger than or equal to $\alpha$ — is invariant under rotations fixing $\theta_0$, hence reflects the symmetry of the distribution $H$ about $\theta_0$.
In contrast, parallel to the angular Mahalanobis depth of [@Leyetal2014], the CDD provides symmetric depth regions of this form for any $H$, i.e, irrespectively of the fact that $H$ is rotationally symmetric or not. This follows from the comments at the end of Section \[secdef\].
[**(Continuity)**]{} \[thcontinuity\] Assume that the distance $d$ is continuous; if $d=d_\delta$, then this is equivalent to assuming that $\delta:[-1,1]\to {\mathbb{R}}^+$ is continuous. Let $H$ be a distribution on $\mathcal{S}^{q-1}$. Then, (i) the mapping $\theta\mapsto D_{d}(\theta, H)$ is continuous on $\mathcal{S}^{q-1}$; (ii) there exists $\theta_{d}(H)\in\mathcal{S}^{q-1}$ such that $D_{d}(\theta_d(H),H)=\sup_{\theta\in\mathcal{S}^{q-1}} D_{d}(\theta,H)$.
Note that the continuity result in Theorem \[thcontinuity\](i) holds without any assumption on $H$, hence will also hold in the empirical case. Theorem \[thcontinuity\](ii) guarantees the existence of a $D_{d}$-deepest point $\theta_{d}(H)$. The deepest point (or collection of deepest points) typically depends on the distance $d$ adopted. For the CDD, the deepest point is the spherical mean, provided that ${\rm E}_H[W]\neq 0$, whereas the deepest point for the ADD is the spherical median of [@Fis1985], which reduces to the circular median ([@MarJup2000], p. 30) in dimension $q=2$. This is in line with the Euclidean case where deepest points typically depend on the depth considered and may be multivariate medians (e.g., Tukey’s halfspace or Liu’s simplicial deepest points) or mean vectors (e.g., the zonoid of [@KosMos1997] or the moment-based Mahalanobis deepest points).
The deepest point may not be unique; for the uniform distribution on $\mathcal{S}^{q-1}$, for instance, any rotation-invariant distance-based depth will be constant over the sphere (this readily follows from Theorem \[thmrotainv\]). This lack of unicity also holds in the Euclidean case, where the barycentre of the collection $\mathcal{C}$ of deepest points is often taken as its unique representative; for most depths, it then follows from the convexity of the depth regions (which guarantees convexity of $\mathcal{C}$) that this barycentre indeed has maximal depth. It is interesting to note that directional depths are fundamentally different in this respect, as no such convexity arguments can be used. The particular nature of the sample space may induce depth regions that are even disconnected. This may occur for some multimodal distributions $H$; an example is given in Section \[secillu\]. In contrast, note that, for $D_{\cos}$, the collection of deepest points is either $\left\{\mu_{H}\right\}$, when ${\rm E}_H[W]\neq 0$ , or $\mathcal{S}^{q-1}$, when ${\rm E}_H[W]=0$, and hence it is always spherically convex.
It is desirable that if the distribution $H$ on $\mathcal{S}^{q-1}$ has an “indisputable" location centre $\theta_0$, then the deepest point $\theta_d(H)$ is unique and coincides with $\theta_0$. The following theorem provides such a Fisher consistency result.
[**(Fisher consistency under monotone rotational symmetry)**]{} \[thFishconsist\] Assume that the rotation-invariant distance $d=d_\delta$ is based on a monotone strictly decreasing function $\delta:[-1,1]\to {\mathbb{R}}^+$. Assume that the distribution $H$ on $\mathcal{S}^{q-1}$ admits a density of the form $x\mapsto c_{q,h} h(x'\theta_0)$ for some $\theta_0\in\mathcal{S}^{q-1}$ and some monotone strictly increasing function $h:[-1,1]\linebreak \to {\mathbb{R}}^+$. Then, $\theta\mapsto D_{d_{\delta}}(\theta, H)$ is a monotone strictly increasing function of $\theta'\theta_0$, so that $\theta\mapsto D_{d_{\delta}}(\theta, H)$ is uniquely maximized at $\theta_0$.
Theorem \[thFishconsist\] ensures that the ADD-, CDD-, and ChDD-deepest points are equal and coincide with the modal location $\theta_0$ of $H$ in case the latter admits a density of the form given in the theorem. The monotonicity result entails that, irrespective of the distance $d_\delta$ used, the depth regions are of the form $\{\theta\in\mathcal{S}^{q-1}: \theta'\theta_0\geq c\}$.
In this setup, the maximal depth, $\max_{\theta\in\mathcal{S}^{q-1}} D_{d_\delta}(\theta,H)$, measures the concentration of $H$, as showed in the following theorem.
[**(Maximal depth as a concentration measure)**]{} \[propconcentr\] Assume that the rotation-invariant distance $d=d_\delta$ is based on a monotone strictly decreasing function $\delta:[-1,1]\to {\mathbb{R}}^+$. Assume that the distribution $H_\kappa$ on $\mathcal{S}^{q-1}$ admits the density $x\mapsto c_{q,\kappa,h} h(\kappa x'\theta_0)$ for some $\theta_0\in\mathcal{S}^{q-1}$ and some monotone strictly increasing and differentiable function $h:{\mathbb{R}}\to {\mathbb{R}}^+$ such that $
t \mapsto t\,\frac{d}{dt}\log h(t)
$ is monotone strictly increasing. Then the maximal depth $D_{d_{\delta}}(\theta_0, H_\kappa)$ is a strictly increasing function of $\kappa$.
In Theorem \[propconcentr\], $\kappa$ plays the role of a concentration parameter; typically, the larger $\kappa$, the more concentrated the probability mass is about the modal location $\theta_0$. Since the maximal depth is a strictly increasing function of $\kappa$, it is itself a concentration (or spread) measure. Note that the assumption that $t \mapsto t\,\frac{d}{dt}\log h(t)
$ is monotone strictly increasing holds in particular if $h$ is $\log$-convex, so that the result applies for von Mises–Fisher (vMF) distributions that are obtained for $h\left(u\right)=\exp\left(u\right)$. While Theorem \[propconcentr\] restricts to rotationally symmetric distributions, the maximal cosine distance depth $\max_{\theta\in\mathcal{S}^{q-1}} D_{\cos}(\theta,H)=1+\|E_{H}[W]\|$ is, irrespective of $H$, related to the “spherical variance" ([@MarJup2000], p. 164), that is, to the mean resultant length $\|E_H[W]\|$ of $W$.
We conclude this section by stating a property showing that the proposed depths may inherit anti-symmetry properties of the distances on which they are based. More precisely, we have the following result which is restricted to rotationally-invariant distances, although a similar result can be stated for an arbitrary distance $d$.
[**(Anti-symmetry)**]{} \[thskewsym\] Assume that the rotation-invariant distance $d=d_\delta$ is based on a function $\delta:[-1,1]\to {\mathbb{R}}^+$ that is anti-symmetric about $0$, i.e., $\delta(-t)+\delta(t)=\delta(-1)$. Let $H$ be a distribution on $\mathcal{S}^{q-1}$. Then,
1. $\theta\mapsto D_{d_{\delta}}(\theta, H)$ is anti-symmetric on $\mathcal{S}^{q-1}$ in the sense that $$D_{d_{\delta}}(-\theta, H)=d_{\delta}^{\sup}-D_{d_{\delta}}(\theta,H);$$
2. If $\theta_0$ has maximal depth, then $-\theta_0$ has minimal depth.
The arc length and cosine distances are based on anti-symmetric functions $\delta$, but the chord distance is not. If $\delta$ is anti-symmetric, then an antipodally symmetric distribution $H \in \mathcal{S}^{q-1}$, under which $H(-B) = H(B)$ for any measurable set $B$ on $\mathcal{S}^{q-1}$ , leads to a depth function $\theta\mapsto D_{d_{\delta}}(\theta,H)$ that is constant. This is another property contrasting sharply with the Euclidean case, where no distribution will provide a constant depth function. To show why the claim on the constancy holds true, consider an arbitrary measurable set $B\subset\mathcal{S}^{q-1}$ such that $\mathcal{S}^{q-1}=(-B)\cup B$ and $(-B)\cap B=\emptyset$. Then, using the antipodal symmetry of $H$ and the antisymmetry of $\delta$, we obtain $$\begin{aligned}
D_{d_{\delta}}(\theta,H)
&=&
\delta(-1)
-
\int_{-B}
\delta(\theta'w)
\,
dH(w)
-
\int_{B}
\delta(\theta'w)
\,
dH(w)
\\[2mm]
&= &
\delta(-1)
-
\int_{B}
\delta(-\theta'w)
\,
dH(w)
-
\int_{B}
\delta(\theta'w)
\,
dH(w)
\\[2mm]
&= &
\delta(-1)
-
\int_{B}
\delta(-1)
\,
dH(w)
=
\frac{ \delta(-1)}{2}
\cdot\end{aligned}$$ An interesting question is whether or not antipodal symmetry of $H$ is also a necessary condition for the constancy of $\theta\mapsto D_{d_{\delta}}(\theta,H)$ with an anti-symmetric function $\delta$. While [@LiuSin1992] proved that this is indeed the case for the ADD in dimension $q=2$ under the assumption that $H$ admits a density, it is not the case for any $\delta$ function. For instance, for the CDD, it directly follows from (\[cosinexplic\]) that $\theta\mapsto D_{\cos}(\theta, H)$ is constant if and only if $E_{H}[W]=0$, which shows that antipodal symmetry is not a necessary condition for the constancy of $D_{\cos}$.
Illustrations {#secillu}
=============
This short section illustrates the theoretical results of the previous section for three empirical distributions on the circle $\mathcal{S}^1$; we restrict to the circle to allow for a visual comparison of the various depths. Denoting as $H^{\rm vMF}_{\alpha,\kappa}$ the vMF distribution on $\mathcal{S}^1$ with modal location $\theta=(\cos\alpha,\sin\alpha)'$ and concentration $\kappa$, the three empirical distributions considered are associated with a random sample of size $n=500$ from each of the following distributions: $H_1=H^{\rm vMF}_{\pi,2}$ (unimodal case), $H_2=\frac{1}{2} H^{\rm vMF}_{\frac{3\pi}{4},5}+\frac{1}{2} H^{\rm vMF}_{\frac{5\pi}{4},5}$ (bimodal symmetric case), $H_3=\frac{1}{2} H^{\rm vMF}_{\frac{5\pi}{9},7}+\frac{1}{2} H^{\rm vMF}_{\frac{13\pi}{9},17}$ (bimodal asymmetric case).
For each of the resulting empirical distributions $H_{\ell n}$, $\ell=1,2,3$, Figure \[Figillu\] provides plots of the distance-based depths ADD, CDD and ChDD, as well as the competing angular simplicial depth (ASD) and angular Tukey depth (ATD). The ASD and ATD were computed through the packages and , respectively. The distance-based depths were computed by means of functions written by the authors. Simulated data and their graphical representations were obtained through the package ([@LunAgo2013]), which is a standard reference to work with data on the unit circle.
![Plots of the depth mapping $\alpha\mapsto D({\cos \alpha \choose \sin \alpha},H_{\ell n})$, for the distance-based depths ADD, CDD and ChDD, as well as the angular simplicial depth (ASD) and angular Tukey depth (ATD), and the empirical distributions $H_{\ell n}$, $\ell=1,2,3$ described in Section \[secillu\] (for easier visualization, depth values were actually multiplied by 1.5 for distance-based depths, by 1 for the ASD, and by 0.5 for the ATD). Deepest points are maked by a black dot. The parent density is also plotted in each case. []{data-label="Figillu"}](Figures/Fig1.pdf){width="100.00000%"}
For $H_{1n}$, all distance-based depth functions are monotonically strictly decreasing from their deepest point ($\approx \pi$) and do so in a symmetric way, which is in accordance with Theorems \[thmrotainv\] and \[thFishconsist\]. These depths functions are also continuous; see Theorem \[thcontinuity\]. In contrast, the ATD is constant outside the interval of length $\pi$ centered at its deepest point, which holds for any distribution on the circle ([@LiuSin1992 Proposition 4.6.]), and both the ASD and ATD are piecewise constant functions. The center-outward rankings provided by the ASD and ATD therefore yield many ties and are more rough than those given by distance-based depths. For the symmetric bimodal distribution $H_{2n}$, all depth functions are unimodal, hence fail to capture the bimodality of the distribution, which is not a problem since depths are not density measures but rather centrality measures. In contrast with the Euclidean case, some directional depths may exhibit multimodality, as it is the case for the ChDD for the distribution $H_{3n}$, where modes are more separated than in $H_{2n}$; (\[cosinexplic\]) entails that the CDD will never exhibit such a multimodal pattern. In this last example, the depth functions reflect the asymmetry of the distribution and do not identify the same deepest point; in particular, the CDD is maximized at the spherical mean, whereas the ADD is maximized at the circular median ([@MarJup2000], p. 20), and so are the ASD and ATD.
Asymptotic and robustness properties {#sec:DistrProperties}
====================================
In this section, we present asymptotic results for the distance-based depths introduced in Definition \[defclass\] and for the corresponding deepest points, as well as a robustness result regarding the breakdown point of these. We start with a Glivenko-Cantelli-type result.
[**(Uniform almost sure consistency)**]{} \[thunifconsistency\] Let $d$ be a bounded and continuous distance on $\mathcal{S}^{q-1}$ and $H$ be a distribution on $\mathcal{S}^{q-1}$. Denote as $H_n$ the empirical distribution associated with a random sample of size $n$ from $H$. Then $$\sup_{\theta\in\mathcal{S}^{q-1}}
\big|D_{d}(\theta, H_n)-D_{d}(\theta, H)\big|
\to
0$$ almost surely as $n\to\infty$.
This result implies that we may explore empirically the properties of $D_d(\theta,H)$ by considering the corresponding sample depth function $D_d(\theta,H_n)$ for a large $n$. This justifies a posteriori the illustration of Theorem \[thFishconsist\] in the previous section. The following asymptotic normality is a direct result of the central limit theorem.
[**(Asymptotic normality of sample depth)**]{} \[thasnormD\] Let $d$ be a bounded distance on $\mathcal{S}^{q-1}$ and $H$ be a distribution on $\mathcal{S}^{q-1}$. Denote as $H_n$ the empirical distribution associated with a random sample of size $n$ from $H$. Then as $n\to\infty$, $
\sqrt{n}(D_{d}(\theta, H_n)-D_{d}(\theta, H))
$ converges weakly to the normal distribution with mean zero and variance ${\rm Var}_H[d(\theta,W)]$.
We turn to asymptotic and robustness results for deepest points. The following strong consistency result requires that the deepest point is uniquely defined, as it is in Theorem \[thFishconsist\].
[**(Almost sure consistency of the deepest point)**]{} \[thdeepestconsistency\] Let $d$ be a bounded and continuous distance on $\mathcal{S}^{q-1}$ and $H$ be a distribution on $\mathcal{S}^{q-1}$. Assume that the deepest point $\theta_d(H)$ is unique. Denote as $H_n$ the empirical distribution associated with a random sample of size $n$ from $H$, and let $\theta_d(H_n)$ be an arbitrary deepest point with respect to $H_n$. Then $$\theta_d(H_n)
\to
\theta_d(H)$$ almost surely as $n\to\infty$.
Constructing confidence zones for $\theta_d(H)$ requires the availability of the asymptotic distribution of $\theta_d(H_n)$. Since $\theta_d(H_n)$ is an $M$-estimator for a location parameter on $\mathcal{S}^{q-1}$, its asymptotic distribution can easily be obtained from the results of [@KoCha1993], at least under rotationally symmetric distributions. We do not pursue in this direction here.
Since deepest points are commonly used as robust location estimators, it is natural to investigate their robustness, and we therefore end this section by deriving a result on their breakdown point (BDP). In the directional setup considered, the classical BDP concept ([@Hametal1986], pp. 97-98) is not suitable, and we adopt the directional concept of [@LiuSin1992] defining the BDP of the (more generally, of a) deepest point $\theta_d(H)$ as the infimum of $\varepsilon$ such that, for some contaminating distribution $G$ on $\mathcal{S}^{q-1}$, $-\theta_d(H)$ is a deepest point of $D_d(\theta, H_{\epsilon})$ with $H_{\epsilon}:= \left(1 - \epsilon \right) H + \epsilon G$. The following result extends to an arbitrary distance $d$ the lower bound result obtained in [@LiuSin1992] for the arc length distance.
[**(Breakdown point of deepest points)**]{} \[thmbdp\] Let $d$ be a bounded distance on $\mathcal{S}^{q-1}$ and $H$ be a distribution on $\mathcal{S}^{q-1}$. Let $\theta_d(H)$ be a deepest point of $D_{d}(\theta,H)$. Then the breakdown point of $\theta_d(H)$ is larger than or equal to $(D_{d}(\theta_d(H),H) - D_{d}(-\theta_d(H),H))\linebreak/(2d^{\rm sup})$.
To investigate how the distance $d$ affects the lower bound, we consider the important case of vMF distributions. If $H^{\rm vMF}_{q,\theta_0,\kappa}$ denotes the vMF($\theta_0,\kappa$) distribution on $\mathcal{S}^{q-1}$, then, for a rotation-invariant distance $d_\delta$ that is decreasing in the sense of Theorem \[thFishconsist\], we have $\theta_{d_\delta}(H^{\rm vMF}_{q,\theta_0,\kappa})=\theta_0$ and $$D_{d_\delta}(\pm\theta_0,H^{\rm vMF}_{q,\theta_0,\kappa})
=
d^{\rm sup}_\delta
-
\frac{
\int_{-1}^1
\delta(\pm v)
(1-v^2)^{(q-3)/2}
\exp(\kappa v)
\,
dv
}
{\int_{-1}^1
(1-v^2)^{(q-3)/2}
\exp(\kappa v)
\,
dv
}
,$$ which allows us to evaluate the lower bound from Theorem \[thmbdp\].
Figure \[BDPfig\] plots this lower bound as a function of $\kappa$ for various dimensions $q$ and for the ADD, CDD and ChDD. Clearly, irrespective of the dimension and the distance, the lower bound is arbitrarily small for arbitrarily small values of $\kappa$ and goes to 50% as $\kappa$ goes to infinity. The lower bound decreases as the dimension $q$ increases. More importantly, for vMF distributions, the CDD-deepest point, namely the spherical mean, provides a larger lower bound than the ADD- and CHDD-deepest ones do.
Applications {#secSimu}
============
We present two applications, which are related to spherical location estimation and supervised classification.
Spherical location estimation {#secSimusub1}
-----------------------------
Depth functions find applications in robust statistics, with the deepest point considered as a robust location estimator.
For this reason, we conducted a simulation study to investigate the efficiency and robustness properties of the deepest points associated with the proposed distance-based depths, and to compare them with those of the competing ASD- and ATD-deepest points We start with efficiency properties. For any combination of a dimension $q\in\{3,5\}$, a sample size $n\in\{25,50,100\}$ and a concentration $\kappa\in\{5,10\}$, we generated $M=500$ independent random samples of size $n$ from the distribution $H^{\rm vMF}_{q,\theta,\kappa}$, where $\theta=e_q$ is the last vector of the canonical basis of ${\mathbb{R}}^q$. For each estimator $\hat{\theta}$ of $\theta$ considered, this leads to estimates $\hat{\theta}_{1},\ldots,\hat{\theta}_{M}$. Figure \[Efficiencyfig\] provides boxplots of the resulting squared errors $${\rm SE}_{m}
=
\|\hat{\theta}_{m}- \theta\|^2
=
2 (1 - \hat{\theta}_{m}' \theta)
,
\qquad
m=1,\ldots, M,
\label{SEdef}$$ and indicates the resulting mean square errors ${\rm MSE} = (1/M) \sum_{m=1}^M {\rm SE}_{m}$. The computational burden for the ASD- and ATD-deepest points is so prohibitive that these were considered for dimension $q=3$ only.
Results indicate that, in dimension $q=3$, the estimators associated with distance-based depths slightly dominate their ATD competitor and outperform their ASD one. As expected, the CDD-deepest point, that is the maximum likelihood estimator in the distributional setup considered, is in most cases the most efficient estimator. In dimension $q=5$, where the ASD/ATD estimators could not be computed, the distance-based depths perform similarly. On the other hand, while the CDD estimator slightly dominates at all sample sizes in dimension $q = 3$, it dominates only at the largest considered sample size in dimension $q=5$.
We now turn to the investigation of robustness properties for which we restricted to dimension $q=3$. For any combination of a contamination level $\varepsilon\in\{0,0.05,0.10\}$ and a concentration $\kappa\in\{5,10\}$, we generated $M=500$ independent random samples of size $n=100$ from the contaminated distributions $(1-\varepsilon)H^{\rm vMF}_{q,\theta,\kappa}+\varepsilon \Delta_{\theta_r}$, $r=1,2$, where $\theta$ is set as $e_q$, $\theta_1=e_{q-1}$, $\theta_2=-\theta$, $\Delta_{\psi}$ denotes the point mass distribution at $\psi$. Hence, $r=1,2$ refers to contamination at an orthogonal point to $\theta$ and at the antipodal point to $\theta$, respectively. In each sample, the deepest points of the same five depths as in Figure \[Efficiencyfig\] were computed. The resulting boxplots of squared errors ${\rm SE}_m$ for $m=1,\ldots,M$ and the mean squared errors (${\rm MSE}$) are provided in Figure \[Robustnessfig\].
The results show that the estimators associated with distance-based depths enjoy good robustness properties. In particular, irrespective of the contamination level $\varepsilon$ and the type of contamination, the ADD, CDD and ChDD estimators outperform the ASD one in terms of robustness. The domination over the ATD estimator is less significant.
Supervised classification {#secSimusub2}
-------------------------
Classification has been one of the most successful applications of statistical depth in the last decade, both for multivariate and functional data. While some proposals were based on the use of local depth concepts ([@PaiVanB2013]) or a depth-based version of kNN classification ([@PaiVanB2012]), the dominant solution finds its source in the *max-depth approach* of [@GhoCha2005B] that has later been refined by [@Lietal2012]. To the best of our knowledge, depth-based classification for directional data has not been considered in the literature. In this section, we show that the max-depth approach also applies for directional data and that, in conjunction with the proposed distance-based depths, it provides classifiers on the hypersphere that dominate ASD/ATD-based ones and that can be applied in higher dimensions as well.
Consider the spherical classification problem where independent random samples $W_{1i}$, $i=1,\ldots,n_1$ and $W_{2i}$, $i=1,\ldots,n_2$, respectively, come from distributions $H_1$ and $H_2$ on $\mathcal{S}^{q-1}$, and one is given the task to classify a point $w(\in\mathcal{S}^{q-1})$ as arising from $H_1$ (“population 1") or from $H_2$ (“population 2"). Denoting as $H_{\ell n_\ell}$ the empirical distribution associated with $W_{\ell i}$, $i=1,\ldots,n_\ell$ ($\ell=1,2$), the max-depth classifier associated with a depth $D$ classifies $w$ into population 1 if $D(w,H_{1n_1})>D(w,H_{2n_2})$, and population 2 otherwise; if $D(w,H_{1n_1})=D(w,H_{2n_2})$, then the classification decision is based on the flip of a fair coin.
To investigate the finite-sample performances of such classifiers, we consider the Monte Carlo algorithm that was performed for dimensions $q=2$ and $q=10$. Denoting as $e_j$ the $j$th vector in the canonical basis of ${\mathbb{R}}^q$ and using the notations $H^{\rm vMF}_{\alpha,\kappa}$ and $H^{\rm vMF}_{q,\theta_1,\kappa}$ from Sections \[secillu\] and \[sec:DistrProperties\], respectively, we considered the following three distributional setups:
- Setup A involves the vMF distributions $H_1=H^{\rm vMF}_{\frac{\pi}{4},5}$ and $H_2=H^{\rm vMF}_{\frac{3\pi}{4},5}$ for $q=2$, and $H_1=H^{\rm vMF}_{q,e_1,5}$ and $H_2=H^{\rm vMF}_{q,e_q,5}$ for $q=10$; Setup A therefore involves distributions differing through the modal location only.
- In Setup B, $H_1=H^{\rm vMF}_{\frac{\pi}{3},2}$ and $H_2=H^{\rm vMF}_{\frac{2\pi}{3},5}$ for $q=2$, and $H_1=H^{\rm vMF}_{q,e_q,2}$ and $H_2=H^{\rm vMF}_{q,(\cos \frac{\pi}{6})e_{q-1}+(\sin \frac{\pi}{6})e_q,5}$ for $q=10$; in this setup, distributions differ through location and concentration.
- Setup C involves discrimination between the vMF distribution $H_1=H^{\rm vMF}_{\frac{3\pi}{4},4}$ and the mixture distribution $H_2=\frac{1}{2}H^{\rm vMF}_{0,4}+\frac{1}{2}H^{\rm vMF}_{\frac{\pi}{2},4}$ for $q=2$, and $H_1=H^{\rm vMF}_{q,(\cos \frac{7\pi}{4})e_{q-1}+(\sin \frac{7\pi}{4})e_q,4}$ and $H_2=\frac{1}{2}H^{\rm vMF}_{q,e_{q-1},4}+\frac{1}{2}H^{\rm vMF}_{q,e_q,4}$ for $q=10$.
For each setup and each $q$, we generated $M=250$ independent training samples of size $n_{\rm train}=200$ and test samples of size $n_{\rm test}=100$ by sampling randomly from $\frac{1}{2}H_1+\frac{1}{2}H_2$. In replication $m \in \{1,\ldots,250\}$, this associates with any depth $D$ on $\mathcal{S}^{q-1}$ the misclassification rate $p_m(D)=N_m(D)/n_{\rm test}$, where $N_m(D)$ is the number of observations in the $m$th test sample that were misclassified by the max-depth classifier associated with $D$ when based on the $m$th training sample. Figure \[Classifig\] provides the boxplots, for several depths $D$, of the resulting $M=250$ misclassification rates. As in Section \[secSimusub1\], the depths considered are the ADD, CDD, ChDD, ASD and ATD; again, computational issues prevented to consider the ASD and ATD in dimension $q=10$.
Results indicate that distance-based depth classifiers dominate in most cases their counterparts based on the ASD/ATD. It is only in Setup $C$ that the ASD/ATD classifiers seem to slightly improve over the ADD and CDD classifiers. In all cases, the classifier based on the ChDD is the best classifier. Most importantly, in higher dimensions, the computational burden for the ASD/ATD is such that only the distance-based depth classifiers can be used.
Discussion {#secfinal}
==========
In the Euclidean multivariate setup, statistical depth has allowed to tackle in a nonparametric and robust way diverse problems, including location/scatter estimation, two-sample hypothesis testing, supervised classification, etc. While depths in the spherical setup, such as the ASD and ATD, were proposed more than two decades ago, the concept has not made its way to applications. Arguably, the reasons are that these depths are, even for moderate dimensions, very computationally intensive and that it is challenging to derive their asymptotic properties.
The class of distance-based depths for directional data defined in this work clearly improve on this. These depths were showed to be computable in higher dimensions, and asymptotic results can be obtained by using standard $M$-estimation techniques. For small dimensions, where distance-based depths as well as the ASD/ATD can be evaluated, we showed through simulations that inference procedures based on the former compete equally or even dominate those based on the latter. In high dimensions, only distance-based depths can be used for directional data, which makes them of potential interest for applications involving high-dimensional spherical problems, such as those encountered in magnetic resonance, gene expression, or text mining; see, among others, [@Dry2005], [@banerjee2003generative], and [@Banetal2005].
Perspectives for future research are rich and diverse. Obviously, it would be of interest to investigate how distance-based depths can tackle the problems considered in the aforementioned high-dimensional applications. More generally, irrespective of the dimension, it would be desirable to develop depth-based inference procedures in various setups, including two-sample hypothesis testing and supervised classification. Finally, the present work also raised some theoretical questions of interest. For instance, in dimension $q=2$, the arc distance depth is constant if and only if the underlying distribution $H$ is antipodal, whereas the cosine distance depth is constant if and only if $H$ has zero mean. In view of this, it is natural to wonder what property of $H$ is characterized by constancy of the chord distance depth. The question can be raised on the circle with $q=2$ or for a general dimension $q>2$. Such characterization results are of interest since they obviously provide the basis for universally consistent tests of the corresponding properties.
As announced in Section \[sec:StructProperties\], we prove the following result for the sake of completeness.
\[prodistinv\] Let $d$ be a rotation-invariant distance on $\mathcal{S}^{q-1}$. Then there exists a function $\delta:[-1,1]\to{\mathbb{R}}^+$ such that $d(\theta,\psi)=\delta(\theta'\psi)$.
Proof of Proposition \[prodistinv\]. For any $\theta,\psi\in\mathcal{S}^{p-1}$, let $\psi_\theta=(\psi-(\psi'\theta)\theta)/\|\psi-(\psi'\theta)\theta\|$ and denote as $\Gamma_{\theta,\psi}$ an arbitrary $q\times (q-2)$ matrix such that $O_{\theta,\psi}=(\theta \vdots \psi_\theta \vdots \Gamma_{\theta,\psi})$ is orthogonal (if $q=2$, then we simply consider $O_{\theta,\psi}=(\theta \vdots \psi_\theta)$). Since $d$ is rotation-invariant, we have $d(\theta,\psi)=d(O_{\theta,\psi}'\theta,O_{\theta,\psi}'\psi)=d(e_1,O_{\theta,\psi}'\psi)$, where $e_1$ stands for the first vector of the canonical basis of ${\mathbb{R}}^q$. The result then follows from the fact that $
O_{\theta,\psi}'\psi
=
(
\theta'\psi , (1-(\theta'\psi)^2)^{1/2},0,\ldots,0
)'
$ depends on $\theta$ and $\psi$ through $\theta'\psi$ only. [$\square$]{}
Proof of Theorem \[thmrotainv\]. Using the notation introduced in the theorem, we have that $D_{d_{\delta}}(O\theta, H_O)
=\delta(-1)-E_{H_O}[\delta((O\theta)'W)]
=\delta(-1)-E_{H}[\delta((O\theta)'OW)]
=\delta(-1)
\linebreak
-E_{H}[\delta(\theta'W)]
=D_{d_{\delta}}(\theta, H)$. [$\square$]{}
Proof of Theorem \[thcontinuity\]. (i) Since the function $w\mapsto d(\theta,w)$ is continuous in $w$ for any $\theta\in\mathcal{S}^{q-1}$ and is bounded, uniformly in $\theta$, by the integrable function $w\mapsto d^{\sup}$, the continuity of $$\theta\mapsto D_{d}(\theta, H)
=
d^{\sup}
-
\int_{\mathcal{S}^{q-1}}
d(\theta,w)
\,
dH(w)$$ results from Corollary 2.8.7(i) in [@Bog2007]. (ii) The result follows from the fact that a continuous function on a compact domain attains its maximal value. [$\square$]{}
Proof of Theorem \[thFishconsist\]. Since the distribution $H$ is rotationally symmetric about $\theta_0$, Theorem \[thmrotainv\] implies that $D_{d_{\delta}}(\theta, H)$ depends on $\theta$ only through $\theta'\theta_0$. Consider then an arbitrary geodesic path $t\mapsto \theta_t$ from $\theta_0$ to $\theta_1=-\theta_0$. The monotonicity assumption on $h$ readily implies that, for any $s\in[-1,1]$, the function $t\mapsto P_{H}[ \theta_t'W \geq s]$ is monotone strictly decreasing. Since $$\begin{aligned}
E_H[\delta(\theta_t' W)]
&=&
\int_{0}^{\delta(-1)} z\, \frac{d}{dz} P_{H}[ \delta(\theta_t'W) \leq z] \,dz
\\[2mm]
&=&
\delta(-1) - \int_{0}^{\delta(-1)} P_{H}[ \delta(\theta_t'W) \leq z] \,dz
\\[2mm]
&=&
\delta(-1) - \int_{0}^{\delta(-1)} P_{H}[ \theta_t'W \geq \delta^{-1}(z)] \,dz
,\end{aligned}$$ it follows that $$\label{ahahah}
D_{d_{\delta}}(\theta_t, H)
=
\delta(-1)-E_H[\delta(\theta_t' W)]
=
\int_{0}^{\delta(-1)} P_{H}[ \theta_t'W \geq \delta^{-1}(z)] \,dz$$ is strictly decreasing in $t$. This establishes the result. [$\square$]{}
Proof of Theorem \[propconcentr\]. First note that for any $s$, $$\label{tqhz1}
P_{H_\kappa}[\theta_0'W\geq s]
=
\frac{
\int_{s}^1
(1-v^2)^{(q-3)/2}
h(\kappa v)
\,
dv
}
{\int_{-1}^1
(1-v^2)^{(q-3)/2}
h(\kappa v)
\,
dv
}$$ (see, e.g., [@PaiVer17b]), which provides $$\label{tqhz2}
\frac{P_{H_\kappa}[\theta_0'W\geq s]}{1-P_{H_\kappa}[\theta_0'W\geq s]}
=
\frac{
\int_{s}^1
(1-v^2)^{(q-3)/2}
h(\kappa v)
\,
dv
}
{\int_{-1}^s
(1-v^2)^{(q-3)/2}
h(\kappa v)
\,
dv
}
\cdot$$ Differentiation with respect to $\kappa$ yields $$\begin{aligned}
\lefteqn{
\frac{d}{ds}
\frac{P_{H_\kappa}[\theta_0'W\geq s]}{1-P_{H_\kappa}[\theta_0'W\geq s]}
}
\\[2mm]
& &
\hspace{3mm}
=
\frac{
\int_{s}^1
\int_{-1}^s
[
v
\dot h(\kappa v)
h(\kappa u)
-
u
\dot h(\kappa u)
h(\kappa v)
]
((1-u^2)(1-v^2))^{(q-3)/2}
\,
du
dv
}
{
(
\int_{-1}^s
(1-v^2)^{(q-3)/2}
h(\kappa v)
\,
dv
)^2
}
\cdot\end{aligned}$$ Since $t \mapsto t\,\frac{d}{dt}\log h(t)=t \dot{h}(t)/h(t)$ is strictly increasing, this derivative is strictly positive at any $\kappa$, so that the lefthand side of (\[tqhz2\]), hence also that of (\[tqhz1\]), is a monotone strictly increasing function of $\kappa$. The result then follows from the identity $
D_{d_{\delta}}(\theta_0, H_\kappa)
=
\int_{0}^{\delta(-1)} P_{H_\kappa}[ \theta_0'W \geq \delta^{-1}(z)] \,dz
$; see (\[ahahah\]). [$\square$]{}
Proof of Theorem \[thskewsym\]. (i) The anti-symmetry of $\delta(\cdot)$ readily yields $
D_{d_{\delta}}(-\theta, H)
+
D_{d_{\delta}}(\theta, H)
=
2 \delta(-1) - E_{H}[d_\delta(-\theta,W)+d_\delta(\theta,W)]
=
2 \delta(-1) - E_{H}[\delta(-\theta' W)+\delta(\theta' W)]
\linebreak
=
\delta(-1)
$, which establishes the result. (ii) Ad absurdum, assume that $-\theta_0$ does not have minimal depth, so that there exists $\theta_1\in\mathcal{S}^{q-1}$ with $D_{d_{\delta}}(\theta_1, H)<D_{d_{\delta}}(-\theta_0, H)$. Then Part (i) of the result implies that $D_{d_{\delta}}(-\theta_1, H)>D_{d_{\delta}}(\theta_0, H)$, which contradicts the fact that $\theta_0$ has maximal depth. [$\square$]{}
Proof of Theorem \[thunifconsistency\]. The result directly follows from Theorem 16(a) in [@Fer1996]. [$\square$]{}
Proof of Theorem \[thasnormD\]. The result trivially follows from applying the central limit theorem to the expression $
\sqrt{n}(D_{d}(\theta, H_n)-D_{d}(\theta, H))
=
-n^{-1/2} \sum_{i=1}^n (d(\theta,W_i)- {\rm E}_H[d(\theta,W)])
$. [$\square$]{}
Proof of Theorem \[thdeepestconsistency\]. In view of Theorem \[thunifconsistency\], the result is a corollary of Theorem 2.12 and Lemma 14.3 in [@Kos2008]. [$\square$]{}
Proof of Theorem \[thmbdp\]. From Lemma 2.3 in [@str1985], we obtain that, for any $\theta\in\mathcal{S}^{q-1}$, $
| D_{d}(\theta,H_\varepsilon)-D_{d}(\theta,H) |
$$
$$
$$
$$
=
\varepsilon (E_G[d(\theta,W)]-E_H[d(\theta,W)])
\leq
\varepsilon d^{\rm sup} d_1(H,G)
,
$ where $d_1(H,G)$ denotes the variational distance between $H$ and $G$. Lemmas 2.4 and 2.5(i) in [@str1985] then yield that, still for any $\theta\in\mathcal{S}^{q-1}$, $
| D_{d}(\theta,H_\varepsilon)-D_{d}(\theta,H) |
\leq
\varepsilon d^{\rm sup}
.
$ The result readily follows. [$\square$]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In 1988, I. Beck introduced the notion of a zero-divisor graph of a commutative rings with $1$. There have been several generalizations in recent years. In particular, in 2007 J. Coykendall and J. Maney developed the irreducible divisor graph. Much work has been done on generalized factorization, especially $\tau$-factorization. The goal of this paper is to synthesize the notions of $\tau$-factorization and irreducible divisor graphs in domains. We will define a $\tau$-irreducible divisor graph for non-zero non-unit elements of a domain. We show that by studying $\tau$-irreducible divisor graphs, we find equivalent characterizations of several finite $\tau$-factorization properties.\
**2010 AMS Subject Classification:** 13A05, 13E99, 13F15, 5C25
address: |
Reinhart Center\
Viterbo University\
900 Viterbo Drive\
La Crosse, WI 54601
author:
- Christopher Park Mooney
title: Generalized Irreducible Divisor Graphs
---
Introduction
============
In this article, $D$ will denote an integral domain. We will always assume that all rings have an identity which is not zero. We will use $G=(V,E)$ to denote a graph with $V$ the set of vertices and $E$ the set of edges. In 1988, I. Beck in [@Beck], introduced for a commutative ring $R$, the notion of a zero-divisor graph $\Gamma(R)$. The vertices of $\Gamma(R)$ are the set of zero-divisors, and there is an edge between $a,b \in Z(R)$ if $ab=0$. This has been studied and developed by many authors including, but not limited to D.D. Anderson, D.F. Anderson, A. Frazier, A. Lauve, P.S. Livingston, and M. Naseer in [@andersonzdg; @davidanderson; @Livingston].\
There have been several generalizations and extensions of this concept. In this paper, we focus on the notion of an irreducible divisor graph first formulated by J. Coykendall and J. Maney in [@Coykendall]. Instead of looking exclusively at divisors of zero in a ring, they restrict to a domain $D$ and choose any non-zero, non-unit $x\in D$ and study the relationships between the irreducible divisors of $x$. In [@Axtellidgzd; @Coykendall], M. Axtell, N. Baeth, J. Coykendall, J. Maney, and J. Stickles present several nice results about factorization properties of domains based on their associated irreducible divisor graphs. M. Axtell and J. Stickles have also studied irreducible divisor graphs in commutative rings with zero-divisors in [@Axtellidgd].\
Unique factorization domains are well known and widely studied. Lesser known are other finite factorization properties that domains might possess which are weaker than UFDs. These come in the form of half factorization domains or half factorial domains (HFDs), finite factorization domains (FFDs) and bounded factorization domains (BFDs). See [@anderson90] for more information on the developments in the theory of factorization in integral domains. More recently, these concepts have been further generalized by way of $\tau$-factorization in several papers, especially by D.D. Anderson and A. Frazier in [@Frazier] as well as the author in [@Mooney; @Mooney2]. In this paper, we seek to take the notion of $\tau$-factorization and apply it to irreducible divisor graphs. We will find that many equivalent characterizations of $\tau$-finite factorization properties given in the aforementioned papers can be given by studying $\tau$-irreducible divisor graphs.\
Section Two gives the necessary preliminary background information and definitions from the study of irreducible and zero-divisor graphs as well as $\tau$-factorization. In Section Three, we define the $\tau$-irreducible divisor graph of a domain $D$ with a fixed relation $\tau$. We give a few examples of $\tau$-irreducible divisor graphs, especially in comparison with the irreducible divisor graphs of [@Axtellidgd; @Coykendall]. In Section Four, we prove several theorems illustrating how $\tau$-irreducible divisor graphs give us another way to characterize various $\tau$-finite factorization properties domains may possess as defined in [@Frazier].
Preliminary Definitions
=======================
Irreducible Divisor Graph Definitions
-------------------------------------
\
We begin with some definitions from J. Coykendall and J. Maney [@Coykendall]. Let $Irr(D)$ be the set of all irreducible elements in a domain $D$. We will let $\overline{Irr}(D)$ be a (pre-chosen) set of coset representatives of the collection $\{a U(D) \mid a \in Irr(D)\}$. Let $x\in D^{\#}$ have a factorization into irreducibles. The irreducible divisor graph of $x \in D^{\#}$, will be the graph $G(x)=(V,E)$ where $V=\{a\in \overline{Irr}(D) \mid a \mid x\}$, i.e. the set of irreducible divisors of $x$ up to associate. Given $a_1, a_2 \in \overline{Irr}(D)$, $a_1a_2 \in E$ if and only if $a_1a_2 \mid x$. Furthermore, $n-1$ loops will be attached to $a$ if $a^n \mid x$. If arbitrarily large powers of $a$ divide $x$, we allow an infinite number of loops. They define the *reduced irreducible divisor graph* of $x$ to be the subgraph of $G(x)$ which is formed by deleting all the loops and denote it as $\overline{G}(x)$.\
A *clique* will refer to a simple (no loops or multiple edges), complete (all vertices are pairwise adjacent) graph. A clique on $n \in {\mathbb{N}}$ vertices will be denoted $K_n$. We will call a graph $G$ a *pseudoclique* if $G$ is a complete graph having some number of loops (possibly zero). This means a clique is still considered a pseudoclique.\
Let $G$ be a graph, possibly with loops, and let $a\in V(G)$. We have two ways of counting the degree of this vertex. We define *deg*$(a):= \left|\{a_1 \in V(G) \mid a_1 \neq a, a_1a \in E(G)\}\right|$, i.e. the number of distinct vertices adjacent to $a$. Suppose a vertex $a$ has $n$ loops. We define *degl*$(a):=n+deg(a)$, the sum of the degree and the number of loops. Given $a,b \in V(G)$, we define $d(a,b)$ to be the shortest path between $a$ and $b$. If no such path exists, i.e. $a$ and $b$ are in disconnected components of $G$, or the shortest path is infinite, then we say $d(a,b)= \infty$. We define Diam$(G):=$sup$(\{d(a,b) \mid a,b \in V(G) \}).$
$\tau$-Factorization Definitions
--------------------------------
\
Let $D$ be a domain. Let $D^*=D-\{0\}$, let $U(D)$ be the set of units of $D$, and let $D^{\#}=D^*-U(D)$ be the non-zero, non-units of $D$. We will say $a$ and $b$ are associates if $a=\lambda b$ for some $\lambda \in U(D)$, and denote this $a \sim b$. Let $\tau$ be a relation on $D^{\#}$, that is, $\tau \subseteq D^{\#} \times D^{\#}$. We will always assume further that $\tau$ is symmetric. Let $a$ be a non-unit. A factorization of the form $a=\lambda a_1 \cdots a_n$ is said to be a *$\tau$-factorization* if $a_i \in D^{\#}$, $\lambda \in U(D)$ and $a_i \tau a_j$ for all $i\neq j$. If $n=1$, then this is said to be a *trivial $\tau$-factorization*. Each $a_i$ is said to be a *$\tau$-factor*, or that $a_i$ *$\tau$-divides* $a$, written $a_i \mid_\tau a$.\
We say that $\tau$ is *multiplicative* (resp. *divisive*) if for $a,b,c \in D^{\#}$ (resp. $a,b,b' \in D^{\#}$), $a\tau b$ and $a\tau c$ imply $a\tau bc$ (resp. $a\tau b$ and $b'\mid b$ imply $a \tau b'$). We say $\tau$ is *associate preserving* if for $a,b,b'\in D^{\#}$ with $b\sim b'$, $a\tau b$ implies $a\tau b'$. We define a *$\tau$-refinement* of a $\tau$-factorization $\lambda a_1 \cdots a_n$ to be a factorization of the form $$(\lambda \lambda_1 \cdots \lambda_n) \cdot b_{11}\cdots b_{1m_1}\cdot b_{21}\cdots b_{2m_2} \cdots b_{n1} \cdots b_{nm_n}$$ where $a_i=\lambda_i b_{i1}\cdots b_{im_i}$ is a $\tau$-factorization for each $i$. This is slightly different from the original definition in [@Frazier] where no unit factor was allowed. One can see they are equivalent when $\tau$ is associate preserving. We then say that $\tau$ is *refinable* if every $\tau$-refinement of a $\tau$-factorization is a $\tau$-factorization. We say $\tau$ is *combinable* if whenever $\lambda a_1 \cdots a_n$ is a $\tau$-factorization, then so is each $\lambda a_1 \cdots a_{i-1}(a_ia_{i+1})a_{i+2}\cdots a_n$.\
We pause briefly to give some examples of particular relations $\tau$.
Let $D$ be a domain and let $\tau=D^{\#}\times D^{\#}$.
This yields the usual factorizations in $D$ and $\mid_{\tau}$ is the same as the usual divides. Moreover, $\tau$ is multiplicative and divisive (hence associate preserving).
Let $D$ be a domain and let $\tau=\emptyset$.
For every $a\in D^{\#}$, there is only the trivial factorization. Furthermore, all $\tau$-divisors of $a$ are associate to $a$. If $b\mid{_\tau} a$, then the factorization is forced to be of the form $a=\lambda b=\lambda(\lambda^{-1}(a))$ for $\lambda \in U(D)$. That is, $a\sim b$. Again $\tau$ is both multiplicative and divisive (vacuously).
Let $D$ be a domain and let $S$ be a non-empty subset of $D^{\#}$. Let $\tau=S\times S$. Define $a\tau b \Leftrightarrow a,b\in S$.
In this case, $\tau$ is multiplicative (resp. divisive) if and only if $S$ is multiplicatively closed (resp. closed under non-unit factors). A non-trivial $\tau$-factorization is (up to unit factors) a factorization into elements from $S$. Some examples of nice sets $S$ might be the set of primes or irreducibles, then a $\tau$-factorization is a prime decomposition or an atomic factorization respectively.
Let $D$ be a domain and let $a \tau b$ if and only if $(a,b)=D$
In this case we get the comaximal factorizations studied by S. McAdam and R. Swan in [@Mcadam]. More generally, as in J. Juett in [@Juettcomax], we could let $\star$ be a star-operation on $D$ and define $a\tau b \Leftrightarrow (a,b)^{\star}=D$, that is $a$ and $b$ are $\star$-coprime or $\star$-comaximal.
Let $a\in D^{\#}$. As in [@Frazier], we will say $a$ is *$\tau$-irreducible* or a *$\tau$-atom* if factorizations of the form $a=\lambda (\lambda^{-1} a)$ are the only $\tau$-factorizations of $a$. Then $D$ is said to be *$\tau$-atomic* if every $a\in D^{\#}$ has a $\tau$-factorization $a=\lambda a_1\cdots a_n$ with $a_i$ being $\tau$-atomic for all $1\leq i \leq n$. We will call such a factorization a *$\tau$-atomic-factorization*. We say $D$ satisfies *$\tau$-ascending chain condition on principal ideals ($\tau$-ACCP)* if for every chain $(a_0) \subseteq (a_1) \subseteq \cdots \subseteq (a_i) \subseteq \cdots$ with $a_{i+1} \mid_{\tau} a_i$, there exists an $N\in {\mathbb{N}}$ such that $(a_i)=(a_N)$ for all $i>N$.\
A domain $D$ is said to be a *$\tau$-unique factorization domain ($\tau$-UFD)* if (1) $D$ is $\tau$-atomic and (2) for every $a \in D^{\#}$ any two $\tau$-atomic factorizations $a=\lambda_1 a_1 \cdots a_n = \lambda_2 b_1 \cdots b_m$ have $m=n$ and there is a rearrangement so that $a_i$ and $b_i$ are associate for each $1 \leq i \leq n$. A domain $D$ is said to be a *$\tau$-half factorization domain ($\tau$-HFD)* if (1) $D$ is $\tau$-atomic and (2) for every $a \in D^{\#}$ any two $\tau$-atomic-factorizations have the same length. A domain $D$ is said to be a *$\tau$-finite factorization domain ($\tau$-FFD)* if for every $a \in D^{\#}$ there are only a finite number of $\tau$-factorizations up to rearrangement and associate. A domain $D$ is said to be a *$\tau$-weak finite factorization domain ($\tau$-WFFD)* if for every $a \in D^{\#}$, there are only finitely many $b\in D$ such that $b$ is a $\tau$-divisor of $a$ up to associate. A domain $D$ is said to be a *$\tau$-irreducible divisor finite domain (idf-domain)* if for every $a \in D^{\#}$, there are only finitely many $\tau$-atomic $\tau$-divisors of $a$ up to associate. A domain $D$ is said to be a *$\tau$-bounded factorization domain ($\tau$-BFD)* if for every $a \in D^{\#}$, there exists a natural number $N(a)$ such that for any $\tau$-factorization $a=\lambda a_1 \cdots a_n$, $n \leq N(a)$.\
We have the following set of relationships between the above finite factorization properties from [@Frazier], where $\nabla$ indicates the relationship requires $\tau$ to be refinable and associate preserving. $$\xymatrix{
& \tau \text{-HFD} \ar@{=>}^{\nabla}[dr] & & & \\
\tau\text{-UFD} \ar@{=>}[ur] \ar@{=>}^{\nabla}[dr] & & \tau\text{-BFD} \ar@{=>}[r]^{\nabla}& \tau\text{-ACCP} \ar@{=>}^{\nabla}[r]& \tau\text{-atomic}\\
& \tau\text{-FFD} \ar@{=>}[ur]\ar@{=>}[r] &\tau\text{-WFFD}\ar@{=>}[dl] & & \\
&\tau\text{-atomic }\tau\text{-idf domain}\ar@{=>}^{\nabla}[u]& & & }$$
The $\tau$-Irreducible Divisor Graph and Examples
=================================================
Let $D$ be a domain and let $\tau$ be a symmetric and associate preserving relation on $D^{\#}$. Let *$Irr_\tau(D)$* be the collection of $\tau$-irreducible elements. We will let $\overline{Irr}_\tau (D)$ be fixed, pre-chosen coset representatives of the cosets $\{a U(D) \mid a \in Irr_\tau(D)\}$. Given an element $a\in D^{\#}$ with a $\tau$-atomic factorization, we define the *$\tau$-irreducible divisor graph* of $x$ to be $G_\tau (x)=(V,E)$ with $V=\{a\in \overline{Irr}_\tau(D) \mid a \mid_{\tau} x\}$, and given $a_1, a_2 \in \overline{Irr}_{\tau}(D)$, $a_1a_2 \in E$ if and only if there is a $\tau$-factorization of the form $x=\lambda a_1 a_2 \cdots a_n$. Furthermore, $n-1$ loops will be attached to the vertex corresponding to $a$ if there is a $\tau$-atomic factorization of the form $x=\lambda a \cdots a a_1 \cdots a_n$ where $a$ occurs $n$ times. Again, if this occurs for arbitrarily large powers of $a$, we allow the possibility of an infinite number of loops. We then define the *reduced $\tau$-irreducible divisor graph* of $x$ to be the subgraph of $G_\tau(x)$ formed by removing all loops, and denote it $\overline{G_\tau(x)}$. If there are no $\tau$-atoms in $D$, then $D$ is said to be a *$\tau$-antimatter domain*. In this case, we define $G_\tau (x)=\overline{G_\tau (x)}=\emptyset$ for all $x \in D^{\#}$.\
\[ex: atom\] Let $D$ be a domain. Suppose $\tau = \emptyset$.
By letting $\tau=\emptyset$, we have eliminated all non-trivial $\tau$-factorizations. This has the effect of making every non-zero, non-unit a $\tau$-atom. This means for every $x \in D^{\#}$, $G_\tau(x)=(\{x\}, \emptyset).$ The $\tau$-irreducible divisor graph of $x$ consists of the single vertex, $x$ itself (or whichever associate of $x$ chosen initially in $\overline{Irr_\tau(D)}$).
Let $D$ be a domain. Suppose $\tau = D^{\#} \times D^{\#}$.
In this case, we are looking at the usual factorizations in $D$. This means $a\mid b$ if and only if $a \mid_\tau b$. Every $\tau$-factorization is a usual factorization and conversely. Moreover, $x\in D^{\#}$ is $\tau$-atomic if and only if $x$ is atomic. This results in the irreducible divisor graphs being identical. Hence, we have $G_\tau(x)=G(x)$ and $\overline{G_\tau(x)}= \overline{G}(x)$ as defined in Coykendall and Maney [@Coykendall].
(Inspired by [@Coykendall Example 2.4]) Let $D={\mathbb{Q}}[x^2, x^3]$ and let $f(x)=x^8-x^9$ and consider the relation $\tau$ defined by $g(x) \tau g'(x)$ if and only if $deg(g(x))=deg(g'(x))$.
The irreducible factorizations of $f(x)$ are $$f(x)=x^8-x^9=x^2\cdot x^2 \cdot x^2 \cdot (x^2-x^3) = (x^3-x^4)\cdot x^2 \cdot x^3=(x^2 - x^3) \cdot x^3 \cdot x^3.$$ It is clear that only the last factorization above is a $\tau$-factorization. It is the only factorization in which all of the factors have the same degree. Every irreducible is certainly $\tau$-irreducible, but the question is: are there any additional $\tau$-atomic elements which $\tau$-divide $f(x)$? For instance, $g(x)=x^5$ is $\tau$-atomic since the only non-trivial factorization up to associates and rearrangement is $g(x)=x^2 \cdot x^3$ is thrown out due to the factors having different degrees. Any $\tau$-factorization of $f(x)$ must have $\tau$-factors of the same degree. This amounts to finding a proper partition of $deg(f(x))=9$ into parts of equal size. This is done only by nine degree $1$ parts, or three degree $3$ parts. Since all non-units in $D$ have degree at least $2$, the former is not possible in this ring. This leaves only one possible $\tau$-atomic factorization up to associates and rearrangement: $$f(x)=(x^2 - x^3) \cdot x^3 \cdot x^3.$$ We show both $G(f(x))$ and $G_\tau(f(x))$ below in Figure 1 for the sake of comparison.
![(a) $G(f(x))$, (b) $G_\tau (f(x))$[]{data-label="fig:Irreducible divisor graphs"}](irreducibledivisorgraphsfigure1.eps)
This yields the following corresponding reduced irreducible divisor graphs $\overline{G}(f(x))$ and $\overline{G_\tau (f(x))}$ in Figure 2.
![(a) $\overline{G}(f(x))$, (b) $\overline{G_\tau (f(x))}$[]{data-label="fig:Reduced Irreducible divisor graphs"}](irreducibledivisorgraphsfigure2.eps)
This above partition argument shows that any polynomial of prime degree $p$ is a $\tau$-atom. The only way to non-trivially partition a prime integer $p$ into equal parts is $p$ parts of degree $1$ each. But all non-unit elements have degree at least $2$. Hence there can only be the trivial factorization. Thus we know the $\tau$-irreducible divisor graph of any polynomial of prime degree, $p(x)$ consists of the single vertex, $p(x)$ (or possibly whichever associate of $p(x)$ was taken in $\overline{Irr_\tau(D)}$). It is perhaps worth noting that the converse is false as conveniently demonstrated above by $x^4-x^3$ being irreducible and hence $\tau$-irreducible, despite having composite degree.
The $\tau$-Irreducible Divisor Graph and $\tau$-Finite Factorization Properties
===============================================================================
We first observe that the $\tau$-irreducible divisor graph yields a new equivalent characterization of a $\tau$-irreducible element of [@Frazier]. This formalizes something we observed in Example \[ex: atom\].
\[thm: atom\] Let $D$ be a domain and $\tau$ a symmetric, associate preserving relation on $D^{\#}$. If $D$ is $\tau$-atomic, then a non-unit $x$ is $\tau$-irreducible if and only if $G_\tau(x) \cong K_1$, the complete graph on a single vertex which is some associate of $x$.
If $x$ is $\tau$-irreducible, then the only $\tau$-factorizations are of the form $x=\lambda (\lambda^{-1}x)$, hence the only $\tau$-divisors of $x$ are associates of $x$. Moreover, $x$ is a non-zero, non-unit, so $x^2$ does not divide $x$, hence there are no loops. Thus, $G_\tau (x)$ is a single isolated vertex generated by whichever associate of $x$ was chosen in $\overline{Irr_\tau(D)}$. Conversely, suppose $x \in R$ is a non-unit such that $G_\tau(x) \cong K_1$. We suppose $x$ were not $\tau$-irreducible. Then there is a non-trivial $\tau$-atomic factorization of the form $x=\lambda a_1 \cdots a_n$ with $n \geq 2$. This yields $a_1, a_2 \in V(G_\tau(x))$, but there is only one vertex and no loops in $G_\tau (x)$, contradicting the hypothesis and completing the proof.
\[thm: accp\] Let $D$ be a domain and $\tau$ a symmetric, refinable and associate preserving relation on $D^{\#}$. If $D$ is $\tau$-atomic such that for all $x \in D^{\#}$, and for all $a \in V(G_\tau(x))$, degl$(a) < \infty$, then $D$ satisfies $\tau$-ACCP.
Suppose $D$ did not satisfy $\tau$-ACCP. Then there exists a chain of principal ideals $(x_1) \subsetneq (x_2) \subsetneq (x_3) \subsetneq \cdots$ such that $x_{i+1} \mid_\tau x_i$. Say
$$\label{1} x_i = \lambda_i x_{i+1}\cdot a_{i1} \cdots a_{in_i}$$
is a $\tau$-factorization for each $i$. Because $D$ is $\tau$-atomic and $\tau$ is refinable and associate preserving, we may replace each $a_{ij}$ with a $\tau$-atomic factorization. This allows us to assume each factor in Equation is a $\tau$-atom. Since $\tau$ is associate preserving, we may assume further that each $a_{ij} \in \overline{Irr}_\tau(D)$. Because $\tau$ is refinable,
$$\label{2}x_1 = \lambda_1 x_{2}\cdot a_{11} \cdots a_{1n_1}=(\lambda_1 \lambda_2) x_{3}\cdot a_{21} \cdots a_{2n_2}\cdot a_{11} \cdots a_{1n_1}= \cdots$$
are all $\tau$-factorizations with $a_{ij}$ $\tau$-atomic. Because $x_i \subsetneq x_{i+1}$, we must have in Equation that $n_i \geq 1$ or else $x_i \sim x_{i+1}$. This means the factorizations in each iteration of Equation increase in length. If $\{a_{ij}\}$ is infinite, then $a_{11}$ has an infinite number of adjacent vertices in $G_\tau(x_1)$, i.e $degl(a_{11}) \geq deg(a_{11})=\infty$. Otherwise, if $\{a_{ij}\}$ is finite, then one of the $a_{i_0j_0}$ for some $i_0$ and $j_0$ occurs an infinite number of times. Hence degl$(a_{i_0j_0})=\infty$ in $G_\tau(x_1)$ since arbitrarily high powers of $a_{i_0j_0}$ $\tau$-divide $x_1$. This is a contradiction and $D$ must satisfy $\tau$-ACCP as desired.
\[thm: ufd\]Let $D$ be a domain and $\tau$ a symmetric and associate preserving relation on $D^{\#}$. If $D$ is $\tau$-atomic, then the following are equivalent.\
(1) $D$ is a $\tau$-UFD.\
(2) $G_\tau(x)$ is a pseudoclique for every $x\in D^{\#}$.\
(3) $\overline{G_\tau(x)}$ is a clique for every $x\in D^{\#}$.\
(4) $G_\tau (x)$ is connected and Diam$(G_\tau (x))=1$ for every $x\in D^{\#}$.\
(5) $\overline{G_\tau(x)}$ is connected and Diam$(\overline{G_\tau(x)})=1$ for every $x\in D^{\#}$.\
(6) $G_\tau(x)$ is connected for every $x\in D^{\#}$.\
(7) $\overline{G_\tau(x)}$ is connected for every $x\in D^{\#}$.
\(1) $\Rightarrow$ (2) Let $x\in D^{\#}$. Let $x=\lambda a_1 \cdots a_m$ be the unique $\tau$-atomic factorization of $x$ up to rearrangement and associates. Because $\tau$ is associate preserving, one may adjust the unit factor in front if necessary, to assume without loss of generality $a_i\in \overline{Irr}(D)$ for each $i$. After rearrangement, we may take the first $n$ $\tau$-factors to be distinct up to associate, and assume the last $m-n$ $\tau$-factors are repeated associates of $a_1, \ldots, a_n$. Since these are the only irreducible $\tau$-divisors of $x$, $V(G_\tau(x))=\{a_1, \ldots, a_n\}$. It is clear that given any two vertices, $a_i, a_j \in V(G_\tau(x))$, $a_ia_j \in E(G_\tau(x))$, because $a_i \tau a_j$ and $a_i a_j \mid_\tau x$. If there are repeated associates in the $\tau$-factorization, say $x= \lambda a_1^{e_1} \cdots a_n^{e_n}$ with $n \leq m$ is a factorization with $e_1 + e_2 + \ldots + e_n=m$. Then if $e_i >1$, we have $a_i \tau a_i$ and $a_ia_i \mid_{\tau} x$, so we have $e_i - 1$ loops attached to $a_i$. Hence $G_\tau (x)$ is a pseudoclique.\
It is immediate from the definition of the reduced $\tau$-irreducible divisor graph that (2) $\Leftrightarrow$ (3) and $(6) \Leftrightarrow (7)$. Furthermore, $(2)\Leftrightarrow (4)$ and $(3)\Leftrightarrow (5)$ respectively since a connected graph is complete if and only if the diameter is $1$.\
If $G_\tau(x)$ is complete it is certainly connected, so (2) $\Rightarrow$ (6). It now suffices to show that (6) $\Rightarrow$ (1). The following is a modification of the proof of [@Axtellidgd Theorem 2.1]. Let $\mathcal{A}$ be the set of all non-zero, non-units which admit at least two distinct $\tau$-atomic factorizations. We show $\mathcal{A}=\emptyset$. Suppose otherwise and let $n:=min_{x\in \mathcal{A}}\{k \mid x=\lambda a_1 \cdots a_k \text{ is a } \tau\text{-atomic factorization}\}$. Clearly, $n\geq 2$. Let $y\in \mathcal{A}$ such that $y=\lambda a_1 \cdots a_n$. Then there is a distinct $\tau$-atomic factorization $y=\mu b_1 \cdots b_t$ with $t \geq n$. There is a path in $G_\tau(x)$ connecting $b_1$ and $a_1$, so without loss of generality we may as well have picked at $\tau$-atomic factorization such that $b_1$ and $a_1$ are actually adjacent in $G_\tau(x)$. If $a_i$ and $b_j$ were associates for any $1 \leq i \leq n$, $1 \leq j \leq t$, then $\frac{y}{a_i}=\lambda'a_1 \cdots \widehat{a_i} \cdots a_n = \mu' b_1 \cdots \widehat{b_k} \cdots b_t$ provides two distinct $\tau$-atomic factorizations of a non-zero, non-unit but contradicts the minimality of $n$. Since we chose $a_1$ and $b_1$ to be adjacent, we know there is a $\tau$-atomic factorization of the form $y=\gamma a_1b_1 c_1 \cdots c_m=\lambda a_1 \cdots a_n$. But this again yields two distinct $\tau$-atomic factorizations of $\frac{y}{a_1}$: $$\frac{y}{a_1}=\gamma b_1 c_1 \cdots c_m=\lambda a_2 \cdots a_n$$ again contradicting the minimality of $n$. Hence, $\mathcal{A}$ must be empty as desired, completing the proof.
\[thm: ffd\] Let $D$ be a domain and $\tau$ a symmetric, refinable and associate preserving relation on $D^{\#}$. If $D$ is $\tau$-atomic, then consider the following statements.\
(1) $G_\tau(x)$ is finite for every $x \in D^{\#}$.\
(2) $\overline{G_\tau(x)}$ is finite for every $x \in D^{\#}$.\
(3) $D$ is a $\tau$-irreducible divisor finite domain.\
(4) $D$ is a $\tau$-weak finite factorization domain.\
(5) $D$ is a $\tau$-finite factorization domain.\
(6) For all $x\in D^{\#}$, degl$(a)< \infty$ for all $a \in V(G_\tau(x))$.\
(7) For all $x\in D^{\#}$, deg$(a)< \infty$ for all $a \in V(G_\tau(x))$.\
We have (1)-(5) are equivalent, (3) $\Rightarrow$ (6) $\Rightarrow$ (7), and if we assume further that $\tau$ is reflexive, then (7) $\Rightarrow$ (3) and all are equivalent.
We begin by showing (1)-(5) are equivalent. $(1) \Leftrightarrow (2)$ and (3) $\Leftrightarrow$ (1) are immediate from the definitions. (5) $\Rightarrow$ (4) is clear since every every $\tau$-divisor up to associate must appear as a $\tau$-factor in one of the finitely many $\tau$-factorizations. (4) $\Rightarrow$ (3) every $\tau$-atomic divisor up to associate is certainly among the $\tau$-divisors.\
We need only prove (3) $\Rightarrow$ (5). We modify the proof of [@anderson90 Theorem 5.1]. Let $a\in D^{\#}$. Let $a_1, a_2, \ldots a_n$ be the collection of all non-associate $\tau$-irreducible divisors of $a$. Because $D$ is $\tau$-atomic and $\tau$ is refinable and associate preserving, given any $\tau$-factorization, we could $\tau$-refine it into a $\tau$-atomic factorization. In this way, every $\tau$-factorization corresponds to a $\tau$-factorization of the form $$a=\lambda a_1^{s_1} \cdots a_n^{s_n}$$ with $0 \leq s_i$. If we can show that the set of factorizations of this form is finite, then every $\tau$-factorization occurs as some grouping of these $\tau$-atomic factors.\
We suppose first that for each $i$, there is a bound $N_i$ such that every $\tau$-refinement of a $\tau$-factorization as above has $0 \leq s_i \leq N_i$. Then this set is finite, with $N_1 \cdot N_2 \cdots N_n$ elements in it. So now we must have some $s_i$ which is unbounded, say $s_1$ is unbounded. Then for each $k \geq 1$, we can write $$x=\lambda_k a_1^{s_{k1}} \cdots a_n^{s_{kn}}$$ $\tau$-factorizations with $s_{k1} < s_{k2} < s_{k3} < \cdots$. Suppose that in the set of factorizations, $\{ s_{ki} \}$ is bounded for $1 < i \leq n$. Then there are only finitely many choices for $s_{k2}, \ldots, s_{kn}$, we must have $s_{k2}=s_{j2}, s_{k3}=s_{j3}, \ldots, s_{kn}=s_{jn}$ for some $j >k$. But then we have $$x=\lambda_j a_1^{s_{j1}} \cdots a_n^{s_{jn}}=\lambda_k a_1^{s_{k1}} \cdots a_n^{s_{kn}}.$$ Since $D$ is a domain, we can cancel leaving $\lambda_j a_1^{s_{j1}}=\lambda_k a_1^{s_{k1}}$ with $s_{j1} > s_{k1}$, a contradiction.\
This means some set $\{ s_{ki} \}$ is unbounded for a fixed $i$ with $1 < i \leq n$. Without loss of generality suppose it is $i=2$. We continue in this manner until we get subsequences with $s_{11} < s_{21} < s_{31} < \cdots$ and $s_{21} < s_{22} < s_{32} < \cdots.$ But this means we have $$x=\lambda_1 a_1^{s_{11}} \cdots a_n^{s_{1n}} = \lambda_2 a_1^{s_{21}} \cdots a_n^{s_{2n}}$$ with $s_{1i} < s_{2i}$ for each $i$, a contradiction. Thus this set must be finite, and the proof is complete.\
(5) $\Rightarrow$ (6) $\Rightarrow$ (7) is immediate.\
We now suppose in addition that $\tau$ is reflexive and show (7) $\Rightarrow$ (3). Suppose there is a $a\in D^{\#}$ such that $\{a_i \}_{i\in I}$ is an infinite collection of non-associate $\tau$-atomic divisors of $a$. Suppose for each $i \in I$, that $a= \lambda_i a_i a_{i1} \cdots a_{in_i}$ are the given $\tau$-factorizations showing $a_i$ is a $\tau$-atomic divisor of $a$. But then since $\tau$ is reflexive, we have $a \tau a$, so $$a^2=a\cdot a = \left(\lambda_i a_i a_{i1} \cdots a_{in_i}\right) \left(\lambda_j a_j a_{j1} \cdots a_{jn_j}\right)$$ is a $\tau$-factorization for any choice of $i,j \in I$. Because $\tau$-is refinable and associate preserving, we have $$a^2= (\lambda_i\lambda_j) a_i a_j a_{i1} \cdots a_{in_i} a_{j1} \cdots a_{jn_j}$$ is a $\tau$-factorization, showing $a_i$ and $a_j$ are adjacent in $G_\tau (a^2)$, for any choice of $i,j\in I$. Thus when we fix $i=i_0$, and let $j$ range over all possible choices of $a_j \in \{a_i \}_{i\in I}$, we see $a_{i_0}a_j \in E(G_\tau (a^2))$ for all $j \in I$, $j \neq i_0$. Hence deg$(a_{i_0})=\infty$ in $G_\tau(a^2)$.
Let $D$ be a domain and $\tau$ a symmetric, refinable and associate preserving relation on $D^{\#}$. If $D$ is $\tau$-atomic and if $G_\tau (x)$ (resp. $\overline{G_\tau(x)}$) is connected for every $x \in D^{\#}$, then $G_\tau (x)$ (resp. $\overline{G_\tau(x)}$) is a finite pseudoclique for every $x \in D^{\#}$. Furthermore, $\overline{G_\tau(x)}=K_n$ for some finite $n$.
If $G_\tau (x)$ is connected for every $x\in D^{\#}$, then by Theorem \[thm: ufd\] $D$ is a $\tau$-UFD which implies $D$ is a $\tau$-atomic $\tau$-irreducible divisor finite ring. Theorem \[thm: ffd\] shows $G_\tau(x)$ has a finite number of vertices. Moreover, by Theorem \[thm: ufd\], each of these vertices is adjacent to every other vertex, i.e. $G_\tau(x)$ is a finite a pseudoclique. The last statement is immediate.
The following result is well known and was proven in [@Frazier], but we include it since the $\tau$-irreducible divisor graph results yield a nice proof.
Let $D$ be a domain and $\tau$ a symmetric, refinable and associate preserving relation on $D^{\#}$. If $D$ is a $\tau$-atomic, $\tau$-irreducible divisor finite domain, then $D$ satisfies $\tau$-ACCP.
Theorem \[thm: ffd\] shows for an associate preserving and refinable $\tau$, a $\tau$-atomic $\tau$-idf domain has the property that for all $x\in D^{\#}$, $degl(a)<\infty$ for any vertex $a\in V(G_\tau(x)).$ By Theorem \[thm: accp\], this shows $D$ satisfies $\tau$-ACCP.
Acknowledgment {#acknowledgment .unnumbered}
==============
I would like to thank the referee for their helpful comments and careful reading which have improved the quality of this article. This research was conducted as a research fellow under the supervision of Professor Daniel D. Anderson while at The University of Iowa.
[9]{} D.D. Anderson, D.F. Anderson, and M. Zafrullah, *Factorization in Integral Domains*, J. Pure Appl. Algebra., **69** (1990), 1-19.
D. D. Anderson and Andrea M. Frazier, *On a general theory of factorization in integral domains*, Rocky Mountain J. Math., **41** (2011), 663–705.
D.D. Anderson and M. Naseer, *Beck’s coloring of a commutative ring*, J. Algebra, **159** (1993), 500–514.
D.F. Anderson, A. Frazier, A. Lauve, and P.S. Livingston, *The zero-divisor graph of a commutative ring. [II]{}*, Ideal theoretic methods in commutative algebra ([C]{}olumbia,[MO]{}, 1999), Lecture Notes in Pure and Appl. Math., **220**, Dekker, New York, (2001), 61–72.
D.F. Anderson and P.S. Livingston, *The zero-divisor graph of a commutative ring*, J. Algebra, **217** (1999), 434–447.
M. Axtell, N. Baeth, and J. Stickles, *Irreducible divisor graphs and factorization properties of domains*, Comm. Algebra, **39** (2011), 4148–4162.
M. Axtell and J. Stickles, *Irreducible divisor graphs in commutative rings with zero-divisors*, Comm. Algebra, **36** (2008), 1883–1893.
I. Beck, *Coloring of commutative rings*, J. Algebra, **116** (1988), 208–226.
J. Coykendall and J. Maney, *Irreducible divisor graphs*, Comm. Algebra, **35** (2007), 885–895.
J. Juett, *Generalized comaximal factorization of ideals*, J. Algebra, **352** (2012), 141–166.
S. McAdam and R. Swan, *Unique comaximal factorization*, J. Algebra, **276** (2004), 180–192.
C.P. Mooney, *Generalized factorization in commutative rings with zero-divisors*, Houston J. Math., to appear.
C.P. Mooney, *Generalized u-factorization in commutative rings with zero-divisors*, Rocky Mountain J. Math., to appear.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The exchange of two correlated pions or kaons provides the main part of the intermediate-range attraction between two baryons. Here, a dynamical model for correlated two-pion and two-kaon exchange in the baryon-baryon interaction is presented, both in the scalar-isoscalar ($\sigma$) and the vector-isovector ($\rho$) channel. The contribution of correlated $\pi\pi$ and $K{\overline{K}}$ exchange is derived from the amplitudes for the transition of a baryon-antibaryon state ($B{\overline{B''}}$) to a $\pi\pi$ or $K{\overline{K}}$ state in the pseudophysical region by applying dispersion theory and unitarity. For the $B{\overline{B''}}\to \pi\pi, K{\overline{K}}$ amplitudes a microscopic model is constructed, which is based on the hadron-exchange picture. The Born terms include contributions from baryon-exchange as well as $\rho$-pole diagrams. The correlations between the two pseudoscalar mesons are taken into account by means of $\pi\pi$-$K{\overline{K}}$ amplitudes derived likewise from a meson-exchange model, which is in line with the empirical $\pi\pi$ data. The parameters of the $B{\overline{B''}}\to \pi\pi, K{\overline{K}}$ model, which are related to each other by the assumption of $SU(3)$ symmetry, are determined by the adjustment to the quasiempirical $N{\overline{N}}\to \pi\pi$ amplitudes in the pseudophysical region. It is found that correlated $K{\overline{K}}$ exchange plays an important role in the $\sigma$-channel for baryon-baryon states with non-vanishing strangeness. The strength of correlated $\pi\pi$ plus $K{\overline{K}}$ exchange in the $\sigma$-channel decreases with the strangeness of the baryon-baryon system becoming more negative. Due to the admixture of baryon-exchange processes to the $SU(3)$-symmetric $\rho$-pole contributions the results for correlated $\pi\pi$-exchange in the vector-isovector channel deviate from what is expected in the naive $SU(3)$ picture for genuine $\rho$-exchange. In present models of the hyperon-nucleon interaction contributions of correlated $\pi\pi$ and $K{\overline{K}}$ exchange are parametrized for simplicity by single $\sigma$ and $\rho$ exchange. Shortcomings of this effective description, e.g. the missing long-range contributions, are pointed out by comparison with the dispersiontheoretic results.'
author:
- |
A. Reuber, K. Holinde, H.-C. Kim[^1], and J. Speth\
Institut für Kernphysik (Theorie),\
Forschungszentrum Jülich GmbH,\
D-52425 Jülich, Germany
title: 'CORRELATED $\pi\pi$ AND $K{\overline{K}}$ EXCHANGE IN THE BARYON-BARYON INTERACTION'
---
Introduction
============
The study of the role of strangeness degrees of freedom in low energy nuclear physics is of high current interest since it should lead to a deeper understanding of the relevant strong interaction mechanisms in the non-perturbative regime of QCD. For example, the system of a strange baryon (hyperon $Y$) and a nucleon ($N$) is in principle an ideal testing ground to investigate the importance of $SU(3)_{\rm
flavor}$ symmetry for the hadronic interactions. This symmetry is obviously broken already by the different masses of hadrons sitting in the same multiplet. However the important question arises whether (on the level of hadrons) it is broken not only kinematically but also dynamically, e.g. in the values of the coupling constants at the hadronic vertices. The answer cannot be given at the moment since the present empirical information about the $YN$ interaction is too scarce and thus prevents any definite conclusions. Hopefully the situation will be improved by experiments of elastic $\Sigma^\pm p$, $\Lambda
p$, and even $\Xi p$, currently performed at KEK [@Ieiri; @Imai].
Existing meson exchange models of the $YN$ interaction assume for the hadronic coupling constants at least $SU(3)$ symmetry, in case of models A and B of the Jülich group [@Holz] even $SU(6)$ of the static quark model. This symmetry requirement provides relations between coupling constants of a meson multiplet to the baryon current, which strongly reduce the number of free model parameters. Specifically, coupling constants at the strange vertices are then connected to nucleon-nucleon-meson coupling constants, which in turn are fixed between close boundaries by the wealth of empirical $NN$ scattering information. All $YN$ interaction models can reproduce the existing empirical $YN$ scattering data. Therefore at present the assumption of SU(3) symmetry for the coupling constants is not in conflict with experiment.
However, the treatment of the scalar-isoscalar meson sector, which provides the intermediate range baryon-baryon interaction, is conceptionally not very convincing so far. The one-boson-exchange models of the Nijmegen group start from the existence of a broad scalar-isoscalar $\pi\pi$ resonance ($\epsilon$-meson, $m_\epsilon=760\,MeV$, $\Gamma_\epsilon=640\,MeV$), which is hidden in the experiment (e.g. $\pi N\to\pi\pi N$) under the strong $\rho^0$-signal and can therefore not be identified reliably. For practical reasons the exchange of this broad $\epsilon$-meson is then approximated by the exchange of a sum of two mesons with sharp mass $m_1$ and $m_2$; the smaller mass is around $500\,MeV$ and thus corresponds to the phenomenological $\sigma$-meson in conventional OBE-models. The $\epsilon$-meson is then treated as $SU(3)$ singlet (model D [@NijII]) or as member of a full nonet of scalar mesons (model F [@NijIII] and NSC [@NijIV]). In the $SU(3)$ framework the $\epsilon$-coupling strength is equal for all baryons in model D, while it depends on four open $SU(3)$ parameters in models F and NSC, which are adjusted to $NN$ and $YN$ scattering data. In the latest Nijmegen model NSC [@NijIV] the scalar meson nonet includes apart from the $\epsilon$ the isoscalar $f_0(975)$, the isovector $a_0(980)$ and the strange mesons $\kappa$, which the authors identify [@deSwMaui] with scalar $q^2\bar q^2$-states predicted by the MIT bag model [@MIT]. This interpretation is however doubtful, at least for the $f_0(975)$ and $a_0(980)$. According to a recent theoretical analysis of the $\pi\pi$, $\pi\eta$, and $K{\overline{K}}$ system [@Janssen] in the meson exchange framework the $f_0(975)$ is a $K{\overline{K}}$ molecule bound in the $\pi\pi$ continuum, while $a_0(980)$ is dynamically generated by the $K{\overline{K}}$ threshold. Thus both mesons do not appear to be genuine quark model resonances, with the consequence that $SU(3)$ relations should not be applied to these mesons. (The non-strange members of the scalar nonet are expected to be at higher energies).
In the Bonn potential [@MHE] the intermediate range attraction is provided by uncorrelated (Fig. \[fig:1\_2\]a,b) and correlated (Fig. \[fig:1\_2\]c) $\pi\pi$ exchange processes with $NN$, $N\Delta$ and $\Delta\Delta$ intermediate states. It is known from the study of the $\pi\pi$ interaction that the $\pi\pi$ correlations are important mainly in the scalar-isoscalar and vector-isovector channel. The Bonn potential includes such correlations, however only in a rough way, namely in terms of sharp mass $\sigma'$ and $\rho$ exchange. One disadvantage of such a simplified treatment is that this parametrization cannot be transported into the hyperon sector in a well defined way. Therefore in the $YN$ interaction models of the Jülich group [@Holz], which start from the Bonn $NN$ potential, the coupling constants of the fictitious $\sigma'$-meson at the strange vertices ($\Lambda\Lambda\sigma'$, $\Sigma\Sigma\sigma'$) are essentially free parameters. In view of the little empirical information about the $YN$ interaction this feature is not satisfactory. This is especially true for an extension of the $YN$ models to baryon-baryon channels with strangeness $S=-2$. So far there is no empirical information about these channels (apart from some data on $\Xi$- and $\Lambda\Lambda$-hypernuclei). Still there is a large interest in these channels initiated by the prediction of the H-dibaryon by Jaffe [@Jaffe]. The H-dibaryon is a deeply bound 6-quark state with the same quark content as the $\Lambda\Lambda$ system ($uuddss$) and with $^1S_0$ quantum numbers. For the experimental search it is important to know whether conventional deuteron-like $\Lambda\Lambda$ states exist. An analysis of possible $S=-2$ bound states in the meson exchange framework could provide valuable information in this regard, but requires a coupled channels treatment of $\Lambda\Lambda$, $\Sigma\Sigma$, and $N\Xi$ channels. An extension of the Jülich $YN$ models to those channels is only of minor predictive power since the strength of the important $\Xi\Xi\sigma'$ vertex is completely undetermined and cannot be fixed by empirical data.
These problems can be overcome by an explicit evaluation of correlated $\pi\pi$ exchange processes in the various baryon-baryon channels. A corresponding calculation has been done for the $NN$ case (Fig. \[fig:1\_2\]c) [@Kim]. Starting point was a fieldtheoretic model for both the $N{\overline{N}}\to\pi\pi$ Born amplitudes and the $\pi\pi$-$K{\overline{K}}$ interaction [@Lohse]. With the help of unitarity and dispersion relations the amplitude for the correlated $\pi\pi$ exchange in the $NN$ interaction has been determined, showing characteristic discrepancies to $\sigma'$ and $\rho$ exchange in the (full) Bonn potential.
For the correct description of the $\pi\pi$ interaction in the scalar-isoscalar channel the coupling to the $K{\overline{K}}$ channel is essential, which is obvious from the interpretation of the $f_0(975)$ as a $K{\overline{K}}$ bound state. Apart from the $\pi\pi$-$K{\overline{K}}$ interaction model the $K{\overline{K}}$ channel is not considered in Ref. [@Kim], i.e. the coupling of the kaon to the nucleon is not taken into account. In fact, this approximation is justified in the $NN$ system [@Durso]; it is however not expected to work in channels involving hyperons.
The aim of the present paper is a microscopic derivation of correlated $\pi\pi$ as well as $K{\overline{K}}$ exchange processes in the various baryon-baryon channels with $S=0, -1, -2$ (Fig. \[fig:1\_1\]). The $K{\overline{K}}$ channel is treated on an equal footing with the $\pi\pi$ channel in order to determine reliably the influence of $K{\overline{K}}$ correlations. Our results replace the phenomenological $\sigma'$ and $\rho$ exchange in the Bonn $NN$ and Jülich $YN$ models by correlated processes and in this way eliminate undetermined model parameters (e.g. $\sigma'$ coupling constants). Corresponding interaction models thus have more predictive power and should make a sensible treatment of $S=-2$ baryon-baryon channels possible.
The formal treatment is similar to that of Refs. [@Kim; @ALV; @LinSerot] dealing with correlated $\pi\pi$ exchange in the $NN$ interaction. Due to the inclusion of the $K{\overline{K}}$ channel and different baryon masses (e.g. in the $N\Lambda$ channel) generalizations are however required at some places. Starting point is a field-theoretic model for the baryon-antibaryon ($B{\overline{B'}}$) $\to\pi\pi,\,K{\overline{K}}$ Born amplitudes in the $J^P=0^+,1^-$ channels. Besides various baryon exchange terms the model includes in complete consistency to the $\pi\pi$-$K{\overline{K}}$ interaction model [@Lohse; @Pearce_piN] also a $\rho$-pole term (cf. Fig. \[fig:1\_3\]). These Born amplitudes are analytically continued into the pseudophysical region below the $B{\overline{B'}}$-threshold. The solution of a covariant scattering equation with full inclusion of $\pi\pi-K{\overline{K}}$ correlations yields the $B{\overline{B'}}\to\pi\pi,\,K{\overline{K}}$ amplitudes in the pseudophysical region. In the $N{\overline{N}}\to\pi\pi$ channel these amplitudes are then adjusted to quasiempirical information [@Hoehler2; @Dumbrajs], which has been obtained by analytic continuation of $\pi N$ and $\pi\pi$ data. With the assumption of $SU(6)$ symmetry for the coupling constants a parameter-free description of the other particle channels can then be achieved.
Via unitarity relations the products of $B{\overline{B'}}\to\pi\pi,\,K{\overline{K}}$ amplitudes fix the singularity structure of the baryon-baryon amplitudes for $\pi\pi$ and $K{\overline{K}}$ exchange. Assuming analyticity for the amplitudes dispersion relations can be formulated for the baryon-baryon amplitudes, which connect physical amplitudes in the $s$-channel with singularities and discontinuities of these amplitudes in the pseudophysical region of the $t$-channel processes. With a suitable subtraction of uncorrelated contributions, which are calculated directly in the $s$-channel and therefore guaranteed to have the correct energy behavior, we finally obtain the amplitudes for correlated $\pi\pi$ and $K{\overline{K}}$ exchange in the baryon-baryon system.
In the next chapter we describe the underlying formalism which is used to derive correlated $\pi\pi$ and $K{\overline{K}}$ exchange potentials for the baryon–baryon amplitudes. Furthermore we present our microscopic model for the required $B{\overline{B'}}\to\pi\pi,\,K{\overline{K}}$ amplitudes. Sect.3 contains our results and also a comparison with those obtained from other models. The paper ends with some concluding remarks.
Formalism
=========
Kinematics and amplitudes {#sec:kap3_1}
-------------------------
The kinematics of a two-body scattering process $A+B \to C+D$ (cf.Fig. \[fig:3\_1\_1\]) is uniquely determined by the 4-momenta $p_A,p_B,p_C,p_D$ of the particles. Taking into account the on-mass-shell relations ($p_X^2=M_X^2, X=A,\ldots,D$) and the conservation of the total 4-momentum ($p_A+ p_B = p_C + p_D$) only two independent Lorentz-scalars can be built out of these momenta. For these Lorentz-scalars one usually introduces the three Mandelstam variables $$\begin{array} {rcccl}
s&=&(p_A+p_B)^2&=&(p_C+p_D)^2 \quad,\\ t&=&(p_C-p_A)^2&=&(p_B-p_D)^2
\quad,\\
u&=&(p_D-p_A)^2&=&(p_B-p_C)^2 \quad,
\end{array}$$ which are related by $$s+t+u=M_A^2 +M_B^2 +M_C^2 +M_D^2 \equiv \Sigma \quad.
\label{eq:3_1_0}$$
By crossing the scattering process $A+B\to C+D$ is closely related to two other processes, as indicated by Fig. \[fig:3\_1\_1\]: $$\begin{array} {lcl@{\qquad}c@{\qquad}lcl@{\qquad\qquad}l}
A,p_a& +& B,p_B &\to& C,p_C& +& D,p_D & \mbox{`$s$-channel'} \;, \\
A,p_A& +& {\overline{C}},-p_C &\to& {\overline{B}},-p_B& +& D,p_D &
\mbox{`$t$-channel'}\;, \\ A,p_A &+& {\overline{D}},-p_D &\to& C,p_C& +&
{\overline{B}},-p_B & \mbox{`$u$-channel'}\;.
\end{array}$$ Here, the channels are named according to the Mandelstam variable which denotes the squared total energy in the center-of-mass (c.m.) system.
For the $s$-channel process the particle 4-momenta in the c.m. system read: $$p_A=\left( \begin{array}{c} E_A \\ \vec p_s \end{array} \right),\quad
p_B=\left( \begin{array}{c} E_B \\ -\vec p_s \end{array}
\right),\qquad p_C=\left( \begin{array}{c} E_C \\ \vec q_s \end{array}
\right),\quad p_D=\left( \begin{array}{c} E_D \\ -\vec q_s \end{array}
\right)$$ with $E_X=\sqrt{M_X^2 + \vec p^{\;2}_X}$. The modulus of the relative momentum $\vec p_s$ ($\vec q_s$) of the initial (final) state can be expressed in terms of $s$: $$\begin{aligned}
{\vec p_s}\,^2 &=& { \left[ s - (M_A+M_B)^2 \right] \; \left[ s -
(M_A-M_B)^2 \right]
\over 4s}\quad,
\nonumber \\
{\vec q_s}\,^2 &=& { \left[ s - (M_C+M_D)^2 \right] \; \left[ s -
(M_C-M_D)^2 \right]
\over 4s}\quad.
\label{eq:3_1_1}\end{aligned}$$ The scattering angle $\vartheta_s = {\;\hbox{$\rlap{)}\!\!\!<$}\,}(\vec p_s,\vec q_s)$ is related to the Mandelstam variables by $$\cos\vartheta_s= { s (t-u) + (M_A^2 - M_B^2) (M_C^2 - M_D^2)
\over 4s |\vec p_s| |\vec q_s|}
\label{eq:3_1_2}$$
For the $t$-channel process $A+{\overline{C}} \to {\overline{B}}+D$ the c.m. 4-momenta of the particles are $$p_A=\left( \begin{array}{c} E_A \\ \vec p_t \end{array} \right),\quad
-p_C=\left( \begin{array}{c} E_C \\ -\vec p_t \end{array}
\right),\qquad p_D=\left( \begin{array}{c} E_D \\ \vec q_t \end{array}
\right),\quad -p_B=\left( \begin{array}{c} E_B \\ -\vec q_t
\end{array} \right).$$ The analogue of Eqs. \[eq:3\_1\_1\] for the modulus of the relative momenta $\vec p_t$, $\vec q_t$ and the scattering angle $\vartheta_t =
{\;\hbox{$\rlap{)}\!\!\!<$}\,}(\vec p_t,\vec q_t)$ now reads: $$\begin{aligned}
{\vec p_t}\,^2 &=& { \left[ t - (M_A+M_C)^2 \right] \; \left[ t -
(M_A-M_C)^2 \right]
\over 4t}\quad,
\nonumber \\
{\vec q_t}\,^2 &=& { \left[ t - (M_B+M_D)^2 \right] \; \left[ t -
(M_B-M_D)^2 \right]
\over 4t}\quad,
\label{eq:3_1_2a}
\\
\cos\vartheta_t &=& { t (u-s) + (M_A^2 - M_C^2) (M_D^2 - M_B^2)
\over 4t |\vec p_t| |\vec q_t|}\quad.
\label{eq:3_1_3}\end{aligned}$$
Instead of the particle 4-momenta, usually the total 4-momentum and the following three linear combinations are used to characterize the kinematics of a two-body scattering process: $$\Delta\equiv p_C-p_A=p_B-p_D, \quad P\equiv{{1\over2}}(p_A+p_C),
\quad Q\equiv{{1\over2}}(p_B+p_D) \quad.
\label{eq:3_1_3a}$$ The scalar products of these three momenta can again be expressed in terms of the Mandelstam variables
$$\begin{array}{rcl@{\qquad\qquad}rcl}
\Delta^2&=&t\quad,&
P\cdot\Delta&=&{{1\over2}}(M_C^2 - M_A^2)\quad, \\ P^2&=&{{1\over2}}(M_A^2 +
M_C^2)-t/4 \quad,& Q\cdot\Delta&=&{{1\over2}}(M_B^2 - M_D^2) \quad, \\
Q^2&=&{{1\over2}}(M_B^2 + M_D^2)-t/4 \quad,& P\cdot Q&=&(s-u)/4 \quad.
\end{array}
\label{eq:3_1_3d}$$
In covariant field-theory the scattering amplitude $T$ for a general process with $n_i$ ($n_f$) particles in the initial (final) state ${\> \vert \, i \, \rangle}$ (${\> \vert \, f \, \rangle}$) is related to the (unitary) $S$ matrix by $${\langle \, f \, \vert \>}S{\> \vert \, i \, \rangle} = {\langle \, f \, \vert \>}{i \,\rangle} - i {(2\pi)^4\over \sqrt{N_f N_i}}
\delta^{(4)}(P_f-P_i) {\langle \, f \, \vert \>} T{\> \vert \, i \, \rangle}\quad,
\label{eq:3_1_10}$$ where $P_i$ ($P_f$) denotes the total 4-momentum in the initial (final) state. The factors $N_x$ ($x=i,f$) are given by $$N_x = (2\pi)^{3n_x} \prod_{j=1}^{n_x} {2E_j\over (2 M_j)^{b_j}}\quad,$$ where the j-th particle of channel $x$ has mass $M_j$, momentum $\vec
p_j$, energy $E_j=(M_j^2 + \vec p_j\,^2)^{1/2}$ and spin $s_j$. (Spin and isospin quantum numbers are suppressed for the moment.) The exponent $b_j$ is given by $$b_j=\left\{
\begin{array}{cl}
0 & \mbox{if\ } 2s_j \mbox{\ even}, \\ 1 & \mbox{if\ } 2s_j \mbox{\
odd}.
\end{array}
\right.$$
For a two-body scattering process $A+B\to C+D$ Eq. \[eq:3\_1\_10\] reads $$\begin{aligned}
\lefteqn{
{\langle \, C p_C, D p_D \, \vert \>}S{\> \vert \, Ap_A,Bp_B \, \rangle}\; =\; {\langle \, C p_C, D p_D \, \vert \>}
{Ap_A,Bp_B \,\rangle}}
\nonumber \\
&& - \; {i\over (2\pi)^2}
\mbox{$ \sqrt{(2M_A)^{b_A} (2M_B)^{b_B}(2M_C)^{b_C}(2M_D)^{b_D}\over 16
E_A E_B E_C E_D}$}
\delta^{(4)}(p_C+p_D-p_A-p_B)
\nonumber\\
&&
\qquad\qquad \qquad\qquad {\langle \, C p_C, Dp_D \, \vert \>}T{\> \vert \, Ap_A,Bp_B \, \rangle}.
\label{eq:3_1_11}\end{aligned}$$
Particles with spin (and helicity $\lambda_X$) are described in the helicity basis according to the conventions of Jacob and Wick [@JW]. By separating off the helicity spinors $u_{X}(\vec
p_X,\lambda_X)$ of the particles from the scattering amplitudes one obtains the transition matrix ${\cal M}$. If, for instance, all four particles of the process $A+B\to C+D$ are spin-1/2 baryons, the transition matrix ${\cal M}$ is a $16\times16$ matrix in spinor space and is defined by $$\begin{aligned}
\lefteqn{
{\langle \, C \vec p_C \lambda_C, D \vec p_D \lambda_D \, \vert \>}
T {\> \vert \, A \vec p_A \lambda_A, B \vec p_B \lambda_B \, \rangle} = } \nonumber \\
&&
{\overline{u}}_{C}(\vec p_C,\lambda_C) {\overline{u}}_{D}(\vec p_D,\lambda_D)
{\cal M}_{AB \to CD}(P,Q) u_{A}(\vec p_A,\lambda_A) u_{B}(\vec
p_B,\lambda_B)\quad.
\label{eq:3_1_11a}\end{aligned}$$
Now the transition matrix ${\cal M}$ can be constructed as a linear combination of the so-called [*kinematic covariants*]{} ${\cal O}_i$, which are like ${\cal M}$ operators in spinor space. $${\cal M}(P,Q) = \sum_i c_i(s,t) {\cal O}_i(P,Q)\quad.
\label{eq:3_1_12}$$ The ${\cal O}_i$ are built up from the Dirac $\gamma$-matrices and the momenta $P$ and $Q$ (cf. Eq. \[eq:3\_1\_3a\]) in such a way that their matrix elements are Lorentz-invariant quantities. The [*invariant amplitudes*]{} $c_i(s,t)$ are Lorentz-scalars.
The number of independent kinematic covariants for a given scattering process corresponds to the number of independent helicity amplitudes and is determined by the dimension of the spinor space and invariance principles for the underlying interaction. For the scattering of four spin-1/2 baryons ($A+B\to C+D$) there are in general eight independent kinematic covariants. For the elastic scattering ($A+B\to A+B$) their number is reduced to six due to time reversal invariance.For the ‘superelastic’ scattering of four identical particles ($A+A\to A+A$) the number of independent kinematic covariants is further reduced to five due to the symmetry under particle exchange.
The set of kinematic covariants is not unique. However, for the forthcoming it is essential that the ${\cal O}_i$ are chosen in such a way that the invariant amplitudes $c_i(s,t)$ do not contain any kinematic singularities, but only ‘physical’ singularities demanded by unitarity. In the case of four spin-1/2 particles this condition is fulfilled by the set of eight covariants given in Ref. [@ScadJon]. This set is based on the so-called Fermi-covariants $S,P,V,A,T$: $$\begin{array}{lcccl@{\quad\qquad}lcccl}
S &\equiv& {\cal O}_S &=& {{\rm 1 \>\! \llap{I}}}_4 \otimes {{\rm 1 \>\! \llap{I}}}_4 \quad,& T
&\equiv&{\cal O}_T &=& {{1\over2}}\sigma_{\mu\nu} \otimes
\sigma^{\mu\nu}\quad,
\\
P &\equiv&{\cal O}_P &=& \gamma_5 \otimes \gamma_5\quad, & && {\cal
O}_6 &=& {{\rm 1 \>\! \llap{I}}}_4 \otimes \gamma_\mu P^\mu \,-\, \gamma_\mu Q^\mu
\otimes {{\rm 1 \>\! \llap{I}}}_4\quad,
\\
V &\equiv&{\cal O}_V &=& \gamma_\mu \otimes \gamma^\mu \quad,& &&
{\cal O}_7 &=& \gamma_5 \otimes \gamma_5 \gamma_\mu P^\mu\quad,
\\
A &\equiv&{\cal O}_A &=& \gamma_5 \gamma_\mu \otimes
\gamma_5\gamma^\mu\quad, & && {\cal O}_8 &=& \gamma_5 \gamma_\mu
Q^\mu \otimes \gamma_5\quad.
\label{eq:3_1_13}
\end{array}$$ These covariants are of the form ${\cal O}_i(P,Q) = {\cal
O}^{(1)}_i(Q)\otimes {\cal O}^{(2)}_i(P)$ and the matrix elements have to be evaluated according to $$\begin{aligned}
\lefteqn{
{\overline{u}}_{C}(\vec p_C,\lambda_C) {\overline{u}}_{D}(\vec p_D,\lambda_D)
{\cal O}_i(P,Q) u_{A}(\vec p_A,\lambda_A) u_{B}(\vec p_B,\lambda_B)=
} \nonumber \\ &&\!\!\!
\left[ {\overline{u}}_{C}(\vec p_C,\lambda_C)
{\cal O}^{(1)}_i(Q)u_{A}(\vec p_A,\lambda_A)\right] \;
\left[ {\overline{u}}_{D}(\vec p_D,\lambda_D) {\cal O}^{(2)}_i(P)u_{B}
(\vec p_B,\lambda_B)\right].\end{aligned}$$
Dispersion relations for baryon-baryon amplitudes of $\pi\pi$ and $K{\overline{K}}$ exchange {#sec:kap3_4}
--------------------------------------------------------------------------------------------
For a general two-body scattering process the physical regions of the Mandelstam variables for the $s$-, $t$- and $u$-channel reaction are non-overlapping. Therefore the transition matrices in the three channels can be interpreted as independent branches of one operator ${\cal M}$ defined in the various kinematic regions. In the $t$-channel ($A+{\overline{C}}\to{\overline{B}}+D$), for instance, the scattering amplitude is related to the invariant amplitudes by $$\begin{aligned}
\lefteqn{
{\langle \, {\overline{B}} -\vec p_B \lambda_B, D \vec p_D \lambda_D \, \vert \>}
T {\> \vert \, A \vec p_A \lambda_A, {\overline{C}} -\vec p_C \lambda_C \, \rangle} = }
\nonumber \\ &&
\hspace{-.6cm}\sum_{i=1}^8 c_i(s,t)
\left[{\overline{v}}_{C}(-\vec p_C,\lambda_C) {\cal O}^{(1)}_i(Q)u_{A}
(\vec p_A,\lambda_A)\right]
\left[{\overline{u}}_{D}(\vec p_D,\lambda_D){\cal O}^{(2)}_i(P)v_{B}
(-\vec p_B,\lambda_B)\right].
\nonumber \\\end{aligned}$$ By introducing the concept of analyticity the invariant amplitudes in the three channels become closely related: The invariant amplitudes $c_i(s,t)$ are supposed to be analytic functions (except for the physical singularities) in the whole complex $st$-plane. Therefore, if all these physical singularities are known the $c_i(s,t)$ can be deduced at any point in the complex Mandelstam plane by the formulation of dispersion integrals.
The singularity structure of the invariant amplitudes is completely determined by the unitarity of the $S$-matrix. In terms of the scattering amplitude unitarity of the $S$-matrix is expressed as $$i\left[
{\langle \, f \, \vert \>}T{\> \vert \, i \, \rangle} - {\langle \, f \, \vert \>}T^\dagger{\> \vert \, i \, \rangle}
\right]
=
\sum_n {(2\pi)^4\over N_n} \delta^{(4)}(P_i-P_n) {\langle \, f \, \vert \>}T^\dagger{\> \vert \, n \, \rangle}
{\langle \, n \, \vert \>}T{\> \vert \, i \, \rangle}\quad,
\label{eq:3_3_1b}$$ where $P_n$ and $P_f=P_i$ denote the total 4-momenta. The summation in Eq. \[eq:3\_3\_1b\] is to be understood over all [*physical*]{} states $n$, i.e. states which are energetically accessible for the system with energy $P_i^0$.
The singularities of the baryon-baryon ($A+B\to C+D$) amplitudes generated by (correlated) $\pi\pi$ and $K{\overline{K}}$ exchange are most easily derived using the unitarity relation for the $t$-channel reaction $A+{\overline{C}}\to D+{\overline{B}}$. For this, one has to restrict the summation in Eq. \[eq:3\_3\_1b\] to physical $\pi\pi$ and $K{\overline{K}}$ states; i.e., contributions of single-meson poles or of heavier two-meson and multi-meson (e.g. $3\pi$, $4\pi$) channels are disregarded. In the c.m. system ($P_i=P_f=(\sqrt{t},\vec 0)$) the summation over the states $n$ then reads $$\begin{aligned}
\lefteqn{
\sum_n {(2\pi)^4\over N_n} \delta^{(4)}(P_i-P_n) {\> \vert \, n \, \rangle}{\langle \, n \, \vert \>}
=}
\nonumber \\
&& ={1\over (2\pi)^2}
\sum_{\mu\bar\mu=\pi\pi,K\bar K} N_{\mu{\overline{\mu}}}
\int d^3k
\delta(\sqrt{t}-2\omega_\mu(k)) {1\over 4\omega_\mu(k)^2}
{\> \vert \, \mu{\overline{\mu}} , \vec k \, \rangle} {\langle \, \mu{\overline{\mu}} , \vec k \, \vert \>}
\nonumber \\
&& ={1\over 32\pi^2}
\sum_{\mu\bar\mu=\pi\pi,K\bar K} N_{\mu{\overline{\mu}}}
\sqrt{t-4m_\mu^2 \over t} \theta(t-4m_\mu^2)
\int d^2\hat k_{\mu\bar\mu}
{\> \vert \, \mu{\overline{\mu}} , \vec k_{\mu\bar\mu} \, \rangle}
{\langle \, \mu{\overline{\mu}} , \vec k_{\mu\bar\mu} \, \vert \>}\end{aligned}$$ with $\omega_\mu(k)=\sqrt{m_\mu^2+k^2}$ and the on-shell momentum $k_{\mu\bar\mu}= \sqrt{t/4 - m_\mu^2}$. The symmetry factor $N_{\mu{\overline{\mu}}}$ is introduced in order to obtain the correct phase space in case of identical particles: $$N_{\mu\bar\mu} =
\left\{ \begin{array}{r@{\quad\mbox{if}\quad}l}
1/2 & \mu{\overline{\mu}}=\pi\pi \\ 1 &\mu{\overline{\mu}}= K{\overline{K}}
\end{array}
\right.
\label{eq:3_3_1a}$$
Now, the kinematic covariants ${\cal O}_i$ in Eq. \[eq:3\_1\_13\] have been chosen in such a way [@ScadJon] that the left-hand side of Eq. \[eq:3\_3\_1b\] yields the imaginary part of the invariant amplitudes, i.e.$$\begin{aligned}
\lefteqn{
i{\langle \, D {\overline{B}}, \vec q\, \lambda_D \lambda_B \, \vert \>} T-T^\dagger {\> \vert \, A
{\overline{C}}, \vec p\, \lambda_A \lambda_C \, \rangle} = }
\nonumber \\
&&\hspace{-.5cm} -2 \sum_i {\sl Im}\left[ c_i (s,t)\right]
{\overline{v}}_{C}(-\vec p,\lambda_C) \; {\overline{u}}_{D}(\vec q,\lambda_D)
{\cal O}_i(P,Q) u_{A}(\vec p,\lambda_A) v_{B}(-\vec q,\lambda_B)\, .\end{aligned}$$ Hence, the unitarity relation for the helicity amplitudes of the process $A+{\overline{C}}\to D+{\overline{B}}$ becomes in the c.m. system: $$\begin{aligned}
\lefteqn{
\sum_i {\sl Im}\left[ c_i (s,t)\right]
{\overline{v}}_{C}(-\vec p,\lambda_C) \; {\overline{u}}_{D}(\vec q,\lambda_D)
{\cal O}_i(P,Q) u_{A}(\vec p,\lambda_A) v_{B}(-\vec q,\lambda_B) = }
\nonumber \\ && =-{1\over 64\pi^2}
\sum_{\mu\bar\mu=\pi\pi,K\bar K} N_{\mu{\overline{\mu}}}
\sqrt{t-4m_\mu^2 \over t} \theta(t-4m_\mu^2)
\nonumber \\
&&
\quad\int d^2\hat k_{\mu\bar\mu}
{\langle \, \mu{\overline{\mu}} , \vec k_{\mu\bar\mu} \, \vert \>}
T{\> \vert \, D {\overline{B}}, \vec q\, \lambda_D \lambda_B \, \rangle}^*
{\langle \, \mu{\overline{\mu}} , \vec k_{\mu\bar\mu} \, \vert \>} T
{\> \vert \, A {\overline{C}}, \vec p \,\lambda_A \lambda_C \, \rangle}\quad .
\label{eq:3_3_3}\end{aligned}$$
From the beginning, the unitarity relations \[eq:3\_3\_1b\], \[eq:3\_3\_3\] are just defined above the kinematic threshold of the process $A+{\overline{C}}\to D+{\overline{B}}$, that is for $t \geq t_0\equiv
\max\{M_A+M_C,M_B+M_D\}$. However, they can be continued analytically into the pseudophysical region ($4m_\pi^2\leq t \leq t_0$) as discussed in Ref. [@Kamp]. Below the kinematic threshold the baryon-antibaryon momenta become imaginary. According to Ref. [@Kamp] the unitarity relation for the $S$-matrix, which can be written symbolically as $\left[S(p)\right]^* S(p)=1$, has to be continued to complex momenta by $\left[S(p^*)\right]^* S(p)=1$. As explained in Ref. [@BrownJack], for imaginary momenta this is equivalent to evaluating the expression $\left[S(p)\right]^*S(p)=1$ with real dummy variables for the momenta and replacing these dummy variables at the very end (after having performed all complex conjugations) by the imaginary momenta.
The right-hand side of Eq. \[eq:3\_3\_3\] obviously vanishes below the $\pi\pi$ threshold at $t=4m_\pi^2$. Since for the processes considered here the left-hand cuts, which are due to unitarity constraints for the $u$-channel process, do not extend up to $t=4m_\pi^2$ (for fixed $s$ lying inside the physical $s$-channel region), the invariant amplitudes $c_i(s,t)$ are real-analytic functions of $t$, i.e. $c_i(s,t^*)=c_i(s,t)^*$. According to Eq. \[eq:3\_3\_3\] and $$2i {\sl Im}\left[ c_i (s,t+i\epsilon)\right]=c_i(s,t+i\epsilon)-
c_i(s,t-i\epsilon) \quad,$$ $c_i(s,t)$ has a branch cut along the real $t$-axis extending from $t=4m_\pi^2$ to $t=+\infty$. Corresponding statements hold for the $K{\overline{K}}$ branch cut.
For c.m. energies below $1\,GeV$ the $\pi\pi$ interaction is dominated by the $JI=00,11$ partial waves [@Lohse]. At low transfered momenta, relevant in low-energy baryon-baryon scattering, correlations between two exchanged pions (kaons) are therefore only considerable when the exchanged $\pi\pi$ ($K{\overline{K}}$) system is in a state with relative angular momentum $J=0$ and isospin $I=0$ (‘$\sigma$-channel’) or $J=1$ and $I=1$ (‘$\rho$-channel’) [@Kim]. In our approach only the correlated part of two-pion and two-kaon exchange is evaluated by dispersiontheoretic means. The uncorrelated part has to be calculated directly in the $s$-channel in order to include all $t$-channel partial waves and to guarantee the correct energy dependence of this contribution. In the $\Lambda N$ channel, for instance, the iterative two-pion exchange with an $N\Sigma$ intermediate state becomes complex above the $N\Sigma $ threshold. This behavior cannot be reproduced with a single-variable dispersion relation in the $t$-channel which will be applied to the correlated contribution (see below).
Hence, the following dispersiontheoretic considerations can be limited to the $\sigma$ and $\rho$ channel of $\pi\pi$ ($K{\overline{K}}$) exchange. For this, the $B{\overline{B'}}\to \mu{\overline{\mu}}$ amplitudes on the right-hand side of the unitarity relation \[eq:3\_3\_3\] are decomposed into partial waves [@GW]. Choosing the coordinate system so that $\vec p$ points along the $\hat
z$-axis and $\vec q$ lies in the $\hat x \hat z$-plane the partial wave decomposition gives $$\begin{aligned}
\lefteqn{
\int d^2\hat k_{\mu\bar\mu}
{\langle \, \mu{\overline{\mu}} , \vec k_{\mu\bar\mu} \, \vert \>} T
{\> \vert \, D {\overline{B}}, \vec q\, \lambda_D \lambda_B \, \rangle}^*
{\langle \, \mu{\overline{\mu}} , \vec k_{\mu\bar\mu} \, \vert \>} T
{\> \vert \, A {\overline{C}}, \vec p \,\lambda_A \lambda_C \, \rangle} =
}
\nonumber \\
&&\!\!\!\!\!
\sum_J {2J+1 \over 4\pi}
d^J_{\lambda_A-\lambda_C,\lambda_D-\lambda_B}(\cos\vartheta_t)
{\langle \, \mu{\overline{\mu}} \, \vert \>} T^J (t) {\> \vert \, D {\overline{B}}, \lambda_D \lambda_B \, \rangle}^*
{\langle \, \mu{\overline{\mu}} \, \vert \>} T^J (t) {\> \vert \, A {\overline{C}}, \lambda_A \lambda_C \, \rangle},
\nonumber \\
\label{eq:3_4_3}\end{aligned}$$ where the on-shell momenta $p=p(t)$ and $q=q(t)$ (see Eq. \[eq:3\_1\_2a\]) are suppressed as arguments of the partial wave decomposed $T$ matrix elements. Note that the right-hand side depends on the Mandelstam variable $s$ only via the angle $\vartheta_t=\vartheta_t(s,t)$ (see Eq. \[eq:3\_1\_3\]) between $\vec p$ and $\vec q$. Now, by restricting the sum over $J$ in Eq. \[eq:3\_4\_3\] to $J=0$ ($J=1$) the contribution $c_i^{(J)}(s,t)=
c^\sigma_i (s,t)$ ($c^\rho_i (s,t)$) of the $\pi\pi$ and $K{\overline{K}}$ intermediate states to the discontinuity of the invariant amplitudes $c_i(s,t)$ in the $\sigma$ ($\rho$) channel can be isolated. From the unitarity relation \[eq:3\_3\_3\] we obtain $$\begin{aligned}
\lefteqn{
\sum_i {\sl Im}\left[ c^{(J)}_i (s,t)\right]
{\overline{v}}_{C}(-\vec p,\lambda_C) \; {\overline{u}}_{D}(\vec q,\lambda_D)
{\cal O}_i(P,Q) u_{A}(\vec p,\lambda_A) v_{B}(-\vec q,\lambda_B) = }
\nonumber \\ && =\sum_{\mu\bar\mu=\pi\pi,K\bar K}
d^J_{\lambda_A-\lambda_C,\lambda_D-\lambda_B}(\cos\vartheta_t)
H^J_{\mu{\overline{\mu}}} (t) \nonumber \\ && \hspace{2.7cm}
{\langle \, \mu{\overline{\mu}} \, \vert \>} T^J (t) {\> \vert \, D {\overline{B}}, \lambda_D \lambda_B \, \rangle}^*
{\langle \, \mu{\overline{\mu}} \, \vert \>} T^J (t) {\> \vert \, A {\overline{C}}, \lambda_A \lambda_C \, \rangle}
\quad,
\label{eq:3_4_4}\end{aligned}$$ with the abbreviation $$H^J_{\mu{\overline{\mu}}} (t) \equiv -{2J+1 \over 256\pi^3}
N_{\mu{\overline{\mu}}}
\sqrt{t-4m_\mu^2 \over t} \theta(t-4m_\mu^2)\quad.$$ Eq. \[eq:3\_4\_4\] is a system of linear equations for the discontinuities ${\sl Im}\left[ c^{(J)}_i (s,t)\right]$. Its solution provides the discontinuities as linear combinations of the following products of $B{\overline{B'}}\to\mu{\overline{\mu}}$ helicity amplitudes: $$F_{\lambda_D\lambda_B,\lambda_A\lambda_C}^{J}\equiv
\sum_{\mu\bar\mu=\pi\pi,K\bar K} H^J_{\mu{\overline{\mu}}} (t)
{\langle \, \mu{\overline{\mu}} \, \vert \>} T^J (t) {\> \vert \, D {\overline{B}}, \lambda_D \lambda_B \, \rangle}^* \;
{\langle \, \mu{\overline{\mu}} \, \vert \>} T^J (t) {\> \vert \, A {\overline{C}}, \lambda_A \lambda_C \, \rangle} \quad.$$ From the symmetry properties of the $B{\overline{B'}}\to\mu{\overline{\mu}}$ helicity amplitudes, which are due to the parity invariance of the underlying strong interaction (see e.g. Ref. [@JW]), $${\langle \, \mu{\overline{\mu}} \, \vert \>}T^J(k,q){\> \vert \, B{\overline{B'}}, \lambda\lambda' \, \rangle}
= {\langle \, \mu{\overline{\mu}} \, \vert \>}T^J(k,q){\> \vert \, B{\overline{B'}},-\lambda -\lambda' \, \rangle}\quad,$$ the following relations can be deduced: $$F_{\lambda_D\lambda_B,\lambda_A\lambda_C}^{J}=
F_{(-\lambda_D)(-\lambda_B),\lambda_A\lambda_C}^{J}=
F_{\lambda_D\lambda_B,(-\lambda_A)(-\lambda_C)}^{J}=
F_{(-\lambda_D)(-\lambda_B),(-\lambda_A)(-\lambda_C)}^{J}\quad,$$ Therefore only four linear independent $F_{\lambda_D\lambda_B,\lambda_A\lambda_C}^{J}$ exist for $J>0$. For $J=0$ there is only one independent $B{\overline{B'}}\to\mu{\overline{\mu}}$ helicity amplitude, ${\langle \, \mu{\overline{\mu}} \, \vert \>}V^J(k,q){\> \vert \, B{\overline{B'}}, ++ \, \rangle}$, and consequently only one independent $F_{\lambda_D\lambda_B,\lambda_A\lambda_C}^{J}$. Using the explicit representation of helicity spinors in Appendix \[sec:AnhA\] the evaluation of the matrix elements of the kinematic covariants ${\cal O}_i$ in Eq. \[eq:3\_4\_4\] is straightforward.
In case of unequal baryon masses $M_B\neq M_{B'}$ the analytic structure of the $B{\overline{B'}}\to\mu{\overline{\mu}}$ helicity amplitudes is much more involved compared to the case when the baryon masses are equal. This complicates an analytic continuation of the amplitudes to the pseudophysical region. The $\Lambda{\overline{\Sigma}}$ channel is the only channel where this problem arises. In order to facilitate an easy handling of the expressions we treat this channel approximately by setting the mass of the $\Lambda$ and of the $\Sigma$ equal to the average mass $(M_\Lambda+M_\Sigma)/2$. This approximation is justified for two reasons: First the mass difference between $\Lambda$ and $\Sigma$ hyperon is small ($77 MeV$) on the baryonic scale. Second, since we evaluate the uncorrelated part directly in the $s$-channel the important energy dependence of this contribution is not affected by our approximation. The correlated contribution on the other hand is known to have a rather smooth energy dependence which is supposed not to change drastically by our approximation.
In the following we restrict ourselves therefore to the case where $$M_A=M_C\equiv M \quad, \qquad M_B=M_D\equiv M' \quad.
$$
Solving the system of linear equations \[eq:3\_4\_4\] yields that in the $\sigma$ channel ($J=0$) all discontinuities ${\sl Im}\left[
c^\sigma_i (s,t)\right]$ vanish except for the scalar component $$\begin{aligned}
&&{\sl Im}\left[ c^\sigma_S (t)\right] = {M' M \over q p}
F_{++,++}^{J=0}\quad,
\nonumber
\\
&&{\sl Im}\left[ c^\sigma_P (s,t)\right]= {\sl Im}\left[ c^\sigma_V
(s,t)\right]= {\sl Im}\left[ c^\sigma_A (s,t)\right]= {\sl Im}\left[
c^\sigma_T (s,t)\right]= {\sl Im}\left[ c^\sigma_6 (s,t)\right]=0 \quad,
\nonumber \\
\label{eq:3_4_5}\end{aligned}$$ In contrast, in the $\rho$ channel only the axial-vector component vanishes:
$$\begin{array}{lcl}
{\sl Im}\left[ c^\rho_S (s,t)\right]&=&\!\!\!\!\!\!\! -
\begin{array}[t]{lcl}
{1\over2} g(t) \sqrt{t} q p \cos\vartheta_t (s,t) & [& -2\sqrt{t}(M'
+ M) (2 F_{++,++}^{J=1} + F_{+-,+-}^{J=1} )
\\ &&
+\sqrt{2}(t + 4M' M) (F_{++,+-}^{J=1} + F_{+-,++}^{J=1})],
\end{array}
\\
{\sl Im}\left[ c^\rho_P (s,t)\right]&=&
\begin{array}[t]{lcl}
g(t) q p \cos\vartheta_t (s,t) &[& (M' + M)(t - 4M' M)
F_{+-,+-}^{J=1} \\ && -2\sqrt{2} \sqrt{t} (p^2 F_{++,+-}^{J=1} + q^2
F_{+-,++}^{J=1}) ] \quad,
\end{array}
\\
{\sl Im}\left[ c^\rho_V (t)\right]&=&
\begin{array}[t]{lcl}
g(t) \sqrt{t} &[& -\sqrt{t} (M' + M)(t/4- M'^2 - M^2 + M' M)
F_{+-,+-}^{J=1}
\\ &&
+2\sqrt{2} M' M (p^2 F_{++,+-}^{J=1} + q^2 F_{+-,++}^{J=1}) ] \quad,
\end{array}
\\
{\sl Im}\left[ c^\rho_A (s,t)\right]&=&0 \quad,
\\
{\sl Im}\left[ c^\rho_T (t)\right]&=&
\begin{array}[t]{lcl}
{1\over4} g(t) t &[& (M' + M)(t - 4M' M) F_{+-,+-}^{J=1} \\ &&
-2\sqrt{2} \sqrt{t}(p^2 F_{++,+-}^{J=1} + q^2 F_{+-,++}^{J=1}) ]
\quad,
\end{array}
\\
{\sl Im}\left[ c^\rho_6 (t)\right]&=&
\begin{array}[t]{lcl}
g(t) \sqrt{t} &[& (M'^2 - M^2)\sqrt{t} F_{+-,+-}^{J=1} \\
&&-2\sqrt{2} (p^2 M' F_{++,+-}^{J=1} - q^2 M F_{+-,++}^{J=1}) ]\quad,
\end{array}
\end{array}
\nonumber \\
\label{eq:3_4_6}$$
with $$g(t)\equiv {M' M\over 2 q^2 p^2 (M' + M) t}\quad.$$ The discontinuity of the pseudoscalar and the tensor component are obviously related by $${\sl Im}\left[ c^\rho_P (s,t)\right]= {4q p \cos\vartheta_t \over t}
{\sl Im}\left[ c^\rho_T (t)\right]
\stackrel {(\ref{eq:3_1_3})}{=} {u-s\over t}{\sl Im}
\left[ c^\rho_T (t)\right]
\quad.
\label{eq:3_lindep}$$ Consequently only four of the five nonvanishing discontinuities are linear independent in alignment with the number of independent $F_{\lambda_D\lambda_B,\lambda_A\lambda_C}^{J=1}$.
Except for poles (corresponding to single-particle exchange) and cuts the invariant amplitudes $c_i^{(J)}(s\; \mbox{fixed},t)$ are real-analytic functions of $t$. Therefore, fixed-$s$ dispersion relations can be formulated for the $c_i^{(J)}(s,t)$. Since we want to restrict the dispersiontheoretic evaluation to the contribution of (correlated) $\pi\pi$ and $K{\overline{K}}$ exchange to the baryon-baryon amplitudes we take into account only those singularities which are generated by $\pi\pi$ and $K{\overline{K}}$ intermediate states, namely the discontinuities of Eqs. \[eq:3\_4\_5\]–\[eq:3\_4\_6\] due to the $\pi\pi$ and $K{\overline{K}}$ unitarity cut (‘right-hand cut’). The left-hand cuts, which are due to unitarity constraints for the $u$-channel reaction, can be neglected in the baryon-baryon channels considered here, since they start at large, negative $t$-values (from which they extend to $-\infty$) and are therefore far away from the physical region relevant for low-energy $s$-channel processes. For identical baryons (e.g. $NN\to NN$) this is only true if the dispersion relations are applied only to the direct baryon-baryon amplitude and the antisymmetrization of the amplitudes is not taken into account from the very beginning but just for the final $s$-channel amplitudes. Otherwise, crossing of the exchange diagram would result in a $u$-channel cut starting at $u=4m_\pi^2$ which could not be neglected in the dispersion integrals [@LinSerot].
In this work the $B{\overline{B'}}\to\mu{\overline{\mu}}$ amplitudes, which enter in Eq. \[eq:3\_4\_4\], are derived from a microscopic model which is based on the hadron-exchange picture (see Sect. \[sec:kap4\]). Of course, this model has a limited range of validity: for energies far beyond $t'_{max}\approx100m_\pi^2$ it cannot provide reliable results. The dispersion integral for the invariant amplitudes extending in principle along the whole $\pi\pi$ right-hand cut has therefore to be limited to an upper bound ($t'_{max}$). In addition left-hand cuts and unphysical cuts introduced for instance by the form factor prescription of the microscopic model for the $B{\overline{B'}}\to\mu{\overline{\mu}}$ amplitudes are neglected. Because of these approximations of the exact expressions, which are necessary in order to obtain a solution of the physical problem, the formulation of either a dispersion relation or a subtracted dispersion relation might lead to different results for the amplitudes although both should be mathematically equivalent. However, the ambiguity which dispersion relation to choose can be avoided by demanding that the analytic structure of the resulting $c_i^{(J)}(s,t)$ should agree as far as possible with the expressions for sharp $\sigma$ and $\rho$ exchange in the baryon-baryon interaction.
The transition amplitude ${\cal M}^\sigma$ for the exchange of a scalar $\sigma$ meson with mass $m_\sigma$ between two $J^P=1/2^+$ baryons $A$ and $B$ follows from the interaction Lagrangians ($X=A,B$) $${\cal L}_{XX\sigma}(x)= g_{XX\sigma} {\overline{\psi}}_{X}(x) \psi_{X}(x)
\phi_\sigma(x)$$ (See Appendix \[sec:AnhA\] for the hadronic field operators.) The result is $${\cal M}^\sigma(t)=g_{AA\sigma} g_{BB\sigma} { F^2_\sigma(t)\over t -
m_\sigma^2} {\cal O}_S \quad,
\label{eq:3_5_sig}$$ where a form factor $F_\sigma(t)$ has to be applied at each vertex since the exchanged $\sigma$ meson is far away from its mass-shell. This form factor is parametrized in the conventional monopole form $$F_\sigma(t)={\Lambda_\sigma^2 - m_\sigma^2 \over \Lambda_\sigma^2 -
t}\quad.
\label{eq:3_5_2}$$ with a cutoff mass $\Lambda_\sigma$ assumed to be uniform for both vertices.
The construction of the amplitude for $\rho$ exchange in the transition $A+B \to C+D$ (with $M_A=M_C\equiv M$ and $M_B=M_D\equiv M'$) starts from the interaction Lagrangians $$\begin{aligned}
{\cal L}_{XY\rho}(x)=&& g_{XY\rho} {\overline{\psi}}_{X}(x) \gamma_\mu
\psi_Y(x) \phi^\mu_\rho(x)
\nonumber \\
&+& {f_{XY\rho} \over 4 M_N} {\overline{\psi}}_{X}(x) \sigma_{\mu\nu}
\psi_Y(x) (\partial^\mu \phi^\nu_\rho(x) - \partial^\nu
\phi^\mu_\rho(x))
\nonumber \\
&& (+ \mbox{\rm h.c., if } X\neq Y)
\label{eq:BBvcoup2}\end{aligned}$$ with $(XY)=(AC),(BD)$. According to the conventional Feynman rules, using Eq. \[eq:3\_1\_3d\] and the generalized Gordon decomposition [@HolzDipl] for the spinors of two Dirac particles $X$ and $Y$, $${\overline{u}}_{X}(\vec p\,',\lambda')
\left[(M_X+M_{Y}) \gamma^\mu - (p'+p)^\mu
-i \sigma^{\mu\nu} (p'-p)_\nu \right]
u_Y(\vec p,\lambda)=0 \quad,$$ the transition amplitude ${\cal
M}^\rho$ comes out to be $$\begin{aligned}
{\cal M}^\rho(P,Q)= {-1 \over t - m_\rho^2}
& \Biggl[ &
{f_{AC\rho}f_{BD\rho}\over 4 M_N^2} (s-u)\;{\cal O}_S
\nonumber \\ &&
+ G_{AC\rho} G_{BD\rho} \;{\cal O}_V
\nonumber \\ &&
+ {G_{AC\rho} f_{BD\rho}-G_{BD\rho} f_{AC\rho} \over 2M_N}\; {\cal O}_6
\nonumber \\ &&
+ {G_{AC\rho} f_{BD\rho}+G_{BD\rho} f_{AC\rho} \over 2M_N}\; {\cal P}_2
\Biggr] \quad,
\label{eq:AnhG1}\end{aligned}$$ with $$G_{XY\rho} \equiv
g_{XY\rho} + {M_X+M_{Y}\over 2 M_N}
f_{XY\rho}$$ and the 4-momenta $P$ and $Q$ defined in Eq. \[eq:3\_1\_3a\]. ${\cal P}_2$ is one of the so-called perturbative covariants introduced in Ref. [@ALV] as $${\cal P}_2= -{{\rm 1 \>\! \llap{I}}}_4 \otimes \gamma_\mu P^\mu \,-\,
\gamma_\mu Q^\mu \otimes {{\rm 1 \>\! \llap{I}}}_4
\quad.$$ ${\cal P}_2$ can be expanded in terms of the ${\cal O}_i$ of Eq. \[eq:3\_1\_13\] since they form a complete set of kinematic covariants. For the given baryon masses one finds $${\cal P}_2 = {1\over 2(M+M')}
\left[
(u-s)({\cal O}_S + {\cal O}_P)
-4MM'{\cal O}_V
+ t {\cal O}_T
+2(M-M') {\cal O}_6
\right]\quad.$$ After replacing ${\cal P}_2$ in Eq. \[eq:AnhG1\] the final result for ${\cal M}^\rho$ reads $$\begin{aligned}
{\cal M}^\rho(P,Q)= {-F^2_\rho(t)\over t - m_\rho^2}
&
\Biggl\{&
{g_{AC\rho}f_{BD\rho}+ f_{AC\rho}g_{BD\rho}\over 4M_N M_{tot} } (u-s)
\; {\cal O}_S
\nonumber \\&&
\left[
g_{AC\rho}g_{BD\rho}+ g_{AC\rho}f_{BD\rho} {M'^2 \over M_N M_{tot} } +
f_{AC\rho}g_{BD\rho} {M^2 \over M_N M_{tot} }
\right] {\cal O}_V
\nonumber \\&&
\left(
g_{AC\rho}f_{BD\rho}{M'\over M_N M_{tot}} - f_{AC\rho}g_{BD\rho} {M_A
\over M_N M_{tot} }
\right) {\cal O}_6
\nonumber \\ &&
{G_{AC\rho}f_{BD\rho}+ f_{AC\rho}G_{BD\rho}\over 4M_N M_{tot} }
\left[(u-s)\, {\cal O}_P \;+ t \,{\cal O}_T \;\right]
\Biggr\}
\nonumber \\
\label{eq:3_5_rho}\end{aligned}$$ with $M_{tot}=M+M'$ and the form factor $F_\rho(t)$ parametrized according to Eq. \[eq:3\_5\_2\].
By comparison of the discontinuities in the $\sigma$ and $\rho$ channel in Eqs. \[eq:3\_4\_5\], \[eq:3\_4\_6\] with the transition amplitudes for sharp $\sigma$ and $\rho$ exchange it follows that $c_S^\sigma,c_V^\rho$ and $c_6^\rho$ obey an unsubtracted dispersion relation, $$c^{(J)}_i (s,t) = {1\over\pi} \int_{4m_\pi^2}^{t'_{max}} { {\sl
Im}\left[ c^{(J)}_i (s,t')\right] \over t'-t} dt' \quad.
\label{eq:3_4_1}$$ The tensor component of sharp $\rho$ exchange is proportional to $t$ (cf. \[eq:3\_5\_rho\]). In order to generate this factor $t$ also for the tensor component of correlated $\pi\pi$ and $K{\overline{K}}$ exchange a subtracted dispersion relation (subtraction point $t_0$ and subtraction constant $c_T^\rho(s,t_0)=0$) is assumed for the invariant amplitude $c_T^\rho(s,t)$: $$c^\rho_T (s,t) = {t\over\pi} \int_{4m_\pi^2}^{t'_{max}} { {\sl
Im}\left[ c^\rho_T (s,t')\right]/t' \over t'-t} dt' \quad.
\label{eq:3_4_2a}$$ Similarly, the $u-s$ dependence of the (pseudo-)scalar component of sharp $\rho$ exchange (cf. Eq. \[eq:3\_5\_rho\]) can be reproduced by assuming a subtracted dispersion relation for $ c^\rho_S (s,t)$ and $c^\rho_P (s,t)$ with $t_0(s)=\Sigma-2s$ and $c^\rho_{S,P} (s,t_0(s)) = 0$: $$c^\rho_{S,P} (s,t) = {u-s\over\pi} \int_{4m_\pi^2}^{t'_{max}} { {\sl
Im}\left[ c^\rho_{S,P} (s,t')\right]/(u'-s) \over t'-t} dt' \quad,
\label{eq:3_4_2b}$$ where according to Eq. \[eq:3\_1\_0\] $s+t+u=s+t'+u'=\Sigma$.
Baryon-baryon interaction arising from correlated $\pi\pi$ and $K{\overline{K}}$ exchange
-----------------------------------------------------------------------------------------
The invariant amplitudes constructed in the preceding Section using dispersion theory still contain the uncorrelated contributions of $\pi\pi$ and $K{\overline{K}}$ exchange. Investigating the problem of baryon-baryon scattering requires the knowledge of the on-shell scattering amplitude $T$, which is usually obtained as a solution of a scattering equation $T=V+VGT$ that iterates the interaction kernel $V$. In general, $V$ contains besides other terms one-pion and one-kaon exchange contributions as well as the contribution $V_{2\pi}$ from two-pion and two-kaon exchange. But iterating $\pi$ and $K$ exchange in the second order term $VGV$ also generates two-pion and two-kaon exchange contributions to the scattering amplitude. In order to avoid double counting these ‘iterative’ contributions therefore have to be left out from the dispersiontheoretically calculated $V_{2\pi}$. As stated above we even go beyond this and subtract all uncorrelated contributions from $V_{2\pi}$. By this the dispersiontheoretic calculations can be restricted to the $\sigma$ and $\rho$ channel (since only there significant correlations occur in the kinematic region considered), whereas the uncorrelated contributions are evaluated in the $s$-channel and therefore contain all $t$-channel partial waves.
In order to eliminate the uncorrelated contributions from $V_{2\pi}$ we determine the discontinuities ${\sl Im} [c^{(J)}_{i,Born}(s,t)]$ generated from the $B{\overline{B'}}\to\mu{\overline{\mu}}$ Born amplitudes $V^J$ (i.e., no $\pi\pi-K{\overline{K}}$ correlations included) using as before the unitarity relation \[eq:3\_4\_4\] (with $T^J$ replaced by $V^J$) and subtract them finally from the full discontinuities ${\sl Im}[c^{(J)}_{i}(s,t)]$. (The contributions of the $\rho$-pole diagram to the $B{\overline{B'}}\to\mu{\overline{\mu}}$ Born amplitudes must not be subtracted since the corresponding $s$-channel processes are not included explicitly in $V$.) Hence, for the invariant amplitudes of [*correlated*]{} $\pi\pi$ and $K{\overline{K}}$ exchange, $\tilde c^{(J)}_{i}(s,t)$, the (unsubtracted) dispersion relation \[eq:3\_4\_1\] has to be modified to $$\tilde c^{(J)}_{i} (s,t) = {1\over\pi} \int_{4m_\pi^2}^{t'_{max}}
{ \rho^{(J)}_i (s,t') \over t'-t} dt'\quad,\qquad
(J)=\sigma,\rho\quad,
\label{eq:3_5_1}$$ where the spectral function $\rho^{(J)}_i $ is given by $$\rho^{(J)}_i (s,t') \equiv
{\sl Im}\left[ c^{(J)}_i (s,t')\right] - {\sl Im}\left[
c^{(J)}_{i,Born} (s,t')\right]\quad.
\label{eq:3_5_1a}$$ Corresponding expressions hold for the subtracted dispersion relations \[eq:3\_4\_2a\] and \[eq:3\_4\_2b\].
Now the baryon-baryon helicity amplitudes arising from correlated $\pi\pi$ and $K{\overline{K}}$ exchange can be evaluated according to Eqs. \[eq:3\_1\_11a\], \[eq:3\_1\_12\] $$\begin{aligned}
\lefteqn{
{\langle \, C D , \vec q\,\lambda_C \lambda_D \, \vert \>}
V^{(J)}_{2\pi} {\> \vert \, A B, \vec p \,\lambda_A\lambda_B \, \rangle} =}
\nonumber \\
&&
{\overline{u}}_{C}(\vec q,\lambda_C)\; {\overline{u}}_{D}(-\vec q,\lambda_D)\;
{\cal V}^{(J)}_{2\pi} (P,Q)\; u_{A}(\vec p,\lambda_A) \; u_{B}(-\vec
p,\lambda_B)
\nonumber\end{aligned}$$ with $${\cal V}^{(J)}_{2\pi} (P,Q) \equiv \sum_i \tilde c_i^{(J)}(s,t)
{\cal O}_i
(P,Q)
\qquad \left( (J)=\sigma,\rho \right)\quad.
\label{eq:3_5_0}$$ The partial wave decomposition of these matrix elements then proceeds as usual (see e.g. Ref. [@Holz]).
Of course, when iterating the baryon-baryon interaction kernel in a scattering equation $V_{2\pi,corr}$ has to be known off-shell. However, dispersion theory applies only to on-shell amplitudes and does not provide any information on the off-shell behavior of the amplitudes. Therefore an arbitrary prescription for the off-shell extrapolation of $V_{2\pi,corr}$ has to be defined [@Kim], which is certainly a drawback of the dispersiontheoretical derivation of this potential. Nevertheless the characteristic features of correlated $\pi\pi$ and $K{\overline{K}}$ exchange like the strength of $V_{2\pi,corr}$ in the various baryon-baryon channels can already be discussed by means of the unique on-shell amplitudes. Therefore we postpone the discussion of how to extrapolate $V_{2\pi,corr}$ off-shell to a subsequent work.
The dispersiontheoretic amplitudes for correlated $\pi\pi$ and $K{\overline{K}}$ exchange (Eqs. \[eq:3\_4\_5\], \[eq:3\_4\_6\]) have been constructed in such a way that their operator structure agrees as far as possible with sharp $\sigma$ and $\rho$ exchange \[eq:3\_5\_sig\], \[eq:3\_5\_rho\]. Therefore our results for the correlated exchange can be parametrized in terms of $\sigma$ and $\rho$ exchange; i.e., the products of coupling constants for $\sigma$ and $\rho$ exchange are replaced by effective coupling strengths $G^{(J)}(s,t)$, which contain the full $s$- and $t$-dependence of the dispersiontheoretic results. In the $\sigma$ channel this gives for the elastic baryon-baryon process $A+B\to A+B$ $$g_{AA\sigma}g_{BB\sigma} \quad\longrightarrow \quad G_{AB\to
AB}^\sigma (t)= {(t-m_\sigma^2)\over F^2_\sigma(t)} \tilde
c^\sigma_S(t) = {(t-m_\sigma^2)\over\pi F^2_\sigma(t)}
\int_{4m_\pi^2}^{t'_{max}} { \rho^\sigma_S(t') \over t'-t} dt'
\label{eq:3_effccsig}$$ Note that sharp $\sigma$ exchange (Eq. \[eq:3\_5\_sig\]) would correspond to a spectral function $$\rho^{\sigma}_S (s,t') = - g_{AA\sigma} g_{BB\sigma}
\delta( t' - m_\sigma^2)$$ (except for form factors). This suggests to interpret the spectral function as a function that denotes the strength of an exchange process depending on the invariant mass of the exchanged system (here: $\pi\pi$, $K{\overline{K}}$).
By comparing the coefficients of the kinematic covariants $ {\cal O}_i$ in Eqs. \[eq:3\_5\_rho\] and \[eq:3\_5\_0\] we obtain in the $\rho$-channel: $$\begin{aligned}
g_{AB\rho}g_{BD\rho} &\!\to \!& ^{VV}\!G_{AB\to CD}^\rho (t)=
{(t-m_\rho^2)\over F^2_\rho(t)} \left[ 4MM'{\tilde c^\rho_S(s,t)\over
u-s} - \tilde c^\rho_V(t) +(M'-M) \tilde c^\rho_6(t)\right],
\nonumber\\
g_{AB\rho}f_{BD\rho} &\!\to\!& ^{VT}\!G_{AB\to CD}^\rho (t)=
-{(t-m_\rho^2)\over F^2_\rho(t)} M_N
\left[ 4M {\tilde c^\rho_S(s,t)\over u-s}
+ \tilde c^\rho_6(t)\right],
\nonumber\\
f_{AB\rho}g_{BD\rho} &\!\to\!& ^{TV}\!G_{AB\to CD}^\rho (t)=
-{(t-m_\rho^2)\over F^2_\rho(t)} M_N
\left[ 4M' {\tilde c^\rho_S(s,t)\over u-s}
- \tilde c^\rho_6(t)\right],
\nonumber\\
f_{AB\rho}f_{BD\rho} &\!\to\! & ^{TT}\!G_{AB\to CD}^\rho (t)=
{(t-m_\rho^2)\over F^2_\rho(t)} 4M_N^2 {\tilde c^\rho_S(s,t)- \tilde
c^\rho_P(s,t)\over u-s}.
\label{eq:3_effccrho}\end{aligned}$$
Obviously the effective coupling strengths do not depend on $s$ but only on $t$. This is only possible by choosing the subtracted dispersion relation \[eq:3\_4\_2b\] for $\tilde c^\rho_{S,P}(s,t)$, since the integrand of the dispersion integral becomes independent of $s$ (${\sl Im}[c^\rho_{S,P}(s,t')]\propto \cos\vartheta_t(s,t')\propto
u'-s$). Therefore $\tilde c^\rho_{S,P}(s,t)$ depends on $s$ only by the factor $(u-s)$, which cancels out exactly when calculating the effective coupling strengths. It should be emphasized that the parametrization of $V_{2\pi,corr}$ discussed here does (so far) not contain any approximations.
### Isospin-Crossing {#sec:kap3_6}
Up to now, in favor of a clear representation of the dispersiontheoretical calculation, we have suppressed isospin degrees of freedom. The isospin structure of the $B{\overline{B'}}\to\mu{\overline{\mu}}$ amplitudes will be discussed in the next Sections when the microscopic model for these amplitudes is presented. However, when ‘crossing’ is applied to the baryon-baryon ($s$-channel) and baryon-antibaryon ($t$-channel) amplitudes it has to be kept in mind that the total isospin in the various channels is constructed from different combinations of particle isospins. The total isospin $I_s$ of the $s$-channel process $A+B\to C+D$ is composed out of $[I_A\otimes I_B]_{I_s}$ or $[I_C\otimes I_D]_{I_s}$ and that of the $t$-channel process $A+{\overline{C}} \to D+{\overline{B}}$ out of $[I_A\otimes I_{\bar C}]_{I_t}$ or $[I_{D}\otimes I_{\bar B}]_{I_t}$.
Consequently, besides the analytic continuation of the invariant amplitudes in $s$ and $t$ the recoupling of the particle isospins has to be taken into account when crossing amplitudes. Therefore, the isospin amplitudes $T_s^{AB\to CD}(I_s)$ and $T_t^{A{\overline{C}} \to D{\overline{B}}}(I_t)$ of the $s$- and $t$-channel processes being independent of the isospin projections $m_s$ and $m_t$ are linearly related: $$T_s^{AB\to CD}(I_s)= \sum_{I_t}
X_{AB,CD}(I_s,I_t)
T_t^{A{\overline{C}} \to D{\overline{B}}}(I_t)\quad,$$ where $X_{AB,CD}(I_s,I_t)$ is the so-called isospin-crossing matrix. Note that our isospin-crossing matrix differs from the $\tilde X_{AB,CD}(I_s,I_t)$ introduced in Refs. [@MartSpear; @Neville] since their $t$-channel process (${\overline{D}}+B\to C+{\overline{A}}$) differs from the one in Sect. \[sec:kap3\_1\]: $$X_{AB,CD}(I_s,I_t) = (-1)^{I_B+I_D+I_t-2I_s}\tilde X_{BA,DC}(I_s,I_t)
\quad.$$ As shown in Ref. [@MartSpear] the isospin-crossing matrix $\tilde X_{AB,CD}(I_s,I_t)$ can be expressed in terms of Clebsch-Gordan coefficients $$\begin{aligned}
\lefteqn{
\tilde X_{AB,CD}(I_s,I_t)= }
\nonumber \\ &&
\eta_A \eta_D \!\! \sum_{m_A,m_B, \atop m_C,m_D,m_t}\!
(-1)^{I_A+I_D+m_A+m_D}
{\langle \, I_AI_Bm_Am_B \, \vert \>}{I_sm_s \,\rangle}
{\langle \, I_CI_Dm_Cm_D \, \vert \>}{I_sm_s \,\rangle}
\nonumber \\ && \hspace{5.5cm}
{\langle \, I_CI_Am_C(-m_A) \, \vert \>}{I_tm_t \,\rangle}
{\langle \, I_DI_B(-m_D)m_B \, \vert \>}{I_tm_t \,\rangle}
\nonumber \\
\label{eq:3_6_1}\end{aligned}$$ with $m_s$ ($|m_s|\leq I_s$) being arbitrary. For a particle $A$ with isospin $I_A$ and isospin projection $m_A$ the particle state ${\> \vert \, A \, \rangle}$ and the isospin state ${\> \vert \, Im \, \rangle}_A$ might differ in sign. The isospin state of the antiparticle ${\overline{A}}$ is generated by applying the $G$-parity operator ${\cal G}$ to ${\> \vert \, Im \, \rangle}_{A}$ [@MartSpear]: $${\> \vert \, Im \, \rangle}_{\bar A}=\eta_A {\cal G} {\> \vert \, Im \, \rangle}_{A}\quad.
\label{eq:3_6_1b}$$ With the phase convention used here for the $SU(3)$ field operators (e.g. when calculating the isospin factors of the $B{\overline{B'}}\to\mu{\overline{\mu}}$ amplitudes in the next Section) the phase $\eta_A$, which is independent of $m_A$, comes out to be [@Neville] $$\eta_A= (-1)^{I_A-Y_A/2}\quad,
\label{eq:SU3phase}$$ where $Y$ denotes the hypercharge of particle $A$. Note that this phase convention differs from the one used in Ref. [@MartSpear]. For the baryon-baryon processes considered in this work the isospin-crossing matrices are tabulated in Tab. \[tab:3\_1\].
By the partial wave decomposition the amplitude of correlated $\pi\pi$ and $K{\overline{K}}$ exchange is separated into the contributions of the $\sigma$- ($I_t=0$) and $\rho$-channel ($I_t=1$). Suppressing the spin-momentum dependence the isospin amplitude can be written as $$T_s(I_s)= X(I_s,0) T_t^\sigma + X(I_s,1) T_t^\rho \quad.$$ The column $X(I_s,0)$ ($X(I_s,1)$) of the isospin-crossing matrix agrees except for a constant factor $F_\sigma$ ($F_\rho$) with the isospin factors for $t$-channel exchange of a $\sigma$ ($\rho$) meson in the corresponding $s$-channel process. Conventionally these constant factors $F_\sigma$ and $F_\rho$, which are also tabulated in Tab. \[tab:3\_1\], are extracted from the isospin-crossing matrix and put into the spectral functions \[eq:3\_5\_1a\], so that the isospin factors for the $s$-channel potential $V_{2\pi}$ agree with the isospin factors of $\sigma$ and $\rho$ exchange.
A microscopic model for the $B{\overline{B'}} \to \pi\pi,K{\overline{K}}$ transition amplitudes {#sec:kap4}
-----------------------------------------------------------------------------------------------
In the preceding Sections we have outlined the dispersiontheoretic calculation of correlated $\pi\pi$ and $K{\overline{K}}$ exchange in the baryon-baryon interaction starting from amplitudes for the transition of a baryon-antibaryon ($B{\overline{B'}}$) state to two pions ($\pi\pi$) or a kaon and a antikaon ($K{\overline{K}}$). These amplitudes have to be known in the so-called pseudophysical region, i.e.for energies below the $B{\overline{B'}}$ threshold. However, in case of the process $N{\overline{N}}\to \pi\pi$, these amplitudes can be derived from empirical data by analytic continuation of $\pi N$ and $\pi\pi$ scattering amplitudes, which are extracted from scattering data, into the pseudophysical region [@Hoehler2; @Hoehler; @Nielsen; @Nielsen2]. Corresponding analyses for the transitions $Y{\overline{Y'}}\to \pi\pi,
K{\overline{K}}$ are out of sight since the required empirical information (e.g. $\pi \Lambda$ scattering data) does not exist. An evaluation of correlated $\pi\pi$ and $K{\overline{K}}$ exchange in the $YN$ or $YY$ interaction therefore necessitates the construction of a microscopic model for the $B{\overline{B'}} \to \pi\pi,K{\overline{K}}$ amplitudes. This model can be tested against the quasiempirical information for the $N{\overline{N}}\to
\pi\pi$ amplitudes and then has to be extrapolated to the other channels of interest. In addition, only the use of a microscopic model for the $B{\overline{B'}} \to \pi\pi,K{\overline{K}}$ amplitudes allows a consistent treatment of medium modifications of the baryon-baryon interaction.
The microscopic model presented in the following is a generalization of the hadron-exchange model for the $N{\overline{N}}\to \pi\pi$ transition amplitudes of Ref. [@Pearce_piN], where it was applied to the analysis of correlated two-pion exchange in the $\pi N$ interaction. A main feature of the model presented here is the completely consistent treatment of its two components, namely the $B{\overline{B'}}\to\pi\pi, K{\overline{K}}$ Born amplitudes and the $\pi\pi-K{\overline{K}}$ correlations. Both components are derived in field theory from an ansatz for the hadronic Lagrangians.
The amplitudes for the processes $B{\overline{B'}} \to \pi\pi,K{\overline{K}}$ are obtained from a scattering equation which can be written in operator form as $$\mbox{\bf T} = \mbox{\bf V} + \mbox{\bf T}\mbox{\bf G}\mbox{\bf V}\quad,
\label{eq:skizz_1}$$ where the scattering amplitude $\mbox{\bf T}$, the Born amplitude $\mbox{\bf V}$ and the Greens operator $\mbox{\bf G}$ are operators in channel space of $B{\overline{B'}}$, $\pi\pi$ and $K{\overline{K}}$.
For large $t'$, contributions of the spectral functions of correlated $\pi\pi$ and $K{\overline{K}}$ exchange to the dispersion integrals \[eq:3\_4\_1\]–\[eq:3\_4\_2b\] values are suppressed by the denominator $1/(t'-t)$ because in the physical $s$-channel $t{\leq}0$. Since the unitarity cuts of the $B{\overline{B'}}$ states start far above the branch points of the $\pi\pi$ and $K{\overline{K}}$ cuts the contribution of the $B{\overline{B'}}$ Greens function $G_{B{\overline{B'}}}$ to the $B{\overline{B'}}\to\pi\pi, K{\overline{K}}$ scattering amplitudes in Eq. \[eq:skizz\_1\] and to the spectral functions can be neglected. For the same reason the coupling to other mesonic channels like the $\rho\rho$ channel can be renounced. Of course, by these approximations the range of validity of the microscopic model for the $B{\overline{B'}}\to\mu{\overline{\mu}}$ ($\mu{\overline{\mu}}=\pi\pi,K{\overline{K}}$) amplitudes is limited to low $t'$ values. Therefore instead of integrating along the whole $\pi\pi$ unitarity cut up to infinity the upper bound of the dispersion integrals, $t'_{max}$, is set to a value somewhere below the $\rho\rho$ threshold at $t'\approx
120m_\pi^2$ so that convergence of the integrals is achieved.
Taking into account only $\pi\pi$ and $K{\overline{K}}$ intermediate states, i.e. neglecting $G_{B{\overline{B'}}}$, we obtain for the two components of Eq. \[eq:skizz\_1\] which give the $B{\overline{B'}}\to \mu{\overline{\mu}}$ amplitudes:
$$\left( \begin{array}{c}
T_{B{\overline{B'}}\to\pi\pi} \\ T_{B{\overline{B'}}\to K{\overline{K}}}
\end{array} \right)
=
\left( \begin{array}{c}
V_{B{\overline{B'}}\to\pi\pi} \\ V_{B{\overline{B'}}\to K{\overline{K}}}
\end{array} \right)
+
\left( \begin{array}{cc}
T_{\pi\pi\to\pi\pi} & T_{K{\overline{K}}\to\pi\pi} \\ T_{\pi\pi\to
K{\overline{K}}} & T_{K{\overline{K}}\to K{\overline{K}}}
\end{array} \right)
\left( \begin{array}{c}
G_{\pi\pi}\; V_{B{\overline{B'}}\to\pi\pi} \\ G_{K{\overline{K}}}\; V_{B{\overline{B'}}\to
K{\overline{K}}}
\end{array} \right)\, .
\label{eq:skizz_2}$$
In Fig. \[fig:4\_0a\] this scattering equation for the $B{\overline{B'}} \to\pi\pi,K{\overline{K}}$ amplitudes is represented symbolically.
The correlations $T_{\pi\pi/K{\overline{K}}\to\pi\pi/K{\overline{K}}}$ are generated with a realistic model [@Pearce_piN] of the $\pi\pi-K
{\overline{K}}$ interaction. This meson-exchange model is a modified version of the so-called Jülich $\pi\pi$ model [@Lohse]. The Born amplitudes of this model for the elastic $\pi\pi$ and $K{\overline{K}}$ channels as well as for the transition $\pi\pi\to K{\overline{K}}$ are shown in Fig. \[fig:4\_5\]. Besides the $t$-channel (and in case of $\pi\pi\to\pi\pi$ also $u$-channel) exchanges of the vector mesons $\rho,\omega,\phi,K^*$ the $s$-channel exchanges (pole graphs) of the $\rho$, the scalar-isoscalar $\epsilon$ and the isoscalar tensor meson $f_2$ are taken into account. The potentials derived from Fig. \[fig:4\_5\] are iterated in a coupled-channel calculation according to the prescription of Blankenbecler and Sugar [@BbS]. The free parameters of the $\pi\pi-K{\overline{K}}$ model of Ref. [@Pearce_piN] were adjusted to the empirical $\pi\pi$ phase shifts and inelasticities. A good agreement with the empirical data was achieved(cf. Fig. \[fig:4\_6\]).
The scattering equation \[eq:skizz\_2\] is likewise solved using the ansatz [@BbS] of Blankenbecler and Sugar (BbS) to reduce the 4-dimensional Bethe-Salpeter equation to a 3-dimensional equation which simplifies the calculation while retaining unitarity.
The total conserved 4-momentum $P$ and the relative 4-momentum $k'$ of the intermediate $\mu'{\overline{\mu'}}=\pi\pi, K{\overline{K}}$ state are expressed in terms of the particle 4-momenta $k'(1)$ and $k'(2)$ by $$P=k'(1) + k'(2) \quad, \qquad k' = (k'(1) - k'(2))/2
\quad.
\label{eq:bbsc}$$ Corresponding relations hold for the relative 4-momenta $q$ of the initial $B{\overline{B'}}$ and $k$ of the final $\mu{\overline{\mu}}=\pi\pi, K{\overline{K}}$ state. In the center-of-mass system we have $P=(\sqrt{t'},\vec 0)$ with $\sqrt{t'}=E_B(q)+E_{B'}(q)$, $\vec k' =\vec k'(1) = - \vec k'(2)$ and, if the particles in the initial and final state are on their mass-shell, $$q_0=(E_B(q) - E_{B'}(q))/2
\quad,\qquad
k_0=k'_0=0
\label{eq:bbsm}$$ with $ E_B(q):=\sqrt{ \vec q\,^2 + M_B^2}$.
Now, according to the prescription of Blankenbecler and Sugar, the relativistic two-particle Greens function, $G_{\mu'{\overline{\mu}}'}$, is replaced by a 3-dimensional Greens function $g_{\mu'{\overline{\mu}}'}\propto \delta(k'_0)$, which respects unitarity. Due to the $\delta$ function $\delta(k'_0)$, which sets the two intermediate particles (being of equal mass) equally off-mass-shell ($k'_0(1)=k'_0(2)$), the integration over $k'_0$ can be carried out immediately and the so-called Blankenbecler-Sugar equation for the helicity amplitudes is obtained: $$\begin{aligned}
\lefteqn{
{\langle \, \mu\bar\mu, \vec k \, \vert \>} T(t) {\> \vert \, B{\overline{B'}},\vec q,\lambda_1
\lambda_2 \, \rangle}
=
{\langle \, \mu\bar\mu, \vec k \, \vert \>} V(t) {\> \vert \, B{\overline{B'}},\vec q,\lambda_1
\lambda_2 \, \rangle}}
\nonumber \\
&& +{1 \over (2\pi)^3 } \sum_{\mu' \bar{\mu}'=\pi\pi,K\bar{K}}
N_{\mu' \bar{\mu}'}
\int d^3k' {T_{ \mu' \bar{\mu}' \to \mu\bar\mu }(\vec k,\vec k';t)
{\langle \, \mu'\bar\mu', \vec k' \, \vert \>} V (t) {\> \vert \, B{\overline{B'}},\vec q,\lambda_1
\lambda_2 \, \rangle}
\over \omega_{k'} (t- 4\omega^2_{k'} + i \epsilon) }\quad,
\nonumber \\
\label{eq:bbsma}\end{aligned}$$ where $N_{\mu' \bar{\mu}'}$ is defined as in Eq. \[eq:3\_3\_1a\] and $T_{ \mu'\bar{\mu}'\to\mu\bar\mu}\equiv{\cal M}_{\mu'\bar{\mu}'\to
\mu\bar\mu}$.
Because of the rotational invariance of the underlying interactions the BbS equation can be decomposed into partial waves [@GW]: $$\begin{aligned}
\lefteqn{
{\langle \, \mu\bar\mu \, \vert \>} T^J(k,q;t) {\> \vert \, B{\overline{B'}},\lambda_1 \lambda_2 \, \rangle}
=
{\langle \, \mu\bar\mu \, \vert \>} V^J(k,q;t) {\> \vert \, B{\overline{B'}},\lambda_1 \lambda_2 \, \rangle}}
\nonumber \\
&& + {1 \over (2\pi)^3 }
\sum_{\mu' \bar{\mu}'} N_{\mu' \bar{\mu}'} \int {k'}^2 dk'
{T^J_{ \mu' \bar{\mu}' \to \mu\bar\mu }(k,k';t)
{\langle \, \mu'\bar\mu' \, \vert \>} V^J(k',q,t) {\> \vert \, B{\overline{B'}},\lambda_1 \lambda_2 \, \rangle}
\over \omega_{k'} (t- 4\omega^2_{k'} + i \epsilon)}.
\nonumber \\
\label{eq:bbsn}\end{aligned}$$ Here, $q,k,k'$ denote the modulus of the corresponding 3-momenta.
The extrapolation of the model for the $N{\overline{N}} \to \pi\pi$ amplitudes to the other particle channels is made under the assumption that the hadronic interactions are, except for the particle masses, $SU(3)_{flavor}$ symmetric. That means that the coupling constants at the various hadronic vertices are related to each other by $SU(3)$ relations. In this way it comes out that all free parameters of the model for the $B{\overline{B'}} \to \mu{\overline{\mu}}$ amplitudes can be fixed by adjusting the $N{\overline{N}} \to \pi\pi$ amplitudes to the quasiempirical data [@Hoehler2; @Dumbrajs] (see Sect. \[sec:5\_1\]).
### $B{\overline{B'}} \to \pi\pi,K{\overline{K}}$ Born amplitudes {#sec:bornsp}
In our model, the $B{\overline{B'}} \to \mu{\overline{\mu}}$ transition potentials are build up from a $\rho$-pole diagram and all diagrams in which a baryon out of the $J^P=1/2^+$ octet or the $J^P=3/2^+$ decuplet is exchanged. Of course only those diagrams are considered which respect the conservation of isospin and strangeness. As an example Fig. \[fig:4\_0b\] shows the Born amplitudes for the transition $\Sigma{\overline{\Sigma}}\to\pi\pi,K{\overline{K}}$. The particles occuring in this model are listed in Tab. \[tab:4\_0\] together with their masses and their basic quantum numbers.
In order to start from a maximum $SU(3)$ symmetry our model differs slightly from the one presented in Ref. [@Pearce_piN] for the $N{\overline{N}}\to\pi\pi$ amplitudes. Here, we include the exchange of the $Y^* \equiv\Sigma(1385)$ in the $N{\overline{N}} \to K{\overline{K}}$ transition potential, which was neglected in Ref. [@Pearce_piN]. In addition, the form factors at the hadronic vertices are chosen identical within a given $SU(3)$ multiplet.
Starting point for the derivation of the various Born amplitudes are the following interaction Lagrangians, which are characterized by the $J^P$ quantum numbers of the hadrons involved. For simplicity, we suppress here the isospin or $SU(3)$ dependence of the Lagrangians. $$\begin{array} {ll}
{{1\over 2}}^+ \otimes {{1\over 2}}^+ \otimes 0^-: &
{\cal L}_{B'Bp}(x)= {f_{B'Bp} \over m_{\pi^+}}
{\overline{\psi}}_{B'}(x) \gamma_5\gamma^\mu \psi_B(x) \partial_\mu
\phi_p(x)
\quad (+ \;\mbox{\rm h.c., if } B'\neq B)\,, \\ \\
{{3\over 2}}^+ \otimes {{1\over 2}}^+ \otimes 0^-:&
{\cal L}_{DBp}(x)= {f_{DBp} \over m_{\pi^+}} {\overline{\psi}}_B (x)
(g_{\mu\nu} + x_\Delta \gamma_\mu \gamma_\nu) \psi^\mu_D(x)
\partial^\nu \phi_p(x)
\quad + \mbox{\rm h.c.}\,,\\ \\
{{1\over 2}}^+ \otimes {{1\over 2}}^+ \otimes 1^-:&
{\cal L}_{B'B\rho}(x)=
\begin{array}[t]{l}
g_{B'B\rho} {\overline{\psi}}_{B'}(x) \gamma_\mu
\psi_B(x) \phi^\mu_\rho(x)\\
+ {f_{B'B\rho} \over 4 M_N} {\overline{\psi}}_{B'}(x) \sigma_{\mu\nu}
\psi_B(x) (\partial^\mu \phi^\nu_\rho(x) - \partial^\nu
\phi^\mu_\rho(x))
\\ \\
(+ \mbox{\rm h.c., if } B'\neq B) \,,
\end{array} \\ \\
0^- \otimes 0^- \otimes 1^-: &
{\cal L}_{p'p\rho}(x)= g_{p'p\rho} \phi_{p'} (x) \partial_\mu \phi_p
\phi^\mu_\rho(x) \quad (+ \mbox{\rm h.c., if } p'\neq p) \,.
\\
\end{array} \nonumber \\
\label{eq:Lag}$$
For the conventions used for the field operators and the Dirac $\gamma$-matrices, see Appendix \[sec:AnhA\]. The Lagrangian ${\cal L}_{DBp}$ includes an off-shell part, which is proportional to $x_\Delta$. As the parameter $A$, which occurs in the free Lagrangian [@Read] and then in the non-pole part of the propagator of a spin-3/2 particle (cf. Eq. \[eq:AnhARS\]), the parameter $x_\Delta$ characterizing the strength of the off-shell part of the $DBp$ coupling is not determined from first principles. However, it is known [@NEK] that fieldtheoretic amplitudes derived with this most general ansatz depend only on a certain combination of $A$ and $x_\Delta$, namely $${1+4x_\Delta \over 2A+1} =: 1+4Z\quad.
\label{eq:bornspa}$$ It follows that different pairs of ($x_\Delta,A$) values, which give the same value of $Z$, describe the same interaction theory. Therefore, without restricting the general validity of our results, we can set $A=-1$ (i.e. omitting the non-pole part of the spin-3/2 propagator) and select the interaction theory (characterized by $Z$) through $x_\Delta$, which is finally adjusted to the quasiempirical data for $N{\overline{N}}\to\pi\pi$.
In order to account for the extended structure of hadrons the vertex functions resulting from the Lagrangians \[eq:Lag\] are modified by phenomenological form factors. For the baryon-exchange processes these form factors are parametrized in the usual multipole form $$F_{X}(p^2) = \left( {n_{X} \Lambda^2_{X} - M_X^2 \over n_{X}
\Lambda^2_{X} - p^2 }
\right)^{n_{X}}\quad,
\label{eq:BexFF}$$ where $p$ denotes the 4-momentum and $M_X$ the mass of the exchanged baryon $X$. The two parameters, the so-called cutoff mass $\Lambda_X$ and the power $n_X$, are chosen uniquely for all $BB'p$ ($\Lambda_8$, $n_8$) and for all $BDp$ ($\Lambda_{10}$, $n_{10}$) vertices in order to keep the number of parameters of our model as low as possible. The dependence of the form factor \[eq:BexFF\] on the power $n_X$ is quite weak [@Pearce_piN]. We choose $n_8=1$ and $n_{10}=2$. Finally, $\Lambda_8$ and $\Lambda_{10}$ are adjusted to the quasiempirical data for the $N{\overline{N}}\to\pi\pi$ amplitudes in the pseudophysical region.
For the $\rho$-pole diagram we parametrize the form factor at the $\mu\mu\rho$ vertex in the same way as is done in the model of the $\pi\pi$-$K{\overline{K}}$ interaction in Ref. [@Pearce_piN]: $$\begin{aligned}
F_{\mu\mu\rho}(\vec k\,^2) &=& \left( {n_{\mu\mu\rho}
\Lambda^2_{\mu\mu\rho} + (m^{(0)}_\rho)^2 \over n_{\mu\mu\rho}
\Lambda^2_{\mu\mu\rho} + 4\omega_\mu^2(\vec k\,^2) }
\right)^{n_{\mu\mu\rho}}
\label{eq:rhoFF}\end{aligned}$$ where $\vec k$ is the relative 3-momentum of the two pseudoscalar mesons and $k_{\mu\bar\mu}^2 (t) = t/4 - m_\mu^2$ is the squared on-shell momentum of the $\mu{\overline{\mu}}$ state.
For the dispersiontheoretic calculation of correlated $\pi\pi$ and $K{\overline{K}}$ exchange the $B{\overline{B'}}\to \mu{\overline{\mu}}$ Born amplitudes are evaluated only for $B{\overline{B'}}$ states being on their mass-shell. Therefore, there is no need for a form factor at the $BB'\rho$ vertex as to assure convergence of the scattering equation \[eq:bbsn\]. Therefore, we disregard this form factor, again in order to keep the parameters of the model as low as possible.
Now, taking into account that the $B{\overline{B'}}$ state is on mass-shell and that due to the Blankenbecler-Sugar condition (Eq. \[eq:bbsm\]) the energy component of the relative momentum of the $\mu{\overline{\mu}}$ state always vanishes ($k_0=0$) the 4-momenta at the external legs of the $B{\overline{B'}}\to\mu{\overline{\mu}}$ Born diagrams read in the center-of-mass system: $$q_B=\left(
\begin{array}{c}
E_B(q) \\ \vec q
\end{array}
\right)
,\quad\!\!\!\!
q_{\bar B'}=\left(
\begin{array}{c}
E_{B'}(q) \\ -\vec q
\end{array}
\right)
,\quad k(\mu)=\left(
\begin{array}{c}
\sqrt{t'} \\ \vec k
\end{array}
\right)
,\quad\!\!\!\! k({\overline{\mu}})=\left(
\begin{array}{c}
\sqrt{t'} \\ -\vec k
\end{array}
\label{eq:4born30}
\right).$$ According to the usual Feynman rules [@BjDr] (for the various propagators, see also Appendix \[sec:AnhA\]) we obtain for the spin-momentum parts of the $B{\overline{B'}}\to \mu{\overline{\mu}}$ Born amplitudes ${\cal V}_{ B{\overline{B'}} \to \mu {\overline{\mu}}}(\vec k,\vec q;t)$:
- Exchange of a baryon $X$ with $j^P={{1\over 2}}^+$ and momentum $p=q-k$ $${\cal V}^X_{ B{\overline{B'}} \to \mu {\overline{\mu}}}(\vec k,\vec q;t') =
\left(
-{f_{XB'\mu}\over m_{\pi^+}} \gamma_5 \gamma_\lambda k^\lambda({\overline{\mu}})
\right)
{-(\not\!p + M_X) \over p^2 - M_X^2 }
\left(
-{f_{XB\mu}\over m_{\pi^+}} \gamma_5 \gamma_\nu k^\nu(\mu)
\right)\quad,
\label{eq:4born32}$$
- Exchange of a baryon $X$ with $J^P={{3\over 2}}^+$ and momentum $p=q-k$ $$\begin{aligned}
\lefteqn{
{\cal V}^X_{ B{\overline{B'}} \to \mu {\overline{\mu}}}(\vec k,\vec q;t') =
\left[
-{f_{XB'\mu}\over m_{\pi^+}} (g_{\lambda\nu} +
x_\Delta\gamma_\nu\gamma_\lambda) k^\nu({\overline{\mu}})
\right]
S^{\lambda\rho}_X(p,A=-1)} \hspace{5.cm}
\nonumber \\
&&
\left[
-{f_{XB\mu}\over m_{\pi^+}} (g_{\rho\sigma} +
x_\Delta\gamma_\rho\gamma_\sigma) k^\sigma(\mu)
\right]\quad,
\label{eq:Del_exch}\end{aligned}$$
- $\rho$-pole graph with bare mass $m^{(0)}_\rho$ $$\begin{aligned}
\lefteqn{
{\cal V}^\rho_{ B{\overline{B'}} \to \mu {\overline{\mu}}}(\vec k,\vec q;t') =
\left(
i g^{(0)}_{BB'\rho} \gamma_\lambda + {f^{(0)}_{BB'\rho} \over 2M_N}
\sigma_{\lambda\nu} P^\nu
\right)
{ g^{\lambda\sigma} - P^\lambda P^\sigma/(m^{(0)}_\rho)^2
\over t' - (m^{(0)}_\rho)^2 }}\hspace{7cm}
\nonumber \\
&& \!\!\!\! \left[ -i g^{(0)}_{\mu\mu\rho} (k(\mu)-k({\overline{\mu}}))_\sigma
\right]\,.
\label{eq:4born33}\end{aligned}$$
These expressions can be further simplified by introducing the momenta given in Eq. \[eq:4born30\] and contracting the $\gamma$-matrices. Finally, the corresponding $B{\overline{B'}} \to \mu {\overline{\mu}}$ helicity amplitudes are obtained by applying $ {\cal V}_{B{\overline{B'}} \to \mu\bar\mu}(\vec k,\vec q;t)$ to the Dirac helicity spinors of the baryons (cf. Eqs. \[eq:app2a\] and \[eq:app3a\] in the Appendix): $${\langle \, \mu\bar\mu, \vec k \, \vert \>} V(t) {\> \vert \, B{\overline{B'}},\vec q ,\lambda_1
\lambda_2 \, \rangle} =
{\overline{v}}_{B'}(-\vec q,\lambda_2) {\cal V}_{B{\overline{B'}} \to \mu\bar\mu}
(\vec k,\vec q;t)
u_{B}(\vec q,\lambda_1)\quad.
\label{eq:born_heli}$$ The final results for the helicity amplitudes are summarized in Appendix \[sec:AnhC\].
In our model the coupling constants at the various hadronic vertices are related to each other by $SU(3)$ arguments. The $SU(3)$ relations together with the isospin factors of the various Born amplitudes are given in Appendix \[sec:borniso\].
As discussed in the previous chapters, for the dispersiontheoretic calculation of correlated $\pi\pi$ and $K{\overline{K}}$ exchange the $B{\overline{B'}} \to \pi\pi,K{\overline{K}}$ have to be known in the pseudophysical region, i.e. for energies $t'$ below the $B{\overline{B'}}$ threshold ($4m_\pi^2 \leq t' < (M_B + M_{B'})^2 $). Therefore, after having derived the analytic expressions for these Born amplitudes in the physical region ($\sqrt{t'} >M_B+M_{B'}$) they have to be continued analytically as functions of $t'$ (and $s$) into the pseudophysical region. For this all energy-dependent quantities occuring in the expressions for the Born amplitudes (cf. Appendix \[sec:AnhC\]) have to be expressed as functions of $t'$ and $s$.
If we adopt the approximation introduced in Sect. \[sec:kap3\_4\], namely that the masses of the baryon and of the antibaryon are equal ($M_B=M_{B'}$), the square of the relative 4-momentum and the one-particle energies of the $B{\overline{B'}}$ state are given in the center-of-mass system for physical values of $t$ by $$\begin{array}{rcl}
q^2(t) &=& t/4 \, - M_B^2 \quad,\\ E_B=E_{B'} &=& \sqrt{t}/2\quad.
\end{array}$$ The analytic continuation of these relations to the pseudophysical region is obvious. Note that if we would have allowed $M_B$ and $M_{B'}$ to be different the corresponding relations would look more involved: $$\begin{array}{rcl}
q^2(t) &=& {
\left[t-(M_B+ M_{B'})^2\right]
\left[t-(M_B- M_{B'})^2\right]
\over 4t}\quad,
\\
E_{B} (t) &=& {(E_{B} + E_{B'})\over2} + {(E_{B}^2 - E_{B'}^2)\over
2(E_{B} + E_{B'})}
\\
&=& {t+M_B^2 - M_{B'}^2 \over 2\sqrt{t}} \quad,
\\ \\
E_{B'} (t) &=&{t+M_{B'}^2 - M_{B}^2 \over 2\sqrt{t}}\quad.
\end{array}
\label{eq:cont_3}$$
Baryon-exchange diagrams in which the mass $M_X$ of the exchanged baryon is sufficiently smaller than the mass $M_B=M_{B'}$ of the external baryons (e.g. $N$-exchange in $\Lambda{\overline{\Lambda}}\to K{\overline{K}}$ or $\Lambda$-exchange in $\Sigma{\overline{\Sigma}}\to\pi\pi$ do not satisfy a Mandelstam representation as was already pointed out in Ref. [@Riska] for the latter example. The nonvalidity of the Mandelstam representation becomes obvious when extrapolating the corresponding Born amplitude in Eq. \[eq:4born32\] to the pseudophysical region. For given $t< 4(M_B^2 - M_X^2)$, the propagator of baryon $X$, $$\begin{aligned}
\left[ p^2 - M_X^2\right]^{-1}
&=& -\left[ t/4 -M_B^2 + M_X^2 + {\vec k}^2 - 2 \vec q \cdot \vec k
\right]^{-1}\quad,\end{aligned}$$ then acquires a singularity at $cos\theta=0 $ ($\vec q \cdot \vec k$ imaginary) and the following off-shell momentum of the $\mu{\overline{\mu}}$ state $${\vec k}^2 = M_B^2 - M_X^2 - t/4 >0
\quad.$$ Since this problem would hinder an evaluation of correlated $\pi\pi$ and $K{\overline{K}}$ exchange and we do not see at present any proper solution, we eliminate the problem by the following approximation: In all baryon-exchange diagrams in which the mass of the exchanged baryon, $M_X$, is smaller than the mass of the external baryons, $M_B$, $M_X$ is increased by hand to $M_B$. Again, since the uncorrelated contributions of two-pion and two-kaon exchange (e.g. iterative two-pion exchange in the $\Sigma N$ channel with a $\Lambda N$ intermediate state ) are evaluated explicitly in the $s$-channel, these contributions are not affected by our approximation and thus have the correct energy dependence. Approximations are only made in the correlated part which have a much weaker energy dependence than the uncorrelated contributions.
Results and Discussion {#sec:results}
======================
Determination of free parameters {#sec:5_1}
--------------------------------
During the construction of the microscopic model for the $B{\overline{B'}}
\to\mu{\overline{\mu}}$ amplitudes it proved to be essential to restrict the number of free parameters as much as possible, which can then be fixed by adjusting the model predictions to the quasiempirical $N{\overline{N}}\to\pi\pi$ amplitudes. Only in this way one can hope to obtain a reasonable description of the other baryon-antibaryon channels, for which no empirical data exist.
As shortly outlined in Sect. \[sec:kap4\] the $\pi\pi-K{\overline{K}}$ interaction model has been developed independently before, with all parameters adjusted to fit the existing $\pi\pi$ scattering phase shifts. Therefore coupling constants and form factors at the $\pi\pi\rho^{(0)}$ and $KK\rho^{(0)}$ vertices occurring in the $\rho^{(0)}$ pole terms are already determined. Assuming that the bare $\rho$ meson couples universally to the isospin current we can also fix all vector couplings $g^{(0)}_{BB'\rho}$ to the baryonic vertices. Corresponding tensor couplings $f^{(0)}_{BB'\rho}$ will be related by $SU(3)$ symmetry, with two parameters remaining, namely the coupling constant $f^{(0)}_{NN\rho}$ and the $F/(F+D)$ ratio $\alpha_v^m$.
The coupling of the pseudoscalar mesons $\pi$ and $K$ to the octet baryons is in the framework of $SU(3)$ symmetry likewise determined by two parameters, the $F/(F+D)$ ratio $\alpha_p$ and the coupling constant $g_{NN\pi}$. (For the latter we will take throughout $g_{NN\pi}^2/4\pi= 14.3$). There is an additional freedom since $SU(3)$ symmetry can be either assumed for the pseudoscalar coupling constants $g_{BB'p}$ or the pseudovector ones $f_{BB'p}$, which are related by $$g_{B'Bp} = f_{B'Bp} \;{M_B + M_{B'}\over m_{\pi^+} } \quad.
\label{eq:ps_pv}$$ These two possibilities are not equivalent since, because of $SU(3)$ breaking of baryon masses, both sets of coupling constants cannot be $SU(3)$ symmetric at the same time. In this work we will assume $SU(3)$ symmetry for the pseudoscalar couplings, but we will also check the influence of the alternative possibility.
Finally, under the assumption of $SU(3)$, the couplings $f_{BDp}$ of $\pi$ and $K$ to the transition current between baryon octet and decuplet are determined by only one parameter, $f_{N\Delta\pi}$. We will use $f^2_{N\Delta\pi}/4\pi = 0.36$ in the following.
In addition we have form-factor parameters at the hadronic vertices. In order to keep their number small we assume that the cutoff masses $\Lambda_{BX\mu}$ are independent of the exchanged baryon $X$ within one $SU(3)$ multiplet. Consequently we have two additional parameters: $\Lambda_8$ if $X$ is a member of the baryon octet and $\Lambda_{10}$ if it is in a decuplet. The power $n_8=1$ and $n_{10}=2$ in the form-factor [*ansatz*]{} are sufficient to ensure convergence of the scattering equation. Finally, $x_\Delta$ (Eq. \[eq:Lag\]) characterizing the off-shell part of the $(3/2)^+\otimes (1/2)^+\otimes
0^-$ coupling is treated as a free parameter in our $B{\overline{B'}}\to\mu{\overline{\mu}}$ model.
In order to reduce the number of parameters further we even assume $SU(6)$ symmetry, which fixes $\alpha_p$ and $\alpha_v^m$ to be 0.4. Thus we are left with four free parameters $f^{(0)}_{NN\rho}$, $x_\Delta$, $\Lambda_8$ and $\Lambda_{10}$, which have been fixed by adjusting our theoretical predictions for the $N{\overline{N}}\to\pi\pi$ amplitudes to the quasiempirical results of Höhler and Pietarinen [@Hoehler2; @Dumbrajs] given in the form of Frazer-Fulco amplitudes $f_\pm^J(t)$, which, up to kinematic factors, correspond to the partial wave decomposed helicity amplitudes of Sect. \[sec:kap3\_4\], i.e. $$\begin{aligned}
f^J_+(t)&=&-{1\over 16 \pi^2} {qM_N\over (qk)^J }
{\langle \, \pi\pi \, \vert \>} T^J(t) {\> \vert \, N{\overline{N}},++ \, \rangle}\times F^J \quad,
\nonumber \\
f^J_-(t)&=&-{1\over 8 \pi^2} {qM_N\over (qk)^J \sqrt{t} }
{\langle \, \pi\pi \, \vert \>} T^J(t) {\> \vert \, N{\overline{N}},+- \, \rangle}\times F^J\quad,\end{aligned}$$ where $k$ and $q$ are the on-shell momenta of pions and nucleons. The factors $F^J = - 1/\sqrt{6}\, (-1/2)$ for $J=0\,(1)$ are due to the transition from isospin amplitudes used in this work to the Frazer-Fulco amplitudes, which are defined in isospin space as coefficients of the independent isospin operators $\delta_{\alpha\beta}$ and ${{1\over2}}[\tau_\alpha,
\tau_\beta]$. Since ${\langle \, \pi\pi \, \vert \>} T^{J=0}(t) {\> \vert \, N{\overline{N}},+- \, \rangle}$ vanishes identically we have only one amplitude, $f_+^0$, in the $\sigma$ channel whereas in the $\rho$ channel we have both $f_+^1$ and $f_-^1$.
Fig. \[fig:5\_2\] shows the predictions of our microscopic model for $f_+^0$, $f_+^1$ and $f_-^1$, in comparison to the quasiempirical results of Ref. [@Hoehler2] in the pseudophysical region $t\geq4m_\pi^2$; Table \[tab:5\_param\] contains the chosen parameter values. (Note that the present model differs somewhat from our former model [@Pearce_piN], e.g. by the inclusion of $Y^*$ exchange; therefore the values differ slightly from those given in Ref. [@Pearce_piN]). With only four parameters we obtain a very satisfactory reproduction of the quasiempirical data, especially in the $\rho$ channel. Some discrepancies occur in the $\sigma$ channel, which however have only small influence on final results for the correlated $\pi\pi$ exchange in the $s$-channel reactions ($NN$, $\pi N$), as also discussed below. Furthermore it should be kept in mind that in this channel the quasiempirical information is plagued with considerable uncertainties.
The $B{\overline{B'}}\to\mu{\overline{\mu}}$ transition potentials {#sec:5_2}
------------------------------------------------------------------
Having fixed all parameters in our microscopic model for the $B{\overline{B'}}\to\mu{\overline{\mu}}$ amplitudes we will now first look at the transition potentials (Born amplitudes) in the various baryon-antibaryon channels. Figs. \[fig:5\_nnborn\] and \[fig:5\_ssborn\]$B{\overline{B'}}\to\pi\pi,\,
K{\overline{K}}$ amplitudes show the contributions of the various baryon-exchange processes to the $N{\overline{N}}\to\pi\pi,K{\overline{K}}$ and $\Sigma{\overline{\Sigma}}\to\pi\pi,K{\overline{K}}$ Born amplitudes above the $\mu{\overline{\mu}}$ thresholds, i.e. for $t\geq 4m_\mu^2$. Contributions of the $\rho^{(0)}$ pole terms are not shown since they possess a singularity at the bare rho mass $m_{\rho}^{(0)}$. This pole will be regularized only after iteration and coupling to the full $\pi\pi$ amplitude and then leads to the resonance structure at the physical rho mass $m_\rho$.
The notation of the (Born) amplitudes follows that of the Frazer-Fulco amplitudes, i.e. $$\begin{aligned}
V^J_+(t)&=&{\langle \, \mu{\overline{\mu}} \, \vert \>} V^J(t) {\> \vert \, B{\overline{B'}},++ \, \rangle}\quad,
\nonumber \\
V^J_-(t)&=&{\langle \, \mu{\overline{\mu}} \, \vert \>} V^J(t) {\> \vert \, B{\overline{B'}},+- \, \rangle}\quad.\end{aligned}$$ The partial wave decomposed Born amplitudes are either even or odd functions of the baryonic relative momenta (cf. Appendix \[sec:AnhC\]). Since the latter are imaginary in the pseudophysical region the Born amplitudes being odd functions (e.g. $V^0_+$) become likewise imaginary.
Apart from coupling constants and isospin factors the strengths of the various contributions are strongly determined by mass ratios. Exchange baryons with the same mass $M_X$ as the outer baryon-antibaryon pairs (according to our approximation introduced in Sect. \[sec:bornsp\] this case includes also those exchange baryons whose physical mass is lower than that of the outer baryons) produce, in the $\pi\pi$ amplitudes $V_+^0$ and $V_+^1$, a typical structure at the $\pi\pi$ threshold. The strong rise of the amplitudes, which acquire a finite value at $t=4m_\pi^2$, is a direct signal of the so-called left-hand cut, which is generated by the singularity of the corresponding $u$-channel pole graph. This cut starts just below the $\pi\pi$ threshold and extends from $t_0=4m_\pi^2(1-m_\pi^2/M_X)$ along the real axis to $-\infty$.
Obviously, in the $\sigma$ channel, the various pieces interfere constructively, whereas in the $\rho$ channel also destructive interferences occur. The contributions generated by a spin-3/2 exchange baryon have opposite sign in $V^1_+$ and $V_-^1$ whereas both amplitudes are almost equal when a spin-1/2 baryon is exchanged.
The $N{\overline{N}}\to \pi\pi$ Born amplitudes in Fig. \[fig:5\_nnborn\] built up by nucleon and $\Delta$ exchange are noticeably larger than the $N{\overline{N}}\to K{\overline{K}}$ Born amplitudes, which are suppressed because of the high mass of the exchange baryons $\Lambda$, $\Sigma$, $Y^*$. Due to different mass ratios this is no longer true in the hyperon-antihyperon channels. For instance, in the $\Sigma{\overline{\Sigma}}$ channel shown in Fig. \[fig:5\_ssborn\] the coupling via $\Delta$ exchange to $K{\overline{K}}$ even dominates, due to the large coupling constant ($f^2_{\Sigma\Delta K}=
f^2_{N\Delta\pi}$) and the small mass difference between $\Sigma$ and $\Delta$ ($M_\Delta-M_\Sigma\approx 39\
MeV$). Therefore already from these results it is to be expected that correlated $K{\overline{K}}$ exchange processes play a minor role in the $NN$ system but are important for the interactions involving hyperons.
The $B{\overline{B'}}\to \mu{\overline{\mu}}$ transition amplitudes {#sec:Kap53}
--------------------------------------------------------------------
The $B{\overline{B'}}\to \mu{\overline{\mu}}$ helicity amplitudes, $T^J_\pm (t)$, which are obtained from the solution of the BbS scattering equation \[eq:bbsn\], consist of Born terms and pieces containing meson-meson correlations.The latter part fixes the phase of the $B{\overline{B'}}\to \mu{\overline{\mu}}$ amplitudes and generates the discontinuities of $T^J_\pm (t)$ along the unitarity cut. It contains $\pi\pi$ and $K{\overline{K}}$ correlations in the form of $\pi\pi$, $K{\overline{K}}$ amplitudes. Corresponding predictions from our field-theoretic model, for the scalar-isoscalar on-shell amplitude $T^{J=0,I=0}(t)$ are shown in Fig. \[fig:5\_tmumu\]. By comparison of the $\pi\pi$ amplitude with the $\pi\pi$ phase shifts (cf. Fig. \[fig:4\_6\]) we can see immediately that the vanishing of the real part of the amplitude at $t=37 m_\pi^2$ and $t=71m_\pi^2$ corresponds to the phase shifts acquiring the value of 90 resp. 270 degrees, respectively. The phase value of 180 degrees (and the corresponding vanishing of the amplitude) occurs just below the $K{\overline{K}}$ threshold ($t=50.48 m_\pi^2$). Remarkably, there is a steep rise of the imaginary part of the $\pi\pi$ amplitude at low energies, which has to vanish at the $\pi\pi$ threshold for unitarity reasons. We mention that, in the $NN$ system, this part provides the dominant contribution to correlated $\pi\pi$ exchange. It should be noted that both the real and imaginary part of the $K{\overline{K}}\to
K{\overline{K}}$ amplitude (which in Fig. \[fig:5\_tmumu\] are shown only in the physical region above the $K{\overline{K}}$ threshold) acquire a non-vanishing value at the $K{\overline{K}}$ threshold due to the open $\pi\pi$ channel.
Figs. \[fig:5\_tnn\] and \[fig:5\_tss\] show the resulting $B{\overline{B'}}\to \pi\pi$ amplitudes for $B{\overline{B'}}=N{\overline{N}}, \Sigma{\overline{\Sigma}}$. Besides the predictions of the full model (solid line) the figures also show the effect which is obtained when neglecting the $B{\overline{B'}}\to K{\overline{K}}$ Born amplitudes in the $\sigma$ channel (dashed line) and restricting oneself to the $\rho$-pole amplitudes in the $\rho$ channel (dash-dotted line). The $B{\overline{B'}}\to K{\overline{K}}$ amplitudes (dotted lines) are only shown for $J=0$ since for $J=1$ as exemplified for the $\Sigma{\overline{\Sigma}}$ channel they provide almost negligible contributions, mainly because of the weak $K{\overline{K}}$ interaction in that partial wave. Note that in the $\sigma$ channel, due to the imaginary Born amplitudes, the role of the real and imaginary part of $T^0_+$ are in a certain sense interchanged: ${\sl Im} [T^0_+]$ contains the Born amplitudes whereas the discontinuity along the unitarity cut is contained in ${\sl Re}[T^0_+]$.
Obviously, in all channels, the structure of the $\pi\pi$ amplitude of Fig. \[fig:5\_tmumu\] can be recognized in those $B{\overline{B'}}\to\pi\pi$ amplitudes $T^0_+$, which have been evaluated without the $B{\overline{B'}}\to K{\overline{K}}$ Born amplitudes (dashed curves). The steep rise of ${\sl Im}[T^0_+]$ at the $\pi\pi$ threshold originates from the analogous behavior of the Born amplitudes (cf. Figs. \[fig:5\_nnborn\] and \[fig:5\_ssborn\]). As follows from the unitarity relation for the $B{\overline{B'}}\to\pi\pi$ amplitudes below the $K{\overline{K}}$ threshold ($t_{thr}\approx 50.48 m_\pi^2$) the discontinuity of the $B{\overline{B'}}
\to\pi\pi$ amplitudes (contained in ${\sl Re}[T^0_+]$) must vanish at $t\approx 50.43 m_\pi^2$ since there $T^{J=0,I=0}=0$. Furthermore, those amplitudes $T^0_+$, which have been evaluated without the $K{\overline{K}}$ Born amplitudes, vanish at an additional, though for each channel different position in the energy region below the $K{\overline{K}}$ threshold. Interestingly this vanishing of the amplitudes is lost in the hyperon-antihyperon channels when the $K{\overline{K}}$ Born amplitudes are included; in case of the $N{\overline{N}}\to\pi\pi$ channel on the other hand the resulting full amplitude still vanishes, at a slightly different position ($t\approx 50.2m_\pi^2$). Note that the quasiempirical $N{\overline{N}}\to\pi\pi$ amplitude $f^0_+$ in Fig. \[fig:5\_2\] has a comparable structure, however the amplitude vanishes already at a somewhat smaller value of $t\approx43m_\pi^2$. Fig. \[fig:5\_thr\] also clearly shows the cusp structure of the amplitudes at the $K{\overline{K}}$ threshold, which is however much stronger in the hyperon-antihyperon channels.
The $B{\overline{B'}}\to K{\overline{K}}$ Born amplitudes do not play any role in the $N{\overline{N}}\to\pi\pi$ channel, as noted already in Refs. [@Durso; @Pearce_piN]. In the hyperon-antihyperon channels on the other hand, we have already at small $t$-values important contributions to the $B{\overline{B'}}\to \pi\pi$ amplitudes $T^0_+$, which lead to an enhancement of the amplitudes below the $K{\overline{K}}$ threshold and to a noticeably different energy dependence.
The amplitudes $T^1_\pm$ in the $\rho$ channel are dominated by the resonance structure in the region of the physical $\rho$ mass at $t=30.4 m^2_\pi$. We stress that by multiplication with the $\pi\pi$ correlations in the scattering equation also the baryon-exchange Born terms lead to resonant contributions. However, in all channels, the discontinuity of the amplitudes mainly arises (to at least $60\%$) from both $\rho$ pole terms and those contributions generated by them in the scattering equation (dash-dotted curves in Figs. \[fig:5\_tnn\] and \[fig:5\_tss\]).
As shown in Fig. \[fig:5\_tss\] for the $\Sigma{\overline{\Sigma}}\to K{\overline{K}}$ transition, the $B{\overline{B'}}\to K{\overline{K}}$ on-shell amplitudes (dotted curves) are completely unimportant in the $\rho$ channel. On the other hand, in the $\sigma$ channel, the $ B{\overline{B'}}\to K{\overline{K}}$ amplitudes can acquire large values especially near the $K{\overline{K}}$ threshold, due to the corresponding behavior of the $K{\overline{K}}$ amplitude $T^{J=0,I=0}$ (cf. Fig. \[fig:5\_tmumu\]), which lead to non-negligible contributions to the spectral functions.
The spectral functions {#sec:5_4}
----------------------
Based on the $B{\overline{B'}}\to\mu{\overline{\mu}}$ amplitudes in the pseudophysical region determined in the last section we can now in principle evaluate the spectral functions $\rho^{\sigma,\rho}_i $ (defined in Eq. \[eq:3\_5\_1\]) of correlated $\pi\pi$ and $K{\overline{K}}$ exchange for all baryon-baryon channels containing the octet baryons $N,\Lambda,\Sigma$ and $\Xi$. In this work however we restrict ourselves to those channels in which experimental information is available at present or in the near future. Besides the baryon-baryon channels with strangeness $S=0$ and $S=-1$, for which numerous ($NN$) and scarce ($N \Lambda$, $N\Sigma$) scattering data exist, we will also consider the $S=-2$ channels $\Lambda\Lambda$, $\Sigma\Sigma$ and $N \Xi$, which are relevant both for the description of $\Lambda\Lambda$- resp. $\Xi$-hypernuclei and for the study of the $H$-dibaryon [@Jaffe] predicted in the ($S=-2$, $J=I=0$) channel.
Fig. \[fig:5\_5\_1\]a (Fig. \[fig:5\_5\_1\]b) shows the spectral function $\rho^{\sigma}_S(BB')$ in the $\sigma$ channel predicted by the full microscopic model for the $S=0,-1$ $BB'$ channels $NN,N\Lambda,N\Sigma$ ($S=-2$ channels $\Lambda\Lambda,\Sigma\Sigma,N\Xi$). Due to isospin conservation the $\pi\pi$ resp. $K{\overline{K}}$ exchange contributes to the $N\Lambda$ and $\Lambda\Lambda$ interaction only in the $\sigma$ channel and to the $N\Lambda-N\Sigma$ transition only in the $\rho$ channel. Up to near the $K{\overline{K}}$ threshold at $t=50.48 m^2_\pi$ $\rho^{\sigma}_S$ is negative in all channels. Therefore, as expected, correlated $\pi\pi$ and $K{\overline{K}}$ exchange provides attractive contributions in all channels. Especially at small $t$-values, which determine the long-range part of the correlated exchanges and therefore yield for low-energy processes in the $s$-channel the main contributions to the dispersion integral, the spectral function $\rho^{\sigma}_S(NN)$ is by about a factor of 2 larger than the results for $\rho^{\sigma}_S(N\Lambda)$ and $\rho^{\sigma}_S(N\Sigma)$, which have about the same size. Due to the sizable contributions of the $K{\overline{K}}$ Born amplitudes to the hyperon-antihyperon amplitudes found already in the last section, $\rho^{\sigma}_S(N\Lambda)$ as well as $\rho^{\sigma}_S(N\Sigma)$ show a noticeably different $t$-dependence than $\rho^{\sigma}_S(NN)$. Below the $K{\overline{K}}$ threshold the spectral functions are broadened to larger $t$-values, therefore the overall range of this correlated exchange is reduced. This is even more true for the $S=-2$ channels $\Lambda\Lambda$ and $\Sigma\Sigma$ in Fig. \[fig:5\_5\_1\]b since there the hyperon-antihyperon amplitudes enter quadratically. The effect of the $B{\overline{B'}}\to K{\overline{K}}$ Born amplitudes is once more shown in Fig. \[fig:5\_5\_1b\] for the $NN$ and $N\Sigma$ channel. Whereas they have small influence on $\rho^{\sigma}_S(NN)$ they provide important contributions to $\rho^{\sigma}_S(N\Sigma)$ already near the $\pi\pi$ threshold.
Fig. \[fig:5\_5\_2\] shows the corresponding spectral functions in the $\rho$ channel ($\rho^\rho_S$, $\rho^\rho_V$, $\rho^\rho_T$ and $\rho^\rho_6$) for the various particle channels, which are non-vanishing. (Note that $\rho^\rho_P$ depends linearly on the four functions shown, cf.Eq. \[eq:3\_lindep\], and is therefore not shown). $\rho^\rho_6$ contributes only in $BB'$ channels with different masses and therefore vanishes in the $NN$ and $\Sigma\Sigma$ channel. According to Eq. \[eq:3\_4\_6\] $\rho^\rho_S(s,t)$ depends on $s$ through the factor $\cos\vartheta_t(s,t)$. In order to enable a comparison of the results for $\rho^\rho_S(s,t)$ in the various particle channels we have throughout set $s$ on the threshold of the $s$-channel process in question, i.e. $s=(M_B+M_{B'})^2$. Since the $\rho$ channel is dominated by the resonant pieces in the $\pi\pi$ channel and the $K{\overline{K}}$ channel does not play any role the $\rho^\rho_i(BB')$ possess almost the same $t$-dependence in the various particle channels.
For the $ NN$ channel, the spectral functions for correlated $\pi\pi$ and $K{\overline{K}}$ exchange can be derived either from the $N{\overline{N}}\to\pi\pi$ amplitudes of our microscopic model or, alternatively, from the quasiempirical results obtained in Refs. [@Hoehler2; @Dumbrajs]. As before we have then to subtract the uncorrelated pieces evaluated from the microscopic model for the $N{\overline{N}}\to\pi\pi$ Born amplitudes. Therefore the results derived from the quasiempirical amplitudes depend on parameters (e.g. $g_{NN\pi}$) of the microscopic model.
In Fig. \[fig:5\_5\_4\] we show the $NN$ spectral functions in the $\sigma$ as well as in the $\rho$ channel obtained from our microscopic model (solid line) and the quasiempirical amplitudes (dashed line). As expected already from the comparison of the amplitudes in Sect. \[sec:5\_1\] the results agree quite well in the $\rho$ channel. On the other hand some discrepancies occur in the $\sigma$ channel. At smaller $t$-values, which determine the long-range part of the correlated exchanges, the theoretical model yields somewhat more attraction than the quasiempirical results. Larger discrepancies occur at higher $t$-values, which are however of minor relevance for the correlated exchange in the $NN$ interaction. Namely, above $t\approx30 m_\pi^2$ the correlated exchanges are of shorter range than $\omega$ exchange, which generates the strong repulsive inner part of the $NN$ potential (as well as of the other baryon-baryon potentials). Therefore the short-ranged parts of the correlated $\pi\pi$ exchanges are completely masked by the repulsive $\omega$ exchange and have a small influence only on $NN$ observables. Furthermore one has to realize that the quasiempirical results obtained by extrapolation of data from the physical region of the $s$- and $u$-channel into the pseudophysical region of the $t$-channel have considerable uncertainties.
The potential of correlated $\pi\pi$ and $K{\overline{K}}$ exchange
--------------------------------------------------------------------
Using the spectral functions of the last chapter we can now evaluate the dispersion integrals, Eqs. \[eq:3\_4\_1\]-\[eq:3\_4\_2b\], in order to obtain the invariant amplitudes in the $s$-channel. Eq. \[eq:3\_5\_0\] then provides the (on-shell) baryon-baryon interaction due to correlated $\pi\pi$ and $K{\overline{K}}$ exchange.
### The effective coupling constants {#sec:3_5_1}
The results will be presented in terms of the effective coupling strengths $G^{\sigma,\rho}_{AB\to CD}(t)$, which have been introduced in Eqs. \[eq:3\_effccsig\], \[eq:3\_effccrho\] to parametrize the correlated processes by (sharp mass) $\sigma$ and $\rho$ exchange. We stress once more that this parametrization does not involve any approximations as long as the full $t$-dependence of the effective coupling strengths are taken into account. The parameters of $\sigma$ resp. $\rho$ exchange (mass of the exchanged particle: $m_\sigma$, $m_\rho$; cutoff mass: $\Lambda_\sigma$, $\Lambda_\rho$) are chosen to have the same values in all particle channels. $m_\sigma$ and $m_\rho$ have been set to the values used in the Bonn-Jülich models of the $NN$ [@MHE] and $YN$ [@Holz] interactions, i.e.$m_\sigma=550\ MeV$, $m_\rho=770\ MeV$. The cutoff masses have been chosen such that the coupling strengths in the $S=0, -1$ baryon-baryon channels vary only weakly with $t$. The resulting values ($\Lambda_\sigma=2.8\ GeV$, $\Lambda_\rho=2.5\ GeV$) are quite large compared to the values of the phenomenological parametrizations used in Refs. [@MHE; @Holz] and thus represent very hard form factors. Note that, unless stated otherwise, the upper limit $t'_{max}$ in the dispersion integrals is put to $120 m_\pi^2$. Fig. \[fig:5\_6\_1\] shows the effective coupling strengths $G^{\sigma }_{AB}(t)$ in the baryon-baryon channels considered here, as function of $-t$. With the exception of $G^{\sigma }_{\Sigma\Sigma}(t)$ (dash-dotted curve) the effective $\sigma$ coupling strengths vary only weakly with $-t$, which proves that $550\ MeV$ is a realistic choice for the $\sigma$ mass. In the $\Sigma\Sigma$ channel the suitable mass lies somewhat higher due to the strong coupling to the $K{\overline{K}}$ channel generated by $\Delta$ exchange, cf. Fig. \[fig:5\_ssborn\]. If we compare the relative strengths of effective $\sigma$ exchange in the various baryon-baryon channels we observe the same features observed already for the spectral functions: The scalar-isoscalar part of correlated $\pi\pi$ and $K{\overline{K}}$ exchange is in the $ NN$ channel about twice as large as in both $YN$ channels and by a factor 3-4 larger than in the $S=-2$ channels.
For sharp mass $\sigma$ exchange the following equation holds $$G^{\sigma }_{NN\to NN}(t) G^{\sigma }_{\Sigma\Sigma\to\Sigma\Sigma}(t)
=
[G^{\sigma }_{N\Sigma\to N\Sigma}(t)]^2\quad.$$ In other words, the three processes are determined by two coupling constants $g_{\sigma NN} \equiv \sqrt{G^{\sigma }_{NN\to NN}}$ and $g_{\sigma\Sigma\Sigma}\equiv\sqrt{G^{\sigma }_{\Sigma\Sigma\to
\Sigma\Sigma}}$, such that $G^{\sigma }_{N\Sigma\to N\Sigma}(t)=g_{\sigma NN}
g_{\sigma\Sigma\Sigma}$. For correlated exchanges this is not necessarily true anymore; here we have in general $$G^{\sigma }_{NN\to NN}(t) G^{\sigma }_{\Sigma\Sigma\to\Sigma\Sigma}(t)
\geq
[G^{\sigma }_{N\Sigma\to N\Sigma}(t)]^2 \quad,$$ which is just a consequence of the Schwarz inequality relation $$\; |\int f(t) g(t) dt |^2 \leq \int f(t)^2 dt \times \int g(t)^2 dt
\quad.$$ The equality holds if both functions have the same $t$-dependence. This is roughly fulfilled for $NN$ and $\Lambda\Lambda$, but not for $\Sigma\Sigma$; therefore the equality approximately holds in the first but not in the second case. Consequently, in general, effective vertex couplings for correlated exchanges are not well defined since they might take different values in different baryon-baryon channels.
Tab. \[tab:5\_6\_10\] contains the coupling strengths at $t=0$. Besides the results for our full model for correlated $\pi\pi$ and $K{\overline{K}}$ exchange the table includes results obtained when neglecting the $B{\overline{B}}\to K{\overline{K}}$ Born amplitudes. Obviously the inclusion of these amplitudes provides only $15\%$ of the coupling strengths in the $NN$ channel; it is much more important in the channels with strangeness, in fact providing the dominant part in the $S=-2$ channels. Furthermore the table contains results obtained when uncorrelated contributions involving spin-1/2 baryons only are subtracted from the discontinuities of the invariant baryon-baryon amplitudes in order to avoid double counting in case a simple OBE-model is used in the $s$-channel. For the full Bonn $NN$ model contributions involving spin-3/2 baryons have to be subtracted too (as done in general in this work) since corresponding contributions are already treated explicitly in the s-channel. Obviously processes involving spin-3/2 baryons increase the ‘true’ correlated contribution by about $30\%$ in all channels.
In the $\rho$-channel the spectral functions are dominated by the resonant contributions in the region of the $\rho$ resonance. Therefore the effective coupling strengths $^{ij}G^\rho_{AB\to CD}(t)$ $(ij=VV,VT,TV,
TT)$ vary even more slowly with $t$ than those in the $\sigma$-channel. Because of this weak $t$-dependence it is for the moment sufficient to consider only the values of coupling strengths at $t=0$. They are shown for the $NN$ channel in Tab. \[tab:5\_6\_11\], for $YN$ in Tab. \[tab:5\_6\_12\] and in Tab. \[tab:5\_6\_13\] for the $S=-2$ baryon-baryon channels.
Note that for the equal (unequal) mass case the results are given in terms of 3 (4) coupling strengths. For equal masses the present description in terms of correlated $\pi\pi$ exchange is more involved compared to sharp mass $\rho$-exchange since the latter can be characterized by two parameters only, the vector and tensor coupling constant. Also, as in the scalar channel, there is no definite relation between coupling strengths in the various channels so that vertex coupling constants cannot be uniquely extracted but depend on the channel chosen. (Thus it is not surprising that our $\rho NN$ coupling strengths, which are determined in the $NN$ system, are not fully consistent with the vector and tensor coupling constants derived by Höhler and Pietarinen [@Hoehler2; @Dumbrajs] from $\pi N$ scattering, though both calculations agree of course qualitatively.)
Tables \[tab:5\_6\_11\]-\[tab:5\_6\_13\] include results obtained when only the $\rho$-pole terms are considered in the $B{\overline{B'}}\to\mu{\overline{\mu}}$ Born amplitudes. In this case the effective coupling strengths are still $SU(3)$ symmetric, i.e. they roughly fulfill the relations $$\begin{array}{l@{\qquad}l}
g_{\Sigma\Lambda\rho}=0\quad, &
f_{\Sigma\Lambda\rho}={2\sqrt{3}\over 5} \left(g_{NN\rho}+f_{NN\rho}
\right)\quad, \\
g_{\Sigma\Sigma\rho}=2g_{NN\rho}\quad,&
f_{\Sigma\Sigma\rho}=-{2\over5}\left(3g_{NN\rho}-2f_{NN\rho}\right)
\quad,\\
g_{\Xi\Xi\rho}= g_{NN\rho}\quad,&
f_{\Xi\Xi\rho}= -{1\over5}\left(6g_{NN\rho}+f_{NN\rho}\right)\quad.\\
\end{array}$$ with $\alpha_v^e=1$ and $\alpha_v^m=0.4$. (We remind the reader that we have chosen the bare couplings to exactly obey $SU(3)$ symmetry.) Obviously the influence of ($SU(3)$ broken) baryon masses on the $\rho$-pole contributions is small, probably because our calculations are performed in the pseudophysical region, far below the baryon-baryon thresholds. As expected however the effective coupling strengths of the complete calculation do not respect $SU(3)$ symmetry anymore, due to the sizable influence of the (non-pole) baryon-exchange processes.
Doing again a restricted subtraction of spin-1/2 baryon contributions only (suitable for an OBE model) we now do not have a unique trend for the change of coupling strengths in all channels, as found before in the scalar case. In case of the vector coupling strength ($VV$) in the $NN$ system the restricted subtraction leads even to a smaller value. This is not too surprising if one realizes that the $\rho$-vector coupling strengths are obtained from differences of approximately equal spectral functions so that well-controlled changes in the spectral functions can lead sometimes to large modifications of coupling strengths, in arbitrary direction.
At this point we would like to make a remark about the sensitivity of our results to the upper limit in the dispersion integral, $t'_{max}$. It is in fact quite small: Lowering $t'_{max}$ from the generally used value of $120m_\pi^2$ to $80m_\pi^2$ the effective $\sigma$-coupling strengths are increased by less than $10\%$; in the $\rho$-channel the variations are even smaller. Moreover, ratios of coupling strengths in the various channels are practically unchanged.
Two points remain to be addressed:
(i)
: For the couplings of baryons to pseudoscalar mesons we have assumed in our calculations $SU(3)$ symmetry to be realized for pseudoscalar-type coupling, cf.Sect. \[sec:bornsp\]. Results are different when the same symmetry is assumed for couplings of pseudovector type, since (apart from the $NN\pi$ coupling) $BB'\mu$ couplings are then increased by a factor $(M_B+M_B')/2M_N$ leading to much stronger $\sigma$-coupling strengths in the strange baryon-baryon channels, see Table \[tab:5\_6\_4\].
(ii)
: We have assumed for the $F/(F+D)$ ratios $\alpha_p=\alpha_v^m=0.4$ predicted by the static quark model. If we change $\alpha_p$ by about $10\%$ (to 0.45) the resulting changes in the effective $\sigma$-coupling strengths are much less than $10\%$. The reason is that part of the Born amplitudes are increased while others are decreased, so that the total effect is quite small. The situation is completely different for $\alpha_v^m$, which determines the bare tensor couplings $f_{BB'\rho}^{(0)}$ in the $\rho$ pole graph. The same increase of $\alpha_v^m$ (to 0.45) leads to strong modifications in the $\rho$ coupling strengths, especially for the tensor-tensor part.
### Comparison with other models
In the $NN$ channel we can also determine the effective $\sigma$- and $\rho$-coupling strengths from the quasiempirical $N{\overline{N}}\to \pi\pi$ amplitudes [@Hoehler2; @Dumbrajs]. Corresponding spectral functions have already been discussed before. We now show in Fig. \[fig:5\_6\_2\] the results for the product of effective coupling strengths and form factors obtained from our microscopic model and the quasiempirical amplitudes [@Hoehler2; @Dumbrajs], in comparison to values used in the (full) Bonn potential [@MHE]. Since the quasiempirical amplitudes are available only up to $t'_{max}=50m_\pi^2$, a corresponding cutoff is used in the dispersion integral. If we use the same lower cutoff also for our model corresponding results essentially agree with those obtained for the quasiempirical amplitudes. Obviously the slightly stronger increase of the spectral function $\rho^\sigma_S$ of the microscopic model at small $t'$ roughly compensates for the larger maximum of the quasiempirical spectral function at $t'\approx 20m_\pi^2$. Furthermore we obtain a considerable reduction of strength in the $\sigma$-channel from inclusion of the repulsive contributions above $t'=50 m_\pi^2$ whereas in the $\rho$-channel such pieces have only a very small influence. Especially in the $\rho$-channel our results are considerably larger than the values used in the Bonn potential. The reason is the form factor with $\Lambda_{NN\rho}=1.4\,GeV$ used in the Bonn potential, which reduces the strength at $t=0$ by as much as $50\%$. One final point remains to be addressed: Our present results differ considerably (by up to $30\%$) from our former calculations based on a different microscopic model for the $N{\overline{N}}\to\pi\pi$ amplitudes. The reason for these discrepancies (in the spectral functions and effective coupling strengths) is that the subtraction of the uncorrelated terms from the discontinuities is model-dependent. Both models differ in the parametrization of $\Delta$ exchange (\[eq:Del\_exch\]); furthermore the model of Ref. [@Kim] does not include the $\rho$-pole term in the transition amplitude. Still both models provide a similar description of the quasiempirical data.
The average size of the effective coupling strengths is only a rough measure of the strength of correlated $\pi\pi$ and $K{\overline{K}}$ exchange in the various particle channels. The precise energy dependence of the correlated exchange as well as its relative strength in the different partial waves of the $s$-channel reaction is determined by the spectrum of the exchanged invariant masses, i.e. the spectral functions, leading to a different $t$-dependence of the effective coupling strengths.
Fig. \[fig:5\_6\_3\] shows the on-shell $NN$ potentials in spin-singlet states with angular momentum $L=0,2$, and 4, which are generated by the scalar-isoscalar part of correlated $\pi\pi$ and $K{\overline{K}}$ exchange. As expected it is attractive throughout. Slight differences occur between the potentials derived from the microscopic model for the $B{\overline{B'}}\to\mu{\overline{\mu}}$ amplitudes and those determined from the quasiempirical $N{\overline{N}}\to\pi\pi$ amplitudes, which can be traced to differences in the spectral function $\rho^\sigma_S(t)$ (cf. Fig. \[fig:5\_5\_4\]): For small $t'$ the microscopic input is larger; therefore the corresponding potential in high partial waves (which is dominated by the small-$t'$ behavior) is by about $20\%$ larger than the quasiempirical result. In the $^1S_0$ partial wave, on the other hand, medium and short ranged exchange processes characterized by larger $t'$-values contribute. In this region the microscopic amplitudes are considerably weaker; furthermore they contain the repulsive contributions above the $K{\overline{K}}$ threshold (cf. Fig. \[fig:5\_5\_1\]). Consequently the resulting potential is somewhat less attractive in the $^1S_0$ partial wave. In agreement with Ref. [@Kim] our present results (evaluated either from the microscopic model or the quasiempirical amplitudes) are stronger than $\sigma'$-exchange of the full Bonn potential. The difference is especially large in high partial waves since $\sigma'$-exchange, which corresponds to a spectral function proportional to $\delta(t'-m^2_\sigma)$, does not contain the long-range part of the correlated processes. Indeed if we parametrize our results derived from the microscopic model by $\sigma$-exchange as before (Sect. \[sec:3\_5\_1\]) but use for the effective coupling strength $G^\sigma_{NN\to NN}(t)$ the constant value at $t=0$ we obtain rough agreement with our unapproximated result in the $^1S_0$ partial wave but underestimate it considerably in high partial waves. Obviously the replacement of correlated $\pi\pi$ and $K{\overline{K}}$ exchanges by an exchange of a sharp mass $\sigma$ meson with $t$-independent coupling strength cannot provide a simultaneous description of low and high partial waves.
It is interesting to compare our results for the effective $\sigma$- and $\rho$- coupling strengths in baryon-baryon channels with non-vanishing strangeness with those used in the hyperon-nucleon interaction models of the Nijmegen [@NijII; @NijIII; @NijIV] and Jülich [@Holz] groups. These are (with the exception of the Jülich model B which will not be considered in the following) OBE models, i.e.$\sigma$ (and $\rho$) exchange effectively include uncorrelated processes involving the $\Delta$-isobar. Therefore we have to use for comparison dispersion-theoretic results in which only uncorrelated processes involving spin–1/2 particle intermediate states have been subtracted. Table \[tab:5\_6\_6\] shows the relative coupling strengths in the different baryon-baryon channels for various models, in the $\sigma$-channel. Apart from the Nijmegen model D [@NijII], in which the scalar $\epsilon$-meson is treated as an $SU(3)$ singlet and therefore couples with the same strength to all channels, the interaction is by far strongest in the $NN$ channel for all remaining models, and it becomes weaker with increasing strangeness. Obviously, the Nijmegen soft core model [@NijIV] is nearest to the dispersion-theoretic predictions. Table \[tab:5\_6\_6b\] shows the analogous results in the $\rho$-channel for the vector-vector ($VV$) and tensor-tensor ($TT$) components. There are sizable differences between the effective coupling constants from correlated exchange and the coupling constants of the OBE-models as well as among the OBE-models themselves. The latter models assume the vector coupling to the isospin current to be universal, which fixes its relative strength in the different particle channels (apart from form factors included in the Jülich model): In the $N\Sigma$ channel it is twice as large as in the $NN$ channel and vanishes for the transition $N\Lambda\to N\Sigma$. The correlated exchange result deviates strongly, which is another manifestation that $SU(3)$ symmetry does not hold for correlated exchanges, even if it is assumed for the bare $\rho BB'$ couplings present in our microscopic model.
Summary and Outlook
===================
An essential part of baryon-baryon interactions is the strong attraction of medium range, which in one-boson-exchange models is parametrized by an exchange of a fictitious scalar-isoscalar meson with a mass of about $500 MeV$. In extended meson exchange models this part is naturally generated by two-pion-exchange processes. Besides uncorrelated processes correlated terms have to be considered in which both pions interact during their exchange; in fact these terms provide the main contribution to the intermediate-range interaction.
In the scalar-isoscalar channel of the $\pi\pi$ interaction the coupling to the $K{\overline{K}}$ channel plays a strong role, which has to be explicitly included in any model meant to be realistic for energies near and above the $K{\overline{K}}$ threshold. As kaon exchange is an essential part of hyperon-nucleon interactions a simultaneous investigation of correlated $\pi\pi$ and $K{\overline{K}}$ exchanges is clearly suggested. In this work we have therefore derived the correlated $\pi\pi$ as well as $K{\overline{K}}$ exchange contributions in various baryon-baryon channels. Starting point of our calculations was a microscopic model for the transition amplitudes of the baryon-antibaryon system ($B{\overline{B'}}$) into two pseudoscalar mesons ($\pi\pi$, $K{\overline{K}}$) for energies below the $B{\overline{B'}}$ threshold. The correlations between the two mesons have been taken into account by means of $\pi\pi-K{\overline{K}}$ amplitudes (determined likewise fieldtheoretically [@Lohse; @Pearce_piN]), which provide an excellent reproduction of empirical $\pi\pi$ data up to $1.3 GeV$. With the help of unitarity and dispersion-theoretic methods we have then determined the baryon-baryon amplitudes for correlated $\pi\pi$ and $K{\overline{K}}$ exchange in the $J^P=0^+$ ($\sigma$) and $J^P=1^-$ ($\rho$) $t$-channel.
In the $\sigma$-channel the strength of correlated $\pi\pi$ and $K{\overline{K}}$ exchange decreases with the strangeness of the baryon-baryon channels becoming more negative. In the $NN$ channel the scalar-isoscalar part of correlated exchanges is by about a factor of 2 stronger than in both hyperon-nucleon channels ($\Lambda N$, $\Sigma N$) and by a factor 3 to 4 stronger than in the $S=-2$ channels ($\Lambda \Lambda$, $\Sigma\Sigma$, $N\Xi$). The influence of $K{\overline{K}}$ exchange is strong in baryon-baryon channels with non-vanishing strangeness while it is small in the $NN$ channel. This feature can be traced to different coupling constants and isospin factors and especially to the different masses involved in the various baryon-antibaryon channels.
The role of correlated $K{\overline{K}}$ exchange is small in the $\rho$-channel. Here the correlations are dominated by the (genuine) $\rho$-resonance in the $\pi\pi$ interaction. Among the various $B{\overline{B'}} - \pi\pi,
K{\overline{K}}$ Born amplitudes the direct coupling of the $\rho$-resonance to the baryons in the form of a $\rho$-pole graph provides the dominant contribution to correlated exchange.
It turns out that our results depend only slightly on the upper limit (cutoff) introduced in the dispersion integral. Some uncertainty results from applying $SU(3)$ resp. $SU(6)$ relations to either pseudoscalar or pseudovector $\pi$ and $K$ coupling constants. Note that the same problem occurs already in OBE-models of the hyperon-nucleon interaction. Ultimately it has to be decided by comparison with experiment which procedure is to be preferred. Moreover, if instead of $SU(6)$ symmetry $SU(3)$ symmetry is assumed only, the results for correlated exchanges depends on the $F/(F+D)$ ratios $\alpha_p$ and $\alpha_v^m$. While the dependence on $\alpha_p$ is only weak variation of$\alpha_v^m$ leads to noticeable changes in the model predictions for the correlated exchange in the $\rho$-channel. Also here a final decision about the correct choice of $\alpha_p$ and $\alpha_v^m$ can be made only by comparison with experiment. Again these parameters occur already in OBE hyperon-nucleon models. Therefore no new parameters are introduced when including correlated $\pi\pi$ and $K{\overline{K}}$ exchange in baryon-baryon interaction models. On the contrary, the elimination of single $\sigma$ and $\rho$ exchange reduces the number of free parameters and thus enhances the predictive power of corresponding interaction models.
Our results can be represented in terms of suitably defined effective coupling strengths. It turns out that the resulting values in the various baryon-baryon channels are not connected by $SU(3)$ relations. For example, although we have even assumed $SU(6)$ symmetry for the coupling strength of the bare $\rho$ to the baryon current sizable baryon exchange processes destroy this symmetry in the final effective couplings. Consequently the assumption of $SU(3)$ symmetry for single $\sigma$- and $\rho$-exchange is not supported by our findings.
With this model constructed in the present work it is now possible to take correlated $\pi\pi$ and $K{\overline{K}}$ exchange reliably into account in the various baryon-baryon channels. Especially in channels in which only little empirical information exists the elimination of phenomenological $\sigma$- and $\rho$-exchange considerably enhances the predictive power of baryon-baryon interaction models. Clearly the inclusion of correlated exchange in existing interaction models (e.g. the Bonn $NN$ potential [@MHE] and the Jülich $YN$ models [@Holz]) requires readjustment of free model parameters to the empirical data. Having fixed these parameters in the $NN$ and $YN$ channel the interaction model can then be extended parameter-free to other baryon-baryon channels with strangeness $S=-2$ using $SU(3)$ arguments for the genuine couplings. In this way in the frame of the Bonn-Jülich models the possibility arises for the first time to make sensible statements about the existence of bound baryon-baryon states with strangeness $S=-2$, which should be of some importance regarding the analysis of H-dibaryon experiments.
[**[Appendix]{}**]{}
Conventions {#sec:AnhA}
===========
As far as possible the conventions in this work are chosen in accordance with Ref. [@BjDr]. The helicity spinors $u(\vec p,\lambda)$ ($v(\vec p,\lambda)$) of a Dirac particle of spin 1/2 and mass $M$ are solutions of the free Dirac equation in momentum space for positive (negative) energy and helicity $\lambda$ $$\begin{aligned}
(\not{\!p} - M) u(\vec p,\lambda) &=& 0\quad,
\nonumber \\
(\not{\!p} + M) v(\vec p,\lambda) &=& 0\quad,\end{aligned}$$ with the 4-momentum $p^\mu$ ($p^0 \equiv E_p = + \sqrt{M^2 + \vec p\,^2}$) and $\not\!{ p} \equiv { p}^\mu \gamma_\mu$. With the phase convention of Ref. [@Hipp] the helicity spinors read ($\lambda=\pm 1/2$) $$\begin{aligned}
u(\vec p,\lambda) &=& \sqrt{\epsilon_p \over 2 M}
\left( \begin{array}{c}
1 \\
2\lambda |\vec p\,| / \epsilon_p
\end{array} \right)
{\> \vert \, \lambda \, \rangle} \quad ,
\nonumber \\
v(\vec p,\lambda) &=& \sqrt{\epsilon_p \over 2 M}
\left( \begin{array}{c}
- |\vec p\,| / \epsilon_p \\
2\lambda
\end{array} \right)
{\> \vert \, -\lambda \, \rangle} \quad,
\label{eq:app2a}\end{aligned}$$ where $\epsilon_p \equiv M + E_p$ They are normalized according to $$\begin{aligned}
{\overline{u}}(\vec p,\lambda) u(\vec p,\lambda') &=& \delta_{\lambda\lambda'}
\quad,
\nonumber \\
{\overline{v}}(\vec p,\lambda) v(\vec p,\lambda') &=&
-\delta_{\lambda\lambda'} \quad,\end{aligned}$$ with ${\overline{u}}(\vec p,\lambda) =
u^\dagger (\vec p,\lambda) \gamma_0$.
If $\vec p$ lies in the $xz$-plane and encloses the polar angle $\theta$ with the $\check{e}_z$-axis the eigenstates ${\> \vert \, \lambda \, \rangle} $ of the helicity operator, $${\vec\sigma \over 2} \cdot {\vec p \over |\vec p\,| } {\> \vert \, \lambda \, \rangle}
=
\lambda {\> \vert \, \lambda \, \rangle}\quad,$$ are related to the Pauli spinors $\chi_m$ by $${\> \vert \, \lambda \, \rangle} =
\exp \left( -{i\over2} \sigma_2 \theta \right) \chi_\lambda
\quad.$$ where $$\vec \sigma\cdot \check{e}_z \; \chi_{\pm{1\over2}}
= \pm \chi_{\pm{1\over2}}
\quad.$$
The same phase convention as in Ref. [@Hipp] is adopted for the helicity spinors \[eq:app2a\]; namely, the particle and antiparticle spinors are related by the charge conjugation $\cal C$ [@BjDr]: $$v(\vec p,\lambda)= {\cal C} {\overline{u}}^T (\vec p,\lambda)
=i \gamma_2 {u}^* (\vec p,\lambda)\quad.$$
Helicity eigenstates of two-particle systems in the center-of-mass frame are built as a product of two helicity spinors according to the phase convention of Jacob and Wick [@JW]. In case of two spin-1/2 particles 1 and 2, the helicity spinors \[eq:app2a\] are used for particle 1 (momentum $\vec p\,$) and for particle 2 (momentum $-\vec p\,$) the following spinors are used: $$\begin{aligned}
u(-\vec p,\lambda_2) &=& \sqrt{\epsilon_p \over 2 M}
\left( \begin{array}{c}
1 \\
2\lambda_2 |\vec p\,| / \epsilon_p
\end{array} \right)
{\> \vert \, \lambda_2 \, \rangle} \quad ,
\\
v(-\vec p,\lambda_2) &=& \sqrt{\epsilon_p \over 2 M}
\left( \begin{array}{c}
|\vec p\,| / \epsilon_p \\
-2\lambda_2
\end{array} \right)
{\> \vert \, -\lambda_2 \, \rangle} \quad,
\label{eq:app3a}\end{aligned}$$ where $${\> \vert \, \lambda_2 \, \rangle} = \exp \left( -{i\over2} \sigma_2 \theta \right)
\chi_{-\lambda_2}\quad.$$
The hadronic field operators can be expanded in momentum space solutions of the corresponding equation of motion. For Dirac particles ($J^P={{1\over 2}}^+$) and (pseudo)scalar particles ($J^P=0^\pm$), see Ref. [@BjDr]. For spin-1 and spin-3/2 particles the field operators $\phi^\mu$ and $\psi^\mu$, respectively, then read: $$\phi^\mu(x)={1\over (2\pi)^{3/2}}
\sum_\lambda \int d^3k
{1 \over \sqrt {2 \omega_k}}
\epsilon^\mu(\vec k,\lambda)
\left[
a(\vec k,\lambda) e^{-i k \cdot x}
+
a^\dagger (\vec k,\lambda) e^{i k \cdot x}
\right]$$ $$\psi^\mu(x)={1\over (2\pi)^{3/2}}
\sum_\Lambda \int d^3p
\sqrt{M \over E_p}
\left[
b(\vec p,\Lambda) u^\mu(\vec p,\Lambda) e^{-i p \cdot x}
+
d^\dagger (\vec p,\Lambda) v^\mu(\vec p,\Lambda) e^{i p \cdot x}
\right]$$ with the polarization vector $\epsilon^\mu(\vec k,\lambda)$, cf. [@Lurie]. Note that the hermitian conjugated component of the spherical operator $a$ is related to the component of the hermitian operator $a^\dagger$ by [@Edmonds]: $$a(\vec k,\lambda)^\dagger = (-1)^\lambda a^\dagger (\vec k,-
\lambda)\quad.
\label{eq:app4a}$$ The Rarita-Schwinger spinors [@Lurie] $ u^\mu(\vec p,\Lambda)$ and $v^\mu(\vec p,\Lambda)$ are solutions of the Rarita-Schwinger equation $$\begin{array}{l@{\qquad}rcl@{\qquad\qquad}rcl}
&(\not{\!p} - M) u^\mu(\vec p,\Lambda) &=& 0,&
(\not{\!p} + M) v^\mu(\vec p,\Lambda) &=& 0 \\
\mbox{and} &&&&&& \\
&\gamma_\mu u^\mu(\vec p,\Lambda) &=& 0,&
\gamma_\mu v^\mu(\vec p,\Lambda) &=& 0.
\end{array}
\label{eq:AnhARS1}$$ The creation and destruction operators of bosons (fermions) follow the usual (anti-)commutator relations.
For spin-1 and spin-3/2 particles [@Read] the Feynman propagators $S_v(k)$ and $S_D^{\mu\nu}(p;A)$, which are derived from the time-ordered product of the corresponding field operators [@BjDr], read in momentum space: $$S_v(k) = {-g^{\mu\nu} + k^\mu k^\nu/m^2 \over k^2 - m^2 +
i\epsilon}
\label{eq:rhoprop}$$ $$\begin{aligned}
S_D^{\mu\nu}(p;A)& =& {\not\! p + M \over p^2 - M^2 + i\epsilon}
\left[
-g^{\mu\nu}
+ {\gamma^\mu \gamma^\nu \over 3}
+{2\over 3 M^2} p^\mu p^\nu
-{p^\mu \gamma^\nu - p^\nu \gamma^\mu \over 3M}
\right]
\nonumber \\ [\baselineskip]
&&+ S^{\mu\nu}_{\rm non-pole} (p;A)
\label{eq:AnhARS}\end{aligned}$$ with the non-pole part $$S^{\mu\nu}_{\rm non-pole} (p;A) =
{1 \over 3 M^2} {A+1 \over 2A+1}
\left[
\gamma^\mu {{(A+1)\over2} \not\!p + A M \over 2A+1} \gamma^\nu
-(p^\mu \gamma^\nu + p^\nu \gamma^\mu )
\right]\quad.$$ Without restricting the general validity of the results the parameter $A$ is set in our work to $A=-1$ so that the non-pole part vanishes (see Sect. \[sec:bornsp\]).
Matrix elements of $B \overline{B'} \to \mu\overline{\mu}$ Born amplitudes {#sec:AnhC}
===========================================================================
In order to evaluate the helicity amplitudes \[eq:born\_heli\] of the $B \overline{B'} \to \mu{\overline{\mu}}$ $
(\mu{\overline{\mu}}=\pi\pi,\;K{\overline{K}})$ Born terms \[eq:4born32\]-\[eq:4born33\] the coordinate system is conveniently chosen such that the relative momentum $\vec q$ of the $B \overline{B'}$ state points along the $z$-axis and the relative momentum $\vec k$ of the two pseudoscalar mesons $\mu\overline{\mu}$ lies in the $xz$-plane; i.e., the components of the two momenta read $$\vec q=\left(
\begin{array}{c}
0 \\ 0 \\ q
\end{array}
\right)\quad,
\qquad
\vec k=\left(
\begin{array}{c}
k \sin\vartheta \\ 0 \\ k \cos\vartheta
\end{array}
\right)
\quad,$$ with the scattering angle $\vartheta={\;\hbox{$\rlap{)}\!\!\!<$}\,}(\vec p,\vec k\,)$.
As explained in Sect. \[sec:kap4\] the $B \overline{B'} \to
\mu\overline{\mu}$ amplitudes need to be evaluated only for $B\overline{B'}$ states being on their mass-shell: $$\sqrt{t} = E_B +E_{B'} = \sqrt{M_B+q^2} + \sqrt{M_{B'}+q^2}$$ Therefore, in the following, $t$ is suppressed as an argument of the helicity amplitudes.
In case of the $B \overline{B'} \to \pi\pi$ amplitudes one has in principle to take into account also the exchange graph (with the external pion lines exchanged) arising from the symmetrized $\pi\pi$ states. However, this can be simply done by considering the selection rule for the $\pi\pi$ states: $ (-1)^{J+I}=+1$ ($J$: total angular momentum, $I$: total isospin); i.e., multiplying the direct, partial wave decomposed $B \overline{B'} \to \pi\pi$ amplitude by a factor $(1+(-1)^{J+I})$.
In the final results for the $B \overline{B'} \to
\mu\overline{\mu}$ Born amplitudes given below we introduced the following abbreviations: $$\begin{array}{rcl@{\qquad}rcl}
\alpha_+ & \equiv& \epsilon_B + \epsilon_{B'}\;,&
\alpha_- & \equiv & \epsilon_B - \epsilon_{B'}\;, \\
\beta_+ & \equiv & \epsilon_B \epsilon_{B'} + q^2\;, &
\beta_- & \equiv & \epsilon_B \epsilon_{B'} - q^2\;, \\
\end{array}$$ with $\epsilon_B\equiv M_B + E_B$.
Exchange of a $J^P= {1\over2}^+$ baryon
----------------------------------------
$$\begin{aligned}
{\langle \, \mu\overline{\mu},\vec k \, \vert \>} V_X {\> \vert \, B \overline{{B'}},\vec q ++ \, \rangle} &=&
C \{
-q [ M_X (t + 4 k^2) \alpha_+
- q_0 (t - 4 k^2) \alpha_-
+ (t + 4 k^2) \beta_- ]
\nonumber \\
&& +k\cos\theta [ 4 M_X \sqrt{t} \beta_+
+ (t - 4 k^2) \beta_- ]
\nonumber \\
&& +8 k^2 q\cos^2\theta \beta_- \}
\quad,\end{aligned}$$
$${\langle \, \mu\overline{\mu},\vec k \, \vert \>} V_X {\> \vert \, B \overline{{B'}},\vec q +- \, \rangle}=
-C k \sin\theta
\{ - 4 q^2 \sqrt{t} \alpha_+
+ (t - 4 k^2) \beta_+
+ 4 M_X \sqrt{t} \beta_-
+ 8 k q \cos\theta \beta_+ \}
\; ,$$
where $$C:=
- {f_{BX\mu} f_{{B'}X\mu}\over 8 m_\mu^2
\sqrt{\epsilon_B \epsilon_{B'} M_B M_{B'} }}
{F^2_{X}(p^2)
\over q_0^2 - E_X^2} \quad,$$ with $E_X=\sqrt{\vec p\,^2 + M_X^2}$ with $\vec p=\vec q -\vec k$.
Exchange of a $J^P={3\over2}^+$ baryon
---------------------------------------
$$\begin{array}{rrl}
\multispan{3}{
$ {\langle \, \mu\overline{\mu},\vec k \, \vert \>} V_X {\> \vert \, B \overline{{B'}},\vec q ++ \, \rangle}=
$\hfill
}
\\
C \Biggl\{ {1 \over q_0^2 - E_X^2 }
\{ & q & [
M_X A(t,k) \alpha_+
+ q_0 A(t,k) \alpha_-
+ 2 M_X k^2 \sqrt{t} \beta_+
+ A(t,k) \beta_- ]
\\
&+ k\cos\theta & [
8 M_X q^2 k^2 \alpha_+
+ 8 q_0 q^2 k^2 \alpha_-
\\ &&
+ 2 M_X \sqrt{t} (q_0^2 - M_X^2 - q^2 - k^2)
\beta_+
+ ( 8 k^2 q^2 - A(t,k) ) \beta_- ]
\\
&+ 2 k^2 q \cos^2\theta & [
- 2 M_X q^2 \alpha_+
- 2 q_0 q^2 \alpha_-
+ M_X \sqrt{t} \beta_+
- 2 ( q^2 + 2 k^2) \beta_- ]
\\
& + 4 k^3 q^2\cos^3\theta & \beta_- \}
\\
+x_\Delta \{ & -q & [
M_X (t + 4 k^2) \alpha_+
- 2 q_0 t \alpha_- ]
\\
&+ 4 k \cos\theta & [
M_X \sqrt{t} \beta_+
- 2 k^2 \beta_- ]
\\
&+ 8 k^2 q \cos^2\theta &
\beta_- \}
\\
+x_\Delta^2 \{ & -q & [
2 M_X (t + 4 k^2) \alpha_+
- q_0 (t - 4 k^2) \alpha_-
+ (t + 4 k^2) \beta_- ]
\\
&+ k\cos\theta & [
+ 8 M_X \sqrt{t} \beta_+
+ (t - 4 k^2) \beta_- ]
\\
&+ 8 k^2 q \cos^2\theta &
\beta_- \}
\Biggr\}
\end{array}$$
$$\begin{array}{rrl}
\multispan{3}{
${\langle \, \mu\overline{\mu},\vec k \, \vert \>} V_X {\> \vert \, B \overline{{B'}},\vec q +- \, \rangle}=$
\hfill
}
\\
C k \sin\theta
\times
\Biggl\{ {1 \over q_0^2 - E_X^2 } \{ & -& [
2 M_X^2 \sqrt{t} q^2 \alpha_+
+ 2 M_X q_0 q^2 \sqrt{t} \alpha_-
\\&&
- A(t,k) \beta_+
+ 2 M_X \sqrt{t} (q_0^2 - M_X^2 - k^2)
\beta_- ]
\\
&+ 2 k q \cos\theta & [
4 k^2 \beta_+
- M_X \sqrt{t} \beta_- ]
\\
&- 4 k^2 q^2 \cos^2\theta &
\beta_+
\}
\\
+4 x_\Delta \{ &&
2 k^2 \beta_+
- M_X \sqrt{t} \beta_-
\\
& - 2 k q \cos\theta &
\beta_+
\}
\\
+x_\Delta^2 \{ & - & [
- 4 q^2 \sqrt{t} \alpha_+
+ (t - 4 k^2) \beta_+
+ 8 M_X \sqrt{t} \beta_- ]
\\
& - 8 k q \cos\theta &
\beta_+ \}
\Biggr\} \quad,
\end{array}$$
with $$C:=
- {f_{BX\mu} f_{{B'}X\mu}\over 12 m_\mu^2 M_X^2
\sqrt{\epsilon_B \epsilon_{B'} M_B M_{B'} }}
F^2_{X}(p^2) \quad,$$ $$A(t,k) := q_0^2 t - 4 k^2 M_X^2 - M_X^2 t - 4 k^4 \quad.$$
$\rho$-pole graph {#sec:AnhC_rho}
------------------
For the $\rho$-pole graph the $\rho$-propagator and the form factor $F_{\mu\mu\rho}(k^2)$ (cf. Eq. \[eq:rhoFF\]) do not depend on the scattering angle $\theta$. Therefore, the partial wave decomposition can easily be performed analytically. The results are: $$\begin{aligned}
{\langle \, \mu\overline{\mu}, k \, \vert \>} V^J_\rho {\> \vert \, B \overline{{B'}}, q ++ \, \rangle}
&=& \delta_{J1}
C_\rho k
[g^{(0)}_{B{B'}\rho} \beta_-
+ {f^{(0)}_{B{B'}\rho} \over 2 M_N } \sqrt{t}
\beta_+ ],
\\
{\langle \, \mu\overline{\mu}, k \, \vert \>} V^J_\rho {\> \vert \, B \overline{{B'}}, q +- \, \rangle}&=&
\delta_{J1}
\sqrt{2} C_\rho k
[g^{(0)}_{B{B'}\rho} \beta_+
+ {f^{(0)}_{B{B'}\rho} \over 2 M_N } \sqrt{t}
\beta_-],\end{aligned}$$ where $$C_\rho:=
{ 4\pi \over 3}
{ g^{(0)}_{\mu\mu\rho} \over \sqrt{\epsilon_B \epsilon_{B'} M_B M_{B'}
}}
{F_{\mu\mu\rho}(k^2) \over t - (m^{(0)}_\rho)^2}
\quad.$$
$SU(3)$ relations for coupling constants {#sec:borniso}
========================================
The microscopic model for the $B{\overline{B'}}\to \pi\pi, K{\overline{K}}$ amplitudes presented in Sect. \[sec:kap4\] imposes $SU(3)$ symmetry to the coupling constants at the hadronic vertices in order to keep the number of free parameters as low as possible. The $SU(3)$ relations for the coupling constants are derived by constructing an $SU(3)$-invariant interaction Lagrangian from particle field operators, which possess a well-defined behavior under $SU(3)$ transformations. This method is well-established [@deSw] and makes use of the $SU(3)$ Clebsch-Gordan coefficients and the so-called isoscalar factors which are tabulated in Ref. [@deSw]. Therefore, we give in the following only the final results relevant for this work.
The coupling of the pseudoscalar meson octet ($\pi$, $\eta_8$, $K$, ${\overline{K}}$) to the current of the $J^P=1/2^+$ baryon octet ($N$, $\Lambda$, $\Sigma$, $\Xi$) is described by the Lagrangian [@deSw]
$$\begin{array}{rcl@{\!\!\quad}cl}
{\cal L}_{B'Bp}&=\;\;\;& g_{NN\pi} (N^\dagger \vec\tau N) \cdot
\vec\pi & +& g_{NN\eta_8} (N^\dagger N) \eta_8 \\ &+&g_{\Lambda NK}
\left[ (N^\dagger K) \Lambda + \Lambda^\dagger(K^\dagger N) \right] &
+&g_{\Sigma NK} \left[ (N^\dagger\vec\tau K)\cdot\vec\Sigma +
\vec\Sigma^\dagger\cdot(K^\dagger\vec\tau N) \right] \\
&+&g_{\Sigma\Lambda\pi} \left[\vec\Sigma^\dagger\cdot\vec\pi \Lambda
+\Lambda^\dagger \vec\Sigma\cdot\vec\pi\right] &
-&ig_{\Sigma\Sigma\pi}\left(\vec\Sigma^\dagger\times\vec\Sigma\right)
\cdot\vec\pi
\\ &+&g_{\Lambda\Lambda\eta_8}\Lambda^\dagger\Lambda\eta_8 &
+&g_{\Sigma\Sigma\eta_8} \vec\Sigma^\dagger\cdot\vec\Sigma \eta_8 \\
&+&g_{\Xi\Lambda K}\left[ \Lambda^\dagger({\overline{K}}^\dagger \Xi) +
(\Xi^\dagger {\overline{K}}) \Lambda\right] & +&g_{\Xi\Sigma K} \left[
\vec\Sigma^\dagger\cdot({\overline{K}}^\dagger\vec\tau\, \Xi) +
(\Xi^\dagger\vec\tau {\overline{K}})\cdot\vec\Sigma
\right]
\\ &+&g_{\Xi\Xi\pi} (\Xi^\dagger \vec\tau \, \Xi) \cdot \vec\pi &
+&g_{\Xi\Xi\eta_8} (\Xi^\dagger \Xi) \eta_8\quad,
\end{array}
\label{eq:su3_11a}$$
where the coupling constants being of the pion and the kaon are given by
$$\begin{array}{l@{\qquad}l}
g_{NN\pi}=g\quad, & g_{\Lambda NK}=-{1\over\sqrt 3}g (1+2\alpha)\quad,\\
g_{\Sigma NK}= g (1-2\alpha)\quad,& g_{\Sigma\Lambda\pi}={2\over\sqrt 3}
g(1-\alpha)\quad,\\
g_{\Sigma\Sigma\pi}=2g \alpha\quad,g_{\Xi\Lambda K}={1\over\sqrt 3} g
(4\alpha-1)\quad,\\
g_{\Xi\Sigma K}=-g \quad,g_{\Xi\Xi\pi}= -g (1-2\alpha)\quad.\\
\end{array}
\label{eq:su3_11b}$$
Hence, the coupling of $\pi$ and $K$ to the baryon octet is determined by two parameters: the coupling strength $g_8$ and the so-called $F/(F+D)$-ratio $\alpha_p$.
The transition from pseudoscalar to vector mesons can be simply made by the replacement $$\pi\to\rho ,\quad K\to K^* ,\quad \eta\to\phi ,\quad \eta'\to\omega.
\nonumber$$ However, as can be seen from the spin-momentum part of the interaction Lagrangian Eq. \[eq:BBvcoup2\] the vector mesons couple to the baryon octet in two different ways: via the vector and the tensor coupling with coupling constants $g$ and $f$, respectively. Now, it is not clear which coupling constants underly $SU(3)$ symmetry relations. Besides the canonic assumption [@NijII; @NijIV] that the $SU(3)$ relations apply to $g$ and $f$, sometimes [@NijIII; @Holz] the electric ($g$) and the magnetic ($G=g+f$) coupling constants are subject to $SU(3)$. In fact, both assumptions are in a certain sense equivalent [@Reu]. For given $SU(3)$ parameters $(g_1,g_8,\alpha_v^g)$ and $(f_1,f_8,\alpha_v^f)$ the parameters $(G_1,G_8,\alpha_v^m)$ of the magnetic coupling can be chosen such that both coupling schemes lead to the same tensor couplings $f=G-g$.
Extending the $SU(3)$ symmetry to $SU(6)$, predictions for the $F/(F+D)$-ratios of the different multiplet couplings can be derived [@Pais]: $$\alpha_p=0.4 ,\qquad
\alpha_v^e=1 ,\qquad
\alpha_v^m=0.4\quad.
\label{eq:alpha_su6}$$ $\alpha_v^e=1$ corresponds to the usual assumption of universal electric coupling of the $\rho$ meson to the isospin current [@Sakurai], demanding for instance the equality $g_{NN\rho}$ and $g_{KK\rho}$.
The coupling of the pseudoscalar meson octet to the current for the transition between the $J^P=3/2^+$ baryon decuplet ($\Delta$, $Y^*$, $\Xi^*$, $\Omega$) and the $J^P=1/2^+$ baryon octet is given by the Lagrangian [@ReuDiss]:
$$\begin{array}{rcl@{\quad}cl@{\quad}cl}
{\cal L}_{DBp}& =\;& f_{N\Delta\pi} (N^\dagger \vec T
\Delta)\cdot\vec\pi &&&& \\ &+& f_{NY^*K}(N^\dagger\vec\tau K)\cdot
\vec Y^* & +& f_{\Sigma\Delta K}
\vec\Sigma^\dagger\cdot(K^\dagger\vec T\Delta)&& \\ &+& f_{\Lambda
Y^*\pi} \Lambda^\dagger\vec Y^*\cdot\vec\pi & -& i f_{\Sigma Y^*\pi}
\left(\vec\Sigma^\dagger\times\vec Y^*\right)\cdot\vec\pi& +&
f_{\Sigma Y^*\eta_8} \vec\Sigma^\dagger\cdot\vec Y^* \eta_8 \\ &+&
f_{\Lambda \Xi^*K} \Lambda^\dagger({\overline{K}}^\dagger \Xi^*) & +&
f_{\Sigma \Xi^*K} \vec\Sigma^\dagger\cdot({\overline{K}}^\dagger\vec\tau
\Xi^*) & +& f_{\Xi Y^*K} (\Xi^\dagger\vec\tau {\overline{K}})\cdot\vec Y^*
\\ &+& f_{\Xi\Xi^*\pi} (\Xi^\dagger \vec\tau \Xi^*) \cdot \vec\pi &
+& f_{\Xi\Xi^*\eta_8} (\Xi^\dagger \Xi^*) \eta_8&& \\ &+&
f_{\Xi\Omega K} (\Xi^\dagger K) \Omega &&&& \\ &+&\quad h.c.&&&&
\end{array}
\label{eq:su3_22}$$
with the $\pi$ and $K$ coupling constants
$$\begin{array}{l@{\qquad}l@{\qquad}l}
f_{N\Delta\pi}=f\quad, && \\ f_{NY^*K}=-f/\sqrt{6}\quad, &
f_{\Sigma\Delta K}=-f\quad, & \\
f_{\Lambda Y^*\pi}=f/\sqrt{2} \quad,
& f_{\Sigma Y^*\pi}=-f/\sqrt{6} \quad,
& f_{\Lambda \Xi^*K}=f/\sqrt{2}\quad, \\
f_{\Sigma \Xi^*K}=f/\sqrt{6} \quad, & f_{\Xi Y^*K}=-f/\sqrt{6} \quad,
& f_{\Xi\Xi^*\pi}=-f/\sqrt{6}\quad, \\
f_{\Xi\Omega K}=f\quad. &&
\end{array}$$
Obviously, the coupling constants now depend only on one parameter, the coupling strength $f$.
The $SU(3)$ part of the transition amplitudes is obtained by evaluating the matrix elements of the appropriate products of the $SU(3)$-Lagrangians \[eq:su3\_11a\] and \[eq:su3\_22\]. If the transition amplitudes are evaluated between states of definite total isospin the results for the $SU(3)$ part can be separated into the product of the corresponding coupling constants and the so-called isospin factor. For instance, the isospin factor $T_{B_1{\overline{B}}_2\to \mu\mu'}^X (I)$ for the transition of a baryon-antibaryon state $B_1{\overline{B}}_2$ (with $B_1,B_2 \in \{ 8_B\}$) to a pseudoscalar meson $\mu$ and its antiparticle ${\overline{\mu}}$ follows from $$\begin{aligned}
\lefteqn{
T_{B_1{\overline{B}}_2\to \mu\mu'}^X (I) g_{B_1X\mu} g_{B_2X\mu'} \equiv
{\langle \, \mu{\overline{\mu}},Im \, \vert \>} \left( {\cal LL'} \right)_X {\> \vert \, B_1{\overline{B}}_2,
Im \, \rangle}
} \nonumber \\ && =\sum_{m_{B_1},m_{B_2} \atop m_\mu,m_{{\overline{\mu}}}}
{\langle \, I_{B_1} I_{\bar B_2} m_{B_1} m_{\bar B_2} \, \vert \>}{Im \,\rangle}
{\langle \, I_\mu I_{ {\overline{\mu}}} m_\mu m_{{\overline{\mu}}} \, \vert \>}{Im \,\rangle}
\nonumber \\
&&\qquad
{\langle \, \mu I_\mu m_\mu , {\overline{\mu}} I_{ {\overline{\mu}}} m_{{\overline{\mu}}} \, \vert \>}
\left( {\cal LL'} \right)_X
{\> \vert \, B_1 I_{B_1} m_{B_1}, {\overline{B}}_2 I_{\bar B_2} m_{\bar B_2} \, \rangle}\quad,
\label{eq:su3_22a}\end{aligned}$$ with $m$ denoting the $z$-component of the total isospin $I$ and a corresponding notation for the particle isospins $I_{B_1},I_{\bar B_2}, I_\mu,I_{\mu'}$. The index $X$ in $\left( {\cal LL'}\right)_X $ indicates that the two Lagrangians ${\cal L}$ and ${\cal L'}$ are coupled by contracting the field operators of the exchanged baryon isomultiplet $X$ (or of the $\rho$-meson in case of the $\rho$-pole diagram).
The isospin factors obtained in this way for the processes included in the Born terms of the $B{\overline{B'}}\to\mu{\overline{\mu}}$ model (cf.Fig. \[fig:4\_0b\]) are listed in Tab. \[tab:4\_1\]. Note that due to the phase convention \[eq:SU3phase\] introduced in connection with the isospin-crossing matrix in Section \[sec:kap3\_6\] some isospin factors (e.g. those for $N{\overline{N}}\to\pi\pi$) differ in sign from the ones usually used in the literature [@Kim; @Durso].
[100]{}
M. Ieiri, in “Properties and Interactions of Hyperons”, Proceedings of the U.S.-Japan Seminar (Maui, Hawaii, Oktober 1993), Eds.: B.F. Gibson, P.D. Barnes and K. Nakai (World Scientific, Singapore, 1994) S. 17.
K. Imai, in “Properties and Interactions of Hyperons”, Proceedings of the U.S.-Japan Seminar (Maui, Hawaii, Oktober 1993), Eds.: B.F. Gibson, P.D. Barnes and K. Nakai (World Scientific, Singapore, 1994) S. 197.
B. Holzenkamp, K. Holinde and J. Speth, 500(1989)485.
M.M. Nagels, T.A. Rijken and J.J. de Swart, 15(1977)2547.
M.M. Nagels, T.A. Rijken and J.J. de Swart, 20(1979)1633.
P.M.M. Maessen, T.A. Rijken and J.J. de Swart, 40(1989)2226.
J.J. de Swart, P.M.M. Maessen and T.A. Rijken, in “Properties and Interactions of Hyperons”, Proceedings of the U.S.-Japan Seminar (Maui, Hawaii, Oktober 1993), Eds.: B.F. Gibson, P.D. Barnes and K. Nakai (World Scientific, Singapore, 1994) S. 37.
R.L. Jaffe, 15(1977)267; 15(1977)281.
G. Janssen, B.C. Pearce, K. Holinde, and J. Speth, 52(1995)2690.
R. Machleidt, K. Holinde and Ch. Elster, 149(1987)1.
R.L. Jaffe, 38(1977)195; 38(1977)617(E).
H.-C. Kim, J.W. Durso and K. Holinde, 49(1994)2355.
D. Lohse, J.W. Durso, K. Holinde and J. Speth, 516(1990)513.
J.W. Durso, A.D. Jackson and B.J. Verwest, 345(1980)471.
D. Amati, E. Leader and B. Vitale, 17(1960)68.
W. Lin and B.D. Serot, 512(1990)637.
C. Schütz, K. Holinde, J. Speth, B.C. Pearce and J.W. Durso, 51(1995)1374.
G. Höhler, F. Kaiser, R. Koch and E. Pietarinen, "\`Handbook of Pion-Nucleon Scattering"4, Physics Data 12-1, Fachinformationszentrum, Karlsruhe, 1979.
O. Dumbrajs et al., “Compilation of Coupling Constants and Low-Energy Parameters”, 216(1983)277.
M. Jacob and G.C. Wick, 7(1959)404.
M.D. Scadron and H.F. Jones, 173(1968)1734.
N.G. van Kampen, 89(1953)1072.
G.E. Brown and A.D. Jackson, “The Nucleon-Nucleon Interaction”, North-Holland, Amsterdam, 1976.
M.L. Goldberger and K.M. Watson, “Collision Theory”, Wiley, New York, 1964.
B. Holzenkamp, Spezielle Berichte der Kernforschungsanlage Jülich, Nr. 315 (1985).
A.D. Martin and T.D. Spearman, “Elementary Particle Theory”, North-Holland, Amsterdam, 1970.
D.E. Neville, 160(1967)1375.
G. Höhler and E. Pietarinen, 95(1975)210.
H. Nielsen, J.L. Petersen and E. Pietarinen, 22(1970)525.
H. Nielsen and G.C. Oades, 49(1972)586.
R. Blankenbecler and R. Sugar, 142(1966)1051.
B.J. Read, 52(1973)565.
L.M. Nath, B. Etemadi and J.D. Kimel, 3(1971)2153.
J.D. Bjorken and S.D. Drell, “Relativistische Quantenfeldtheorie”, Bibliographisches Institut, Mannheim, 1984.
D.O. Riska, 40(1972)177; 56(1973)445.
T. Hippchen, Spezielle Berichte der Kernforschungsanlage Jülich, Nr. 494 (1989).
D. Lurie, “Particles and Fields”, Interscience Publishers, New York, 1968.
A.R. Edmonds, “Angular Momentum in Quantum Mechanics”, Princeton University Press, Princeton, 1957.
J.J. de Swart, 35(1963)916.
A. Reuber, K. Holinde and J. Speth, 570(1994)543.
A. Pais, 38(1966)215.
J.J. Sakurai, 11(1960)1.
A. Reuber, Berichte des Forschungszentrums Jülich, Nr. 3117 (1995).
$$\begin{array}{|cc|cc||cc|c||c|c|}
\hline
A&B&C&D&I_s&I_t&X(I_s,I_t)& F_\sigma & F_\rho \\
\hline\hline
N &N &N &N &0,1 &0,1 &
\left(
\begin{array}{cc}
1/2 & -3/2 \\
1/2 & 1/2
\end{array} \right)&
1/2&1/2\\
\hline
N &\Lambda&N &\Lambda&1/2 &0 & -1/\sqrt{2}
&-1/\sqrt{2}&\mbox{---}\\
\hline
N &\Sigma &N &\Sigma &1/2,3/2 &0,1 &
\left(
\begin{array}{cc}
1/\sqrt{6} & -1 \\
1/\sqrt{6} & 1/2
\end{array} \right)
&1/\sqrt{6}&1/2 \\
\hline
N &\Sigma&N &\Lambda&1/2 &1 & -\sqrt{3/2}
&\mbox{---}&-1/\sqrt{2}\\
\hline
\Lambda &\Lambda&\Lambda&\Lambda&0 &0 & 1
&1&\mbox{---}\\
\hline
\Sigma &\Sigma &\Sigma &\Sigma &0,1,2 &0,1,2 &
\left(
\begin{array}{ccc}
1/3 &-1 &5/3 \\
1/3 &-1/2 &-5/6 \\
1/3 &+1/2 &1/6 \\
\end{array} \right)
&1/3&1/2\\
\hline
N &\Xi &N &\Xi &0,1 &0,1 &
\left(
\begin{array}{cc}
-1/2 & 3/2 \\
-1/2 & -1/2
\end{array} \right)
&-1/2&-1/2\\
\hline
\end{array}$$
particle mass ($MeV$) $B$ $S$ $J^P$ $I$
------------------------- -------------- ----- ------ --------------- -------------
$\pi$ 139.57 0 $0$ $0^-$ $1$
$K$ 495.82 0 $+1$ $0^-$ ${1\over2}$
$\rho$ 770. 0 $0$ $1^-$ $1$
$N$ 938.919 1 $0$ ${1\over2}^+$ ${1\over2}$
$\Lambda$ 1115.68 1 $-1$ ${1\over2}^+$ $0$
$\Sigma$ 1193.1 1 $-1$ ${1\over2}^+$ $1$
$\Xi$ 1318.1 1 $-2$ ${1\over2}^+$ ${1\over2}$
$\Delta$ 1232. 1 $0$ ${3\over2}^+$ ${3\over2}$
$Y^*\equiv\Sigma(1385)$ 1385. 1 $-1$ ${3\over2}^+$ $ 1$
$\Xi^*$ 1533.4 1 $-2$ ${3\over2}^+$ ${1\over2}$
$\Omega^-$ 1672.5 1 $-3$ ${3\over2}^+$ $0$
: Particles considered in the model for the $B{\overline{B'}}\to\mu{\overline{\mu}}$ Born amplitudes. Given are their masses and the relevant quantum numbers ($B$: baryon number, $S$: strangeness, $J$: spin, $P$: parity, $I$: isospin).[]{data-label="tab:4_0"}
$B{\overline{B'}}\to\mu{\overline{\mu}}$ process $T(0)$ $T(1)$ $T(2)$
-------------------------------------------------- --------------- --------------- ------------- --------
$N{\overline{N}}\to\pi\pi$ $N$ $+\sqrt{6}$ $+2$
$\Delta$ $+\sqrt{8/3}$ $-2/3$
$\rho$ $-2$
$N{\overline{N}}\to K{\overline{K}}$ $\Lambda$ $+1$ $+1$
$\Sigma, Y^*$ $+3$ $-1$
$\rho$ $-2$
$\Lambda{\overline{\Lambda}}\to\pi\pi$ $\Sigma, Y^*$ $-\sqrt{3}$
$\Lambda{\overline{\Lambda}}\to K{\overline{K}}$ $N$ $-\sqrt{2}$
$\Xi, \Xi^*$ $-\sqrt{2}$
$\Sigma{\overline{\Sigma}}\to\pi\pi$ $\Lambda$ $+1$ $+1$ $+1$
$\Sigma, Y^*$ $+2$ $+1$ $-1$
$\rho$ $-2$
$\Sigma{\overline{\Sigma}}\to K{\overline{K}}$ $N$ $+\sqrt{6}$ $-2$
$\Delta$ $+\sqrt{8/3}$ $+2/3$
$\Xi, \Xi^*$ $+\sqrt{6}$ $+2$
$\rho$ $-2$
$\Lambda{\overline{\Sigma}}\to\pi\pi$ $\Sigma, Y^*$ $-\sqrt{2}$
$\rho$ $+\sqrt{2}$
$\Lambda{\overline{\Sigma}}\to K{\overline{K}}$ $N$ $-\sqrt{2}$
$\Xi, \Xi^*$ $+\sqrt{2}$
$\rho$ $+\sqrt{2}$
$\Xi{\overline{\Xi}}\to\pi\pi$ $\Xi, \Xi^*$ $-\sqrt{6}$ $-2$
$\rho$ $+2$
$\Xi{\overline{\Xi}}\to K{\overline{K}}$ $\Lambda$ $-1$ $+1$
$\Sigma, Y^*$ $-3$ $-1$
$\Omega^-$ $-1$ $-1$
$\rho$ $+2$
: Isospin factors $T(I)$ for the various $B{\overline{B'}}\to\mu{\overline{\mu}}$ Born amplitudes. In the column labelled ‘process’ the exchanged baryon ($Y^*\equiv \Sigma(1385)$) or a ‘$\rho$’ in case of a $\rho$-pole diagram is given. []{data-label="tab:4_1"}
$$\begin{array}{|c|cl|}
\hline
\kappa_{\rho}^{(0)} &4.21 & \\
x_\Delta &-0.823 & \\
\Lambda_8 &1779.1 &MeV \\
\Lambda_{10} &1704.7 &MeV \\
\hline
\alpha_{p} &2/5 &(SU(6)\protect\cite{Pais}) \\
\alpha_v^m &2/5 &(SU(6)\protect\cite{Pais}) \\
\hline
\end{array}$$
$$\begin{array}{|r|r|rr|rrr|}
\hline
\multicolumn{7}{|c|}{G^\sigma_{AB\to AB}/4\pi } \\
\hline
&NN&N\Lambda&N\Sigma&\Lambda\Lambda&\Sigma\Sigma&N\Xi \\
\hline
\mbox{full model} &5.87 &2.82 &2.58 &1.52
&1.72 &1.19 \\
\mbox{without $K{\overline{K}}$ Born terms} &5.07 &1.80 &1.06 &0.64
&0.22 &0.35 \\
\mbox{subtractions for OBE model} &7.77 &3.81 &3.15 &2.00
&2.31 &1.52 \\
\hline
\end{array}$$
$$\begin{array}{|r|rcr|}
\hline
\multicolumn{4}{|c|}{^{ij}G^\rho_{NN\to NN}/4\pi\quad (ij=VV,VT,TV,TT) }
\\
\hline
&\multicolumn{3}{c|}{NN}
\\
&VV&VT,TV&TT \\
\hline
\mbox{full model}
&1.00 &5.35 &28.91 \\
\mbox{without baryon exchange}
&0.52 &2.17 &9.13 \\
\mbox{subtraction for OBE model}
&0.33 &5.58 &35.23 \\
\hline
\end{array}$$
$$\begin{array}{|r|rrrr|rrrr|}
\hline
\multicolumn{9}{|c|}{^{ij}G^\rho_{NY\to NY'}/4\pi\quad (ij=VV,VT,TV,TT)
} \\
\hline
&\multicolumn{4}{c|}{N\Sigma}
&\multicolumn{4}{c|}{N\Lambda\to N\Sigma} \\
&VV&VT&TV&TT &VV&VT&TV&TT\\
\hline
\mbox{full model}
&1.64 &1.92 &8.95 &10.15 &-0.15 &3.97 &-0.81 &21.42 \\
\mbox{without baryon exchange}
&1.03 &1.12 &4.34 &4.70 &0.00 &1.86 &0.00 &7.83 \\
\mbox{subtraction for OBE model}
&1.53 &1.52 &9.87 &9.85 &-0.52 &4.54 &-1.61 &25.60 \\
\hline
\end{array}$$
$$\begin{array}{|r|rcr|rrrr|}
\hline
\multicolumn{8}{|c|}{^{ij}G^\rho_{AB\to AB}/4\pi\quad (ij=VV,VT,TV,TT) }
\\
\hline
&\multicolumn{3}{c|}{\Sigma\Sigma}
&\multicolumn{4}{c|}{N\Xi} \\
&VV&VT,TV&TT &VV&VT&TV&TT \\
\hline
\mbox{full model} &2.78 &3.14 &4.02 &0.87 &-1.92
&4.71 &-10.38 \\
\mbox{without baryon exchange} &2.06 &2.23 &2.42 &0.52 &-1.05
&2.17 &-4.43 \\
\mbox{subtraction for OBE model}&3.09 &2.64 &4.38 &0.99 &-2.30
&5.40 &-12.09 \\
\hline
\end{array}$$
$$\begin{array}{|cr|r|rr|rrr|}
\hline
\multicolumn{8}{|c|}{G^\sigma_{AB\to AB}/4\pi ( \alpha_p) } \\
\hline
&\alpha_p &NN&N\Lambda&N\Sigma&\Lambda\Lambda&\Sigma\Sigma&N\Xi \\
\hline
&0.35 &5.86 &2.83 &2.52 &1.49 &1.69 &1.21
\\
\mbox{\sl ps} &0.40 &5.87 &2.82 &2.58 &1.52 &1.72 &1.19
\\
&0.45 &5.90 &2.84 &2.70 &1.59 &1.80 &1.19
\\
\hline
&0.35 &6.00 &3.36 &3.16 &2.04 &2.37 &1.64
\\
\mbox{\sl pv} &0.40 &6.00 &3.31 &3.28 &2.04 &2.46 &1.57
\\
&0.45 &6.00 &3.31 &3.45 &2.10 &2.62 &1.54
\\
\hline
\end{array}$$
$$\begin{array}{|r|rrrrrr|}
\hline
&NN&N\Lambda&N\Sigma&\Lambda\Lambda&\Sigma\Sigma&N\Xi \\
\hline
\pi\pi+ K{\overline{K}} &1 &0.49 &0.41 &0.26
&0.30 &0.19 \\
\hline
\mbox{OBEPT~\protect\cite{MHE} \& J\"ulich A~\protect\cite{Holz}}
&1 &0.45 &0.63 &0.34
&0.66 & \\
\mbox{Nijmegen D~\protect\cite{NijII}} &1 &1 &1 &1
&1 &1 \\
\mbox{Nijmegen F~\protect\cite{NijIII}} &1 &0.74 &0.61 &0.55
&0.37 &0.41 \\
\mbox{Nijmegen SC~\protect\cite{NijIV}} &1 &0.58 &0.45 &0.34
&0.20 &0.10 \\
\hline
\end{array}$$
$$\begin{array}{|r|rr|rr|rr|}
\hline
&\multicolumn{2}{c|}{NN}
&\multicolumn{2}{c|}{N\Sigma}
&\multicolumn{2}{c|}{N\Lambda\to N\Sigma} \\
&VV&TT &VV&TT &VV&TT \\
\hline
\pi\pi+ K{\overline{K}}
&0.28&28.87&1.25&8.07&-0.43&20.97\\
\hline
\mbox{OBEPT~\protect\cite{MHE} \& J\"ulich A~\protect\cite{Holz}}
&0.50&18.57&0.79&8.87&0&9.84 \\
\mbox{Nijmegen D~\protect\cite{NijII}} &
0.35&23.20&0.71&15.51&0&17.82 \\
\mbox{Nijmegen F~\protect\cite{NijIII}} &
0.63&27.34&1.25&28.76&0&14.96\\
\mbox{Nijmegen SC~\protect\cite{NijIV}} &
0.79&14.16&1.59&7.79&0&11.85 \\
\hline
\end{array}$$
\[fig:5\_5\_1\]
[^1]: Present address: Institut für Theoretische Physik II, Ruhr-Universität Bochum, 44780 Bochum, Germany
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the $\chi_{c1} \to \eta \pi^+ \pi^-$ decay, paying attention to the production of $f_0(500)$, $f_0(980)$ and $a_0(980)$ from the final state interaction of pairs of mesons that can lead to these three mesons in the final state, which is implemented using the chiral unitary approach. Very clean and strong signals are obtained for the $a_0(980)$ excitation in the $\eta \pi$ invariant mass distribution and for the $f_0(500)$ in the $\pi^+ \pi^-$ mass distribution. A smaller, but also clear signal for the $f_0(980)$ excitation is obtained. The results are contrasted with experimental data and the agreement found is good, providing yet one more test in support of the picture where these resonances are dynamically generated from the meson-meson interaction.'
author:
- 'Wei-Hong Liang'
- 'Ju-Jun Xie'
- 'E. Oset'
title: '$f_0(500)$, $f_0(980)$ and $a_0(980)$ production in the $\chi_{c1} \to \eta \pi^+\pi^-$ reaction'
---
Introduction
============
The $\chi_{c1} \to \eta \pi^+ \pi^-$ reaction has been measured by the CLEO collaboration in Ref. [@Adams:2011sq] and is presented as the reaction where a cleanest signal for the $a_0(980)$ resonance is seen. Indeed, a neat and strong peak is observed in the $\eta \pi$ invariant mass distribution, peaking around the $K \bar K$ threshold and with the characteristic strong cusp structure of this resonance, as observed in other high statistics experiments [@rubin]. What makes this experiment singular is that the strength of the peak is much bigger than the rest of the distribution at other $\eta \pi$ invariant masses. The complementary $\pi^+ \pi^-$ mass distribution shows a clear contribution from the $f_0(500)$ resonance at lower invariant masses, a dip in the region of the $f_0(980)$ and also a strong peak for the $f_2(1270)$ resonance and of the $f_4(2050)$ at larger invariant masses. The reaction has been remeasured with much more statistics by the BESIII collaboration and is presently under internal discussion. A preliminary view of the results is available in Ref. [@kornicer], where in the region of the $f_0(980)$ a small peak seems to show up followed by a dip around 1070 MeV. This hence constitutes a clear case for a test of the ideas of the unitarized chiral perturbation theory, the chiral unitary approach. In this approach the input from chiral Lagrangians [@gasser] for the meson-meson interaction is used in a coupled channels Bethe Salpeter equation, from where the $f_0(500)$, $f_0(980)$ and $a_0(980)$ resonances emerge [@npa; @kaiser; @markushin; @juanito]. They are dynamically generated from the meson-meson interaction and would qualify as meson-meson molecular states. The same results are obtained using an equivalent unitarizing method, the inverse amplitude method, in Refs. [@ramonet; @rios].
The nature of the low energy scalar mesons has generated a long controversy [@Klempt:2007cp]. Yet, other different approaches which start from a seed of $q \bar q$ for these resonances, also get a large meson-meson component for these states as soon as this seed is coupled to two mesons and the mesons are allowed to interact in a realistic scheme fulfilling unitarity [@vanBeveren:1986ea; @Tornqvist:1995ay; @Fariborz:2009cq; @Fariborz:2009wf]. A thorough recent review on these issues can be found in Ref. [@sigma], presenting theoretical arguments and abundant experimental information that support the picture of the dynamical generation of these resonances and its clear difference from $q \bar q$ states.
The chiral unitary approach not only provides a picture for these resonances, it also allows to make clean predictions for any reaction where these resonances are produced, providing, in the worse of the cases, when not enough dynamical information is available for the process studied, ratios for the production of the different resonances. This is a remarkable property of this approach that is not shared by other theoretical approaches trying to interpret the data. Hence experimental data could easily disprove the model, but so far this has not been the case in spite of the many reactions studied (see a recent review of $B$ and $D$ decays where many such reactions are analyzed and discussed [@weakreview]). Two of the most recent cases are the $B^0$ and $B^0_s$ decays into $J/\psi \pi^+ \pi^-$ measured in Ref. [@Aaij:2011fx] and analyzed in Ref. [@liang] (see Refs. [@Wang:2015uea; @hanhart] for a different approach based on the use of form factors) and the $D^0$ decay into $K^0$ and the $f_0(500)$, $f_0(980)$ and $a_0(980)$ measured in Ref. [@Muramatsu:2002jp] and analyzed in Ref. [@dai] (see Ref. [@robert] for also an approach based on form factors). Yet, the present reaction, with its spectacular signal for $a_0(980)$ production, a large signal for $f_0(500)$ and the small signal for the $f_0(980)$, all seen in the same reaction, is a case that should not be missed to challenge this theoretical approach. The purpose of the present paper is to make the theoretical study of the process along the lines of the chiral unitary approach, to confront the results with the relevant data already existing and eventually predict some features that could also be detected with the coming analysis from the BESIII large statistics experiment.
Formalism
=========
The $\chi_{c1} \to \eta \pi^+ \pi^-$ decay is depicted in Fig. \[fig:FeynmanDiag1\] with the quantum numbers of the different particles.
![$\chi_{c1} \to \eta \pi^+ \pi^-$ process with the quantum numbers of the particles produced. The $\pi^+ \pi^-$ pair combines to $f_0(500), f_0(980)$ and $f_2(1270)$. \[fig:FeynmanDiag1\]](20151210-Fig1.eps)
The $\chi_{c1}$ has $I^G(J^{PC})\equiv 0^+(1^{++})$ and the $\eta$, $0^+(0^{-+})$. The conservation of quantum numbers indicates that the $\pi^+ \pi^-$ pair must have isospin $I=0$, $C$-parity positive and $G$-parity positive. In addition, since the $\chi_{c1}$ has spin 1 we need one unity of spin or angular momentum in the final state. Since neither of the final $\eta$, $\pi^+$, $\pi^-$ has spin, we need to form a scalar with the polarization vector of the $\chi_{c1}$ and a momentum of one of the mesons. We can have a structure like $$\begin{aligned}
\label{eq:V_structure}
V_1 &= A \vec{\epsilon}_{\chi_{c1}} \cdot \vec{p}_{\eta}, \nonumber \\
V_2 &= B \vec{\epsilon}_{\chi_{c1}} \cdot \vec{p}_{\pi^{+}}, \\
V_3 &= C \vec{\epsilon}_{\chi_{c1}} \cdot \vec{p}_{\pi^{-}}. \nonumber\end{aligned}$$ Let us take the first structure of $V_1$. The $\vec{p}_{\eta}$ coupling introduces $L=1$ and forces the $\pi^+ \pi^-$ pair to also have positive parity. With these quantum numbers, the $\pi^+ \pi^-$ pair can be $0^+ (0^{++})$, $0^+ (2^{++})$, and then can produce the resonance $f_0(500)$, $f_0(980)$ and $f_2(1270)$, which are well known resonances.
Let us single out the term of $V_2$ in Eq. (\[eq:V\_structure\]). Now it is the $\pi^+$ the one that carries $L=1$. We write for this term another diagram in Fig. \[fig:FeynmanDiag2\].
![$\chi_{c1} \to \eta \pi^+ \pi^-$ process with the quantum numbers of the particles produced. The $\pi^- \eta$ pair combines to $a_0(980)$ and $a_2(1320)$. \[fig:FeynmanDiag2\]](20151210-Fig2.eps)
Now the $\eta \pi^-$ system must have $I=1$ and positive parity. The angular momentum of $\pi^- \eta$ can be $L'= 0,2$ and then we can have as primary choice $a_0(980)$ production and in the analysis of Ref. [@Adams:2011sq] they also allow $a_2(1320)$ formation.
As we can see, we can produce $f_0(500)$, $f_0(980)$ and $a_0(980)$ in the same reaction and there is still one more symmetry that we must consider and which, together with the ingredients of the chiral unitary approach, will allow us to establish the connection between the production of any of them in this reaction. The symmetry that we invoke is SU(3) symmetry. Since the $\chi_{c1}$ is a $c\bar c$ state, with respect to the $u,d,s$ quarks, it behaves as a neutral system, it is a scalar of SU(3). Thus, we must construct a scalar of SU(3) with three pseudoscalar mesons. This means that we will inevitably mix the $\eta \pi^+ \pi^-$ with other three meson states that can appear in the $\chi_{c1}$ decay. This will occur at a primary step of the $\chi_{c1}$ decay, but then the mesons will interact in coupled channels and finally produce the $\eta \pi^+ \pi^-$ in a final step.
In order to see the proper combination of three mesons that lead to a SU(3) scalar, we introduce the $q\bar q$ matrix $M$ $$\label{eq:M_matrix}
M=\left(
\begin{array}{ccc}
u\bar u & u \bar d & u\bar s \\
d\bar u & d\bar d & d\bar s \\
s\bar u & s\bar d & s\bar s \\
\end{array}
\right)= \left(\begin{array}{c}
u \\ d \\ s
\end{array}\right)
\left(\begin{array}{ccc}
\bar{u} & \bar{d} & \bar{s}
\end{array}\right).$$ This matrix has the property $$\begin{aligned}
\label{eq:M_matrix-2}
&MMM \nonumber\\
&=
\left(\begin{array}{c}
u \\ d \\ s
\end{array}\right)
\left(\begin{array}{ccc}
\bar{u} & \bar{d} & \bar{s}
\end{array}\right)
\left(\begin{array}{c}
u \\ d \\ s
\end{array}\right)
\left(\begin{array}{ccc}
\bar{u} & \bar{d} & \bar{s}
\end{array}\right)
\left(\begin{array}{c}
u \\ d \\ s
\end{array}\right)
\left(\begin{array}{ccc}
\bar{u} & \bar{d} & \bar{s}
\end{array}\right) \nonumber\\
&=
\left(\begin{array}{c}
u \\ d \\ s
\end{array}\right)
\left(\begin{array}{ccc}
\bar{u} & \bar{d} & \bar{s}
\end{array}\right) (\bar u u +\bar d d +\bar s s)^2 \nonumber\\
&=M (\bar u u +\bar d d +\bar s s)^2.\end{aligned}$$ Since $(\bar u u +\bar d d +\bar s s)$ is a SU(3) scalar, then the scalar that we form with the combination of Eq. (\[eq:M\_matrix-2\]) is $$\label{eq:trace}
{\rm Trace} [M (\bar u u +\bar d d +\bar s s)^2] = (\bar u u +\bar d d +\bar s s)^3= {\rm Trace} [MMM].$$ Next we write the matrix $M$ in terms of the pseudoscalar mesons, taking into account the $\eta \eta'$ mixing [@bramon] and we obtain [@dani]
$$\label{eq:phimatrix}
M \to \phi \equiv \left(
\begin{array}{ccc}
\frac{1}{\sqrt{2}}\pi^0 + \frac{1}{\sqrt{3}}\eta + \frac{1}{\sqrt{6}}\eta' & \pi^+ & K^+ \\
\pi^- & -\frac{1}{\sqrt{2}}\pi^0 + \frac{1}{\sqrt{3}}\eta + \frac{1}{\sqrt{6}}\eta' & K^0 \\
K^- & \bar{K}^0 & -\frac{1}{\sqrt{3}}\eta + \sqrt{\frac{2}{3}}\eta' \\
\end{array}
\right).$$
Then, the combination of three mesons that behaves as a SU(3) scalar is given by $$\label{eq:phiphiphi}
{\rm SU(3)[scalar]} \equiv {\rm Trace} (\phi \phi \phi).$$ By performing the algebra involved in Eq. (\[eq:phiphiphi\]) and isolating the $\eta$ term we find the combination $$\label{eq:eta_term}
C_1 :~~ \eta \left( \frac{6}{\sqrt{3}} \pi^+ \pi^-
+ \frac{3}{\sqrt{3}} \pi^0 \pi^0 + \frac{1}{3 \sqrt{3}} \eta \eta \right).$$ Thus, when taking the structure of $V_1$ of Eq. (\[eq:V\_structure\]), apart from a $\eta$ in $P$-wave we shall have a $\pi^+ \pi^-$, $\pi^0 \pi^0$ or $\eta \eta$ produced in the primary step which will undergo final state interaction to produce a $\pi^+ \pi^-$. The $\eta$ will in principle interact with the pions but this would involve a $P$-wave, where the interaction is very weak and negligible in the energy region of interest to us [@Bernard:1991xb]. We shall explicitly take into account the $\pi \pi$ or $\eta \eta$ interaction in $S$-wave [@npa], which will give rise to the $f_0(500)$, $f_0(980)$ resonances. We shall take into account the contribution of the $f_2(1270)$ empirically. The $f_2(1270)$ appears within the chiral unitary approach as a bound state of $\rho \rho$ in $S$-wave [@Molinavec; @gengvec] and decays into $\pi \pi$ in $D$-waves. This resonance gives a small contribution in the $\pi^+ \pi^-$ distribution in the region of the $f_0(500)$ and $f_0(980)$ that we are concerned about, and we take it into account to allow for a proper comparison with the data.
Similarly, if we isolate one pion to carry the $P$-wave, taking for instance the term $V_2$ in Eq. (\[eq:V\_structure\]), then we find the combination $$\label{eq:pi1_term}
C_2 :~~ \pi^+ \left( \frac{6}{\sqrt{3}} \pi^- \eta
+3 K^0 K^- \right)$$ and equivalently the term with $V_3$ in Eq. (\[eq:V\_structure\]) comes with the combination $$\label{eq:pi2_term}
C_3 :~~ \pi^- \left( \frac{6}{\sqrt{3}} \pi^+ \eta
+3 K^+ \bar K^0 \right).$$
Once again, we shall now allow the $\pi \eta$ in each of these combinations to interact in $S$-wave, which will give rise to a big signal of the $a_0(980)$. Note that in the $C_2$ and $C_3$ combinations, the $\pi \eta$ interaction in $P$-wave is negligible, and since the $\pi \pi$ system is necessarily produced in $I=0$, then it can not interact in $P$-wave either (in fact there is no trace of $\rho$ production in the experiment).
In the present process, we shall have the combination of the three structures of Eq. (\[eq:V\_structure\]) and then the primary amplitude will be of the type $$\label{eq:t_amplitue}
t= A~ \vec{\epsilon}_{\chi_{c1}} \cdot \vec{p}_{\eta} + B~ \vec{\epsilon}_{\chi_{c1}} \cdot \vec{p}_{\pi^{+}} + C~ \vec{\epsilon}_{\chi_{c1}} \cdot \vec{p}_{\pi^{-}},$$ and the first thing to note is that there is no interference between these terms. Indeed, the crossed terms in $|t|^2$ after averaging over the polarization of the massive $\chi_{c1}$ state go as $$\begin{aligned}
\label{eq:cross_term}
& {\overline{ \sum}} 2 {\rm Re}(A B^*)\vec{\epsilon}_{\chi_{c1}} \cdot \vec{p}_{\eta}~\vec{\epsilon}_{\chi_{c1}} \cdot \vec{p}_{\pi^{+}} \nonumber \\
=& 2 {\rm Re}(A B^*) \frac{1}{3} \delta_{ij} p_{\eta i} p_{\pi^{+} j} = \frac{2}{3} {\rm Re}(A B^*) \vec{p}_{\eta} \cdot \vec{p}_{\pi^{+}},\end{aligned}$$ which will vanish upon integration over angles in phase space. Thus, for $|t|^2$ we shall have the sum of the squares of each amplitude in Eq. (\[eq:t\_amplitue\]) which are described below.
Next, we must take into account the final state interaction. For the process corresponding to $V_1$ of Eq. (\[eq:V\_structure\]) we can have $\eta \pi^+ \pi^-$ in the final state by considering the $C_1$ combination of Eq. (\[eq:eta\_term\]) as depicted in Fig. \[fig:FeynmanDiag3\].
![Production of $\eta \pi^+ \pi^-$ through tree level (a) or rescattering (b) of $\pi^+ \pi^-$ pair. \[fig:FeynmanDiag3\]](Fig3-20160817.eps)
We will have $$\label{eq:t_eta}
t_{\eta}=\left( \vec{\epsilon}_{\chi_{c1}} \cdot \vec{p}_{\eta} \right) \tilde{t}_{\eta},$$ with $$\label{eq:t_eta_tilde}
\tilde{t}_{\eta}=V_P \left( h_{\pi^+ \pi^-}+\sum_i h_i S_i G_i(M_{\rm inv}) t_{i,\pi^+ \pi^-} \right),$$ where $$\label{eq:h_factor-1}
h_{\pi^+ \pi^-}=\frac{6}{\sqrt{3}},~~~~~~ h_{\pi^0 \pi^0}=\frac{3}{\sqrt{3}},~~~~~~ h_{\eta \eta}=\frac{1}{3\sqrt{3}}$$ are the weights of Eq. (\[eq:eta\_term\]) and $S_i$ are symmetry and combination factors for the identical particles, $$\label{eq:S_factor}
S_{\pi^0 \pi^0}=2\times \frac{1}{2} ~~({\rm for~ two}~ \pi^0);~~~~~S_{\eta \eta}=3! \frac{1}{2} ~~({\rm ~for~ three} ~\eta).$$ The functions $G_i$ and $t_{i, \pi^+ \pi^-}$ are the meson-meson loop functions and scattering amplitudes, which we take from Ref. [@npa] updated in Ref. [@liang; @dai].
Similarly, corresponding to $V_2$ of Eq. (\[eq:V\_structure\]), we would have the mechanism depicted in Fig. \[fig:FeynmanDiag4\].
![Production of $\pi^+ \eta \pi^-$ through tree level (a) and rescattering (b) of $\eta \pi^-$ pair. \[fig:FeynmanDiag4\]](Fig4-20160817.eps)
The amplitude corresponding to the diagrams of Fig. \[fig:FeynmanDiag4\] is given by $$\label{eq:t_pi+}
t_{\pi^+}=\left( \vec{\epsilon}_{\chi_{c1}} \cdot \vec{p}_{\pi^+} \right) \tilde{t}_{\pi^+},$$ with $$\label{eq:t_pi+_tilde}
\tilde{t}_{\pi^+}= V_P \left( h_{\pi^- \eta}+\sum_i h_i G_i(M_{\rm inv}) t_{i,\pi^- \eta} \right),$$ and $$\label{eq:h_factor-2}
h_{\pi^- \eta}=\frac{6}{\sqrt{3}},~~~h_{K^0 K^-}=3.$$ For the process associated to $V_3$ of Eq. (\[eq:V\_structure\]), we would have [^1] $$\label{eq:t_pi-}
t_{\pi^-}=\left( \vec{\epsilon}_{\chi_{c1}} \cdot \vec{p}_{\pi^-} \right) \tilde{t}_{\pi^-},$$ with $$\label{eq:t_pi-_tilde}
\tilde{t}_{\pi^-}=V_P \left( h_{\pi^+ \eta}+\sum_i h_i G_i(M_{\rm inv}) t_{i,\pi^+ \eta} \right),$$ and $$\label{eq:h_factor-3}
h_{\pi^+ \eta}=\frac{6}{\sqrt{3}},~~~h_{\bar K^0 K^+}=3.$$
As mentioned before, the interaction of the meson that comes with the $P$-wave with any of the other two, should proceed in $P$-wave, which is negligible for $\pi \eta$ and zero for $\pi \pi$ which have been created in $I=0$. This makes the interpretation of the signals particularly easy in this case, since they come from either the $\pi \pi$ or $\eta \pi$ interaction in $S$-wave.
The amplitudes for $\pi \pi, K\bar K, \pi\eta$ interaction are taken from Refs. [@liang; @dai], where only the neutral components are considered. Here we also need the charged components, which can easily be obtained using isospin symmetry and we find [@Xie:2015mzp] $$\begin{aligned}
\label{eq:t_structure}
&t_{K^0 K^-, \pi^- \eta} = \sqrt{2} t_{K^+ K^-, \pi^0 \eta}, \nonumber \\
&t_{K^+ \bar K^0, \pi^+ \eta} = \sqrt{2} t_{K^+ K^-, \pi^0 \eta}, \\
&t_{\pi^+ \eta, \pi^+ \eta} = t_{\pi^- \eta, \pi^- \eta}=t_{\pi^0 \eta, \pi^0 \eta}. \nonumber\end{aligned}$$
With all these ingredients we can write the differential mass distribution for $\pi^+ \pi^-$ as $$\label{eq:dGamma_pipi}
\frac{d \Gamma}{d M_{\rm inv}(\pi \pi)}=\frac{1}{(2\pi)^3}\frac{1}{4M_{\chi_{c1}}^2}\frac{1}{3}p_{\eta}^2 p_{\eta} \tilde{p}_{\pi} \left| \tilde{t}_{\eta} \right|^2,$$ where $p_{\eta}$ is the $\eta$ momentum in the $\chi_{c1}$ rest frame $$\label{eq:peta}
p_{\eta}=\frac{\lambda^{1/2}(M_{\chi_{c1}}^2, m_{\eta}^2, M_{\rm inv}^2(\pi \pi))}{2M_{\chi_{c1}}},$$ and $\tilde{p}_{\pi}$ is the pion momentum in the $\pi^+ \pi^-$ rest frame $$\label{eq:ppi_tilde}
\tilde{p}_{\pi}=\frac{\lambda^{1/2}(M_{\rm inv}^2(\pi \pi), m_{\pi}^2, m_{\pi}^2)}{2M_{\rm inv}(\pi \pi)}.$$
For the case of $\pi \eta$ invariant mass we would sum the contributions of $\pi^+ \eta$ and $\pi^- \eta$, which would give the same contribution, hence, the formula for $ \frac{d \Gamma}{d M_{\rm inv}(\pi \eta)}$, to be compared with experiment, will be $$\label{eq:dGamma_pieta}
\frac{d \Gamma}{d M_{\rm inv}(\pi \eta)}=\frac{2}{(2\pi)^3}\frac{1}{4M_{\chi_{c1}}^2}\frac{1}{3}p_{\pi}^2 p_{\pi} \tilde{p}_{\eta} \left| \tilde{t}_{\pi^+} \right|^2,$$ where now $$\label{eq:ppi}
p_{\pi}=\frac{\lambda^{1/2}(M_{\chi_{c1}}^2, m_{\pi}^2, M_{\rm inv}^2(\pi \eta))}{2M_{\chi_{c1}}},$$ $$\label{eq:ppi_tilde2}
\tilde{p}_{\eta}=\frac{\lambda^{1/2}(M_{\rm inv}^2(\pi \eta), m_{\pi}^2, m_{\eta}^2)}{2M_{\rm inv}(\pi \eta)}.$$ The factor $V_P$ is the only unknown quantity in our approach, which provides a global normalization, and it is fitted to the data. Note that the factors $A, B$ and $C$ in Eq. (\[eq:V\_structure\]) are absorbed in factor $V_P$.
In principle we could have summed all the amplitudes and use the general $\frac{d^2 \Gamma}{d M_{\rm inv}(\pi \pi)~d M_{\rm inv}(\pi \eta)}$ formula, integrating over each of them to find the invariant mass distribution of the other pair. In practice, we find it unnecessary for the comparison of our results with data in the relevant region of invariant masses. The reason can be seen in the Dalitz plot that we show in Fig. \[fig:DalitzPlot\]
![Dalitz plot for the $\chi_{c1} \to \eta \pi^+ \pi^-$ decay. \[fig:DalitzPlot\]](DalitzPlot.eps)
If we consider $\frac{d \Gamma}{d M_{\rm inv}(\pi \eta)}$ in the region of the $a_0(980)$, the $\pi \pi$ invariant mass has a range between 500 MeV and 2800 MeV. So its strength is divided over a large range of $\pi \pi$ invariant masses, providing a smooth background in the $\pi \pi$ mass distribution. We will take this into account empirically, following the analysis of Ref. [@kornicer].
There is another element to consider. We have taken the $\pi \eta$ invariant mass from the interacting pair of Fig. \[fig:FeynmanDiag4\](b). Since, one is summing $\pi^+ \eta$, and $\pi^- \eta$ distributions, one would also have to account for the invariant mass distribution of the $\eta$ with the odd pion carrying the $P$-wave. Yet, it is easy to see where this mass distribution goes. Indeed, using the property $$m^2_{12} + m^2_{13} +m^2_{23} = M^2_{\chi_{c1}} + m^2_{\pi}+m^2_{\pi}+m^2_{\eta},$$ taking $m_{23}=980 {\rm MeV}$ and $m_{12}\simeq 1200 {\rm MeV}$ (in the middle of the phase space allowed in the Dalitz plot), we find $m_{13} \simeq 2968 {\rm MeV}$, which is very far away from the region of the $a_0(980)$ and does not disturb the shape and strength of the $a_0(980)$.
Similarly, we will find a large contribution in the $\pi \pi$ invariant mass distribution from the $f_0(500)$ ($\sigma$ meson). Once again, by looking at the Dalitz plot, we see the strength is distributed in a region of $\pi \eta$ invariant masses from 1200 MeV to 3400 MeV, again a large region of invariant masses, but the most important for our discussion is that it does not contribute below the peak of the $a_0(980)$. Thus, the signal for the $a_0(980)$ is clean and easy to interpret, coming basically from the $\pi \eta$ interaction.
Results
=======
In Fig. \[fig:Mpieta\], we show our results for the $\pi \eta$ invariant mass distribution. The parameter $V_P$ has been adjusted to the strength of the experimental preliminary data of BESIII at its peak [@kornicer]. As we can see, both the theory and the experiment show the typical huge cusp form of the $a_0(980)$. The agreement of our results with experiment is quite good, and some missing strength from 1100 MeV on can be clearly attributed to background from other sources. As we mentioned before, one does not see in the experiment much trace of a background below 1000 MeV.
In Fig. \[fig:Mpipi\] we plot the invariant mass distribution for the $\pi^+ \pi^-$, using the same $V_P$ factor determined before. What we see is a relatively large strength for the production of the $f_0(500)$ and a small contribution from the $f_0(980)$. The experiment reflects both, a broad peak in the $f_0(500)$ region, and a rapid increase of the distribution in the region of $980 ~{\rm MeV}$. Our contribution of the $f_0(980)$ is rather sharp, while the experiment has a resolution of 20 MeV, and the raise of $\frac{d \Gamma}{d M_{\rm inv}(\pi \pi)}$ around 980 MeV is not so sharp. We should note that the strength of the $f_0(500)$ at its peak is about 110 events/10 MeV, compared to 560 events/10 MeV of the $a_0(980)$ at its peak. The signal for the $a_0(980)$ is thus quite big. Even integrating the strength of the $a_0(980)$ up to 1200 MeV and the one of the $f_0(980)$ up to 1000 MeV, we find a strength for the $a_0(980)$ almost 2.7 times bigger than that of the $f_0(980)$.
![$\pi \eta$ invariant mass distribution for the $\chi_{c1} \to \eta \pi^+ \pi^-$ decay. Preliminary BESIII data from Ref. [@kornicer]. \[fig:Mpieta\]](Mpieta.eps)
![$\pi \pi$ invariant mass distribution for the $\chi_{c1} \to \eta \pi^+ \pi^-$ decay. Preliminary BESIII data from Ref. [@kornicer]. The dashed line is the theoretical prediction. The solid line adds an empirical background(see text). \[fig:Mpipi\]](Mpipi-NoBGvsLinearBGf2.eps)
It is interesting to recall that the features of the $\pi \pi$ mass distribution are remarkably similar to those of the $J/\psi \to \omega \pi \pi$ reaction measured in Refs. [@Wu:2001vz; @Augustin:1988ja], which was studied along similar lines as here in Refs. [@Meissner:2000bc; @Roca:2004uc].
The features observed here are also similar to those observed in the $\bar B^0 $ decay into $D^0$ and $f_0(500)$, $f_0(980)$ and $a_0(980)$ [@liangBD], yet the relative strength of the structures found is different, and is related to the weight of the different meson-meson components prior to the final state interaction. The fact that one describes all these reactions with this picture, and the chiral unitary approach for the meson-meson interaction, offers support for the picture of these resonances as dynamically generated from the meson-meson interaction. Together with other reactions mentioned in the Introduction, the support for this picture is, indeed, remarkable.
To facilitate the comparison with the data, we have added a background, very similar to the one of Ref. [@kornicer] coming from the $a_0(980)$ peak, and which we have taken linear in the invariant mass for simplicity. In addition, to account for the tail of the $f_2(1270)$, which shows up in Fig. 7 at high invariant masses, we have taken a Breit Wigner shape, with physical mass and width, and adjusted the strength to reproduce the data around 1100-1200 MeV. The agreement with the $\pi \pi$ mass distribution is quite good, with some discrepancy around 1000-1040 MeV. As mentioned above, the data shows a fast raise around 980 MeV as our theory predicts, only that the theoretical raise is sharper than experiment, where data are collected in bins of 20 MeV. On the other hand, the data shows a peak around 1040 MeV that the theory cannot reproduce, even if we convolute the $f_0(980)$ signal with the experimental resolution. The discrepancy is in two experimental points and it would be worth checking whether this is just a fluctuation or a genuine peak. We should note that in Ref. [@Adams:2011sq], the data, with admittedly smaller statistics, one does not see a structure around 1000-1060 MeV like in Ref. [@kornicer].
In any case, the data of Ref. [@kornicer] is also telling us that the strength of the $f_0(980)$ is far smaller than the one of the $f_0(500)$, as the theory predicts. It would be interesting to see what comes out from the final analysis of Ref. [@kornicer], which motivated our work. A partial wave analysis can separate the contribution of the different structures, allowing for a more quantitative comparison with our results.
Conclusions
===========
We have made a study of the $\chi_{c1} \to \eta \pi^+ \pi^-$ reaction, looking at the $\pi^+ \pi^-$ and $\eta \pi$ invariant mass distributions. We have used a simple picture to combine the mesons to give a singlet of SU(3), which corresponds to the $c \bar c$ nature of the $\chi_{c1}$. This gives us the relative weights of three mesons at a primary production step, which can revert into the $\eta \pi^+ \pi^-$ in the final state upon interaction of pairs of mesons in coupled channels. We have used the chiral unitary approach to describe this interaction, which generates the $f_0(500)$, $f_0(980)$ and $a_0(980)$ scalar mesons. The interesting feature of the approach is that, up to a global normalization constant, we are able to construct the $\pi \pi$ and $\eta \pi$ invariant mass distributions and compare with the experimental data available. We observed a prominent signal of the $a_0(980)$ production with a relative strength to the other two resonances much bigger than in other reactions studied previously. We also observed a clear signal for $f_0(500)$ production in the $\pi^+ \pi^-$ mass distribution, and also a clear signal for $f_0(980)$ production, but with much smaller strength. The agreement with experiment is quite good in the two invariant mass distributions up to about 1040 MeV, once a background borrowed from the experiment is implemented in the $\pi \pi$ distribution. We also justified that no background for the $\eta \pi$ distribution was needed in that energy range.
We noted some small discrepancy with the data around 1040 MeV that could be given extra attention in the final analysis of the work of Ref. [@kornicer].
The agreement found in general lines for the shapes and strengths of the $f_0(500)$, $f_0(980)$ and $a_0(980)$ excitation in this reaction adds to the long list of reactions that give support to these resonances as being dynamically generated from the interaction of pseudoscalar mesons.
Acknowledgments {#acknowledgments .unnumbered}
===============
One of us, E. O. wishes to acknowledge support from the Chinese Academy of Science in the Program of Visiting Professorship for Senior International Scientists (Grant No. 2013T2J0012). This work is partly supported by the National Natural Science Foundation of China under Grants No. 11565007, No. 11547307 and No. 11475227. It is also supported by the Youth Innova- tion Promotion Association CAS (No. 2016367). This work is also partly supported by the Spanish Ministerio de Economia y Competitividad and European FEDER funds under the contract number FIS2011-28853-C02-01, FIS2011- 28853-C02-02, FIS2014-57026-REDT, FIS2014-51948-C2- 1-P, and FIS2014-51948-C2-2-P, and the Generalitat Valenciana in the program Prometeo II-2014/068.
[999]{}
G. S. Adams [*et al.*]{} \[CLEO Collaboration\], Phys. Rev. D [**84**]{}, 112009 (2011)
P. Rubin [*et al.*]{} \[CLEO Collaboration\], Phys. Rev. Lett. [**93**]{}, 111801 (2004)
M. Kornicer \[BESIII Collaboration\], AIP Conf. Proc. [**1735**]{}, 050011 (2016). J. Gasser and H. Leutwyler, Nucl. Phys. B [**250**]{}, 465 (1985).
J. A. Oller and E. Oset, Nucl. Phys. A [**620**]{}, 438 (1997) \[Erratum-ibid. A [**652**]{}, 407 (1999)\]. N. Kaiser, Eur. Phys. J. A [**3**]{}, 307 (1998). M. P. Locher, V. E. Markushin and H. Q. Zheng, Eur. Phys. J. C [**4**]{}, 317 (1998). J. Nieves and E. Ruiz Arriola, Nucl. Phys. A [**679**]{}, 57 (2000); Phys. Lett. B [**455**]{}, 30 (1999). J. A. Oller, E. Oset and J. R. Pelaez, Phys. Rev. D [**59**]{}, 074001 (1999) \[Phys. Rev. D [**60**]{}, 099906 (1999)\] \[Phys. Rev. D [**75**]{}, 099903 (2007)\]
J. R. Pelaez and G. Rios, Phys. Rev. Lett. [**97**]{}, 242002 (2006). E. Klempt and A. Zaitsev, Phys. Rept. [**454**]{}, 1 (2007)
E. van Beveren, T. A. Rijken, K. Metzger, C. Dullemond, G. Rupp and J. E. Ribeiro, Z. Phys. C [**30**]{}, 615 (1986)
N. A. Tornqvist and M. Roos, Phys. Rev. Lett. [**76**]{}, 1575 (1996)
A. H. Fariborz, R. Jora and J. Schechter, Phys. Rev. D [**79**]{}, 074014 (2009)
A. H. Fariborz, N. W. Park, J. Schechter and M. Naeem Shahid, Phys. Rev. D [**80**]{}, 113001 (2009)
J. R. Pelaez, arXiv:1510.00653 \[hep-ph\].
E. Oset [*et al.*]{}, Int. J. Mod. Phys. E [**25**]{}, 1630001 (2016) \[arXiv:1601.03972 \[hep-ph\]\].
R. Aaij [*et al.*]{} \[LHCb Collaboration\], Phys. Lett. B [**698**]{}, 115 (2011)
W. H. Liang and E. Oset, Phys. Lett. B [**737**]{}, 70 (2014)
W. F. Wang, H. n. Li, W. Wang and C. D. Lü, Phys. Rev. D [**91**]{}, no. 9, 094024 (2015)
J. T. Daub, C. Hanhart and B. Kubis, JHEP [**1602**]{}, 009 (2016) \[arXiv:1508.06841 \[hep-ph\]\]. H. Muramatsu [*et al.*]{} \[CLEO Collaboration\], Phys. Rev. Lett. [**89**]{}, 251802 (2002) \[Phys. Rev. Lett. [**90**]{}, 059901 (2003)\]
J. J. Xie, L. R. Dai and E. Oset, Phys. Lett. B [**742**]{}, 363 (2015)
J.-P. Dedonder, R. Kaminski, L. Lesniak and B. Loiseau, Phys. Rev. D [**89**]{}, no. 9, 094018 (2014)
A. Bramon, A. Grau and G. Pancheri, Phys. Lett. B [**283**]{}, 416 (1992). D. Gamermann, E. Oset and B. S. Zou, Eur. Phys. J. A [**41**]{}, 85 (2009) V. Bernard, N. Kaiser and U. G. Meissner, Phys. Rev. D [**44**]{}, 3698 (1991).
R. Molina, D. Nicmorus and E. Oset, Phys. Rev. D [**78**]{}, 114018 (2008) L. S. Geng and E. Oset, Phys. Rev. D [**79**]{}, 074009 (2009) J. J. Xie, W. H. Liang and E. Oset, Phys. Rev. C [**93**]{}, no. 3, 035206 (2016)
N. Wu, hep-ex/0104050. J. E. Augustin [*et al.*]{} \[DM2 Collaboration\], Nucl. Phys. B [**320**]{}, 1 (1989). U. G. Meissner and J. A. Oller, Nucl. Phys. A [**679**]{}, 671 (2001) L. Roca, J. E. Palomar, E. Oset and H. C. Chiang, Nucl. Phys. A [**744**]{}, 127 (2004) W. H. Liang, J. J. Xie and E. Oset, Phys. Rev. D [**92**]{}, no. 3, 034008 (2015)
[^1]: The diagrams are similar to those of Fig. \[fig:FeynmanDiag4\].
| {
"pile_set_name": "ArXiv"
} |
\
S. Cecchini$^1$, T. Chiarusi$^1$, G. Giacomelli$^1$, M. Giorgini$^1$, A. Kumar$^{1,4}$, G. Mandrioli$^1$, S. Manzoor$^{1,2,3}$, A. R. Margiotta$^1$, E. Medinaceli$^1$, L. Patrizii$^1$, V. Popa$^{1,5}$, I. E. Qureshi$^{2,3}$, G. Sirri$^1$, M. Spurio$^1$ and V. Togo$^1$
[*1. Phys. Dept. of the University of Bologna and INFN, Sezione di Bologna, Viale C. Berti Pichat 6/2, I-40127 Bologna, Italy\
2. PD, PINSTECH, P.O. Nilore, Islamabad, Pakistan\
3. COMSATS Institute of Information Technology 30, H/8-1, Islamabad, Pakistan\
4. Dept. Of Physics, Sant Longowal Institute of Eng. and Tech., Longowal 148 106, India\
5. Institute of Space Sciences, Bucharest R-077125, Romania*]{}
[**Abstract.**]{} [We present new measurements of the total and partial fragmentation cross sections in the energy range $0.3 \div 10$ A GeV of $^{56}$Fe, $^{28}$Si and $^{12}$C beams on polyethylene, CR39 and aluminum targets. The exposures were made at BNL, USA and HIMAC, Japan. The CR39 nuclear track detectors were used to identify the incident and survived beams and their fragments. The total fragmentation cross sections for all targets are almost energy independent while they depend on the target mass. The measured partial fragmentation cross sections are also discussed. ]{}
Introduction
============
The interaction and propagation of intermediate and high energy heavy ions in matter is a subject of interest in the fields of astrophysics, radio-biology and radiation protection [@1]. An accurate description of the fragmentation of heavy ions is important to understand the effects of the high $Z$ component of Cosmic Rays (CRs) on humans in space [@2] and for shielding in space and in accelerator environments. More recently the interaction and transport of light energetic ions in tissue-like matter became of particular interest in medicine and for hadron therapy of cancer [@4].
When a heavy ion impinges on a target, it undergoes fragmentation processes depending on the impact parameter between the colliding nuclei. The target fragments carry little momentum. At high energies, the projectile fragments travel at nearly the same velocity as the beam ions and have only a small deflection.
The availability of heavy ion beams at the CERN SPS, at BNL (USA) and at the HIMAC (Japan) facilities made possible to investigate the projectile fragmentation on different targets and for different projectile energies. Several authors \[4-10\] have successfully used Nuclear Track Detectors (NTD’s) for systematic measurements of nuclear fragmentation cross sections.
The present study is focused on Fe, Si and C ion interactions in CH$_2$, CR39 $(C_{12}H_{18}O_7)_n$ and Al targets. We used CR39 detectors, which are sensitive for a wide range of charges down to $Z = 6e$ in the relativistic energy region [@11; @12]. NTD’s have been used to search for exotic particles like Magnetic Monopoles and Nuclearites [@25; @27], to study cosmic ray composition [@26] and for environmental studies [@28].
Experimental Procedure
======================
Stacks composed of several CR39 NTD’s, of size $11.5 \times 11.5$ cm$^2$, and of different targets were exposed to 0.3, 1, 3, 5 and 10 A GeV Fe$^{26+}$, 1, 3, 5 A GeV Si$^{14+}$ ions at BNL, 0.41 A GeV Fe$^{26+}$, 0.29 A GeV C$^{6+}$ ions at HIMAC. For these exposures we used the geometry sketched in Fig. \[fig:1\]: three and four CR39 sheets, $\sim 0.7$ mm thick, were placed before and after the target, respectively. The exposures were done at normal incidence, with a density of $\sim 2000$ ions/cm$^2$. After exposures the CR39 foils were etched in 6N NaOH aqueous solution at 70 $^\circ$C for 30 h (in two steps 15h+15h) in a thermostatic water bath with constant stirring of the solution. After etching, the beam ions and their fragments manifest in the CR39 NTD’s as etch pit cones on both sides of each detector foil.
The base areas of the etch-pit cones (“tracks”), their eccentricity and central brightness were measured with an automatic image analyzer system [@17] which also provides their absolute coordinates. A tracking procedure was used to reconstruct the path of beam ions through the front faces of the detector upstream (with respect to the target) foils; a similar tracking procedure was performed through the three measured front faces of downstream CR39 detectors. The average track base area was computed for each reconstructed ion path by requiring the existence of signals in at least two out of three sheets of the detectors. In Fig. \[fig:2\]a,b the average base area distributions for 1 A GeV Si$^{14+}$ and 1 A GeV Fe$^{26+}$ beam ions and their fragments after the CH$_2$ targets are shown.
Total fragmentation cross sections
==================================
The numbers of incident and survived beam ions were determined considering the mean area distributions of the beam peaks before and after the target and evaluating the integral of the gaussian fit of the beam peaks.
The total charge changing cross sections were determined with the survival fraction of ions using the following relation $$\sigma_{tot} = \frac {A_T \ln (N_{in} / N_{out})}{\rho~ t ~N_{Av}}$$ where $A_T$ is the nuclear mass of the target (average nuclear mass in case of polymers: $A_{CH2} = 4.7,~ A_{CR39} = 7.4$); $N_{in}$ and $N_{out}$ are the numbers of incident ions before and after the target, respectively; $\rho$ (g/cm$^3$) is the target density; $t$ (cm) is the thickness of the target and $N_{Av}$ is Avogadro number.
Systematic uncertainties in $\sigma_{tot}$ were estimated to be smaller than $10\%$: contributions arise from the measurements of the density and thickness of the targets, from the separation of the beam peak from the $\Delta Z = Z_{fragment} - Z_{beam} = -1$ fragments (Fig. \[fig:2\]), from fragmentation in the CR39 foils and from the tracking procedure.
The measured total charge changing cross sections are given in the $4^{th}$ column of Table \[table:1\]. Fig. \[fig:3\]a shows the total cross sections of Fe$^{26+}$ projectiles at various beam energies on the CH$_2$ and Al targets. Our results for Si$^{14+}$ and C$^{6+}$ projectiles are given in Table \[table:2\] and are plotted vs energy in Fig. \[fig:3\]b.
The total cross sections are almost energy independent, in agreement with the data from other authors [@6; @21; @23; @24].
[|c|c|c|c|]{} & [**Target**]{} & [**[A$_T$]{}**]{} & [**[$\sigma_{tot}$]{} (mb)**]{}\
[**(A GeV)**]{} & & &\
10 & CH$_2 $& 4.7 & 1147 $\pm$ 97\
10 & CR39 & 7.4 & 1105 $\pm$ 360\
5 & CH$_2 $ & 4.7 & 1041 $\pm$ 130\
5 & CR39 & 7.4 & 1170 $\pm$ 470\
3 & CH$_2 $ & 4.7 & 904 $\pm$ 140\
3 & CR39 & 7.4 & 1166 $\pm$ 67\
1 & CH$_2 $ & 4.7 & 1105 $\pm$ 60\
1 & CR39 & 7.4 & 1113 $\pm$ 176\
1 & Al & 27 & 1870 $\pm$ 131\
0.41 & CH$_2 $ & 4.7 & 948 $\pm$ 54\
0.41 & CR39 & 7.4 & 1285 $\pm$ 245\
0.41 & Al & 27 & 1950 $\pm$ 126\
0.30 & CH$_2 $ & 4.7 & 949 $\pm$ 61\
0.30 & CR39 & 7.4 & 1174 $\pm$ 192\
0.30 & Al & 27 & 2008 $\pm$ 144\
Various theoretical models/formulae for the total fragmentation cross sections were proposed and fitted to the experimental data with different geometrical radii and overlapping parameters [@8]. In Fig. \[fig:3\] our data are compared with the semi-empirical formula [@14] for nuclear cross sections (solid lines)
$$\sigma_{tot} = \pi r_0^2~ (A_P^{1/3} + A_T^{1/3} -b_0)^2
\label{eq:2}$$
where $r_0 = 1.31$ fm, $b_0 = 1.0$, $A_P$ and $A_T$ are the projectile and target mass numbers, respectively. Various authors used different values for the overlap parameter $b_0$ within the interval $0.74 \div 1.3$ \[5-10\].
Figs. \[fig:4\]a,b show the total fragmentation cross sections vs target mass number $A_T$ for Fe$^{26+}$, Si$^{14+}$ and C$^{6+}$ beams of various energies. The solid lines are the predictions of Eq. \[eq:2\], to which we added the electromagnetic dissociation contribution, $\sigma_{EMD} = \alpha Z_T^{\delta}$, with $\alpha = 1.57$ fm$^2$ and $\delta =
1.9$ \[last ref. of [@11]\]. The total fragmentation cross sections increase with increasing target mass number. Part of the increase is due to the effect of electromagnetic dissociation.
The data from other authors [@6; @21; @24; @9] are plotted for comparison and show good agreement with our data, within the experimental uncertainties.
[|c|c|c||c|c|c|]{} &\
& [**Target**]{} & [**[$\sigma_{tot}$]{} (mb)**]{} & [**Energy**]{} & [**Target**]{} & [**[$\sigma_{tot}$]{} (mb)**]{}\
[**(A GeV)**]{} & & & [**(A GeV)**]{} & &\
5 & CH$_2$ & 757 $\pm$ 168 & 0.29 & CH$_2$ & 460 $\pm$ 53\
3 & Al & 1533 $\pm$ 133 & 0.29 & CR39 & 513 $\pm$ 52\
1 & CR39 & 1113 $\pm$ 176 & 0.29 & Al & 1155 $\pm$ 108\
1 & H & 483 $\pm$ 76 & & &\
1 & CH$_2$ & 694 $\pm$ 70 & & &\
1 & C & 1117 $\pm$ 62 & & &\
1 & Al & 1397 $\pm$ 138 & & &\
Partial fragmentation charge changing cross sections
====================================================
If the thickness of the target is small compared to the mean free path of the fragments in that material, the partial fragmentation cross sections can be calculated using the simple relation $$\sigma(Z_i, Z_f) \simeq \frac{1}{Kt} \frac{N_f}{N_i}$$ where $\sigma (Z_i, Z_f)$ is the partial fragmentation cross section of an ion $Z_i$ into the fragment $Z_f$, $K$ is the number of target nuclei per cm$^3$, $t$ is the thickness of the target, $N_i$ is the number of survived ions after the target and $N_f$ is the number of fragments produced with charge $Z_f$. This expression may be valid also for a thick target, assuming that the number of fragments before the target is zero.
For the Fe ions, we observed that fragments are present even before the targets. In this case the partial charge change cross sections have been computed via the relation
$$\sigma_{\Delta Z} = \frac{1}{Kt} \left( \frac{N^f_{out}}{N^p_s} - \frac{N^f_{in}}{N^p_{in}} \right)$$
where $N_{in}^{f}$ and $N_{out}^f$ are the numbers of fragments of each charge before and after the target, and $N_{in}^{p}$ and $N_{s}^p$ are the numbers of incident and survived projectile ions.
The distributions, after the CH$_2$ targets, of the fragments for 1 A GeV Si$^{14+}$ and 1 A GeV Fe$^{26+}$ ions are shown in Figs. \[fig:2\]a,b. The relative partial fragmentation cross sections for $\Delta Z = -1, -2, -3,~..,~ -18$ are given in Table \[table:3\]. The quoted errors are statistical standard deviations; systematic uncertainties are estimated to be about $10\%$. A clear odd-even effect is visible in Fig. \[fig:2\]: the cross sections for the $Z-$even fragments are generally larger than those for the $Z-$odd fragments close by.
[|c|c|c|]{} & [**1 A GeV [Fe$^{26+}$]{}**]{} & [**1 A GeV [Si$^{14+}$]{}**]{}\
-1 & - & 293 $\pm$ 18\
-2 & 338 $\pm$ 11 & 177 $\pm$ 12\
-3 & 285 $\pm$ 11 & 123 $\pm$ 11\
-4 & 252 $\pm$ 10 & 122 $\pm$ 11\
-5 & 249 $\pm$ 10 & 62 $\pm$ 8\
-6 & 197 $\pm$ 9 & 117 $\pm$ 11\
-7 & 168 $\pm$ 8 & 83 $\pm$ 9\
-8 & 132 $\pm$ 7 & 90 $\pm$ 10\
-9 & 175 $\pm$ 8 &\
-10 & 107 $\pm$ 7 &\
-11 & 152 $\pm$ 6 &\
-12 & 105 $\pm$ 8 &\
-13 & 103 $\pm$ 6 &\
-14 & 81 $\pm$ 6 &\
-15 & 80 $\pm$ 6 &\
-16 & 50 $\pm$ 4 &\
-17 & 76 $\pm$ 5 &\
-18 & 86 $\pm$ 6 &\
Conclusions
===========
The total fragmentation cross sections for $^{56}$Fe, $^{28}$Si and $^{12}$C ion beams of $0.3 \div 10$ A GeV energies on polyethylene, CR39 and aluminum targets were measured using CR39 NTD’s [@19].
The total cross sections for all the targets and energies used in the present work do not show any observable energy dependence. There is a dependence on target mass; the highest cross sections are observed for Al targets and this is mainly due to the contribution of electromagnetic dissociation. The present data of total fragmentation cross sections are in agreement with similar experimental data in the literature \[4-10\]. The presence of well separated fragment peaks, see Fig. \[fig:2\], allowed the determination of the partial fragmentation cross sections. On the average the partial cross sections decrease as the charge change $\Delta Z$ increases. The data in Fig. \[fig:2\] and the partial cross sections in Table \[table:3\] indicate a clear $Z$ odd-even effect.
The measured cross section data indicate that passive NTD’s, specifically CR39, can be used effectively for studies of the total and partial charge changing cross sections, also in comparison with active detectors.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work was in part financed by the MIUR PRIN 2004 Program (ex $40\%$), Prot. 2004021217.
We thank the technical staff of BNL and HIMAC for their kind cooperation during the beam exposures. We acknowledge the contribution of our technical staff, in particular of A. Casoni, M. Errico, R. Giacomelli, G. Grandi and C. Valieri. We thank INFN and ICTP for providing fellowships and grants to non-Italian citizens.
[99]{} C. X. Chen et al., Phys. Rev. C49 (1994) 3200. J. W. Wilson et al., Health Phys. 68 (1995) 50.
U. Amaldi, Nucl. Phys. A751 (2005) 409. S. Cecchini et al., Astrop. Phys. 1 (1993) 369 ; Nucl. Phys. A707 (2002) 513.\
H. Dekhissi et al., Nucl. Phys. A662 (2000) 207. W. R. Webber et al., Phys. Rev. C41 (1990) 520.\
P. B. Price and Y. D. He, Phys. Rev. C43 (1991) 835.\
S. E. Hirzebruch et al., Phys. Rev. C46 (1992) 1487; Nucl. Instr. Meth. B74 (1993) 519.\
L. Sihver et al., Phys. Rev. C47 (1993) 1225.\
Y. D. He and P. B. Price, Z. Phys. A348 (1994) 105.\
L. Y. Geer et al., Phys. Rev. C52 (1995) 334.\
G. Iancu et al., Radiat. Meas. 39 (2005) 525.\
T. Toshito et al., Phys. Rev. C75 (2007) 054606. C. Brechtmann et al., Z. Phys. A330 (1988) 407; Phys. Rev. C39 (1989) 2222. C. Zeitlin et al., Phys. Rev. C56 (1997) 388. F. Flesch et al., Radiat. Meas. 34 (2001) 237. C. Zeitlin et al., Nucl. Phys. A784 (2007) 341. A. N. Golovchenko et al., Nucl. Instr. Meth. B159 (1999) 233; Phys. Rev. C66 (2002) 014609. G. Giacomelli et al., Nucl. Instr. Meth. A411 (1998) 41.\
S. Cecchini et al., Radiat. Meas. 34 (2001) 55.\
S. Balestra et al., Nucl. Instr. Meth. B254 (2007) 254. S. Manzoor et al., Nucl. Phys. B Proc. Suppl. 172 (2007) 296.\
G. Giacomelli et al., hep-ex/0702050. M. Ambrosio et al., Eur. Phys. J. C25 (2002) 511. T. Chiarusi et al., Radiat. Meas. 40 (2005) 424. S. Manzoor et al., Nucl. Phys. B Proc. Suppl. 172 (2007) 92. A. Noll et al., Nucl. Tracks Radiat. Meas. 15 (1988) 265. H. L. Bradt, B. Peters, Phys. Rev. 77 (1950) 54. S. Manzoor, Ph.D. Thesis, University of Bologna, Italy, and CIIT, Islamabad, Pakistan (2007).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Evgeny Shaverin
bibliography:
- 'mybibfile.bib'
title: 'A breakdown of a universal hydrodynamic relation in Gauss-Bonnet gravity'
---
Introduction {#S:Intro}
============
Conformal viscous hydrodynamics {#conformal-viscous-hydrodynamics .unnumbered}
-------------------------------
Hydrodynamics is a low energy effective description of many-body systems which are studied at length scales much larger than their mean free path. When only energy and momentum are conserved, this effective description is given by a handful of hydrodynamics fields: the velocity field $ u^{\mu} $ (normalized so that $ u^{\mu}u_{\mu}=-1 $), and the temperature field $ T $. The dynamics of these fields is governed by energy conservation, $$\partial_{\mu}T^{\mu\nu} = 0 \, .$$
In a conformal theory, the dependence of the stress-energy tensor $ T^{\mu\nu} $ on the hydrodynamic fields takes the following form $$\label{E:GeneralTmn}
T^{\mu\nu} = P \left(4 u^{\mu}u^{\nu} + \eta^{\mu\nu}\right) + \tilde{\Pi}^{\mu\nu} \, ,$$ where $ P=P(T) $ is the pressure and $ \tilde{\Pi}^{\mu\nu} $ denotes dissipative terms. The tensor $ \tilde{\Pi}^{\mu\nu} $ depends on gradients of the hydrodynamic variables $ T $ and $ u^{\mu} $. It is convenient to impose on $ \tilde{\Pi}^{\mu\nu} $ the Landau frame condition $ u_\mu \tilde{\Pi}^{\mu\nu} = 0 $. Following [@Baier:2007ix], an expansion of $ \tilde{\Pi}^{\mu\nu} $ in derivatives of $ T $ and $ u^{\mu} $ up to second order is given by $$\label{DissipativeParts}
\tilde{\Pi}^{\mu\nu} = - \eta \sigma^{\mu\nu} + \sum_{i=0}^{4} \lambda_{i} \Sigma^{(i)\mu\nu} \, .$$ The coefficient $ \eta $ in is the [shear]{} viscosity, and the coefficients $ \lambda_i $, $ i=0,\ldots,4 $, are referred to as ‘second order transport coefficients’. The tensors $ \sigma^{\mu\nu} $ and $ \Sigma^{(i)}_{\mu\nu} $ are given by: $$\begin{aligned}
\label{2orderTemrs}
\begin{split}
\sigma^{\mu\nu} &= 2\partial^{\langle\mu}u^{\nu\rangle} \, ,\\
\Sigma^{(0)}_{\mu\nu} &= {}_{\langle}u^{\alpha}\partial_{\alpha}\sigma_{\mu\nu\rangle} + \frac{1}{3} \sigma_{\mu\nu}\partial_{\alpha}u^{\alpha} \, ,\\
\Sigma^{(1)}_{\mu\nu} &= \sigma_{\langle \mu \alpha}\sigma^{\alpha}{}_{\nu\rangle} \,,
\quad
\Sigma^{(2)}_{\mu\nu} = \sigma_{\langle \mu\alpha} \omega^{\alpha}{}_{\nu\rangle} \, , \\
\quad
\Sigma^{(3)}_{\mu\nu} &= \omega_{\langle \mu\alpha} \omega^{\alpha}{}_{\nu\rangle} \, ,
\quad
\Sigma^{(4)}_{\mu\nu} = R_{\langle\mu\nu\rangle} - 2 u^{\alpha} R_{\alpha\langle\mu\nu\rangle\beta} u^{\beta} \,,
\end{split}\end{aligned}$$ with the following definitions: $R_{\mu\nu} $ and $ R_{\alpha\mu\nu\beta} $ are the Ricci and Riemann tensor respectively, the vorticity, $ \omega_{\mu\nu} $, is given by $ \omega_{\mu\nu}=\frac{1}{2}P_{\mu}{}^{\alpha}P_{\nu}{}^{\beta}\left(\partial_{\alpha}u_{\beta}-\partial_{\beta}u_{\alpha}\right) $, and triangular brackets denote a traceless and transverse projection $$\nonumber
A^{\langle \mu \nu \rangle}= \frac{1}{2} P^{\mu\alpha}P^{\nu\beta} \left(A_{\alpha\beta}+A_{\beta\alpha}\right) - \frac{1}{3}P^{\mu\nu}P^{\alpha\beta}A_{\alpha\beta} \, ,$$ where we have defined the projection on the spatial directions $ P^{\mu\nu} = \eta^{\mu\nu}+u^{\mu}u^{\nu} $. We assume that the field theory is living in flat space, where the Ricci and Riemann tensors vanish, therefore $ \Sigma^{(4)}_{\mu\nu} = 0 $. Collecting (\[E:GeneralTmn\]) and (\[DissipativeParts\]) we find that in our setup, $$\label{E:GeneralTmn1}
T^{\mu\nu} = P \left(4 u^{\mu}u^{\nu} + \eta^{\mu\nu}\right) - \eta \sigma^{\mu\nu} + \sum_{i=0}^{3} \lambda_{i} \Sigma^{(i)\mu\nu} \, .$$
In [@Kovtun:2003wp] it was conjectured that for all known physical systems the ratio of the shear viscosity to entropy density is given by $$\label{BOUND}
\frac{\eta}{s} \ge \frac{1}{4 \pi} \, .$$ The bound is found to be saturated in all thermally equilibrated systems with isotropic and homogeneous configuration described by a dual two-derivative gravity action [@Buchel:2003tz; @Kovtun:2004de; @Buchel:2004qq]. The bound has been shown to be violated in the non-isotropic configurations, see [@Erdmenger:2010xm; @Basu:2011tt; @Erdmenger:2011tj; @Rebhan:2011vd; @Jain:2014vka; @Ovdat:2014ipa]. Recent investigations [@Kats:2007mq; @Brigante:2007nu; @Buchel:2008vz] show that in theories dual to higher derivative gravity there is a violation of this bound. In the special case of Gauss-Bonnet gravity, $$\label{S_EGB}
\mathcal{S}_{\rm{EGB}} = -\frac{1}{16 \pi G_5} \int \sqrt{-g} \left( R + \frac{12}{L^2} - \theta L_{GB} \right) d^5x \, ,$$ $$\nonumber
L_{GB} = R_{mnpq}R^{mnpq} - 4 R_{mn}R^{mn} + R^2 \, ,$$ it was shown in [@Brigante:2007nu] that for general values of a Gauss-Bonnet coupling $ \theta $ $$\label{eta/s}
\frac{\eta}{s} = \frac{1}{4 \pi} \left( 1-8\theta \right) \, .$$
When considering the second order transport coefficients, it has been observed that the linear relation $$\label{relation}
-2 \lambda_0 + 4 \lambda_1 - \lambda_2 = 0 \,,$$ seems to be satisfied for a large class of two derivative gravity theories [@Erdmenger:2008rm]. In contrast, the coefficients $\lambda_i$ where also computed perturbatively using kinetic theory where the relation (\[relation\]) seems not to hold.
Surprisingly, Gauss-Bonnet gravity seems not to correct the relation when it is treated perturbatively [@Shaverin:2012kv]. This result has been confirmed and extended in [@Grozdanov:2014kva] to include perturbatively small eight derivative corrections. In what follows we extend these results and compute the expression to all orders in the Gauss-Bonnet coupling. We find that $$\label{FinResult}
- 2 \lambda_0 + 4 \lambda_1 - \lambda_2 = \frac{\pi^{2}T^{2}}{\pi G_{5}}\left[\frac{\sqrt{2}\left(1-\sqrt{1-8\theta}+16\theta\right)\theta}{\left(1+\sqrt{1-8\theta}\right)^{3/2}}\right] \, .$$ Thus, the relation which seems to hold even at the perturbative level in 2, 4 and 8 derivative gravity breaks down at the non perturbative level. We note that the relation has been published before in [@Grozdanov:2014kva] using methods and the non-perturbative expressions for all Gauss-Bonnet transport coefficients published in [@Andrei]. Treating the Gauss-Bonnet coupling $ \theta $ non perturbatively via AdS/CFT may be regarded as a toy model for probing finite $ N $ corrections in a gauge theory [@Brigante:2007nu; @Kats:2007mq; @Buchel:2008vz].
Our work is organised as follows. In the chapter 2 we introduce the action and equations of motion for the Einstein-Gauss-Bonnet gravity, dual to a viscous fluid in the field theory. Using black hole solution to the equations of motion and AdS/CFT dictionary we find an ideal fluid (non-dissipative) part of the energy momentum tensor. In the chapter 3 we introduce derivative corrections to the gravity dual corresponding to the viscous contributions in the field theory side. Solving the equations of motion up to second order in derivatives we reproduce the known result for the shear viscosity over entropy density (\[eta/s\]) and show explicitly the breakdown of the universal relation (\[FinResult\]).
AdS Gauss-Bonnet black holes
============================
The equations of motion for Gauss-Bonnet gravity can be derived from the following action, $$\label{action}
\mathcal{S} = \mathcal{S}_{\rm{EGB}} + S_{\rm{b}}$$ where $ \mathcal{S}_{\rm{EGB}} $ was defined in (\[S\_EGB\]) and $S_{\rm{b}}$ are appropriate counter terms, imposed to make the variational principle well defined [@PhysRevD.15.2752; @Balasubramanian:1999re]. The Roman indices $m,\,n=0,\ldots,4$ refer to bulk (AdS) quantities while Greek indices $\mu,\,\nu = 0,\ldots,3$ refer to boundary quantities. We set $ L=1 $ from now on. The equations of motion that follow from (\[action\]) are given by $$\begin{gathered}
\label{EOM}
R_{mn}-\frac{1}{2}Rg_{mn}-6g_{mn}-\frac{\theta}{2}g_{mn}L_{GB} \\
+2\theta\left(R_{mpql}R_{n}^{\phantom{n}pql}-2R^{pq}R_{mpnq}-2R_{m}^{\phantom{m}q}R_{qn}+RR_{mn}\right)=0 \,.\end{gathered}$$
The black hole solution to is given by $$\label{eq:ansatz0}
ds^2 = -r^2 f(\rho) dt^2 + r^2 \left(dx^i\right)^2 + 2 s(\rho) dt dr \, ,$$ where $ \rho\equiv br $ and $$\begin{aligned}
\label{0th_order_functions}
\begin{split}
s(\rho) & = \sqrt{\frac{1+\sqrt{1-8\theta}}{2}}\equiv S_{0}\\
f(\rho) & = \frac{S_{0}^{2}}{4\theta}\left(1-\sqrt{1-8\theta\left(1-\frac{1}{\rho^4}\right)}\right) \, ,
\end{split}\end{aligned}$$ with $ b $ an integration constant associated with the mass of the black hole [@PhysRevLett.55.2656; @Cai:2001dz].
Under the gauge-gravity duality, the black hole solution is dual to a thermal state of the black hole. The energy momentum tensor of this state can be computed using the standard prescription, [@Brihaye:2008kh; @Dutta:2008gf; @Buchel:2008vz],
\[E:Tmnprescription\] & T\_ = \_[r]{} (\_-\_ +2(3\_-\_) -3\_+\_)
where:
- $ \gamma_{mn} = g_{mn} - N_m N_n $ is the boundary metric
- $ N_n = \delta_n^r / \sqrt{g^{rr}} $ a unit outward vector to the boundary
- $ \mathcal{K}_{mn} = -\frac{1}{2} \left(\nabla_{m} N_n + \nabla_n N_m\right) $ is the extrinsic curvature on the boundary (the covariant derivative here is with respect to the bulk metric $ g_{mn} $)
- $ \mathcal{J}_{mn} = \frac{1}{3} \Big(2 \mathcal{K} \mathcal{K}_{mp}\mathcal{K}^p_{\phantom{p}n} + \mathcal{K}_{ps}\mathcal{K}^{ps} \mathcal{K}_{mn} - 2 \mathcal{K}_{mp}\mathcal{K}^{ps}\mathcal{K}_{sn} - \mathcal{K}^2 \mathcal{K}_{mn}\Big) $
- $\mathcal{K}$ and $\mathcal{J}$ are the trace (using $ \gamma_{mn} $ to contract indices) of $\mathcal{K}_{mn}$ and $\mathcal{J}_{mn}$ respectively
- $ S_{0} = \sqrt{\frac{1+\sqrt{1-8\theta}}{2}}$
Inserting into and using the following relation $$T = \frac{S_0}{b \pi}$$ between the temperature of the dual theory and the horizon parameter $b$, derived in (\[TEMP\]), we find $$\label{zeroth_order}
T^{\mu\nu} = \frac{T^4 \pi^4}{16\pi G_{5} S_0^3} \left(4 u^{\mu}u^{\nu} + \eta^{\mu\nu}\right) \, .$$ The parameter $G_5$ is theory dependent. For instance, in the planar limit of strongly coupled $\mathcal{N}=4$ super Yang Mills one finds $$G_5 = \frac{\pi}{2N^2}\,.$$ Gauss-Bonnet coupling $ \theta $ can be regarded as $ \mathcal{O}(N) $ corrections to the action [@Kats:2007mq; @Buchel:2008vz].
Solving the equations of motion perturbatively
==============================================
In order to compute the transport coefficients of the fluid dual to the theory defined by we use the method of [@Bhattacharyya:2008jc]; The metric (\[eq:ansatz0\]) may be boosted with velocity $ u^{\alpha} = \frac{1}{\sqrt{1-\beta^2}} \left( 1, \vec{\beta} \right) $, resulting in the line element $$\label{eq:ansatz1}
ds^2 = -r^2 f(br) u_{\mu}u_{\nu}dx^{\mu}dx^{\nu} + r^2 P_{\mu\nu} dx^{\mu}dx^{\nu} - 2 s(br) u_{\mu} dx^{\mu} dr \, .$$ We now promote $ u^\mu $ and $ b $ to be a slowly-varying fields, $$\label{coordep}
\beta_i \rightarrow \beta_i(x^{\alpha}) \, , \quad b \rightarrow b(x^{\alpha}) \,.$$ Inserting (\[coordep\]) into (\[eq:ansatz1\]) is in general, not a solution to the equations of motion (\[EOM\]). Therefore we correct (\[eq:ansatz1\]) by adding new terms to the metric. The corrections to the metric may be classified as follows. Let us decompose the metric into scalar, vector and tensor modes of the $SO(3)\subset SO(3,1)$ symmetry under which $u^{\mu}$ is (locally) invariant, i.e.,
\[E:lineelement\] ds\^[2]{} & = r\^[2]{} k u\_u\_dx\^dx\^ +r\^[2]{}P\_dx\^dx\^-2Su\_dx\^dr\
& +r\^[2]{}(u\_V\_+u\_V\_)dx\^dx\^ +r\^[2]{}\_dx\^dx\^
and expand $k$, $S$, $V_{\mu}$ and $\Pi_{\mu\nu}$ in gradients of $ u^{\alpha} $ and $ b $. It is convenient to denote the $n$’th order correction to $k$, $S$, $V_{\mu}$ and $\Pi_{\mu\nu}$ with a superscript $(n)$. For instance,
\[kSVP\]
k & = k\^[(0)]{} + k\^[(1)]{} + k\^[(2)]{} + …= - f() + k\^[(1)]{} + k\^[(2)]{} + …\
S & = S\^[(0)]{} + S\^[(1)]{} + S\^[(2)]{} + …= s() + S\^[(1)]{} + S\^[(2)]{} + …\
V\_ & = V\^[(0)]{}\_ + V\^[(1)]{}\_ + V\^[(2)]{}\_ + …= V\^[(1)]{}\_ + V\^[(2)]{}\_ + …\
\_ & = \^[(0)]{}\_ + \^[(1)]{}\_ + \^[(2)]{}\_ + …= \^[(1)]{}\_ + \^[(2)]{}\_ + …,
where $ f(\rho) $ and $ s(\rho) $ are given in (\[0th\_order\_functions\]). Inserting (\[kSVP\]) into (\[E:lineelement\]) and using the equations of motion (\[EOM\]) we find, at each order in $ n>0 $,
\[E:EOMs\] (S\^[(n)]{})\^ & = \^[(n)]{}\
( \^4 (S\_0\^2-4f()) (k\^[(n)]{}+ S\^[(n)]{} ) )\^ & = \^[(n)]{}\
( \^5 (S\_0\^2-4f()) V\_\^[(n)]{} )\^ & = \_\^[(n)]{}\
( \_\^[(n)]{})\^ & =\_\^[(n)]{} ,
where the quantities on the right-hand side of (\[E:EOMs\]) depend on the known $ 0 $ to $ n-1 $ order metric components. Integrating (\[E:EOMs\]) leads to
\[E:SOL\] S\^[(n)]{} & = - \_\^ \^[(n)]{}(x’)dx’\
k\^[(n)]{} & = - S\^[(n)]{} +\
V\_\^[(n)]{} & = - \_\^ dx +\_\^2 \_\^\
\_\^[(n)]{} & = - \_\^ dx .
The boundary conditions for the integral were chosen to ensure that
1. The boundary metric is flat: $ \lim\limits_{r\rightarrow\infty} ds^2 = r^2 \eta_{\mu\nu} dx^{\mu}dx^{\nu} $, i.e. for $ n\ge1 $:
\_[r]{}k\^[(n)]{} & = 0 , \_[r]{}V\_\^[(n)]{} = 0 , \_[r]{}\_\^[(n)]{} = 0 .
2. The metric is regular at $ r=1 $.
3. $ T^{\mu\nu} $ is in the Landau frame: there are no $ 1/\rho^4 $ dependence in the near-boundary expansion of $ k^{(n)} $ and $ V_{\mu}^{(n)} $ with $ n\ge1 $.
With a perturbative solution to the equations of motion at hand one can use the prescription to compute the energy momentum tensor order by order in the derivative expansion.
First order
-----------
Using (\[kSVP\]) to first order in derivatives, we find (\[E:EOMs\]) with $$\begin{aligned}
\begin{split}
\label{E:sources1}
\mathbf{S}^{(1)} & = 0\\
\mathbf{k}^{(1)} & = b \, \left(\frac{2}{3} S_0 \rho^{3 }\left(S_0^2-4\theta f(\rho)\right)\right)^{\prime} \partial_{\alpha}u^{\alpha}\\
\mathbf{V}_{\mu}^{(1)} & = b \left(S_0 \rho^3 \left(S_0^2-4\theta f(\rho)\right)\right)^{\prime} u^{\alpha}\partial_{\alpha}u_{\mu}\\
\mathbf{P}_{\mu\nu}^{(1)} & = b \left(\frac{- S_0 \rho^3}{S_0^2-4\theta f(\rho)}\right)^{\prime} \sigma_{\mu\nu} \, .
\end{split}\end{aligned}$$ Inserting (\[E:sources1\]) into (\[E:SOL\]) we get
\[1st\_order\_functions\] S\^[(1)]{} & = 0\
k\^[(1)]{} & = b \_u\^\
V\_\^[(1)]{} & = - b u\^ \_u\_\
\_\^[(1)]{} & = b () \_ ,
where $$\nonumber
\pi(\rho) = - \int_{\rho}^{\infty} \frac{\left(S_0^2-4\theta f(x)\right)-S_0^2 x^3}{S_0 x^5 f(x)} \, dx \, .$$
Evaluating (\[E:Tmnprescription\]) by using the first order metric solution and temperature relation (\[TEMP\]), we get $$\label{eta}
T^{\mu\nu}=\frac{T^{4}\pi^{4}}{16\pi G_{5}S_{0}^{3}}\left(4u^{\mu}u^{\nu}+\eta^{\mu\nu}\right)-\frac{T^{3}\pi^{3}\left(1-8\theta\right)}{16\pi G_{5}S_{0}^{3}}\sigma^{\mu\nu} \, .$$ Using we can read the shear viscosity: $$\eta = \frac{T^{3}\pi^{3}\left(1-8\theta\right)}{16\pi G_{5}S_{0}^{3}} \, .$$ Note that using $ s = \frac{dP}{dT} $ together with the relations (\[0th\_order\_functions\]), we recover the result of [@Brigante:2007nu] $$\frac{\eta}{s} = \frac{1}{4\pi} \left(1-8\theta\right) \, .$$
Second order
------------
Inserting the zeroth and first order solutions (\[0th\_order\_functions\]) and (\[1st\_order\_functions\]) into (\[kSVP\]) and solving (\[EOM\]), one finds that the solution to the equations of motion take the form
\[secondordermetricC\]
S & = S\_0 + \_[4]{}()\_[4]{}+\_[5]{}()\_[5]{}\
k & = -f(b r) + b \_u\^ + \_[0]{}()\_[3]{} + \_[1]{}()\_[1]{}\
& +\_[3]{}()\_[3]{}+\_[4]{}()\_[4]{} +\_[5]{}()\_[5]{}\
V\_ & = - b u\^ \_u\_ + \_[4]{}()\_[4]{} +\_[5]{}()\_[5]{}\
& +\_[1]{}()\_[1]{} + \_[2]{}()\_[2]{}+\_[3]{}()\_[3]{}\
\_ & = b () \_ + b\^[2]{}\_[i=0]{}\^[3]{} \_[i]{}() \_\^[(i)]{} ,
where $$\begin{aligned}
\mathbf{s}_{3} & = \frac{1}{b}P^{\alpha\beta}\partial_{\alpha}\partial_{\beta}b,\\
\mathfrak{S}_{1} & =\mathcal{D}u^{\alpha}\mathcal{D}u_{\alpha},\quad
\mathfrak{S}_{3}=\left(\partial_{\mu}u^{\mu}\right)^{2}\\
\mathfrak{S}_{4} & = l_{\mu}l^{\mu},\quad
\mathfrak{S}_{5}=\frac{1}{4}\sigma_{\mu\nu}\sigma^{\mu\nu} \, ,\end{aligned}$$
$$\mathcal{D}\equiv u^{\mu}\partial_{\mu},\quad
l^{\mu}=\epsilon^{\alpha\beta\gamma\mu}u_{\alpha}\partial_{\beta}u_{\gamma},$$
$$\begin{aligned}
\mathbf{v}_{4\,\nu} & =\frac{9}{5}\left[P^{\alpha}{}_{\nu}P^{\beta\gamma}\partial_{\gamma}\partial_{\left(\beta\right.}u_{\left.\alpha\right)}
-\frac{1}{3}P^{\alpha\beta}P^{\gamma}{}_{\nu}\partial_{\gamma}\partial_{\alpha}u_{\beta}\right]
-P^{\mu}{}_{\nu}P^{\alpha\beta}\partial_{\alpha}\partial_{\beta}u_{\mu},\\
\mathbf{v}_{5\,\nu} & =P^{\mu}{}_{\nu}P^{\alpha\beta}\partial_{\alpha}\partial_{\beta}u_{\mu},\\
\mathfrak{V}_{1\,\nu} & =\partial_{\alpha}u^{\alpha}\mathcal{D}u_{\nu},\quad
\mathfrak{V}_{2\,\nu}=\epsilon_{\alpha\beta\gamma\mu}u^{\alpha}\mathcal{D}u^{\beta}l^{\gamma},\quad
\mathfrak{V}_{3\,\nu}=\frac{1}{2}\sigma_{\alpha\nu}\mathcal{D}u^{\alpha}\end{aligned}$$
and $ \sigma_{\mu\nu} $, $ \Sigma_{\mu\nu}^{\left(i\right)} $ and $ \omega_{\mu\nu} $ were defined in (\[2orderTemrs\]).
The functions $ \tilde{\mbox{s}}_{i}(\rho) $, $ \tilde{\mbox{k}}_{i}(\rho) $, $ \tilde{\mbox{w}}_{i}(\rho) $ and $ \tilde{\pi}_{i}(\rho) $ are determined by solving (\[E:SOL\]). We have not written the explicit form of the sources $ \mathbf{S}^{(2)} $, $ \mathbf{k}^{(2)} $ and $ \mathbf{V}_{\mu}^{(2)} $. The sources for the second order contribution to $ \Pi_{\mu\nu} $, $ \mathbf{P}_{\mu\nu}^{(2)}, $ are given in Appendix \[appB\].
Once the functions $ \tilde{\mbox{s}}_{i}(\rho) $, $ \tilde{\mbox{k}}_{i}(\rho) $, $ \tilde{\mbox{w}}_{i}(\rho) $ and $ \tilde{\pi}_{i}(\rho) $ are determined, we can expand them around the asymptotic boundary, $$\begin{aligned}
\label{E:Expansion}
\begin{split}
S & = S_{0}+\frac{b^{2}}{\rho^{2}}\left(\mbox{s}_{4}\,\mathfrak{S}_{4}
+\mbox{s}_{5}\,\mathfrak{S}_{5}\right)\\
k & = -1+\frac{\kappa_{4}}{\rho^{4}}
+b \frac{2S_{0}}{3\rho}\,\partial_{\alpha}u^{\alpha}
+\frac{b^{2}}{\rho^{2}}\left(\mbox{k}_{0}\,\mathbf{s}_{3}
+\mbox{k}_{1}\,\mathfrak{S}_{1}+\mbox{k}_{3}\,\mathfrak{S}_{3}
+\mbox{k}_{4}\,\mathfrak{S}_{4}+\mbox{k}_{5}\,\mathfrak{S}_{5}\right)\\
V_{\mu} & = -b\frac{S_{0}}{\rho}\, u^{\alpha}\partial_{\alpha}u_{\mu}
+\frac{b^{2}}{\rho^{2}}\left(\mbox{w}_{4}\,\mathbf{v}_{4\,\mu}
+\mbox{w}_{5}\,\mathbf{v}_{5\,\mu}+\mbox{v}_{1}\,\mathfrak{V}_{1\,\mu}
+\mbox{v}_{2}\,\mathfrak{V}_{2\,\mu}+\mbox{v}_{3}\,\mathfrak{V}_{3\,\mu}\right)\\
\Pi_{\mu\nu} & = b\left(\frac{\pi_{1}}{\rho}+\frac{\pi_{4}}{\rho^{4}}\right)\sigma_{\mu\nu}
+b^{2}\sum_{i=0}^{3}\left\{ \left(\frac{\tilde{\pi}_{2\, i}}{\rho^{2}}
+\frac{\tilde{\pi}_{4\, i}}{\rho^{4}}\right)\Sigma_{\mu\nu}^{\left(i\right)}\right\} \, ,
\end{split}\end{aligned}$$ where $$\begin{aligned}
\label{E:element4}
\kappa_{4}= \frac{S_{0}^{2}}{2S_{0}^{2}-1},\quad
\pi_{1}=S_{0},\quad
\pi_{4}=-\frac{S_{0}}{4} \left(2S_{0}^{2}-1\right)\end{aligned}$$ are found using the zeroth (\[0th\_order\_functions\]) and first (\[1st\_order\_functions\]) order metric functions. The coefficients $ \pi_{i} $, $ \tilde{\pi}_{2\, i} $ and $ \tilde{\pi}_{4\, i} $ can be obtained by expanding (\[E:SOL\]) near the boundary located at $ \rho \rightarrow \infty $. In particular, we find that $$\begin{aligned}
\label{E:Pi_tilde}
\begin{split}
\tilde{\pi}_{4\, i} &= \frac{S_{0}^{2} \left( 2 S_{0}^{2} - 1 \right)}{24}
\Bigg\{ \frac{6 \mathbf{P}_{i}^{(2)}(t)}{t} + 5 \mathbf{P}_{i}^{(2) \prime}(t) + t \, \mathbf{P}_{i}^{(2) \prime\prime}(t) \\
& \qquad\qquad\qquad\qquad + \int_t^0 \left( \frac{6 \mathbf{P}_{i}^{(2) \prime}(z)}{z} + 6 \mathbf{P}_{i}^{(2) \prime\prime}(z)
+ z \mathbf{P}_{i}^{(2) \prime\prime\prime}(z)\right) dz \Bigg\} \, ,
\end{split}\end{aligned}$$ where each $ \mathbf{P}_{i}^{(2)} $ is a part of the following decomposition $$\label{2orderPIsources}
\mathbf{P}^{(2)}_{\mu\nu} = \mathbf{P}_0^{(2)} \Sigma^{(0)}_{\mu\nu} + \mathbf{P}_1^{(2)} \Sigma^{(1)}_{\mu\nu} + \mathbf{P}_2^{(2)} \Sigma^{(2)}_{\mu\nu} + \mathbf{P}_3^{(2)} \Sigma^{(3)}_{\mu\nu} \, .$$ The explicit form of the source terms $ \mathbf{P}_i^{(2)} $ ($ i=0,\ldots,3 $) appear in the Appendix \[appB\]. Inserting (\[E:Expansion\]) into (\[E:Tmnprescription\]) we find $$\label{E:Tmn}
T^{\mu\nu} = P \left(4 u^{\mu}u^{\nu} + \eta^{\mu\nu}\right) - \eta \sigma^{\mu\nu} + \sum_{i=0}^{3} \lambda_{i} \Sigma^{(i)\mu\nu} \, ,$$ where $ P $ and $ \eta $ were determined in (\[zeroth\_order\]) and (\[eta\]) respectively and $$\label{E:TranspCoeffPrescription}
\lambda_{i} = \frac{\tilde{\pi}_{4\, i}\left(2S_{0}^{2}-1\right)}{4\pi b^2 G_{5}S_{0}} \, .$$
We have not managed to reduce (\[E:Pi\_tilde\]) to a closed form expression. We do note, however, that $$-2 \mathbf{P}_0^{(2)} + 4 \mathbf{P}_1^{(2)} - \mathbf{P}_2^{(2)} = \frac{d}{d\rho}\bm{\mathcal{S}}(\rho) \, ,$$ where $$\begin{aligned}
\label{totalderiv}
\begin{split}
\bm{\mathcal{S}}(\rho) =&
-\frac{8\theta\left(2S_{0}^{4}+S_{0}^{2}\rho^{4}-4\theta\rho^{4}f\left(\rho\right)\right)}
{S_{0}^{2}\left(1-8\theta\right)\rho^{2}\left(S_{0}^{2}-4\theta f\left(\rho\right)\right)}
+\frac{4\left(S_{0}^{2}-S_{0}^{2}\rho^{3}-4\theta f\left(\rho\right)\right)}
{S_{0}\left(S_{0}^{2}-4\theta f\left(\rho\right)\right)}\pi\left(\rho\right) \\
&-\frac{8\theta\rho\left(2S_{0}^{2}-f\left(\rho\right)\right)\left(S_{0}^{2}\left(1+\rho^{3}\right)
-4\theta f\left(\rho\right)\right)}{S_{0}^{3}\left(1-8\theta\right)\left(S_{0}^{2}-4\theta f\left(\rho\right)\right)}\pi'\left(\rho\right) \, .
\end{split}\end{aligned}$$ Plugging (\[totalderiv\]) into (\[E:SOL\]) and then using (\[E:TranspCoeffPrescription\]), we find that $$\nonumber
- 2 \lambda_0 + 4 \lambda_1 - \lambda_2 = \frac{\pi^{2}T^{2}}{\pi G_{5}}\left[\frac{\sqrt{2}\left(1-\sqrt{1-8\theta}+16\theta\right)\theta}{\left(1+\sqrt{1-8\theta}\right)^{3/2}}\right] \,,$$ as advertised in . This is the main result of our paper.
We note in passing that the authors of [@Andrei] have found closed form expressions for the $ \lambda_i $’s. We have compared our integral expression (\[E:Pi\_tilde\]) with their results both numerically and analytically and found excellent agreement. For instance, expanding (\[E:Pi\_tilde\]) at small $ \theta $ we get $$\begin{aligned}
\nonumber
\begin{split}
\lambda_{0} & = \frac{\pi^2 T^2}{32 \pi G_5} \left[2-\log 2 + \left(-21+5 \log 2\right) \theta +
\left(5+\frac{21}{2}\log 2\right) \theta^2 + \mathcal{O}(\theta^3) \right]\\
\lambda_{1} & = \frac{\pi^2 T^2}{32 \pi G_5} \left[1 - 7 \theta + \frac{151}{2} \theta^2 + \mathcal{O}(\theta^3) \right]\\
\lambda_{2} & = \frac{\pi^2 T^2}{32 \pi G_5} \left[2 \log 2 + \left(14-10 \log 2\right) \theta -
\left(28+21\log 2\right) \theta^2 + \mathcal{O}(\theta^3) \right]\\
\lambda_{3} & = \frac{\pi^2 T^2}{32 \pi G_5} \left[- 112 \theta - 80 \theta^2 + \mathcal{O}(\theta^3) \right] \, ,
\end{split}\end{aligned}$$ with $$\nonumber
- 2 \lambda_0 + 4 \lambda_1 - \lambda_2 = \frac{\pi^2 T^2}{32 \pi G_5} 320 \, \theta^2 + \mathcal{O}(\theta^3) \, .$$
Acknowledgements {#acknowledgements .unnumbered}
================
I would like to thank A. Yarom for useful discussions and collaboration. This work was supported by the ISF under grant numbers 495/11, 630/14 and 1981/14, by the BSF under grant number 2014350, by the European commission FP7, under IRG 908049 and by the GIF under grant number 1156/2011.
Hawking black hole temperature
==============================
To find the black-hole temperature we convert our ansatz (\[eq:ansatz0\]) into a non-zero temperature black 3-brane metric. With $ dt \rightarrow dt - \frac{S_0}{r^2 f(br)}dr $, one gets $$ds^2 = -r^2 f(br) dt^2 + \frac{dr^2}{r^2 f(br)/S_0^2} + r^2 \left(dx^i\right)^2 \,.$$ Since the Hawking temperature is completely determined at the vicinity of horizon, we expand the metric functions near $ r=r_h $ (in our case $ r_h = b^{-1} $ with $ f(b r_h) = 0 $) $$\nonumber
ds^2 = -\left. (r^2 f(br))'\right|_{r=r_h} (r-r_h) dt^2 +
\frac{S_0^2 \, dr^2}{\left. (r^2 f(br))'\right|_{r=r_h} (r-r_h)} + r^2 \left(dx^i\right)^2 \, .$$ Defining $ r-r_h = z^2 \frac{\left. (r^2 f(br))'\right|_{r=r_h}}{4 S_0^2} $ and using the zeroth order solution (\[0th\_order\_functions\]), we find $$\begin{aligned}
\nonumber
\begin{split}
ds^2 & = - \frac{4 a^{1/2} S_0^2}{b^2} z^2 dt^2 + dz^2 + r^2 \left(dx^i\right)^2 \, .
\end{split}\end{aligned}$$ After a Wick rotation to Euclidean time $ \tau $, we have (without the last term) a flat space metric in cylindrical coordinates $ ds^2 = dz^2 + z^2 d\varphi^2 $ where $ \varphi = \frac{2 S_0}{b} \tau $. To avoid a conical singularity at $ z=0 $, $ \varphi $ must have a periodicity of $ 2\pi $. Since periodicity of Euclidean time is the inverse temperature, we have $ 2\pi = \frac{2 S_0}{bT} $ and therefore $$\label{TEMP}
T = \frac{S_0}{b \pi} \, .$$
Second order source terms {#appB}
=========================
Source terms for the second order $ \Pi_{\mu\nu}^{(2)} $ metric functions, see (\[2orderPIsources\]) and (\[E:SOL\]).
$$\begin{aligned}
\nonumber
\begin{split}
\mathbf{P}_0^{(2)} & = \frac{S_0^6 \left(24 \, \theta + (1-8 \theta) \rho^4\right)}{\rho^3 \left(S_0^2-4\theta f(\rho)\right)^3}
+ \frac{S_0^5 \left(-40 \, \theta - 3 (1-8\theta)\rho^4\right)}{\rho^2 \left(S_0^2-4\theta f(\rho)\right)^3} \, \pi(\rho) + \frac{-2 S_0 \rho^3}{S_0^2-4\theta f(\rho)} \pi'(\rho)
\end{split}\end{aligned}$$
$$\begin{aligned}
\nonumber
\mathbf{P}_1^{(2)} = &\frac{\rho}{S_0^{2}}\left(\frac{-4\theta}{1-8\theta}
+\frac{-2 S_0^{8}\left(1-8\theta\right)}{\left(S_0^{2}-4\theta f(\rho)\right)^{3}}
+\frac{3 S_0^{4}}{S_0^{2}-4\theta f(\rho)}\right) +\frac{-S_0^{5}\left(40\theta+3\left(1-8\theta\right)\rho^{4}\right)}
{\rho^{2}\left(S_0^{2}-4\theta f(\rho)\right)^{3}} \pi(\rho) \\
&+\frac{2\rho^{3}}{S_0}\left(\frac{-2}{1-8\theta}
+\frac{-S_0^{6}\left(1-8\theta\right)}{\left(S_0^{2}-4\theta f(\rho)\right)^{3}}
+\frac{3S_0^{2}}{S_0^{2}-4\theta f(\rho)}\right)\pi'(\rho) \\
&+\frac{\rho^{5}\left(-16S_0^{8}\theta+f(\rho)\left(S_0^{6}-4\theta f(\rho)\left(S_0^{2}\left(1-16\theta\right)
+4\theta f(\rho)\right)\left(3S_0^{2}-8\theta f(\rho)\right)\right)\right)}{S_0^{2}\left(1-8\theta\right)\left(S_0^{2}
-4\theta f(\rho)\right)^{3}}\left(\pi'(\rho)\right)^{2}\\
&+\frac{\rho^{4}}{S_0}\left(\frac{-1}{1-8\theta}+\frac{S_0^{2}}{S_0^{2}-4\theta f(\rho)}\right)\pi''(\rho)
+\frac{-4\theta\rho^{6}f(\rho)\left(2S_0^{2}-f(\rho)\right)}{S_0^{2}\left(1-8\theta\right)\left(S_0^{2}
-4\theta f(\rho)\right)}\pi'(\rho)\pi''(\rho)\end{aligned}$$
$$\begin{aligned}
\nonumber
\begin{split}
\mathbf{P}_2^{(2)} & = \frac{2 S_0^2 \rho}{S_0^2-4\theta f(\rho)}
+ \frac{2 S_0^5 \left(40 \, \theta + 3 (1-8\theta)\rho^4\right)}{\rho^2 \left(S_0^2-4\theta f(\rho)\right)^3} \, \pi(\rho) + \frac{4 S_0 \rho^3}{S_0^2-4\theta f(\rho)} \pi'(\rho)
\end{split}\end{aligned}$$
$$\begin{aligned}
\nonumber
\begin{split}
\mathbf{P}_3^{(2)} & = \frac{S_0^2 \left(24 \, \theta +(1-8 \theta) \rho^4\right)}{\theta (1-8 \theta) \rho^3}
+ \frac{S_0^6 \, (1-8\theta)\, (3 S_0^2 - 16\theta) \rho}{\theta \left(S_0^2-4\theta f(\rho)\right)^3}
+ \frac{4 S_0^2 \, (6\theta - S_0^2) \rho}{\theta \left(S_0^2-4\theta f(\rho)\right)}
\end{split}\end{aligned}$$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the chemo-dynamical evolution of elliptical galaxies and their hot X-ray emitting gas using high-resolution cosmological simulations. Our Tree N-body/SPH code includes a self-consistent treatment of radiative cooling, star formation, supernovae feedback, and chemical enrichment. We present a series of ${\rm \Lambda}$CDM cosmological simulations which trace the spatial and temporal evolution of heavy element abundance patterns in both the stellar and gas components of galaxies. X-ray spectra of the hot gas are constructed via the use of the [vmekal]{} plasma model, and analysed using XSPEC with the XMM EPN response function. Simulation end-products are quantitatively compared with the observational data in both the X-ray and optical regime. We find that radiative cooling is important to interpret the observed X-ray luminosity, temperature, and metallicity of the interstellar medium of elliptical galaxies. However, this cooled gas also leads to excessive star formation at low redshift, and therefore results in underlying galactic stellar populations which are too blue with respect to observations.'
author:
- 'Daisuke Kawata and Brad K. Gibson'
---
Introduction
============
The hot X-ray emitting gas of elliptical galaxies represents an important interface between galaxies and the intergalactic medium (perhaps even the [*primary*]{} interface). The X-ray halos of ellipticals carry with them two fundamental mysteries:
- their X-ray luminosities are lower than that expected from an extrapolation of the cluster X-ray luminosity-temperature (${\rm L_X}-{\rm T_X}$) relation (e.g. Matsushita et al. 2000).
- their X-ray metallicities are lower than that of the mean stellar iron abundance (the so-called “iron discrepancy” - e.g. Arimoto et al. 1997 - a “discrepancy” in the sense that the halo gas metallicity was expected to exceed that of the stars, since it should bear the pollution of the enrichment from earlier generation of stars - enrichment byproducts that were not locked up into subsequent stellar generations).[^1]
Conversely, the optical properties of ellipticals appear less contentious! The Colour-Magnitude Relation (CMR) and Fundamental Plane provide strong constraints for any elliptical galaxy formation paradigm. We present here our preliminary work aimed ultimately at the construction of successful self-consistent optical $+$ X-ray cosmological chemodynamical simulations of elliptical galaxies.
Methods
=======
In our simulations, the dynamics of collisionless dark matter and stars is calculated using a gravitational Tree N-body code, and the gas component is modeled using Smoothed Particle Hydrodynamics (SPH). We calculate radiative cooling, star formation, chemical enrichment, and supernovae (SNe) feedback, self-consistently, and take into account both Type Ia and Type II SNe. We assume that SNe feedback is released as thermal energy. Details of the code are presented in Kawata (2001) and Kawata & Gibson (2003, in prep).
We have carried out a series of high-resolution simulations within the adopted standard $\Lambda$CDM cosmology ($\Omega_0$=0.3, $\Lambda_0$=0.7, $\Omega_{\rm b}$=0.019$h^{-2}$, $h$=0.7, and $\sigma_8$=0.9). Gas dynamics and star formation are included only within the relevant high-resolution region ($\sim$12 Mpc at $z$=0); the surrounding low-resolution region ($\sim$43 Mpc) contributes to the high-resolution region only through gravity. The mass of individual gas particles in the high-resolution region was $5.9\times10^7$ ${\rm M}_\odot$. We next identified an appropriate elliptical galaxy analog in the high-resolution region, which acts as the focus for this preliminary study. The total virial mass of this target galaxy is $2\times10^{13}$ ${\rm M_\odot}$, similar in size to that of NGC 4472, a bright elliptical galaxy in the Virgo Cluster. The target galaxy is relatively isolated, with only a few low-mass satellites remaining at $z$=0.
Figure \[evol-fig\] shows the morphological evolution of dark matter in the simulation volume, and the evolution of the stellar component in a 200 kpc region centred on the target galaxy. The galaxy forms through conventional hierarchical clustering between redshifts $z$=3 and $z$=1; the morphology has not changed dramatically since $z$=1. Three different radiative cooling and SNe feedback models were considered: Model A is an adiabatic model (i.e. no cooling); Model B includes cooling and weak feedback; Model C mimics Model B, but incorporates stronger feedback (100 times larger thermal energy per supernova).
For all the models, we examine both the resulting X-ray [*and*]{} optical properties, comparing them quantitatively with observation. The gas particles in our simulations carry with them knowledge of the density, temperature, and abundances of various heavy elements in their immediate vicinity. Using the XSPEC [vmekal]{} plasma model, we derive the X-ray spectra for each gas particle, and synthesize them within the assumed aperture (R$\sim$20 kpc). We next generate “fake” spectra with the response function of XMM EPN detector, assuming an exposure time (40 ks) and target galaxy distance (17 Mpc). Finally, our XPSEC fitting provides the X-ray weighted temperatures and abundances of various elements. Conversely, the simulated star particles each carry their own age and metallicity “tag”, which enables us to generate an optical-to-near infrared spectral energy distribution for the target galaxy, when combined with our population synthesis code adopting simple stellar populations of Kodama & Arimoto (1997).
Results and Conclusion
======================
The left panel of Figure \[lxt-fig\] shows the predicted ${\rm L_X-T_X}$ relation for the three models at $z$=0; crosses with error bars represent the observational data from Matsushita et al. (2000). The adiabatic model (Model A) appears incompatible with the data due to its excessive luminosity and low temperature. The inclusion of radiative cooling leads to lower luminosities and higher temperatures - as a result, models with cooling (Models B and C) are (roughly) consistent with the ${\rm L_X-T_X}$ relation of the observed elliptical galaxies. These conclusions are consistent with the analysis of Muanwong et al. (2001). The right panel of Figure \[lxt-fig\] shows the effect of cooling more clearly. In the gas density versus temperature diagrams, the region above than the line corresponds to a parameter space in which the cooling time is shorter than the Hubble time. Cooling ensures the gas within this region is cold (ie. non X-ray emitting), and of low density and high temperature, ensuring that ultimately radiative cooling drives the observed ${\rm L_X-T_X}$ relation.
The left panel of Figure \[lxfecmr-fig\] compares the X-ray weighted iron abundance of our simulations with the observational data of Matsushita et al. (2000). As the adiabatic model (by construction) does not form any stars (not having any cooling!), we show the results only for Models B and C. Both these models show lower gas-phase iron abundance, compared to their stellar abundance, consistent with the low iron abundances observed in the X-ray emitting gas of ellipticals. We find that a large fraction of iron ejected from stars is locked into future generation of stars. Stars preferentially enrich the gas in the central region, where cooling is efficient (right panel of Figure \[lxt-fig\]). The enriched gas can then cool easily and be incorporated into future generations of stars. Consequently, the hot gaseous halo has not been enriched efficiently, leading to a lower X-ray weighted iron abundance.
In summary, our radiative cooling models explain the two X-ray “mysteries” alluded to in Section 1. Having said that, any successful scenario must also explain the optical properties of the underlying stellar component. To this end, we examined the position of our simulated target galaxy in the observed Coma cluster CMR (Bower et al. 1992). We can see immediately that the colours of the resulting stellar components of both Models B and C are inconsistent with the data (being too blue). This inconsistency can be traced to an excessive population of young and intermediate age stars ($<$8 Gyr) which form from successively cooled gas, regardless of the strength of SNe feedback. If the contribution of these young stars was ignored, the resulting colours would match the observed CMR. Therefore, one exotic (if somewhat [*ad hoc*]{}) solution is to “hide” these younger stars within a bottom-heavy initial mass function (IMF) such that they cannot be seen today even if they did exist (e.g. Fabian et al. 1982; Mathews & Brighenti 1999). Another (more plausible) possibility is that extra heating sources, such as intermittent AGN activity, suppress star formation at low redshift. Before suggesting this is the true solution though, we must re-examine the predicted X-ray properties of the simulation end-products after introducing these additional heating sources; we will be pursuing this comparison in a future paper.
Our cosmological chemodynamical code makes it possible to undertake quantitative comparisons between numerical simulations and observational data in both the X-ray and optical regime with minimal assumptions. We find that radiative cooling is required to explain the observed X-ray luminosity, temperature, and metallicity of elliptical galaxies. Unfortunately, the resulting cooled gas also leads to unavoidable overproduction of young and intermediate age stellar populations, at odds with the observational constraints. Although a bottom-heavy IMF is one solution for this problem, extra heating by intermittent AGN activity seems to be more plausible (e.g. Brighenti & Mathews 2002); recent observations are consistent with this latter picture (e.g. Churazov et al. 2001).
Acknowledgements {#acknowledgements .unnumbered}
================
We acknowledge the support of the Australian Research Council through the Large Research Grant Program (A0010517) and Swinburne University through the Research Development Grants Scheme.
[1]{} Arimoto, N., Matsushita, K., Ishimaru, Y., et al., 1997, ApJ, 477, 128 Bower, R.G., Lucey, J..R., & Ellis, R.S., 1992, MNRAS, 254, 589 Brighenti, F., & Mathews, W.G., 2002, ApJL, 574, L11 Buote, D.A., Fabian, A.C., 1998, MNRAS, 296, 977 Churazov E., Brüggen, M., Kaiser, C.R., et al., 2001, ApJ, 554, 261 Edge, A.C., & Stewart, G.C., 1991, MNRAS, 252, 414 Fabian, A.C., Nulsen, P.E.J., & Canizares, C.R., 1982, MNRAS, 201, 933 Kawata, D., 2001, ApJ, 558, 598 Kodama, T., & Arimoto, N., 1997, A&A, 320, 41 Matsushita, K., Ohashi, T., & Makishima, K., 2000, PASJ, 52, 685 Mathews, W.G., & Brighenti, F., 1999, ApJ, 526, 114 Muanwong, O., Thomas, P.A., Kay, S.T., et al., 2001, ApJL, 552, L27 Sakelliou, I. et al., 2002, A&A, 391, 903 Xu, H. et al., 2002, ApJ, 579, 600
[^1]: X-ray iron abundances remain a controversial issue (c.f. Buote & Fabian 1997), although the iron discrepancy appears to hold based upon recent high-resolution XMM RGS observations (Xu et al. 2002; Sakelliou et al. 2002).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '[**A large variety of interacting complex systems are characterized by interactions occurring between more than two nodes. These systems are described by simplicial complexes. Simplicial complexes are formed by simplices (nodes, links, triangles, tetrahedra etc.) that have a natural geometric interpretation. As such simplicial complexes are widely used in quantum gravity approaches that involve a discretization of spacetime. Here, by extending our knowledge of growing complex networks to growing simplicial complexes we investigate the nature of the emergent geometry of complex networks and explore whether this geometry is hyperbolic. Specifically we show that an hyperbolic network geometry emerges spontaneously from models of growing simplicial complexes that are purely combinatorial. The statistical and geometrical properties of the growing simplicial complexes strongly depend on their dimensionality and display the major universal properties of real complex networks (scale-free degree distribution, small-world and communities) at the same time. Interestingly, when the network dynamics includes an heterogeneous fitness of the faces, the growing simplicial complex can undergo phase transitions that are reflected by relevant changes in the network geometry.** ]{}'
author:
- Ginestra Bianconi
- Christoph Rahmede
title: Emergent Hyperbolic Network Geometry
---
INTRODUCTION {#introduction .unnumbered}
============
Simplicial complexes are the many-body generalization of networks [@BA; @SW; @Doro_book; @Newman_book; @Santo; @Laszlo_book] and they can encode interactions occurring between two or more nodes [@Bassett; @flavor; @Kahle; @CQNM; @Emergent; @Newman_hypergraph; @Dima; @Owen]. While networks are formed exclusively by nodes and links, simplicial complexes include higher dimensional simplices i.e. triangles, tetrahedra etc. As such they are fundamental to study a large variety of real complex interacting systems, including brain functional networks [@Bassett], protein interaction networks [@proteins], collaboration networks [@Newman_hypergraph]. Because simplices have a natural topological and geometrical interpretation, simplicial complexes are ideal to investigate the underlying geometry and topology of networks [@perspective; @Bassett; @Vaccarino1; @Vaccarino2] and for these reasons they are extensively used in quantum gravity [@Lee; @CDT1; @CDT2; @Oriti; @Cortes].
One of the fundamental quests of quantum gravity is to describe the emergence of a continuous, finite dimensional space, using pre-geometric models, where space is an emergent property of a network or of a simplicial complex [@pregeometry_review; @CDT1; @CDT2; @graphity]. This fundamental mathematical problem has its relevance also in the field of network theory [@perspective] where one of the major aim of network geometry is to characterize the continuous hidden metric behind the inherently discrete structure of complex networks. In fact, it is believed that most complex networks have a continuous hidden network geometry [@Boguna_Internet; @Boguna_navigability; @Boguna_metabolic; @Boguna_growing; @Boguna_hyperbolic] such that any two connected nodes are also close in the hidden metric [@Kleinberg; @Aste; @Boguna_navigability; @Boguna_Internet]. In this context, there is increasing evidence that the hidden geometry of a large variety of networks including the Internet, airport networks, the brain functional networks, and metabolic networks [@Boguna_Internet; @Boguna_navigability; @Boguna_metabolic; @Boguna_growing] is hyperbolic. Characterizing the hyperbolicity of networks is not only a fundamental theoretical question, but it can also have practical implications as it can be used to improve significantly the navigability on such networks [@Boguna_navigability; @Boguna_Internet]. While the mathematical definition of the curvature of networks is a hot mathematical subject for which different definitions have been given [@perspective; @Yau1; @Yau2; @Gromov; @Jost1; @Jost2], most of the results obtained so far are related to the embeddings of complex networks in hyperbolic spaces [@Aste; @Boguna_navigability; @Boguna_Internet; @Boguna_metabolic; @Boguna_growing].
The underlying assumptions of several models [@Boguna_hyperbolic; @Boguna_growing] of complex hyperbolic networks is that nodes are sprinkled randomly in the hidden hyperbolic metric and links are established according to their hyperbolic distance. Interestingly this type of models can be related to causal sets [@Sorkin] in de Sitter space and they have been used to describe a “network cosmology” [@Network_cosmology]. From the complexity point of view, if we want for example to use this type of models for describing the evolution of the World-Wide-Web, the sprinkling of the nodes in hyperbolic space reflects some distribution of interest of the webpage owners, and links between the webpages are established depending on the similarities between the interests of webpage owners. Although this is a very plausible mechanism for network evolution, it cannot be adopted if we aim at describing the emergence of the underlying hyperbolic geometry as the result of the endogenous dynamics of the network.
Here we will show in the framework of a very simple, stylized model, that the hyperbolic network geometry can be an emergent property of growing simplicial complexes that share the universal properties of complex network structures. Specifically, we will propose a model in which the hidden hyperbolic metric is not causing the network dynamics but it is instead the outcome of the network evolution.
Our model of emergent geometry is based on a growing simplicial complex. Metric spaces satisfy the triangular inequality, therefore a network with non-trivial geometry should include a high clustering coefficient and high density of triangles, ensured by building the network using simplicial complexes. Additionally growing networks have been extensively used as a non-equilibrium framework [@BA; @Doro_book; @Doro_model; @Krapivsky; @BB; @BE] for the emergence of complex statistical properties of networks such as the power-law degree distribution. By extending the well established framework of growing network models [@BA; @Doro_book; @Doro_model; @Krapivsky; @BB; @BE] to simplicial complexes we will provide significant new insights into emergent geometry. Importantly, the emergent hidden geometry of growing simplicial complexes is hyperbolic, i.e. the hyperbolic geometry emerges spontaneously from the evolution of the simplicial complexes. In this way we provide evidence that hyperbolic network geometry emerges from growing simplicial complexes whose temporal evolution is purely combinatorial, i.e. it does not take into account the hidden geometry.
Interestingly the properties of the network geometry change significantly with the dimension of the simplicial complex and the network geometry can be strongly affected by phase transitions occurring when a fitness parameter [@BB; @BE; @CQNM; @flavor] is associated to each face of the simplicial complexes describing intrinsic local heterogeneities.
RESULTS {#results .unnumbered}
=======
We consider simplicial complexes formed by gluing together $d$-dimensional simplices. A $d$-dimensional simplex (or $d$-simplex) is a topological object including the set of $d+1$ nodes and all its subsets. The underlying network structure of a $d$-simplex is constituted by a fully connected network, or a clique, of $d+1$ nodes, such as links ($1$-simplices), triangles ($2$-simplices), tetrahedra ($3$-simplices) etc. The $\delta$-faces of a $d$-dimensional simplex are all the $\delta$-dimensional simplices that can be built by a subset of $(\delta+1)$ of its nodes. For example the faces of a triangle (2-simplex) are its three links (1-simplices) and its three nodes (0-simplices), the faces of a tetrahedron are its four triangular faces (2-simplices), its six links (1-simplices) and its four nodes (0-simplices), etc. As long as we are concerned exclusively with the network properties of simplicial complexes, the use of simplicial complexes is equivalent with the use of hypergraphs and hypernetworks that are recently attracting increasing attention [@Newman_hypergraph; @Coutinho].
The simplicial complexes that we are considering in this paper are constructed by gluing $d$-simplices along their $(d-1)$-faces. To every $(d-1)$-face $\alpha$ of the simplicial complex, (i.e. a link for $d=2$, or a triangular face for $d=3$) we associate an [*incidence number*]{} $n_{\alpha}$ given by the number of $d$-dimensional simplices incident to it minus one. The simplicial complex dynamics is dictated by the following algorithm and depends on a parameter $s=-1,0,1$ called [*flavor*]{}. We start from a single $d$-dimensional simplex, i.e a triangle for $d=2$, a tetrahedron for $d=3$. At each time we add a $d$-dimensional simplex to a $(d-1)$-face $\alpha$. The face $\alpha$ is chosen randomly with probability $\Pi_{\alpha}$ given by \_=. \[prob\] The new $d$-dimensional simplex is induced by a new node and all the nodes of the chosen $(d-1)$-face $\alpha$. For this type of dynamics, the combinatorial condition to obtain a discrete manifold is that $n_{\alpha}$ can take exclusively the values $n_{\alpha}=0,1$.
For $s=-1$ it is possible to attach a simplex only to faces with $n_{\alpha}=0$. In fact for $n_{\alpha}=0$ we have $\Pi_{\alpha}=\frac{1}{\sum_{\alpha'}1+s n_{\alpha'}}$ but for $n_{\alpha}=1$ we have $\Pi_{\alpha}=0$. As a consequence of this, the resulting network is a manifold, with each $(d-1)$-face incident at most to two $d$-dimensional simplicial complexes, i.e. $n_{\alpha}=0,1$. For $s=0$ the $d-$dimensional simplices are attached with uniform probability to any $(d-1)$-face, while for $s=1$ the dynamics follows a generalized preferential attachment and the new simplex is attached to a $(d-1)$-face $\alpha$ proportionally to the number of simplicies already attached to the face , i.e. $1+n_{\alpha}$. Therefore for $s=0$ as for $s=1$ the incidence number $n_{\alpha}$ can take values $n_{\alpha}=0,1,2,3\ldots$.
Simplices are topological objects that can be turned into geometrical entities when we attribute a given length to their links. Here, in order to describe the emergent geometry of our model of growing simplicial complexes, we assume that every simplicial complex is built by simplices that have links of equal length across the entire simplicial complex.
The resulting networks are small world for every flavor $s$ and any dimension $d$ except from the special case $s=-1,d=1$ in which the resulting network is a chain. This implies that the number of nodes in the network $N$ increases exponentially with its diameter $D$, i.e. $N\simeq e^{D}$. Therefore, if all the links have equal length, the hidden geometry of these networks cannot be the one of a Euclidean space of finite Hausdorff dimension $d_E$ because this would imply a power-law scaling $N\simeq D^{d_E}$. As a consequence of this the small-world property suggests that the natural embedding of these networks is hyperbolic. Nevertheless the small-world property might not be sufficient to guarantee an embedding in the hyperbolic space. Here we show that for our class of growing simplicial complexes the hidden geometry, corresponding to the embedding where all the links have the same distance, are the hyperbolic spaces ${\mathbb {H}}^d$, and specifically the Poincaré ball model [@Hyperbolic].
The great advantage of the present class of models with respect to general small-world networks is that their dual is a tree. The dual network can be constructed by associating to every $d$-simplex a node of the dual network, and to every pair of $d$-simplicies sharing a $(d-1)$-face a link of the dual network. Since in our simplicial complex evolution at each time we glue a new $d$-simplex to a $(d-1)$-face, the resulting structure of the dual network is a tree. Taking advantage of this simple structure of the dual the present class of models admits several embeddings in the ${\mathbb{H}}^d$ hyperbolic space model. Between the different possible embeddings, only one embedding can fill the entire space in the asymptotic limit $t\to \infty$. Therefore this embedding defines the emergent geometry of our simplicial complexes.
Let us consider a Poincaré ball model of $\mathbb{H}^d$. The Poincaré ball model includes all the points of the unit ball $B^n=\{{\bf x}\in\mathbb{R}^d: |{\bf x}|<1\}$, with $|\ldots|$ indicating the Euclidean norm. The Poicaré ball model is associated to the hyperbolic metric $d_B$ assigning to each pair of points ${\bf x,y}\in \mathbb{R}^d$ the distance d\_B([**x, y**]{})=1+. Here we identify every $d$-dimensional simplex of our simplicial complex with an ideal simplex of the Poincaré ball model. An ideal simplex has all its nodes at the boundary of the hyperbolic ball, so all the nodes $i$ have a position ${\bf r}_i\in \mathbb{R}^d$ satisfying $|{\bf r}_i|=1$. This allows to have all the nodes of each simplex at equal hyperbolic distance. Note that interestingly this distance is actually infinite but this is the “cost” required for having an embedding that asymptotically in time fills the entire hyperbolic space. In order to fully characterize the hidden geometry of the studied networks, we need also to determine further the position of the ideal nodes at the boundary of the ball. To this end we start from a $d$-dimensional simplex whose ($d+1$) nodes have the same maximum (Euclidean) distance from each other. Therefore the positions ${\bf r}_i$ of the initial $(d+1)$ nodes $i=1,2,\ldots d+1$ satisfy \_[i=1]{}\^[d+1]{}[**r**]{}\_i=[**0**]{}. Each new node $i$ of the network has a position $\bf{r}_i\in \mathbb{R}^d$ at the boundary of the ball determined by the position of its “ancestors”, i.e. the nodes of the face $\alpha$ connected to the new node. In particular the new node $i$ is placed at equal (Euclidean) distance from all the nodes $j$ of the $\alpha$ face to which it is attached,i.e. \_i=. In this way the angular position of the new node is fully determined by the stochastic dynamics of the network (see Figure \[Figure1\] for details).
![ The first steps of the temporal evolution of growing simplicial complex with $N$ nodes is shown here in the emergent hyperbolic space for $d=2$ (panel A) and $d=3$ (panel B). The flavor is $s=-1$.[]{data-label="Figure1"}](hyperbolic_combined.pdf){width="15cm"}
The resulting networks have a rich geometrical structure, which is linked to the mathematics of Farey sequences in $d=2$ [@Farey]. Additionally, the simplicial complexes in dimension $d=3$ are characterized by a boundary with notable geometrical features.The induced geometry on this boundary can be studied by placing the nodes of the network in a $(d-1)$-dimensional space characterized by the angular coordinates of the nodes. The network resulting from the projection on the boundary of the ball $B^n$ is a random Apollonian network [@apollonian; @apollonian2] for $d=3$.
Let us make three important observations related to the geometric nature of the proposed class of growing simplicial complexes.
First of all we note that the hyperbolic nature of the emergent geometry it is a consequence of the assumption that each link of the simplicial complex must have equal length. This assumption implies that distances of different links can be compared. Therefore strictly speaking here the network geometry is actually a consequence of a kind of “proto-geometry” that allows comparison of length of different links. If we allow, instead, to have links of different lengths the curvature of the hidden geometry is not determined and it is even possible to tune the length of the links such that the same simplicial complex can be embedded in a $d$-sphere. For instance this can be achieved simply by taking the embedding on the Poicaré ball model described above, and considering instead of the hyperbolic metric on the ball the Euclidean metric.
Secondly we note that the natural hyperbolic embedding of growing simplicial complexes that we discussed above, works particularly well for flavor $s=-1$ while some caution is required when using this embedding for flavors $s=0$ and $s=1$. In fact, for $s=-1$ the simplicial complexes are manifolds and as a consequence of this, links do not cross and each node of the simplicial complex has a distinct position in the hyperbolic space. However for flavor $s=0$ and $s=1$ the proposed embedding implies that some nodes of the simplicial complex (the nodes that are immediate “descendant” of the same face) acquire the same position in the hyperbolic space. As a result in a geometrical embedding links are effectively weighted. For this reason the growing simplicial complexes with flavor $s=-1$ play a very special role with respect to the other flavors $s=0$ and $s=1$.
Finally we note that for $s=-1$ and $d=3$ the growing simplicial complex model presented here belongs to the class of stacked polytopes that are equivalent to Apollonian packings, whose discrete Lorentzian geometry is raising recent interest in the mathematical community [@apollonian2; @Apollonian_group; @Chen_Hao1; @Chen_Hao2]. In fact, these stacked polytopes in $d=3$ have a symmetry group ${\cal G}$ that is a noncompact discrete subgroup of the Lorentz group SO(3,1)=SL(2,$\mathbb{C}$)/$\mathbb{Z}_2$. Therefore these results provide an additional important insight on the hyperbolic nature of the underlying geometry of the class of models proposed in this paper.
![[ The effect of dimensionality and flavor in the degree distribution $P_d^{[s]}(k)$. The symbols (blue dimonds) indicate the simulation result of a single realization of the growing simplicial complex with dimension $d$ and flavor $s$ with $N=10^5$ nodes. The solid lines indicate the theoretical predictions.]{} []{data-label="Figurepk"}](Pk.pdf){width="16cm"}
The networks resulting from the proposed model of growing simplicial complexes are strongly affected by their dimensionality. In fact they are scale-free for dimension d>(1-s), while for dimension $d=1-s$ they have exponential degree distribution. In particular, the degree distribution $P_d^{[s]}(k)$ of growing simplicial complexes of dimension $d$ and flavor $s$ is given for for $d+s=1$ by (see Methods for details) P\_[d]{}\^[\[s\]]{}(k)&=&()\^[k-d]{}, with $k\geq d$ while for $d+s>1$ it is given by (see Methods for details) P\_[d]{}\^[\[s\]]{}(k)&=&, \[Pksf0\] with $k\geq d.$\
Therefore for $d+s>1$ the degree distribution is scale-free and has a power-law scaling P\_[d]{}\^[\[s\]]{}(k)k\^[-\_d\^[\[s\]]{}]{} for $k\gg 1$, with power-law exponent $\gamma_d^{[s]}$ \_d\^[\[s\]]{}=2+3. Finally for $d+s=0$ the simplicial complexes reduce to chains (see Methods for details).\
In Figure $\ref{Figurepk}$ we show the perfect agreement between the predicted degree distribution and simulation results for $d=1,2,3$ and flavor $s=-1,0,1$. For dimensions $d>1$ these networks display a significant community structure (high modularity) and high average clustering coefficient (see Methods) as most complex networks. The values of the modularity and the clustering coefficient are modulated by the dimension $d$ and the flavor $s$ of the growing simplicial complex (see Table $\ref{table1}$).
The emergent geometry of growing simplicial complexes is strongly affected by phase transitions occurring in the network evolution. In order to show evidence for this statement we study a variation of the model including fitness of the faces of the simplicial complex. First we associate to each node an energy $\epsilon_i$ drawn from a $g(\epsilon)$ distribution. To any $\delta$-face $\alpha$ with $0<\delta<d$ we associate an energy \_=\_[i]{}\_i and a fitness \_=e\^[-\_]{}, where $\beta=1/T>0$ is a global parameter called [*inverse temperature*]{}, such that for $\beta=0$ all the nodes have the same fitness while for $\beta \gg 1$ small differences in energy yield big differences in the fitness of the nodes. The model remains the same with the exception that the probability $\Pi_{\alpha}$ that a $(d-1)$-face $\alpha$ is selected is no longer given by Eq. $(\ref{prob})$ but is given by \_=. \[prob2\] This model displays a structural phase transition at high value of $\beta$. In order to characterize this phase transition we make use of the hyperbolic embedding. While above the phase transition the simplicial complex is growing in all directions of the hyperbolic space and has statistical properties related to quantum statistics [@CQNM; @flavor], below the phase transition there is symmetry breaking and the network evolves asymmetrically in the hyperbolic space. In figure $\ref{Figure2}$ we show a visualization of the model above and below the phase transition for dimension $d=2,3$ showing that also the geometry of the boundary of simplicial complexes in $d=3$ is strongly affected by the geometrical phase transition occurring in the model.
![The hyperbolic emergent network geometry is shown here for the growing simplicial complex with fitness of the nodes and flavor $s=-1$ above and below the phase transition in dimension $d=2$ (panel a-b) and dimension $d=3$ (panel c-d). Panel (e-f) display the projection of the network on the boundary of the $d=3$ hyperbolic space. The energy distribution is uniform over discrete values of the energies of the nodes $0\leq\epsilon_i< 10$. The inverse temperature is $\beta=0.01$ for panels a,c,e, $\beta=50$ for panel b, $\beta=20$ for panels d,f. The number of nodes is $N=200$. The color of the links indicates the different values of their energies.[]{data-label="Figure2"}](figure_condensation_combined.pdf){width="12cm"}
![ The phase transition in the network geometry of the growing simplicial complexes with fitnesses is characterized here by showing $R,\sigma$ and the maximum degree $k_{max}$ as a function of $\beta$. Here simulation results are reported for growing simplicial complexes with flavor $s=-1$, dimension $d=2$ (left panels), and $d=3$ (right panels) and network sizes $N=2500$ (blue dashed line), $5000$ (red dot dashed line), $10000$ (green solid line). All the data are averaged over $500$ realizations.[]{data-label="Figure3"}](transition.pdf){width="9cm"}
In order to numerically study the phase transition occurring in the growing simplicial complexes with fitnesses we define a vector ${\bf R}$ given by =\_i [**r**]{}\_i, where ${\bf r}_i$ is the (Euclidean) position vector of the node $i$ in the Poincaré ball. We study the Euclidean norm $R=|{\bf R}|$ and the standard deviation $\sigma=\sqrt{\frac{1}{N}\sum_i |{\bf r}_i-{\bf R}|^2},$ and the maximum degree $k_{max}$ as a function of the inverse temperature $\beta$. Here we focus on the result for $d=2,3$, flavor $s=-1$ (see Figure \[Figure3\]). As $\beta$ increases across the phase transition, $\sigma$ develops relevant finite size effects and becomes vanishingly small in the large network limit. For the same values of the parameters $R$ approaches one indicating that the simplicial complex grows in a well defined preferential direction. The phase transition in the network geometry for flavor $s=-1$ has these characteristics both for $d=2$ and $d=3$. However the behavior of the maximum degree $k_{max}$ across the phase transition shows major differences between the case $d=2$ and $d=3$, displaying a clear maximum at the transition point for $d=3$. This implies that the transition might affect the degree distribution in different ways depending on the dimension $d$. Interestingly similar transitions are observed for different flavors $s=0,1$.
DISCUSSION {#discussion .unnumbered}
==========
In conclusion, this paper shows that the study of simplicial complexes allows great advances in our understanding of complex networks. In fact by extending the framework of growing complex networks to simplicial complexes our simple model produces networks that display most of the universal properties of complex networks including scale-free degree distribution, small-world properties and significant modular structure. These networks have statistical and geometrical properties that are a function of their dimension $d$ and their flavor $s$ that modulate the values of their modularity and their clustering coefficient. These non-equilibrium models of simplicial complexes are ideal frameworks to show the emergence of the hyperbolic network geometry. Specifically they can explain how real hyperbolic networks might result from purely combinatorial rules that do not take into consideration the hidden metric of the network.This network geometry has a very interesting structure linked to Farey sequences and Apollonian tilings. Additionally these network geometries can undergo relevant changes following phase transitions in the network evolution.
We believe that this paper, showing that growing simplicial complexes give rise to a complex emergent hyperbolic geometry, related to Apollonian packings, is opening new perspectives for understanding the origin of the emergent hyperbolic geometry of complex networks. On the same time, our hope is that further research in this direction could also indicate a path for establishing a cross-fertilization between network theory and quantum gravity.
METHODS {#methods .unnumbered}
=======
Degree distribution of growing simplicial complexes with flavor $s$ {#degree-distribution-of-growing-simplicial-complexes-with-flavor-s .unnumbered}
-------------------------------------------------------------------
In order to derive the degree distribution of the growing simplicial complexes with flavor $s$, we use the master equation approach [@Doro_book; @Newman_book]. It can be easily shown that the average number $\tilde{m}_{d}^{[s]}(k)$ of nodes of degree $k$ that at each time increase their degree by one is given by \_[d]{}\^[\[s\]]{}(k)=, \[pia111\] as long as $d+s\neq0$ , i.e. $(d,s)\neq (1,-1)$ for which the growing simplicial complex reduces to a chain. To derive the exact degree distribution of the simplicial complex, we consider the master equation for the average number of nodes $N_{d,}^{t,[s]}(k)$ that at time $t$ have degree $k$ in a growing $d$ dimensional simplicial complex of flavor $s$. The master equation [@Doro_book; @Newman_book] for $N_{d}^t(k)$ reads N\^[t+1, \[s\]]{}\_[d]{}(k)-N\^[t, \[s\]]{}\_[d]{}(k)=\_[d]{}\^[\[s\]]{}(k-1)N\_[d]{}\^[t,\[s\]]{}(k-1)(1-\_[k,d]{})-\_[d]{}\^[\[s\]]{}(k)N\^[t,\[s\]]{}\_[d]{}(k)+\_[k,d]{} with $k\geq d$.\
Solving this equation we get both exponential and power-law degree distribution. In particular, the degree distribution $P_d^{[s]}(k)$ of growing simplicial complexes of dimension $d$ and flavor $s$ is given for for $d+s=1$ by P\_[d]{}\^[\[s\]]{}(k)&=&()\^[k-d]{}, with $k\geq d$ while for $d+s>1$ it is given by P\_[d]{}\^[\[s\]]{}(k)&=&, \[Pksf0b\] with $k\geq d.$\
Therefore for $d+s>1$ the degree distribution is scale-free and has a power-law scaling P\_[d]{}\^[\[s\]]{}(k)k\^[-\_d\^[\[s\]]{}]{} for $k\gg 1$, with power-law exponent $\gamma_d^{[s]}$ \_d\^[\[s\]]{}=2+3.
Community structure of growing simplicial complexes {#community-structure-of-growing-simplicial-complexes .unnumbered}
---------------------------------------------------
The modularity $M$ evaluates the significance of the community structure of a network. It is defined [@Newman_book] as M=\_[ij]{}(a\_[ij]{}-)(q\_i, q\_j) , where ${\bf a}$ denotes the adjacency matrix of the network, $L$ the total number of links, and $\{q_i\}$, where $q_i=1, 2\ldots Q$, indicates to which community the node $i$ belongs. Finding the network community structure that optimizes modularity is NP hard. One of the most popular greedy algorithms to find the community structure is the generalized Louvain method [@Gen_louvain; @Louvain] that is able to determine a lower bound on the maximum modularity of the network. As shown in Table I of the main text the growing simplicial complexes with $d\geq 2$ are characterized by a large modularity.
Average clustering coefficient of growing simplicial complexes {#average-clustering-coefficient-of-growing-simplicial-complexes .unnumbered}
--------------------------------------------------------------
The local clustering coefficient $C_i$ measures the density of triangles passing through node $i$, i.e. C\_i= . where ${\bf a}$ denotes the adjacency matrix of the network, and where $k_i=\sum_j a_{ij}$ is the degree of node $i$.\
As shown in Table I of the main text the growing simplicial complexes with $d\geq 2$ are characterized by a large clustering coefficient.
Acknowledgments {#acknowledgments .unnumbered}
===============
- G. B. acknowledges interesting discussions with D. Krioukov, M. A. Serrano, L. Smolin and R. Sorkin.
- This work has been partially supported by SUPERSTRIPES Onlus. G.B. was partially supported by the Perimeter Institute for Theoretical Physics (PI). The PI is supported by the Government of Canada through Industry Canada and by the Province of Ontario through the Ministry of Research and Innovation.
Author Contributions Statement {#author-contributions-statement .unnumbered}
==============================
G. B. and C. R. designed the research, conducted the research and wrote the main manuscript text.
Additional Information {#additional-information .unnumbered}
======================
The authors declare that they have no competing financial interests.\
Correspondence to [email protected]\
[99]{} Barabási, A.-L. & Albert, R. Emergence of scaling in random networks. [*Science*]{} [**286**]{}, 509 (2009). Watts, D. J. & Strogatz, S. H. Collective dynamics of ‘small-world’networks. [*Nature* ]{}[ **393**]{}, 440 (1998). Dorogovtsev, S. N. & Mendes, J. F. F. Evolution of networks. [*Advances in physics*]{} [**51**]{}, 1079 (2002). Newman, M. E. J. [*Networks: An introduction*]{}. (Oxford University Press, 2010). Barabási, A.-L. [*Network science*]{}. (Cambridge University Press, 2016). Fortunato, S. Community detection in graphs. [*Physics Reports*]{} [**486**]{}, 75 (2010).
Giusti, C., Ghrist, R. & Bassett, D.S. Two’s company, three (or more) is a simplex: Algebraic-topological tools for understanding higher-order structure in neural data. arXiv preprint arXiv:1601.01704 (2016).
Kahle, M. Topology of random clique complexes. [*Discrete Mathematics*]{} [**309**]{}, 1658 (2009).
Bianconi, G. & Rahmede, C. Network geometry with flavor: from complexity to quantum geometry. [*Phys. Rev. E*]{} [**93**]{}, 032315 (2016).
Bianconi G., & Rahmede, C. Complex quantum network manifolds in dimension $d> 2$ are scale-free. [*Scientific Reports*]{} [**5**]{}, 13979 (2015).
Wu, Z., Menichetti, G., Rahmede, C. & Bianconi, G. Emergent complex network geometry. [*Scientific Reports*]{} [**5**]{}, 10073 (2014).
Ghoshal, G., Zlatić, V., Caldarelli G. & Newman M.E. J. Random hypergraphs and their applications. [*Phys. Rev. E*]{} [**79**]{}, 066118 (2009). Zuev, K., Eisenberg, O. & Krioukov D. Exponential random simplicial complexes. [*Jour. of Phys. A*]{} [**48**]{}, 465002 (2015). Courtney, O. T. & Bianconi, G. Generalized network structures: The configuration model and the canonical ensemble of simplicial complexes. [*Phys. Rev. E*]{} [**93**]{}, 062311 (2016). Wan, C., et al. Panorama of ancient metazoan macromolecular complexes. [*Nature*]{} [**525**]{}, 339 (2015).
Bianconi, G. Interdisciplinary and physics challenges of network theory. [*EPL (Europhysics Letters)*]{} [**111**]{}, 56001 (2015).
Petri, G., Scolamiero, M., Donato, I. & Vaccarino F. Topological strata of weighted complex networks. [*PloS one*]{} [**8**]{}, e66506 (2013). Petri, G., et al. Homological scaffolds of brain functional networks. [*Journal of The Royal Society Interface*]{} [**11**]{}, 20140873 (2014).
Rovelli, C., & Smolin, L. Spin networks and quantum gravity. [*Phys. Rev. D*]{} [**52**]{}, 5743 (1995).
Oriti, D. Spacetime geometry from algebra: spin foam models for non-perturbative quantum gravity. [*Reports on Progress in Physics*]{} [**64**]{}, 1703 (2001).
Ambjørn, J., Jurkiewicz, J. & Loll, R. Reconstructing the universe. [*Phys. Rev. D*]{} [**72**]{}, 064014 (2005). Ambjørn, J., Jurkiewicz, J. & Loll, R. Emergence of a 4D world from causal quantum gravity. [*Phys. Rev. Lett.*]{} [**93**]{}, 131301 (2004). Cortês, M. & Smolin, L. Quantum energetic causal sets. [*Phys. Rev. D*]{} [**90**]{}, 044035 (2014).
Gibbs, P. E. The small scale structure of space-time: A bibliographical review. arXiv preprint hep-th/9506171 (1995). Konopka, T., Markopoulou, F. & Severini, S. Quantum graphity: a model of emergent locality. [*Phys. Rev. D*]{} [**77**]{}, 104029 (2008).
Krioukov, D., et al. Hyperbolic geometry of complex networks. [*Phys. Rev. E*]{} [**82**]{} 036106 (2010). Papadopoulos, F., Kitsak, M., Serrano, M. A: Boguñá, M. & Krioukov, D. Popularity versus similarity in growing networks. [*Nature*]{} [**489**]{}, 537 (2012).
Serrano, M. A., Boguñá, M. & Sagués, F. Uncovering the hidden geometry behind metabolic networks. [*Molecular BioSystems*]{} [**8**]{}, 843 (2012).
Boguñá, M., Krioukov, D. & Claffy, K.C. Navigability of complex networks. [*Nature Physics*]{} [**5**]{}, 74 (2008). Boguñá, M., Papadopoulos, F. & Krioukov, D. Sustaining the internet with hyperbolic mapping. [*Nature Communications*]{} [**1**]{} 62 (2010). Kleinberg, R. Geographic routing using hyperbolic space. [*In INFOCOM 2007. 26th IEEE International Conference on Computer Communications. IEEE,*]{} pp. 1902-1909. IEEE, (2007). Aste, T., Di Matteo, T., & Hyde, S. T. Complex networks on hyperbolic surfaces. [*Physica A*]{} [**346**]{}, 20 (2005). Lin, Y., Lu,L. & Yau, S.-T. Ricci curvature of graphs. [*Tohoku Mathematical Journal*]{} [**63**]{}, 605 (2011). Lin, Y., & Yau, S.-T. Ricci curvature and eigenvalue estimate on locally finite graphs. [*Math. Res. Lett*]{} [**17**]{}, 343 (2010).
Gromov, M. [*Hyperbolic groups.*]{} (Springer, 1987).
Bauer, F., Jost, J. & Liu, S. Ollivier-Ricci curvature and the spectrum of the normalized graph Laplace operator. [*Math. Res. Lett.*]{} [**19**]{} 1185 (2012). Sreejith, R. P., Mohanraj, K., Jost, J., Saucan, E. & Samal, A. Forman curvature for complex networks. [*J. Stat. Mech.*]{} 063206 (2016).
Sorkin, R. D. Causal sets: Discrete gravity.In [*Lectures on quantum gravity*]{}, pp. 305-327 (Springer, 2005). Krioukov, D., et al. Network cosmology. [*Scientific Reports*]{} [**2**]{}, 793 (2012). Dorogovtsev, S. N., Mendes, J. F. F. & Samukhin, A.N. Size-dependent degree distribution of a scale-free growing network. [*Phys. Rev. E*]{} [**63**]{}, 062101 (2001).
Krapivsky, P. L., Redner, S. & Leyvraz, F. Connectivity of growing random networks. [*Phys. Rev. Lett.*]{} [**85**]{}, 4629 (2000). Bianconi, G., & Barabási A.L. Bose-Einstein condensation in complex networks. [*Phys. Rev. Lett.*]{} [**86**]{}, 5632 (2001). Bianconi, G., & Barabási A.-L. Competition and multiscaling in evolving networks. [*EPL (Europhysics Letters)*]{} [**54**]{}, 436 (2001). Coutinho, B. C. Zhou, H.-J. & Yang-Yu L. “Percolations on hypergraphs.” arXiv preprint arXiv:1605.00897 (2016).
Mucha, P. J., Richardson, T., Macon, K., Porter, M.A. & Onnela, J.-P. Community structure in time-dependent, multiscale, and multiplex networks. [*Science*]{} [**328**]{} 876 (2010).
Blondel, V. D., Guillaume, J. L., Lambiotte, R. & Lefebvre, E. Fast unfolding of communities in large networks. [*Jour. Stat. Mech. Theory and Experiment*]{} [**10**]{}, P10008 (2008).
Ratcliffe, J. [*Foundations of hyperbolic manifolds.*]{} Vol. 149. (Springer Science & Business Media, 2006).
Series, C. The modular surface and continued fractions. [*Journal of the London Mathematical Society*]{} [**2**]{}, 69 (1985).
Andrade, Jr J. S., Herrmann, H. J., Andrade, R. F.S. & Da Silva, L. R. Apollonian networks: Simultaneously scale-free, small world, euclidean, space filling, and with matching graphs. [*Phys. Rev. Lett.*]{} [**94**]{}, 018702 (2005). Andrade, R. F.S., & Herrmann, H. J. Magnetic models on Apollonian networks. [*Phys. Rev. E*]{} [**71**]{}, 056131 (2005). Söderberg, B. Apollonian tiling, the Lorentz group, and regular trees. [*Phys. Rev. A*]{} [**46**]{}, 1859 (1992).
Graham, R. et al. Apollonian circle packings: geometry and group theory I. The Apollonian group. [*Discrete & Computational Geometry*]{} [**34**]{}, 547 (2005). Chen, H. [*Ball packings and Lorentzian discrete geometry.*]{} PhD diss., Freie Universität Berlin, (2014). Chen, H. Apollonian ball packings and stacked polytopes. [*Discrete & Computational Geometry*]{} [**55**]{}, 801 (2016).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'For quantum fluids, the role of quantum fluctuations may be significant in several regimes such as when the dimensionality is low, the density is high, the interactions are strong, or for low particle numbers. In this paper we propose a fundamentally different regime for enhanced quantum fluctuations without being restricted by any of the above conditions. Instead, our scheme relies on the engineering of an effective attractive interaction in a dilute, two-component Bose-Einstein condensate (BEC) consisting of thousands of atoms. In such a regime, the quantum spin fluctuations are significantly enhanced (atom bunching with respect to the noninteracting limit) since they act to reduce the interaction energy - a remarkable property given that spin fluctuations are normally suppressed (anti-bunching) at zero temperature. In contrast to the case of true attractive interactions, our approach is not vulnerable to BEC collapse. We numerically demonstrate that these quantum fluctuations are experimentally accessible by either spin or single-component Bragg spectroscopy, offering a useful platform on which to test beyond-mean-field theories. We also develop a variational model and use it to analytically predict the shift of the immiscibility critical point, finding good agreement with our numerics.'
author:
- 'R. N. Bisset'
- 'P. G. Kevrekidis'
- 'C. Ticknor'
title: 'Enhanced quantum spin fluctuations in a binary Bose-Einstein condensate'
---
Introduction
============
Quantum fluctuations are ubiquitous in nature, lying at the heart of a wide variety of physical phenomena ranging from, for example, the Van der Waals force and the Casimir effect, through to Hawking radiation of black holes and the theory of cosmic inflation, providing the initial seed for the large-scale structure of the universe today. Despite their importance, quantum fluctuations are notoriously difficult to describe theoretically and many important questions remain unanswered [@SachdevBook; @Pitaevskii16].
Focusing on quantum fluids with short-range interactions, the role of quantum fluctuations becomes important in several qualitatively distinct regimes. Among the most dramatic of these are the one-dimensional systems, for which quantum fluctuations are so large that long-range order is destroyed [@Mermin1966; @Hohenberg1967]. Another well-known regime is that of high density (and strong interactions) where, for example, in the superfluid phase of liquid $^{4}$He, quantum fluctuations cause a depletion of the underlying Bose-Einstein condensate (BEC) of around 90% [@Dalfovo1999].
Dilute quantum gases have accelerated quantum-fluctuation research in recent years, offering a clean and highly-controllable testbed in which nearly all aspects are tunable, such as the interaction strength, confinement, mass and particle number [@Dalfovo1999; @Leggett2001]. This has allowed a series of innovative experiments to probe the emergence of quantum fluctuations by approaching the high-density regime from the dilute limit [@Shin2008; @Lopes2017; @Navon2010; @Navon2011; @Papp2008A; @Armijo2012; @Altmeyer2007; @Pollack2009; @Chang2016]. More recently, experimental groups were able to create dilute, self-bound droplets with liquid properties using dipolar [@Schmitt2016; @Chomaz2016] and two-component condensates [@Cabrera2017; @Semeghini2017]. Remarkably, these are stabilized against collapse by quantum fluctuations thanks to an almost complete cancellation of the various mean-field contributions [@Petrov2015; @Wachtler2016; @Lima2011; @Bisset2016; @Saito2016].
Quantum fluctuations can also be important for the regime of very small particle number $N$, typically when $N\sim10$. Several groups have proposed using ultra-dilute quantum gases to investigate these in different ways, including: vortex nucleation in slowly rotating traps [@Parke2008; @Dagnino2009]; the transition between Rabi-oscillation and self-trapped regimes in double well potentials [@JuliaDiaz2010]; the generation of macroscopic state superpositions in rotating ring superlattices [@Nunnenkamp2008]; as well as few-boson systems in one dimension [@Garcia2014; @Garcia2014A]. Furthermore, the enhancement of quantum fluctuations near the superfluid to Mott-insulator phase transition occurs for small atom number per lattice site [@Greiner2002; @Jaksch1998].
In this paper, we propose an alternative regime for enhancing quantum fluctuations that does not require low dimensionality, high density, strong interactions or small $N$. Our scheme instead relies on the engineering of effectively attractive interactions for the spin excitations of two-component (binary) BECs in the ground state, while the actual intracomponent and intercomponent contact interactions remain repulsive. By [*spin excitations*]{}, we mean the out-of-phase excitations of the two components. The interactions are effectively attractive in the sense that spin fluctuations act to decrease the interaction energy. Contrary to the case of true attractive contact interactions [@Bradley1995; @Donley2001], the proposed regime is not vulnerable to collapse of the underlying BEC.
Numerous groups have experimentally realized binary BECs by combining different elements [@Thalhammer2008; @McCarron2011; @Lercher2011; @Pasquiou2013; @Wacker2015; @Wang2016], isotopes [@Papp2008; @Sugawa2011; @FerrierBarbut2014], hyperfine states [@Myatt1997; @Hall1998B; @Matthews1999; @Hoinka2012], and even different spin states [@Maddaloni2000]. There has also been considerable theoretical attention, e.g. see [@Ho1996; @Pu1998; @Ohberg1999; @Trippenbach2000; @Svidzinsky2003; @Kasamatsu2004; @Dutton2005; @Mertes2007; @Ronen2008; @Oles2008; @Anderson2009; @Hofmann2014; @Roy2015; @Liu2016]. In the thermodynamic (TD) limit, a binary condensate is miscible for small intercomponent scattering length $a_{12}$, up until $a_{12}^{\rm c,TD} \equiv \sqrt{a_{11}a_{22}}$, above which it phase separates and becomes immiscible [^1]. Interestingly, at finite temperature ($T$), long-wavelength spin fluctuations tend to diverge on approach to the immiscibility phase transition, while at $T=0$ they do not [@Bisset2015; @Bienaime2016]. In fact, they instead approach the noninteracting, quantum shot-noise limit [@Astrakharchik2007; @Klawunn2011]. This occurs because at the transition $a_{12}^{\rm c,TD}$ the interactions have no preference for miscibility over immiscibility, and so the spin fluctuations behave as if the BEC was noninteracting.
We demonstrate that the situation changes dramatically, however, if one enters the regime in which quantum pressure becomes important. Here the critical point $a_{12}^{\rm c}$ is shifted higher than the thermodynamic-limit prediction, i.e. $a_{12}^{\rm c}>a_{12}^{\rm c,TD}$, thus opening a gap [@Navarro2009; @Wen2012]. For such a system, if the intercomponent scattering length lies within this gap, i.e. $a_{12}^{\rm c,TD}<a_{12}<a_{12}^{\rm c}$, then quantum spin excitations experience an effective attractive interaction, in the sense that they act to decrease the interaction energy. We show that the resulting spin fluctuations are then greatly enhanced. An intuitive picture for the physics at play: the trap pins the two components together, forcing miscibility of the ground state solution within the regime where the interactions would otherwise prefer immiscibility. Quantum spin fluctuations are then enhanced as a reward for lowering the interaction energy.
By numerically solving the Gross-Pitaevskii equations (GPE) and Bogoliubov de Gennes (BdG) equations, we demonstrate that such quantum spin fluctuations should be experimentally observable via either spin or single-component Bragg spectroscopy. The advantage of Bragg spectroscopy is that thermal contributions cancel, to first order [@Brunello2001; @Blakie2002], resulting in the observation of the quantum fluctuations. Moreover, we develop a variational model and use this to analytically predict the shift of the critical point, providing a simple means for finding suitable regimes of enhanced quantum spin fluctuations. We find good agreement between the predictions of the variational model and the numerical results.
.
Formalism
=========
System and Parameters {#Sec:System}
---------------------
We investigate a binary condensate for which interspecies interconversion is prohibited and the populations are fixed. The trapping potential is harmonic $V({\mathbf{x}}) = m(\omega_x^2x^2 + \omega_y^2y^2+\omega_z^2z^2)/2$, where $\omega_j$ represents the trapping frequencies and $m$ is the mass, which we take to be the same for both components. We consider a cylindrically symmetric trap, $\omega_x=\omega_y\equiv\omega_\rho$, in the quasi-2D regime, $\omega_z \gg \omega_\rho$, where the $z$ direction is well described by the harmonic oscillator ground state. Scattering is still three-dimensional, i.e. $a_{\alpha\beta}<<a_z$, where $a_{\alpha\beta}$ is the s-wave scattering length between components $\alpha$ and $\beta$, and $a_z = \sqrt{\hbar/ m\omega_z}$. From here on we work with the planar coordinate, ${\boldsymbol\rho}= \{x,y\}$, and results will be given in terms of the radial length $a_\rho = \sqrt{\hbar/m\omega_\rho}$.
The solutions for each component $\psi_\alpha({\boldsymbol\rho})$ (where $\alpha=\{1,2\}$) are obtained by solving the coupled GPEs [@Esry1997; @Pu1998; @Ho1996], $$\label{Eq:GPE}
\left[H_{\rm sp}({\boldsymbol\rho}) + \frac{\hbar^2}{m}\sum_\beta g_{\alpha\beta}n_\beta({\boldsymbol\rho})-\mu_\alpha \right]\psi_\alpha({\boldsymbol\rho}) = \mathbf{0},$$ where the condensate densities $n_\alpha = |\psi_\alpha({\boldsymbol\rho})|^2$ are individually unit-normalized, $\mu_\alpha$ are the corresponding chemical potentials, and the single-particle Hamiltonian is $H_{\rm sp}({\boldsymbol\rho}) = -\hbar^2\nabla^2/2m + V({\boldsymbol\rho},z=0)$. The dimensionless interactions are described by $g_{\alpha\beta}=N_\beta\sqrt{8\pi}a_{\alpha\beta}/a_z$, where $N_\beta$ is the number of atoms in component $\beta$. In this work, we restrict our attention to balanced populations, $N_1=N_2\equiv N/2$, and intracomponent interactions, $g_{11}=g_{22}\equiv g$. However, more general mixtures, along with a physical example, are discussed in Sec. \[Sec:ExpConsid\].
Variational Model {#Sec:Var}
-----------------
In addition to numerically solving the GPEs, we derive analytical predictions for the stationary-state solutions, including a prediction for the miscible-to-immiscible phase transition. To achieve this, motivated by the one-dimensional work of [@Navarro2009], we implement a chirped gaussian ansatz (superscript $\mathcal V$) to approximate the wavefunctions, $$\psi^{\mathcal V}_\alpha({\boldsymbol\rho})=A\exp\left[\frac{-(x\mp B)^2-y^2}{2W^2}\right] \exp\left[i(C \pm Dx+Ex^2)\right],$$ where variational parameters describe the amplitude $A$, component separation $2B$, phase $C$, wavenumber $D$, chirp $E$ and width $W$. The phase is a product of the chemical potential and time, i.e. $C=-\mu_\alpha^\mathcal{V} t/\hbar$.
As detailed in Appendix \[Sec:AppendixAna\], we use the Euler-Lagrange equations to derive equations of motion for the ansatz parameters. We then go on to derive analytical stationary-state predictions for the density, $n_\alpha^\mathcal{V}=|\psi_\alpha^\mathcal{V}|^2$, in the miscible regime \[Eqs. (\[Eq:Amisc\])-(\[Eq:muMisc\])\] and a set of transcendental equations for the immiscible regime \[Eqs. (\[Eq:cp0Imisc\])-(\[Eq:muImisc\])\], both of which we will compare with our numerical results. Finally, the variational prediction for the critical value of the intercomponent interaction strength, above which the system becomes immiscible, takes the form $$g_{12}^{\rm c,\mathcal{V}} = g + 2\pi . \label{Eq:gc}$$ The thermodynamic-limit result $ g_{12}^{\rm c,TD}=\sqrt{g_{11}g_{22}}=g$ is recovered for large particle number or intracomponent scattering length (recall that $g_{\alpha\beta}=N_\beta\sqrt{8\pi}a_{\alpha\beta}/a_z$).
Two-Component Structure Factor {#Sec:SFformalism}
------------------------------
Bragg spectroscopy furnishes a means by which to selectively probe the $T=0$ component of the structure factor. Naturally, this provides a useful avenue for the study of quantum fluctuations by *seeing through* the thermal fluctuations. More specifically, Bragg spectroscopy measures the imaginary part of the response function and this relates to the dynamic structure factor (which will be defined shortly) as $$\mathrm{Im}[\chi({\mathbf{k}_\rho},\omega)] = -\frac{\pi}{\hbar}[S({\mathbf{k}_\rho},\omega)-S(-{\mathbf{k}_\rho},-\omega)] ,$$ where the finite-temperature contributions cancel to leading order [@Brunello2001; @Blakie2002]. Several groups have proposed schemes for the implementation of spin (and density) Bragg spectroscopy in binary condensates [@Rodriguez2002; @Bruun2006; @Carusotto2006; @Baillie2016]. The general idea is to engineer a different coupling of the Bragg lasers to each component. This has been experimentally realized recently by the Vale group [@Hoinka2012].
We find the excitations of the BEC by linearizing about the ground state using a small parameter $\eta$. This amounts to inserting the time-dependent ansatz $$\Psi_\alpha({\boldsymbol\rho}) = \left\{\psi_\alpha({\boldsymbol\rho}) + \eta \left[ u_\alpha({\boldsymbol\rho})e^{-i\omega t} + v_\alpha^*({\boldsymbol\rho})e^{i\omega^*t} \right] \right\} e^{-i\mu_\alpha t/\hbar}$$ into the GPE (Eq. \[Eq:GPE\]) and solving the resulting BdG equations. Further details are provided in Appendix \[Sec:AppendixBog\] (also see Refs. [@Pu1998A; @Timmermans1998; @Ticknor2013; @Ticknor2014]).
The $T=0$ component of the density $S_\mathcal{D}$ [@Zambelli2000] and spin $S_\mathcal{S}$ [@Abad2013; @Symes2014A] structure factors are calculated from the BdG excitations, with each excitation labeled with superscript $\kappa$. Explicitly, we evaluate $$\begin{aligned}
S_{\{\mathcal{D,S}\}} &({\mathbf{k}_\rho},\omega) = \sum_\kappa \Big| \int d^2{\boldsymbol\rho}e^{i{\mathbf{k}_\rho}\cdot{\boldsymbol\rho}} \Big([u^\kappa_1({\boldsymbol\rho})+v^\kappa_1({\boldsymbol\rho})]\psi_1({\boldsymbol\rho}) \notag \\
& \pm [u^\kappa_2({\boldsymbol\rho})+v^\kappa_2({\boldsymbol\rho})]\psi_2({\boldsymbol\rho})\Big) \Big|^2 \delta(\omega-\omega^\kappa) , \label{Eq:SF}\end{aligned}$$ where the $+~(-)$ indicates $S_\mathcal{D}$ ($S_\mathcal{S}$), $\omega^\kappa = \epsilon^\kappa/\hbar$ (for BdG energy $\epsilon^\kappa$) and we have used the planar momentum coordinate ${\mathbf{k}_\rho}= \{k_x,k_y\}$. The corresponding static structure factors relate as $S_{\{\mathcal{D,S}\}}({\mathbf{k}_\rho}) = \int d\omega S_{\{\mathcal{D,S}\}}({\mathbf{k}_\rho},\omega)$.
![Density profiles for (a) the miscible phase for $g_{12}/g_{12}^{\rm c}=0.9$, and the immiscible phase for (b) $g_{12}/g_{12}^{\rm c}=1.1$ and (c) $g_{12}/g_{12}^{\rm c}=1.5$, where $g_{12}^{\rm c}$ is the immiscibility critical point. The first component ($\alpha=1$) is shown in red (gray) while the second component ($\alpha=2$) is black. Column 1 offers a side view ($y=0$) while column 2 consists of top-down views, each exhibiting a single contour at $1/e$ of the peak density. The same two intracomponent interaction strengths are compared in all panels, i.e. $g=\{4,500\}$. The variational predictions $n_\alpha^\mathcal{V}$, shown as thin blue lines, are only plotted for the second component to avoid cluttering. The mixture is balanced, i.e. $g_{11}=g_{22}\equiv g$ and $N_1=N_2$, where $g_{\alpha\beta}=N_\beta\sqrt{8\pi}a_{\alpha\beta}/a_z$, with $N_\beta$ being the number of atoms in component $\beta$, $a_{\alpha\beta}$ is the s-wave scattering length between components $\alpha$ and $\beta$, and $a_{\{\rho,z\}}=\sqrt{\hbar/m\omega_{\{\rho,z\}}}$. \[Fig:DenProfs\]](Fig1){width="3.4in"}
Results
=======
The Immiscibility Phase Transition {#Sec:ResultsPT}
----------------------------------
In this section we investigate how the immiscible critical value $g_{12}^{\rm c}$ shifts to higher values than the thermodynamic-limit prediction $g_{12}^{\rm c,TD}=g$. As will be shown in Sec. \[Sec:ResultsSF\], when the intercomponent interaction strength lies within the resulting gap, i.e. $g_{12}^{\rm c,TD}<g_{12}<g_{12}^{\rm c}$, then the quantum spin fluctuations become greatly enhanced.
In Fig. \[Fig:DenProfs\], we compare our numerical and variational predictions for the density, using two intracomponent coupling strengths $g=\{4,500\}$ and three intercomponent couplings $g_{12}/g_{12}^{\rm c}=$ (a) 0.9, (b) 1.1 and (c) 1.5 [^2]. For the stronger coupling, $g=500$, the density is represented by the flatter, broader distribution in each subplot. In the miscible phase \[Fig. \[Fig:DenProfs\](a)\] the average density of both components is identical, while in the immiscible phase, Figs. \[Fig:DenProfs\](b-c) show that the two components phase separate, forming a domain wall [@Trippenbach2000], with a sharper separation for $g=500$. As can be seen for $g=4$, the variational prediction agrees well with the full numerics for weak interactions, which is the regime of primary interest to this work. For larger interactions, e.g. $g=500$, the system enters the Thomas-Fermi regime where the Gaussian ansatz is less appropriate - this is most apparent in the top-down views in the immiscible phase \[Figs. \[Fig:DenProfs\](b2-c2)\]. While in this regime a variational formulation based on Thomas-Fermi clouds may provide a better approximation, we will not pursue this avenue here.
![Component separation (indicative of immiscibility) versus intercomponent coupling strength $g_{12}$ for various intracomponent couplings: $g = 4$ (dashed), $g=8$ (solid) and $g=500$ (dotted). Full numerical results are displayed in black while the variational results are in blue (gray). The position of component $\alpha$ is calculated according to $\langle x_\alpha\rangle=\int x \, n_\alpha({\boldsymbol\rho})d{\boldsymbol\rho}$, where we assume that the components always separate along the $x$ direction. For the variational model, $|\langle x_2\rangle-\langle x_1\rangle|=2B$. \[Fig:Sepa\]](Fig2){width="3.4in"}
To better characterize the onset of immiscibility, in Figure \[Fig:Sepa\] we plot the component separation as a function of $g_{12}$. Here it can be seen that the separation begins at a well-defined critical value $g_{12}^{\rm c}$ that increases with decreasing $g$. Overall, there is good qualitative agreement of the variational model (blue/gray) with the numerical results (black) for all cases. Deep within the immiscible phase, $g_{12}\gg g_{12}^{\rm c}$, and for small $g$, the variational model does a particularly good job. However, as previously discussed, the quantitative agreement for the spatial condensate profiles deteriorates for large coupling ($g=500$) where the variational model is unable to capture the plateau seen for the numerical solution. This can be understood by noting from Figs. \[Fig:DenProfs\](b-c) that the components are better able to avoid each other for the numerical solution, making it less sensitive to the precise value of $g_{12}$.
![Phase diagram for the immiscibility boundary. The critical intercomponent interaction strength $g_{12}^{\rm c}$ is plotted as a function of the intracomponent interaction strength $g$, where the miscible phase exists in the lower region. Full numerical results are in black, while the analytical prediction \[Eq. (\[Eq:gc\])\] is blue (gray). The vertical dashed lines represent the three interaction strengths that are the focus throughout the paper, i.e. $g = \{4, 8, 500\}$. Inset: magnification of the region where the immiscibility boundary begins to significantly rise above the thermodynamic-limit result $g_{12}^{\rm c,TD}/g=1$. \[Fig:PhaseD\]](Fig3){width="3.3in"}
The immiscibility boundary is summarized by a phase diagram in Fig. \[Fig:PhaseD\], with the numerical result in black and the variational one in blue (gray). For large $g$, the boundary levels off to the thermodynamic-limit prediction $g_{12}^{\rm c,TD}/g=1$. At the other extreme, where $g$ becomes small, the size of the miscible region grows as the role of quantum pressure increases and acts to stabilize the mixed phase. The small-$g$ regime is precisely where the variational model is appropriate. This explains why it captures the increasing trend of $g_{12}^{\rm c}/g$, yet the curve is qualitatively accurate throughout the region of parameters used.
Enhanced Quantum Spin Fluctuations {#Sec:ResultsSF}
----------------------------------
![Spin $S_{\mathcal{S}}(k_x,k_y=0,\omega)$ and density $S_{\mathcal{D}}(k_x,k_y=0,\omega)$ dynamic structure factor for the case of (a-b) uncoupled condensates ($g_{12}=0$), and (c-d) balanced intercomponent and intracomponent interactions ($g_{12}/g=1$), where $g = 500$. \[Fig:DynSF\]](Fig4){width="3.4in"}
{width="7in"}
As we saw in Fig. \[Fig:PhaseD\], the regime of small intracomponent interactions shifts the immiscibility phase transition higher than the thermodynamic-limit prediction, i.e. $g_{12}^{\rm c}>g_{12}^{\rm c,TD}=g$. This opens up a region of forced miscibility, $g<g_{12}<g_{12}^{\rm c}$, within which quantum spin fluctuations are greatly enhanced thanks to an effective attractive interaction [^3]. Bragg spectroscopy provides an especially useful tool for observing quantum fluctuations since it provides access to the $T=0$ component of the structure factor, even at finite $T$. Thus, in this section, we numerically calculate the $T=0$ spin and density structure factors to show how the enhanced quantum fluctuations can be studied in experimentally accessible settings.
### Dynamic structure factor
For reference, we first consider a BEC in the Thomas-Fermi regime, $g\gg1$, where the immiscibility transition occurs near the thermodynamic-limit prediction, $g_{12}^{\rm c}\approx g$, and the effective interactions experienced by the spin excitations are always repulsive. When the two components are uncoupled, $g_{12}=0$, Figs. \[Fig:DynSF\] (a-b) show that the spin (left) and density (right), dynamic structure factors are identical, as expected. On the other hand, a remarkable phenomenon emerges at the phase transition, where $g_{12}=g$. While here the density structure factor \[Fig. \[Fig:DynSF\] (d)\] is much the same as it was for $g_{12}=0$, apart from a slight hardening of the phonon excitations, the spin structure factor \[Fig. \[Fig:DynSF\] (c)\] is qualitatively different. In fact, the spin dispersion relation becomes parabolic. This happens because the spin excitations become effectively noninteracting when $g_{12}=g$ since there is no preference for miscibility over immiscibility.
### BdG energies and the shifted phase transition
We now extend the discussion to beyond the Thomas-Fermi regime to include small $g$, where spin excitations may now experience an effective attractive interaction. The lowest BdG energies as a function of $g_{12}$ are plotted in the first column of Fig. \[Fig:SFandE\], for $g=\{500,8,4\}$, from top to bottom. The phase transition can be identified as where the energy of at least one excitation softens to zero. This occurs at the same point when approaching from either side, consistent with the second-order nature of the phase transition. While in the Thomas-Fermi regime \[Fig. \[Fig:SFandE\](a1)\] the transition occurs at $g_{12}^{\rm c}/g\approx1$, a considerable shift to $g_{12}^{\rm c}/g>1$ can be seen for small $g$ in Figs. \[Fig:SFandE\](b1-c1). The region of effective attractive interactions and enhanced quantum spin fluctuations, $g<g_{12}<g_{12}^{\rm c}$, is schematically shaded with blue (darker in grayscale) color.
### Static structure factor: enhanced quantum fluctuations
Before proceeding, it is worth discussing the relationship between the structure factor and the sign of the interaction for a single-component, uniform system. Consider the Bijl-Feynman formula $S(\mathbf{k})=\epsilon_0(\mathbf{k})/\epsilon_B(\mathbf{k})$, where $\epsilon_0$ is the noninteracting dispersion relation and $\epsilon_B$ is the BdG energy [@Brunello2001]. If an excitation with some $\mathbf{k}$ experiences an effective attractive interaction, then the energy lies below its noninteracting counterpart, thus lifting $S(\mathbf{k})$ above unity as a result of atom bunching with respect to the noninteracting limit.
We present the numerical results for the spin (Fig. \[Fig:SFandE\], column 2) and density (Fig. \[Fig:SFandE\], column 3) static structure factors, restricting our attention to the miscible phase. Consistent with the energy results of the first column, each row in Fig. \[Fig:SFandE\] considers one of $g=\{500,8,4\}$, from top to bottom. The different curves are for various $g_{12}$ and the line styles are chosen to match (within each row) the values marked by vertical lines in column 1. As a benchmark, consider the uncoupled case, $g_{12}=0$, indicated by the green dashed lines. Here, as expected, the spin $S_\mathcal{S}$ and density $S_\mathcal{D}$ structure factors are identical for a given $g$. Increasing $g_{12}$ from zero has the opposite effect on $S_\mathcal{D}$ as it does on $S_\mathcal{S}$. The density structure factor exhibits a reduction, indicating a further suppression of the density fluctuations and an increase of the overall effective interaction. The spin structure factor, on the other hand, markedly increases with increasing $g_{12}$, signaling an increase of the spin fluctuations and a reduction of the effective interaction. In fact, once the intracomponent interaction has reached $g_{12}=g$ (red solid line), $S_\mathcal{S}$ becomes a flat line at unity for all three values of $g$. This indicates that the spin fluctuations are in the shot-noise regime and that the effective interactions have vanished (recall the Bijl-Feynman formula), as we already saw in Fig. \[Fig:DynSF\](c) for $g_{12}=g$. Note that the dip to zero at small momentum is due to the finite size of the system. For $g=\{8,4\}$, a remarkable thing happens since we can have $g_{12}/g>1$ while still being in the miscible phase. In this special regime \[see shading in Figs. \[Fig:SFandE\](b1-c1)\] the effective interaction becomes attractive and spin fluctuations cause bunching with respect to the noninteracting limit, as indicated by the large peak of $S_\mathcal{S}$ in Figs. \[Fig:SFandE\] (b2-c2) (solid black lines). This constitutes an important result of the paper.
It is worth noting that even though quantum fluctuations are large, the condensate depletion need not be, unless of course the system is very close to the immiscibility transition. This can be understood by the following. When interactions are effectively repulsive for a given excitation, the $v_\alpha^\kappa$ \[see Eq. (\[Eq:SF\])\] is approximately equal to $u_\alpha^\kappa$, but with opposite sign. The ensuing cancellation within Eq. (\[Eq:SF\]) results in a suppressed structure factor, implying anti-bunching with respect to the noninteracting limit. For attractive interactions, on the other hand, both $v_i$ and $u_i$ have the same sign, resulting in enhanced fluctuations (bunching). This sign reversal of $v_i$ is reminiscent of the weakly-interacting dipolar roton [@Santos2003], for which the effective interaction is momentum dependent, switching from repulsive to attractive near the roton wavelength [@Bisset2013C; @Blakie2013].
![(a) Density $S_\mathcal{D}$, spin $S_\mathcal{S}$, and one-component static structure factor $S_1$ for $g_{12}/g = 1.51$. The horizontal dashed lines indicate the structure factor peaks. (b) Spin and one-component static structure factor peaks as a function of $g_{12}/g$. The vertical dashed line indicates the interaction strength considered in (a). Note that $g_{12}^{\rm c}/g=1.55$ and $g=8$. \[Fig:SFpeak\]](Fig6){width="3.4in"}
### Single-component Bragg spectroscopy
While spin Bragg spectroscopy should provide the most-direct signal of quantum spin fluctuations, this may not always be the easiest option in experiments. We now demonstrate that it is also possible to study quantum spin fluctuations with single-component Bragg spectroscopy of a binary condensate. The single-component Bragg spectroscopy, corresponding to the single-component structure factor, can be calculated by restricting the integral of Eq. (\[Eq:SF\]) to the appropriate component, giving, for example, $$\begin{aligned}
S_1 &({\mathbf{k}_\rho},\omega) \\
&= \sum_\kappa \Big| \int d^2{\boldsymbol\rho}e^{i{\mathbf{k}_\rho}\cdot{\boldsymbol\rho}} [u^\kappa_1({\boldsymbol\rho})+v^\kappa_1({\boldsymbol\rho})]\psi_1({\boldsymbol\rho})
\Big|^2 \delta(\omega-\omega^\kappa) \notag .\end{aligned}$$ The one-component static structure factor then can be obtained as $S_1({\mathbf{k}_\rho}) = \int d\omega S_1({\mathbf{k}_\rho},\omega)$. The spin (solid), single-component (dot-dashed) and density (dashed) static structure factors are plotted in Fig. \[Fig:SFpeak\](a) for $g_{12}/g=1.51$ and $g=8$. As expected, $S_1$ provides an intermediate measure between the spin and density structure factors; while the peak is somewhat subdued it is still clearly visible. Figure \[Fig:SFpeak\](b) presents the spin (solid) and single-component (dot-dashed) structure factor peaks \[indicated by horizontal lines in Fig. \[Fig:SFpeak\](a)\] as a function of $g_{12}$. Here it can be seen that close to the immiscibility phase transition, which is at $g_{12}^{\rm c}=1.55$, the spin contribution dominates and $S_1$ provides an effective measure of the enhanced quantum spin fluctuations.
Experimental considerations {#Sec:ExpConsid}
===========================
From a qualitative perspective we expect our results to be quite general. Binary condensates can be constructed from a wide variety of mixtures, such as by combining different elements, isotopes, hyperfine states, or even spin states. To realize enhanced quantum spin fluctuations, the key ingredients are for both components to be miscible while, at the same time, $g_{12}>\sqrt{g_{11}g_{22}}$. This is possible if the immiscibility critical point is shifted higher than the thermodynamic-limit prediction $g_{12}^{\rm c,TD} = \sqrt{g_{11}g_{22}}$ (recall Fig. \[Fig:PhaseD\]). Our proposal relies on achieving this by enhancing the role of quantum pressure.
To give a flavor for the different kinds of experimental regimes that could realize enhanced quantum spin fluctuations, consider $^{87}$Rb, for which a mixture can be constructed from the hyperfine states $|F,m_F\rangle=|2,-1\rangle$ and $|1,1\rangle$. For these, there is a Feshbach resonance that can control the interspecies scattering length [@Nicklas2015; @Widera2004; @Erhard2004; @Gross2010]. Let us then take $g=8$ as a representative example for the results in this paper. Since the intraspecies scattering lengths are $a_{11}=a_{22}=100a_0$, then choosing a trap with frequencies $\{\omega_\rho,\omega_z\}=2\pi\times\{1,10\}$ Hz would correspond to $N_\alpha=1030$ atoms per component.
To increase the atom number, one can consider species with smaller scattering lengths and/or masses. Furthermore, it is not necessary for both intraspecies scattering lengths, or even masses, to be equal. A possible complication for the imbalanced case, though, is that inhomogeneous trapping potentials can cause the binary mixture to become immiscible even when $g_{12}<\sqrt{g_{11}g_{22}}$ [@Lee2016], hence shifting $g_{12}^{\rm c}$ in the wrong direction. Fortunately, this negative $g_{12}^{\rm c}$ shift can be reversed by imbalancing the atom numbers in favor of the component with the strongest interactions. In any case, the enhanced quantum pressure in the regime of interest will act to raise $g_{12}^{\rm c}$ as required.
Although the flattened trap geometry considered in this paper has some advantages, such as for [*in situ*]{} imaging of the fluctuations and numerical tractability, our results should be qualitatively extendible to more spherical geometries. In fact, loosening the trap along the tight direction should allow the regime of enhanced quantum fluctuations to extend to larger atom numbers.
Not only are quantum spin fluctuations enhanced on approach to the transition, but so too are thermal fluctuations [@Bisset2015]. Fortunately, though, Bragg spectroscopy provides a window through which one can selectively observe the quantum contribution, even at finite $T$. To appreciate this, we estimate the temperature that would otherwise be required for quantum fluctuations to dominate. The excitations primarily responsible for enhanced fluctuations have energies that soften on approach to the transition and, as can be seen in Figs. \[Fig:SFandE\](b1) and \[Fig:SFandE\](c1), these typically have energies of the same order as the radial confinement energy $\hbar\omega_\rho$. Again choosing $\omega_\rho=2\pi\times 1$ Hz means that without Bragg spectroscopy, for quantum fluctuations to dominate, experiments would otherwise need to be in the very-low-temperature regime, $T< 0.1$ nK. Concerning the validity of the quasi-2D approximation for the candidate regime above, we note that for $g=8$ the chemical potential relative to that of the noninteracting ground state is $\mu/\hbar\omega_z = 0.149$, where $\mu_1=\mu_2\equiv\mu$, thus satisfying the requirement that $\mu/\hbar\omega_z \ll 1$.
Conclusions {#Sec:Conclusions}
===========
In quantum fluids, quantum fluctuations are typically not important unless the dimensionality is low, the density is high or the particle number is very low. We have proposed an alternative scheme for enhanced quantum fluctuations that does not require any of these conditions to be satisfied, but instead relies on engineering an effective attractive interaction for the spin excitations of a two-component BEC. A strength of this approach is that the BEC is not vulnerable to dynamic instability, as would be the case for true attractive interactions.
In addition to numerical calculations, we developed a variational model to corroborate the phase diagram of the shifted immiscibility phase transition. Such a shift of $g_{12}^{\rm c}$ is crucial for reaching the regime of enhanced quantum fluctuations. We find excellent qualitative and good quantitative agreement between the numerical findings and the variational model, and the latter should be useful when scouting for experimentally relevant regimes. We numerically performed BdG calculations to show that the enhanced quantum spin fluctuations should be observable using either spin or single-component Bragg spectroscopy for two-component BECs.
This work opens a number of possibilities for future research. An obvious direction is for experiments to quantitatively test beyond-mean-field theories. Since the size of the quantum fluctuations is tunable, by adjusting $g_{12}$ to approach the immiscibility transition, it would be interesting to see at which point, for example, BdG theory breaks down. It is also an open question as to how quantum fluctuations might shift the immiscibility transition. Moreover, the proposed regime could be useful for quantitative comparisons between beyond-mean-field theories with Monte Carlo simulations. Another intriguing direction would be to go beyond BdG theory to study the quantum critical physics on approach to the immiscibility phase transition.
As a final note, the enhanced fluctuations should also be directly observable via [*in situ*]{} imaging of the spin density ($n_1-n_2$). Such a scheme was theoretically investigated for thermal spin fluctuations in Ref. [@Bisset2015]. However, in this case the quantum and thermal fluctuations will both contribute, and the temperature will need to be quite low to suppress the thermal contribution.\
[**Acknowledgments: $\,$**]{} We thank D. Baillie, T. Bienaimé, P. B. Blakie, R. Carretero-Gonz[á]{}lez, G. Ferrari, S. I. Mistakidis, A. Recati and C. Qu for useful and stimulating discussions. RNB acknowledges the EU QUIC project and Provincia Autonoma di Trento for financial support. PGK acknowledges the support of NSF-DMS-1312856 and PHY-1602994, from the ERC under FP7, Marie Curie Actions, People, International Research Staff Exchange Scheme (IRSES-605096), and the Stavros Niarchos Foundation via the Greek Diaspora Fellowship Program. CT acknowledges support by Los Alamos National Laboratory, which is operated by LANS, LLC for the NNSA of the U.S. DOE and, specifically, Contract No. DEAC52-06NA25396.
Two-component Gaussian ansatz {#Sec:AppendixAna}
=============================
To gain a theoretical handle on the miscibility transition in the quasi-2D setting we apply a chirped Gaussian ansatz (superscript $\mathcal V$) as an approximation to the wavefunctions (similar to what was done in 1D by [@Navarro2009]), i.e., $$\psi^{\mathcal V}_\alpha({\boldsymbol\rho})=A\exp\left[\frac{-(x\mp B)^2-y^2}{2W^2}\right] \exp\left[i(C \pm Dx+Ex^2)\right],$$ where $A$ is the amplitude, $B$ is the component separation and $W$ determines the width. The phase $C$ is is related to the chemical potential and time according to $C=-\mu_\alpha^\mathcal{V} t/\hbar$, while $D$ and $E$ represent the wavenumber and chirp, respectively. The upper of $\mp$ and $\pm$ are for component $\alpha=1$ while the lower are for $\alpha=2$.
We utilize the Euler-Lagrange equations, $$\frac{\partial \mathcal L}{\partial \mathcal{P}_j} - \frac{d}{dt}\left(\frac{\partial \mathcal L}{\partial \dot{\mathcal{P}_j}} \right) = 0 ,$$ with $\mathcal{P}_j$ = {$A$, $B$, $C$, $D$, $E$, $W$}. The Lagrangian takes the form $$\mathcal{L} = \int_{-\infty}^\infty ( L_1 + L_2 + L_{12} ) d^2{\boldsymbol\rho},$$ where $$\begin{aligned}
L_\alpha &= \frac{\hbar^2}{2m} \left( \left|\frac{\partial \psi_\alpha^{\mathcal{V}}}{\partial x}\right|^2 + \left|\frac{\partial \psi_\alpha^{\mathcal{V}}}{\partial y}\right|^2 \right) \\
&+ \frac{m\omega_\rho^2}{2}\left(x^2+y^2 \right)\left|\psi_\alpha^{\mathcal{V}}\right|^2 + \frac{\hbar^2g}{2m} \left|\psi_\alpha^{\mathcal{V}}\right|^4 \\
&+ \frac{i\hbar}{2} \left(\psi_\alpha^\mathcal{V}\frac{\partial{\psi_\alpha^\mathcal{V}}^*}{\partial t} - {\psi^\mathcal{V}_\alpha}^*\frac{\partial\psi_\alpha^\mathcal{V}}{\partial t}\right) ,\end{aligned}$$ and the intercomponent coupling is given by $$L_{12} = \frac{\hbar^2g_{12}}{m}\left|\psi_1^\mathcal{V}\right|^2\left|\psi_2^\mathcal{V}\right|^2.$$ Recall that we have taken the symmetric situation $g_{11}=g_{22}\equiv g$ (and also $\mu_1=\mu_2\equiv \mu$), where $g_{\alpha\beta}=N_\beta\sqrt{8\pi}a_{\alpha\beta}/a_z$ with $N_\beta$ being the number of atoms in component $\beta$, $a_{ij}$ is the scattering length between component $\alpha$ and $\beta$, and $a_z = \sqrt{\hbar/m\omega_z}$ is the confinement length in the tight direction.
Equations of motion {#equations-of-motion .unnumbered}
-------------------
For clarity, we now switch to dimensionless form: $\tilde\mu = \mu/\hbar\omega_
\rho$; $\tilde t = t\omega_\rho$; $\tilde{A}=a_\rho A$; $\tilde{B}=B/a_\rho$; $\tilde{C}=C$; $\tilde{D}=a_\rho D$; $\tilde{E}=a_\rho^2 E$; $\tilde{W}=W/a_\rho$, where $a_\rho = \sqrt{\hbar/m\omega_\rho}$. Simultaneously solving the Euler-Lagrange equations we obtain the equations of motion for the ansatz parameters: $$\frac{d\tilde A}{d\tilde t} = -2 \tilde A \tilde E , \label{Eq:dA_dt}$$ $$\frac{d\tilde B}{d\tilde t} = \tilde D + 2\tilde B \tilde E ,$$ $$\begin{aligned}
\frac{d\tilde C}{d\tilde t} &= \frac{\tilde B^2}{\tilde W^4} - \frac{\tilde D^2}{2} -\frac{1}{\tilde W^2}- \frac{\tilde B^2}{2} -\frac{3\tilde A^2g}{4} + \frac{\tilde A^2 \tilde B^2 g}{2\tilde W^2} \notag \\
& - \tilde A^2e^{-2\tilde B^2/\tilde W^2}g_{12} \left(\frac{3}{4} + \frac{\tilde B^4}{\tilde W^4} \right) , \label{Eq:dC_dt}\end{aligned}$$ $$\frac{d\tilde D}{d\tilde t} = \tilde B - 2 \tilde D \tilde E - \frac{2\tilde B}{\tilde W^4} -\frac{\tilde A^2\tilde Bg}{\tilde W^2} + \frac{2\tilde A^2\tilde B^3g_{12}}{\tilde W^4}e^{-2\tilde B^2/\tilde W^2} ,$$ $$\begin{aligned}
\frac{d\tilde E}{d\tilde t} &= \frac{1}{\tilde W^4} - 1 - 2\tilde E^2 + \frac{\tilde A^2g}{2\tilde W^2} \notag\\
& + \frac{\tilde A^2g_{12}}{2\tilde W^4}e^{-2\tilde B^2/\tilde W^2}(\tilde W^2-2\tilde B^2) ,\end{aligned}$$ $$\frac{d\tilde W}{d\tilde t} = 2\tilde E\tilde W . \label{Eq:dW_dt}$$
Stationary state solutions for the miscible phase {#stationary-state-solutions-for-the-miscible-phase .unnumbered}
-------------------------------------------------
We look for stationary-state solutions by setting the left hand side of the equations of motion \[(\[Eq:dA\_dt\])-(\[Eq:dW\_dt\])\] to zero, with the exception of Eq. (\[Eq:dC\_dt\]) for which the phase continues to evolve steadily according to the chemical potential, $d\tilde C/d\tilde t = -\tilde\mu$. For such a state, one can immediately see that the wavenumber and chirp are zero, i.e. $D_0=E_0=0$.
By definition, the miscible phase has $B_0=0$, and solving the resulting equations produces analytic expressions for the equilibrium amplitude and width, respectively: $$\begin{aligned}
\tilde A_0^2 &= \frac{4}{3(g+g_{12})} \left(2\tilde\mu_0 - \sqrt{\tilde\mu^2_0+3} \right) ,\label{Eq:Amisc} \\
\tilde W_0^2 &= \frac{1}{3}\left(\tilde\mu_0 + \sqrt{\tilde\mu^2_0+3} \right),\end{aligned}$$ where the chemical potential is $$\tilde\mu_0 = \frac{1}{\sqrt{8\pi}} \frac{(3g + 3g_{12} + 4\pi)}{\sqrt{g+g_{12} + 2\pi}} . \label{Eq:muMisc}$$
Stationary state solutions for the immiscible phase {#stationary-state-solutions-for-the-immiscible-phase .unnumbered}
---------------------------------------------------
Finding stationary state solutions for the immiscible phase ($B_0\neq0$) is not so straightforward and the equations of motion, (\[Eq:dA\_dt\])-(\[Eq:dW\_dt\]), reduce to a set of transcendental equations that we solve numerically:
$$\begin{aligned}
0 &= \frac{1}{\tilde W^2_0} + \frac{3\tilde A^2_0g}{4} + \frac{\tilde B^2_0}{2} -\frac{\tilde B^2_0}{\tilde W^4_0} - \frac{\tilde A^2_0\tilde B^2_0g}{2\tilde W^2_0} \notag \\
& + \tilde A^2_0g_{12}e^{-2\tilde B^2_0/\tilde W^2_0} \left(\frac{3}{4} + \frac{\tilde B^4_0}{\tilde W^4_0} \right) - \tilde\mu_0 ~, \label{Eq:cp0Imisc}\end{aligned}$$
$$0 = \tilde B_0 - \frac{2\tilde B_0}{\tilde W^4_0} - \frac{\tilde A^2_0\tilde B_0g}{\tilde W^2_0} + \frac{2\tilde A^2_0 \tilde B^3_0 g_{12}}{\tilde W^4_0} e^{-2\tilde B^2_0/\tilde W^2_0} ,$$
$$0 = \frac{1}{\tilde W^4_0} + \frac{\tilde A^2_0g}{2\tilde W^2_0} - 1 + \frac{\tilde A^2_0g_{12}}{2\tilde W^4_0}e^{-2\tilde B^2_0/\tilde W^2_0} \left(\tilde W^2_0 - 2\tilde B^2_0 \right) ,$$
where the chemical potential is given by $$\tilde\mu_0 = \frac{\pi \tilde A^2_0\tilde W^2_0}{2} \left(\frac{1}{\tilde W^2_0} + \tilde A^2_0g + \tilde B^2_0 + \tilde W^2_0 + \tilde A^2_0 g_{12}e^{-2\tilde B^2_0/\tilde W^2_0} \right). \label{Eq:muImisc}$$
Critical interaction strength {#critical-interaction-strength .unnumbered}
-----------------------------
To find the transition point between miscible and immiscible phases we construct the Jacobian matrix, $\mathcal J_{jk} = \partial \mathcal F_j/\partial \mathcal P_k$, where $\mathcal F_j=d\mathcal P_j/dt$ are the equations of motion given (in dimensionless form) by Eqs. (\[Eq:dA\_dt\])-(\[Eq:dW\_dt\]). The stationary-state solution for the miscible phase, Eqs. (\[Eq:Amisc\])-(\[Eq:muMisc\]), is substituted into the Jacobian matrix and we subsequently calculate its eigenvalues. Finding the interaction strengths at which the appropriate eigenvalue vanishes then gives an analytic prediction for the critical value of $g_{12}$, i.e., $$g_{12}^{c,\mathcal{V}} = g + 2\pi .$$
Two-component Bogoliubov-de Gennes theory {#Sec:AppendixBog}
=========================================
For continuity of notation, we continue to work in the quasi-2D regime with planar coordinates, ${\boldsymbol\rho}= \{x,y\}$, although the generalization to 3D is straightforward. Recall, that in the weakly interacting limit, the condensate components $\psi_\alpha({\boldsymbol\rho})$ (with $\alpha = \{1,2\}$) are computed by solving the coupled GPEs, $$H_\alpha^{GP}\psi_\alpha({\boldsymbol\rho}) \equiv \left[H_{\rm sp}({\boldsymbol\rho}) +C_\alpha({\boldsymbol\rho})-\mu_\alpha \right]\psi_\alpha({\boldsymbol\rho}) = \mathbf{0},$$ with the single particle Hamiltonian $H_{\rm sp}({\boldsymbol\rho}) = -\hbar^2\nabla^2/2m + V({\boldsymbol\rho})$, where $V({\boldsymbol\rho})$ is the trap potential. The interactions are described by $$\begin{aligned}
C_\alpha({\boldsymbol\rho}) &= \sum_\beta \int d^2 {\boldsymbol\rho}^\prime U_{\alpha\beta}({\boldsymbol\rho}-{\boldsymbol\rho^\prime}) n_\beta({\boldsymbol\rho}^\prime) ~, \\
&= \frac{\hbar^2}{m}\sum_\beta g_{\alpha\beta}n_\beta({\boldsymbol\rho}) ~,\end{aligned}$$ where the condensate density $n_\alpha$ is unit-normalized $\int n_\alpha({\boldsymbol\rho}) d^2{\boldsymbol\rho}= \int |\psi_\alpha({\boldsymbol\rho})|^2 d^2{\boldsymbol\rho}= 1$ and we consider the case of contact interactions, i.e. $U_{\alpha\beta}({\boldsymbol\rho}-{\boldsymbol\rho^\prime})=\delta({\boldsymbol\rho}-{\boldsymbol\rho^\prime})\hbar^2g_{\alpha\beta}/m$ with $g_{\alpha\beta}=N_\beta\sqrt{8\pi}a_{\alpha\beta}/a_z$ and $N_\beta$ is the number of atoms in condensate $\beta$.
Excitations are found by linearizing about the ground state. Using the ansatz, $$\Psi_\alpha({\boldsymbol\rho}) = \left\{\psi_\alpha({\boldsymbol\rho}) + \eta \left[ u_\alpha({\boldsymbol\rho})e^{-i\omega t} - v_\alpha^*({\boldsymbol\rho})e^{i\omega^*t} \right] \right\} e^{-i\mu_\alpha t/\hbar} ,$$ and keeping terms up to first-order in the small parameter $\eta$, results in the two-component Bogoliubov de Gennes equations [@Pu1998A; @Timmermans1998; @Ticknor2013], $$\begin{aligned}
\left( \begin{array}{cc}
H^{GP}_\alpha & 0 \\
0 & -H^{GP}_\alpha
\end{array} \right)
\left( \begin{array}{c}
u_{\alpha}^{\kappa} \\ v_{\alpha}^{\kappa}
\end{array} \right)
&+ \sum_{\beta=1,2}\left( \begin{array}{cc}
X_{\alpha\beta}^a & -X_{\alpha\beta}^b \\
{X_{\alpha\beta}^b}^* & -X_{\alpha\beta}^a
\end{array} \right)
\left( \begin{array}{c}
u_{\beta}^{\kappa} \\ v_{\beta}^{\kappa}
\end{array} \right) \notag \\
&= \epsilon^\kappa \label{Eq:BdG}
\left( \begin{array}{c}
u_{\alpha}^{\kappa} \\ v_{\alpha}^{\kappa}
\end{array} \right) ,\end{aligned}$$ where the superscript index $\kappa$ labels each excitation. The condensate exchange operators act on a test function $f$ according to $$\begin{aligned}
(X_{\alpha\beta}^af)({\boldsymbol\rho}) &= \psi_\alpha({\boldsymbol\rho})\int d^2{\boldsymbol\rho^\prime}U_{\alpha\beta}({\boldsymbol\rho}-{\boldsymbol\rho^\prime})\psi_\beta^*({\boldsymbol\rho^\prime})f({\boldsymbol\rho^\prime}) , \notag \\
&= \frac{\hbar^2}{m}\psi_\alpha({\boldsymbol\rho})\psi_\beta^*({\boldsymbol\rho})f({\boldsymbol\rho})g_{\alpha\beta} , \\
(X_{\alpha\beta}^bf)({\boldsymbol\rho}) &= \psi_\alpha({\boldsymbol\rho})\int d^2{\boldsymbol\rho^\prime}U_{\alpha\beta}({\boldsymbol\rho}-{\boldsymbol\rho^\prime})\psi_\beta({\boldsymbol\rho^\prime})f({\boldsymbol\rho^\prime}) , \notag \\
&= \frac{\hbar^2}{m}\psi_\alpha({\boldsymbol\rho})\psi_\beta({\boldsymbol\rho})f({\boldsymbol\rho})g_{\alpha\beta} , \\
({X_{\alpha\beta}^b}^*f)({\boldsymbol\rho}) &= \psi_\alpha^*({\boldsymbol\rho})\int d^2{\boldsymbol\rho^\prime}U_{\alpha\beta}({\boldsymbol\rho}-{\boldsymbol\rho^\prime})\psi_\beta^*({\boldsymbol\rho^\prime})f({\boldsymbol\rho^\prime}) , \notag \\
&= \frac{\hbar^2}{m}\psi_\alpha^*({\boldsymbol\rho})\psi_\beta^*({\boldsymbol\rho})f({\boldsymbol\rho})g_{\alpha\beta} .\end{aligned}$$ Finally, quasiparticle modes are normalized according to $$\sum_{\alpha=1,2}\int d^2{\boldsymbol\rho}\Big(|u_\alpha^\kappa({\boldsymbol\rho})|^2-|v_\alpha^\kappa({\boldsymbol\rho})|^2 \Big) = 1 .$$
We note, that it is important to ensure that all excitations are orthogonal to the ground state [@Morgan2000]; to achieve this, we apply a condensate-projector before and after the exchange operators. We solve Eqs. \[Eq:BdG\] using a spectral basis of non-interacting harmonic oscillator modes, and utilize the Gauss-Hermite quadrature.
[94]{}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 [**]{}, ed. (, ) @noop [**]{} (, ) [****, ()](\doibase 10.1103/PhysRevLett.17.1133) [****, ()](\doibase 10.1103/PhysRev.158.383) [****, ()](\doibase 10.1103/RevModPhys.71.463) [****, ()](\doibase 10.1103/RevModPhys.73.307) [****, ()](\doibase 10.1103/PhysRevLett.101.070404) @noop [ ()]{}, [****, ()](\doibase 10.1126/science.1187582) [****, ()](\doibase
10.1103/PhysRevLett.107.135301) [****, ()](\doibase
10.1103/PhysRevLett.101.135301) [****, ()](\doibase 10.1103/PhysRevLett.108.225306) [****, ()](\doibase 10.1103/PhysRevLett.98.040401) [****, ()](\doibase
10.1103/PhysRevLett.102.090402) [****, ()](\doibase
10.1103/PhysRevLett.117.235303) @noop [****, ()]{} [****, ()](\doibase
10.1103/PhysRevX.6.041039) [****, ()](\doibase 10.1126/science.aao5686) @noop [ ()]{}, [****, ()](\doibase 10.1103/PhysRevLett.115.155302) [****, ()](\doibase 10.1103/PhysRevA.93.061603) [****, ()](\doibase 10.1103/PhysRevA.84.041604) [****, ()](\doibase 10.1103/PhysRevA.94.033619) [****, ()](\doibase 10.7566/JPSJ.85.053001) [****, ()](\doibase 10.1103/PhysRevLett.101.110401) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevA.81.023615) [****, ()](\doibase 10.1103/PhysRevA.77.023622) [****, ()](\doibase 10.1103/PhysRevA.90.063605) @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase
10.1103/PhysRevLett.81.3108) [****, ()](\doibase 10.1103/PhysRevLett.75.1687) @noop [****, ()]{} [****, ()](\doibase
10.1103/PhysRevLett.100.210402) [****, ()](\doibase
10.1103/PhysRevA.84.011603) [****, ()](\doibase
10.1140/epjd/e2011-20015-6) [****, ()](\doibase
10.1103/PhysRevA.88.023601) [****, ()](\doibase 10.1103/PhysRevA.92.053602) [****, ()](http://stacks.iop.org/0953-4075/49/i=1/a=015302) [****, ()](\doibase 10.1103/PhysRevLett.101.040402) [****, ()](\doibase
10.1103/PhysRevA.84.011610) [****, ()](\doibase 10.1126/science.1255380) [****, ()](\doibase
10.1103/PhysRevLett.78.586) [****, ()](\doibase
10.1103/PhysRevLett.81.1539) [****, ()](\doibase 10.1103/PhysRevLett.83.2498) [****, ()](\doibase 10.1103/PhysRevLett.109.050403) [****, ()](\doibase
10.1103/PhysRevLett.85.2413) [****, ()](\doibase 10.1103/PhysRevLett.77.3276) [****, ()](\doibase 10.1103/PhysRevLett.80.1130) [****, ()](\doibase 10.1103/PhysRevA.59.634) [****, ()](http://stacks.iop.org/0953-4075/33/i=19/a=314) [****, ()](\doibase 10.1103/PhysRevA.67.053608) [****, ()](\doibase 10.1103/PhysRevLett.93.100402) [****, ()](\doibase 10.1103/PhysRevA.71.063618) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevA.78.053613) @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevLett.113.095702) [****, ()](\doibase 10.1103/PhysRevA.92.011601) [****, ()](\doibase 10.1103/PhysRevA.93.023628) [****, ()](\doibase 10.1103/PhysRevA.91.053613) [****, ()](\doibase 10.1103/PhysRevA.94.063652) [****, ()](\doibase 10.1103/PhysRevA.76.063616) [****, ()](\doibase 10.1103/PhysRevA.84.033612) [****, ()](\doibase 10.1103/PhysRevA.80.023613) [****, ()](\doibase
10.1103/PhysRevA.85.043602) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevA.65.033602) [****, ()](\doibase 10.1103/PhysRevLett.78.3594) [****, ()](\doibase 10.1103/PhysRevA.66.033601) [****, ()](\doibase 10.1103/PhysRevA.74.033623) [****, ()](http://stacks.iop.org/0953-4075/39/i=10/a=S20) [****, ()](\doibase 10.1103/PhysRevA.93.033607) [****, ()](\doibase 10.1103/PhysRevLett.80.1134) [****, ()](\doibase 10.1103/PhysRevLett.81.5718) [****, ()](\doibase 10.1103/PhysRevA.88.013623) [****, ()](\doibase 10.1103/PhysRevA.89.053601) @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevA.89.053628) [****, ()](\doibase 10.1103/PhysRevLett.90.250403) [****, ()](\doibase 10.1103/PhysRevA.88.043606) [****, ()](\doibase 10.1103/PhysRevA.88.013638) [****, ()](\doibase 10.1103/PhysRevLett.115.245301) [****, ()](\doibase
10.1103/PhysRevLett.92.160406) [****, ()](\doibase 10.1103/PhysRevA.69.032705) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevA.94.013602) [****, ()](http://stacks.iop.org/0953-4075/33/i=19/a=303)
[^1]: We have assumed that the particles of both components have the same mass
[^2]: Note that for the numerical results we numerically determine $g_{12}^{\rm c}$, whereas for the variational results we take $g_{12}^{\rm c}=g_{12}^{{\rm c},\mathcal{V}}$ \[Eq. (\[Eq:gc\])\]
[^3]: The interactions are effectively attractive in the sense that spin fluctuations act to decrease the interaction energy.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Dynamical evolution of quantum mechanical squeezing and entanglement in a two-mode Bose-Einstein condensate (TBEC) with an adiabatic, time-varying Raman coupling is studied by finding analytical expressions for these quantities. In particular, we study the entanglement between the atoms in the condensate as well as that between the two modes. The nature of the enhanced quantum correlations in TBEC is clarified by considering squeezing and entanglement both with and without the nonlinear interaction turned on; it is found that entanglement approaching maximal value can be achieved with the nonlinear interactions present. Somewhat counter-intuitively, greater squeezing is found in the absence of nonlinear interactions. This is due to the collapses and revivals of the TBEC quantum state induced by the nonlinear interactions. In addition, results involving the self-trapping phase state of TBEC indicate potential for creating a dynamically stable, macroscopic entangled quantum state which is relatively robust with respect to atom number fluctuations.'
author:
- 'S. Choi and N. P. Bigelow'
title: 'Quantum squeezing and entanglement in a two-mode Bose-Einstein condensate with time-dependent Josephson-like coupling'
---
Introduction
============
Bose-Einstein condensates (BECs) provide a useful system for investigating matter wave quantum squeezing and entanglement. Squeezing in BECs has already been investigated previously from a number of different perspectives[@squeeze1; @squeeze2; @squeeze3; @squeeze4; @squeeze5]. Squeezed states are quantum states for which no classical analog exists[@WallsNature]. The definition of quantum mechanical squeezing, or the reduction of quantum fluctuations below the standard quantum limit (SQL), derives directly from the Heisenberg Uncertainty Principle: For two arbitrary operators $\hat{A}$ and $\hat{B}$ which obey the commutation relation $[\hat{A},\hat{B}]=\hat{C}$, quantum squeezing exists when one of the variables satisfies the relation[@wallszoller] $$\lbrack \Delta \hat{A}]^{2}<\frac{1}{2}|\langle \hat{C}\rangle |.
\label{inequality}$$ Although the origin of quantum squeezing in many experiments is enhanced quantum correlations, it must be noted that the inequality (\[inequality\]) is [*not*]{} a sufficient proof of enhanced quantum correlations between the particles, as discussed by Kitagawa and Ueda in the context of spin squeezing[@ueda]. A measure of the amount of spin squeezing was defined as[@ueda; @wineland] $$\xi =\frac{2[\Delta J_{\bot }]^{2}}{J}, \label{KUxi}$$ where $[\Delta J_{\bot }]^{2}$ is the variance in the direction orthogonal to the total spin vector for an $N$ spin-$\frac{1}{2}$ particle system. This definition takes into account the effect of quantum correlations as it derives from the fact that a minimum uncertainty state for $N$ elementary spin-$\frac{1}{2}$ particles has spin $J/2$ ($J = N/2$) equally distributed over any two orthogonal components normal to the mean spin direction. This follows from the fact that, in the absence of quantum correlations, the total variance in the normal direction is simply given by the sum of the variances of the individual elementary spins. Any state with uncertainty less than the minimum uncertainty state, i.e. spin squeezed due to enhanced quantum correlations between the particles, therefore gives $\xi <1$.
Quantum entanglement, which is closely related to quantum squeezing, has recently generated a lot of activity amongst researchers, owing to its significant role in the studies of quantum information theory. A large number of theoretical studies exist on generating entangled states using BECs in a variety of different physical settings. Many of these involve two-mode BECs (TBECs) in which atoms in two different hyperfine states are entangled[@twocomp1; @Micheli; @kennedy; @twocomp2; @twocomp3]. In simple terms, a quantum state is said to be entangled when the state cannot be written as a simple product state, the most prominent example being the famous Bell-state for a two particle system: $|\psi \rangle = (|1 \rangle| 0 \rangle + |0 \rangle|1 \rangle)/\sqrt{2}$, where “1” and “0” represent spin “up” and “down” states respectively. From the inseparability criterion for the $N$-particle density matrix, a parameter for the entanglement between atoms in TBECs has been derived[@twocomp1; @Micheli], which has a form similar to Eq. (\[KUxi\]): $$\xi _{{\bf n}}^{2}=\frac{N[\Delta {\bf n}_{1}\cdot \hat{J}]^{2}}{\langle
{\bf n}_{2}\cdot \hat{J}\rangle ^{2}+\langle {\bf n}_{3}\cdot \hat{J}\rangle
^{2}} \label{Sxi}$$ where ${\bf n}_{i}$, $i=1,2,3$ are unit orthogonal vectors, and $\hat{J}$ is the well-known Schwinger angular momentum operator representing a many particle system. It is easy to see that Eq. (\[Sxi\]) may be viewed as a generalization of Eq. (\[KUxi\]).
On the other hand, it has also been argued that quantum entanglement in a TBEC is more meaningful when one considers the system as a bipartite system of two modes, analogous to considering entanglement of electromagnetic modes, as opposed to the photons themselves, as there is no definitive measure for entanglement between three or more subsystems[@Milburn]. In addition, the two modes of a TBEC are clearly distinguishable and experimentally accessible subsystems. In this case, the standard measure of entanglement is the von Neumann entropy of the reduced density operator of either of the subsystems: $$E(\rho_A) = - {\rm Tr} \rho_A \log_2 (\rho_A),$$ where $\rho_A$ is the reduced density operator for subsystem $A$ defined by the partial trace over the other subsystem $B$, $\rho_A = {\rm Tr}_B \rho$. It can be easily shown that for a general bipartite system of TBECs written in the Fock basis $|\psi (t) \rangle = \sum_{m} c_{m} (t) |N - m \rangle_A |m \rangle_B$, $\rho_A = \sum_{m = 0}^{N} |c_m (t)|^2 | m \rangle \langle m |$ and hence the dynamically evolving bipartite entanglement in TBECs can be parameterized by $$E(t) = - \sum_{m = 0}^{N} |c_m (t)|^2 \log_2 |c_m (t)|^2. \label{vonNeumann}$$
In this paper, we calculate analytically dynamical evolution of quantum squeezing as well as quantum entanglement for both types of system decompositions of a TBEC: $N$ particle subsystems based on Eq. (\[Sxi\]) and the subsystems composed of the two modes based on Eq. (\[vonNeumann\]). We shall consider a TBEC with a continuous time-dependent Raman coupling between the two states, with and without nonlinear interactions. A physical realization of such a system already exists; the Josephson-like coupling between the two levels is provided by an external laser in the Rb TBEC of Ref. [@myatt], while a state-dependent magnetic field gradient may be applied to induce Josephson tunneling in the Na spinor system of Ref. [@stenger]. We consider a system with an adiabatically time-varying coupling with an off-resonance atom-light interaction. The TBEC is itself an interesting and important system to study: not only does it display effects similar to that of Josephson junction in superconductors, such as the collapse and revival phenomenon in macroscopic scale, it is also the simplest Bose-Hubbard model with a two site lattice potential for which complete description of entanglement can be given.
The central result of this paper is that we use an exact solution to the time-dependent Schrödinger equation to find analytical expressions for the dynamical evolution of squeezing and entanglement. In particular, we explicitly identify the effect of the nonlinear interactions on squeezing and entanglement. Since one of the comments of Ref. [@Milburn] with regard to the atom-atom entanglement parameterized by Eq. (\[Sxi\]) was that of practicality, we shall restrict our calculations to two different possible initial states for a TBEC, the Dicke state and the phase state. We consider atom-atom entanglement for which clear physical measurement is in principle possible, namely those involving the measurement of the mean and variance of the atom number difference and the relative phase between the two species.
The paper is organized as follows: In Sec. II, we describe, using the Schwinger notation for the angular momentum operators, the TBEC and the exact solution to the time-dependent Schrödinger Equation. In Sec. III, we present our main results: the dynamics of the TBEC system with a time-dependent Josephson-like coupling between the two species and the evolution of quantum squeezing and entanglement for the two types of subsystem decompositions for various values of laser couplings, with and without the nonlinear interaction. We conclude in Sec. IV.
Formalism
=========
Hamiltonian
-----------
We consider atomic BECs in two different hyperfine states trapped in a single trap, with a time-varying Raman coupling between the two levels given by a spatially uniform electromagnetic field. The Hamiltonian for this system, with the annihilation operators for the two distinct states denoted $\hat{a}$ and $%
\hat{b}$ under the two mode approximation is the following: $$\begin{aligned}
\hat{H} & = & \hat{H}_a + \hat{H}_b + \hat{H}_{{\rm int}} + \hat{H}_{{\rm las%
}} \\
\hat{H}_{a} & = & \omega_{a}\hat{a}^{\dagger} \hat{a} + \frac{U_{a}}{2} \hat{%
a}^{\dagger}\hat{a}^{\dagger}\hat{a}\hat{a} \\
\hat{H}_{b} & = & \omega_{b}\hat{b}^{\dagger}\hat{b} + \frac{U_{b}}{2} \hat{b%
}^{\dagger}\hat{b}^{\dagger}\hat{b}\hat{b} \\
\hat{H}_{{\rm int}} & = & \frac{U_{ab}}{2} \hat{a}^{\dagger}\hat{a} \hat{b}%
^{\dagger}\hat{b} \\
\hat{H}_{{\rm las}} & = & \Omega(t) ( \hat{a}^{\dagger}\hat{b}e^{i\varphi
(t)} + \hat{b}^{\dagger}\hat{a} e^{-i\varphi (t)} ).\end{aligned}$$ $\hat{H}_{a}$ and $\hat{H}_{b}$ describe the two condensates undergoing self-interaction while $\hat{H}_{{\rm int}}$ and $\hat{H}_{{\rm las}}$ describe the condensates interacting with each other via collisions and laser-induced interactions respectively. $\hat{H}_{{\rm las}}$ describes a time-dependent coupling, rendering an overall time-dependent Hamiltonian for the system. The Hamiltonian may be rewritten by employing the Schwinger notation for the angular momentum operators, namely, $\hat{J}_x = \frac{1}{2}
( \hat{a}^{\dagger}\hat{b} + \hat{b}^{\dagger}\hat{a})$, $\hat{J}_y = \frac{1%
}{2i} ( \hat{a}^{\dagger}\hat{b} - \hat{b}^{\dagger}\hat{a})$, and $\hat{J}%
_z = \frac{1}{2} ( \hat{a}^{\dagger}\hat{a} - \hat{b}^{\dagger}\hat{b})$ with the Casimir invariant $J^2 = \frac{\hat{N}}{2} (\frac{\hat{N}}{2} + 1)$ where $\hat{N} = \hat{a}^{\dagger} \hat{a} + \hat{b}^{\dagger} \hat{b}$ is the total number operator and is a conserved quantity. Physically, eigenvalues of the operator $\hat{J}_z$ represents the difference in the number of atoms in different hyperfine levels, while $\hat{J}_{x}$ and $\hat{J}_{y}$ takes on the meaning of relative phase between the two species. The $z$ component angular momentum eigenstates, known also as the Dicke states[@bloch; @barnettradmore], are written as $|j, m \rangle$ where $m = -j, \ldots, j$ with $\hat{J}_z |j, m \rangle = m |j, m \rangle$. Here, $j = N/2$ and is the quantum number of angular momentum. The raising and lowering operators are defined in the usual way as $\hat{J}_{\pm} = \hat{J}_{x} \pm i \hat{J}_{y} $ such that $\hat{J}_{\pm }|j,m\rangle =\sqrt{(j\mp m)(j\pm
m+1)}|j,m\pm 1\rangle$. In terms of the angular momentum operators, the Hamiltonian takes the form[@choi; @chen]: $$\hat{H}(t) = \omega_{0} \hat{J}_z + q\hat{J}_z^2 + \Omega(t) \left [ \hat{J}%
_{+}e^{i\varphi (t)} + \hat{J}_{-}e^{-i\varphi (t)} \right ],
\label{Hamiltonian}$$ where $\omega_{0} = \omega_{a} - \omega_{b} + (N-1) (U_{a} - U_{b})/2$, $q =
(U_{a} + U_{b} - U_{ab})/2$. It is noted that $U_{a}(U_{b})$ or $U_{ab}$ may, in principle, be tuned via Feshbach resonance through the application of an external magnetic field[@feshbach]; the factors $\omega_{0}$ and $q$ are consequently adjustable parameters.
Solution to the time-dependent Schrödinger Equation
---------------------------------------------------
An exact solution to the Schrödinger Equation $$i\hbar \frac{d}{dt} |\psi(t)\rangle = \hat{H}(t) |\psi (t) \rangle$$ with the time-dependent Hamiltonian $\hat{H}(t)$ of Eq. (\[Hamiltonian\]) can be given in terms of a time evolution operator $\hat{U}(t)$, $|\psi(t)
\rangle = \hat{U}(t) |\psi(0) \rangle$ [@chen]: $$U(t) = \hat{R}^{\dagger}(t)e^{-iH^{\prime}(t)}\hat{R}(0), \label{U_t}$$ where $\hat{R}$ is a time dependent unitary transformation defined as $$\hat{R}(t)= \exp \left [ \frac{\lambda}{2} (\hat{J}_{-} e^{-i \varphi (t)} - \hat{J}_{+}
e^{i \varphi (t)}) \right ], \label{R_t}$$ and $$\hat{H}^{\prime}(t) = \hat{R}\hat{H}(t)\hat{R}^{\dagger} - i \hat{R} \frac{%
\partial}{\partial t} \hat{R}^{\dagger}. \label{Hprimed}$$ The parameter $\lambda$ in Eq. (\[R\_t\]) is an auxiliary parameter which may be chosen to simplify the transformed Hamiltonian, $\hat{H}^{\prime}$. It can be shown that $\hat{R}$ generates a gauge transformation under which the time-dependent Schrödinger Equation is covariant[@liang]: $$i\hbar \frac{d}{dt} |\psi^{\prime}(t)\rangle = \hat{H}^{\prime}(t)
|\psi^{\prime}(t) \rangle,$$ where the transformed state $|\psi^{\prime}(t) \rangle = \hat{R} |\psi (t)
\rangle$. For our case, $\lambda$ is chosen such that the Hamiltonian is diagonal in the $\hat{J}_z$ representation. With an additional assumption of adiabaticity conditions $d \varphi/dt \approx 0$ and $d \lambda /dt \approx
0 $, and also assuming that the two-photon transition terms proportional to $%
\hat{J}_{+}^2e^{i2\varphi(t)}$ and $\hat{J}_{-}^2e^{-i2\varphi(t)}$ can be neglected [@tunneling], one obtains for the transformed Hamiltonian $$\hat{H}^{\prime}(t) = \sqrt{\omega_0^2 + 4 \Omega^{2}(t)} \hat{J}_z - \frac{q%
}{2} \hat{J}_{z}^2, \label{Hamiltonian2}$$ where $\lambda$ is chosen such that $$\tan \lambda = - \frac{2\Omega(t)}{\omega_0}.$$ The time evolution of any observable $\hat{A}$ is then given in the Heisenberg picture as $\langle \hat{A}(t) \rangle \equiv \langle \psi(0) |
U^{\dagger}(t) \hat{A} U(t) | \psi(0) \rangle$, where $U(t)$ is defined in Eqs. (\[U\_t\]-\[Hprimed\]) and $| \psi(0) \rangle$ is the initial quantum state of the system. We note here that by ignoring the two-photon transitions for the purpose of obtaining analytical solutions, the effect of “two-axis counter-twisting”[@ueda] is not included in our effective Hamiltonian. For concreteness, we shall consider in this paper the case $\varphi (t)=\Delta t$ where $\Delta$ gives the detuning of the laser from the $|A\rangle \rightarrow |B\rangle $ transition between the two species. Also, we shall consider the case $U_{a} = U_{b} > U_{ab}$ and identical trapping potentials for the two species, corresponding to the $|F=1, M_{F} = \pm 1 \rangle$ hyperfine states of Na trapped in an optical dipole trap[@twocomp1]. This implies $q > 0$ as well as $\omega_0 = 0$ i.e. $\lambda = -\pi/2$ in Eq. (\[R\_t\]).
Initial quantum state of TBEC
-----------------------------
We consider in this paper an $SU(2)$ atomic coherent state or a coherent spin state (CSS), $| \theta, \phi \rangle$[@bloch; @barnettradmore] as the initial quantum state for a TBEC. A CSS is a minimum uncertainty state that describes a system with well-defined relative phase between the two species, and provides a good description of TBECs under suitable experimental conditions[@savage]. It is defined mathematically by applying the rotation operator on the extreme Dicke state $%
|j, j \rangle$ or $|j, -j \rangle$. The definition that we use in this paper is: $$|\theta, \phi \rangle = \exp \left [ \frac{\theta}{2} ( \hat{J}_{-}
e^{i\phi} - \hat{J}_{+} e^{-i \phi} ) \right ] |j, j \rangle.
\label{coherent}$$ A CSS $| \theta, \phi \rangle$ is therefore an eigenstate of the spin component in the $(\theta, \phi)$ direction $\hat{J}_{\theta, \phi} \equiv
\hat{J}_{x} \sin \theta \cos \phi + \hat{J}_{y} \sin \theta \sin \phi + \hat{%
J}_{z} \cos \phi$ with eigenvalue $J$ where $\theta$ and $\phi$ denote polar and azimuthal angles. It can be shown, following from Eq. (\[coherent\]), that CSS may be written as a superposition state: $$\begin{aligned}
|\theta, \phi \rangle & = & \sum_{m = -j}^{j} {\cal R}_{m}^{j}(\theta, \phi) | j, m \rangle , \label{coherent2}\end{aligned}$$ where $$\begin{aligned}
{\cal R}_{m}^{j}(\theta, \phi) & = & (C^{2j}_{j + m})^{1/2} \cos^{j+m} \left ( \frac{\theta}{2} \right )\sin^{j-m} \left ( \frac{\theta }{2} \right ) \nonumber \\
& & \times e^{i(j-m)\phi} \label{Rjmtp}\end{aligned}$$ and $C^{n}_{m}$ denotes the combination, $C^{n}_{m} = n!/[(n-m)!m!]$. Further extensive discussions on the properties of the CSS may be found in Refs. [@bloch; @barnettradmore]. In this paper, we shall consider initial CSS’s that may be produced experimentally, $|\theta = 0, \phi = 0 \rangle$, and $|\theta = \pi/2, \phi = 0 \rangle$. The CSS $|\theta = 0, \phi = 0 \rangle$ is simply a Dicke state $|j,j \rangle$ in which all atoms are found in one hyperfine level, and the CSS $|\theta = \pi/2, \phi = 0 \rangle$ is an eigenstate of $\hat{J}_x$ which is a state with a well-defined phase difference $\phi = 0$, and is a phase state of a two-mode boson system[@Castin]. States with different values of $\theta$, although simple to deal with mathematically, are difficult to produce and observe experimentally. For the same reason, we shall restrict our consideration of squeezing to those in the $x$, $y$, and $z$ directions since the variances in the relative phase and atom number are physical quantities clearly measurable without the ambiguity associated with the measurement of variances in other directions.
Results
=======
In this section, we present various results for the two initial states for a TBEC, the Dicke state and the phase state. The coefficient $q$ provides the strength of the scattering interaction between bosons, and its magnitude in relation to the tunnelling coupling $\Omega$ (as we consider $\omega_0 = 0$ in this paper) is an important parameter in determining distinct coupling regimes[@leggett]. Of the three coupling regimes, namely, the Rabi ($q/\Omega \ll 1/N$), the Josephson ($1/N \ll q/\Omega \ll N$) and the Fock ($q/\Omega \gg N$) regimes, it is clear that the regimes of more physical interest are the Rabi and the Josephson regimes, since, in the Fock regime the tunnelling coupling is overwhelmed by the nonlinear interaction term, resulting in no new effects of particular interest. This has been confirmed numerically[@tonel].
In the Rabi regime, the tunnelling coupling dominates with $q/\Omega$ near zero. We shall therefore consider the limiting case of $q = 0$ i.e. the linear regime which not only illustrates the essential physics of the Rabi regime but also demonstrate clearly the effect of nonlinear interaction on squeezing and entanglement. In addition, we shall consider the lower end of the Josephson regime, as it was found that important features rapidly disappear as one nears the Fock regime. The two couplings considered in this paper are therefore $q = 0$ and $q = q_{j}$ with $q_{j}/\Omega = 3/N$ for the Josephson regime, where we choose $N$ to be 400. We also have an additional parameter in our model not considered in some of the previous work, that of a detuning $\Delta$. As will be shown below, this enters mainly as a phase shift which affects the results in a nontrivial manner. To illustrate the effect of this parameter we shall consider detuning of the same magnitude as the $q$ for the Josephson regime, ($\Delta = q_{j}$) as well as a relatively large value of detuning ($\Delta/\Omega = 10q_{j}$). In particular, the results in the linear ($q=0$) regime will help us see the role of $\Delta$ with respect to $q$. The result for $\Delta = 0$ can be easily interpolated from these two values of detunings so will not be explicitly presented.
Dynamics of the macroscopic spin vector
---------------------------------------
We first consider how the quantum state of a TBEC represented by the fictitious spin vector evolves in time by calculating the expectation values $\langle \hat{J}_{\alpha}(t) \rangle$ where $\alpha$ stands for $x$, $y$, and $z$. This later gives us insight into how the squeezing and entanglement in a TBEC dynamically evolves. We evaluate the mean values of spin components based on Eq. (\[U\_t\]), $$\langle \hat{J}_{\alpha} \rangle = \langle \theta, \phi |\hat{R}%
^{\dagger}(0) e^{i\hat{H}^{\prime}} \hat{R}(t) \hat{J}_{\alpha} \hat{R}%
^{\dagger}(t) e^{-i\hat{H}^{\prime}} \hat{R}(0) | \theta, \phi \rangle . \label{J_alpha}$$ In the first instance, this may be calculated by using the commutation relation amongst the spin-$J$ components, namely $[\hat{J}_z,
\hat{J}_{\pm}] = \pm \hat{J}_{\pm}$ and $[\hat{J}_{+}, \hat{J}_{-}] = 2 \hat{%
J}_{z}$, and employing the Baker-Hausdorff theorem[@barnettradmore]. Useful identities are given in Eqs. (\[RJR\]-\[RJmR\]) with $\varphi(t) \equiv \Delta t$: $$\begin{aligned}
\hat{R} \hat{J}_{z} \hat{R}^{\dagger} & = & - \frac{1}{2} ( \hat{J}_{+}e^{i
\Delta t} + \hat{J}_{-}e^{-i\Delta t} ), \label{RJR} \\
\hat{R} \hat{J}_{+} \hat{R}^{\dagger} & = & \frac{1}{2} ( \hat{J}_{+} -
\hat{J}_{-}e^{-2i\Delta t}) + \hat{J}_{z}e^{-i\Delta t} , \label{RJpR}
\\
\hat{R} \hat{J}_{-} \hat{R}^{\dagger} & = & \frac{1}{2} ( \hat{J}_{-} -
\hat{J}_{+}e^{2i\Delta t}) + \hat{J}_{z}e^{i\Delta t} . \label{RJmR}\end{aligned}$$ In principle, it is possible to repeatedly apply the Baker-Hausdorff theorem to obtain a general expression for $\hat{J}_{\alpha}(t) \equiv \hat{U}%
^{\dagger}(t)\hat{J}_{\alpha}\hat{U}(t)$; for the special case of $q=0$, the expressions may, in fact, be simplified further, giving $$\langle \hat{J}_{z}(t) \rangle = \frac{N}{2} \cos \theta \cos [2\Omega
t+\Delta t + \phi] , \label{evJztq0}$$ and $$\begin{aligned}
\langle \hat{J}_{\pm}(t) \rangle &=& \frac{N}{2} e^{\mp i\Delta t} \left [
\cos \left(\frac{\theta - \pi/2}{2} \right ) \right. \nonumber \\
&& \left. \mp i \sin \left(\frac{\theta - \pi/2}{2} \right ) \sin(2\Omega t + \Delta t + \phi) \right ]. \label{evJpmtq0}\end{aligned}$$ It is found however that the expressions quickly become unwieldy especially for $q \neq 0$. The way around the problem is to note an important and useful point that, for the Hamiltonian under consideration, the transformation operator $\hat{R}$ happens to have the same form as the generator of $SU(2)$ atomic coherent states\[Eq. (\[coherent\])\] if one writes $\varphi(t) = -\phi$. In particular, the effect of $\hat{R}$ is to rotate the CSS such that $\hat{R}%
(\theta ^{\prime },\phi ^{\prime })|\theta ,\phi \rangle \rightarrow |\theta
+\theta ^{\prime },\phi - \phi ^{\prime }\rangle $, i.e. $$\hat{R}(t)|\theta ,\phi \rangle =
\sum_{m=-j}^{j} {\cal R}_{m}^{j}(\theta - \pi/2, \phi - \Delta t) |j,m\rangle , \label{Rphi}$$ where ${\cal R}_{m}^{j}(\theta, \phi)$ is defined in Eq. (\[Rjmtp\]). Using the well-known relation for the raising and lowering operators $\hat{J}_{\pm}$, and the fact that $F(\hat{J}_{z})|j,m\rangle =F(m)|j,m\rangle $ where $F$ denotes some analytic function, one can obtain, after some algebra, an analytical expression for $\langle \hat{J}_{z}(t) \rangle$ as a summation: $$\begin{aligned}
\langle \hat{J}_{z}(t)\rangle &=& -\sum_{m=-N/2}^{N/2-1} {\cal D}%
_{1}(\theta,m) \nonumber \\
&& \times \cos \left[2\Omega t - q\left( m+ \frac{1}{2}\right) t +
\Delta t + \phi \right], \label{Jzt2}\end{aligned}$$ where we have defined $$\begin{aligned}
{\cal D}_{\alpha}(\theta,m) & = & C_{N/2+m+1}^{N}\left( \frac{N}{2}+m+1
\right ) \nonumber \\
&& \times \cos ^{2N}\left( \frac{\theta -\pi /2}{2} \right) \nonumber \\
&& \times \tan^{N-2m-\alpha}\left( \frac{\theta -\pi /2}{2}\right).\end{aligned}$$ Following somewhat more involved but identical steps as above, we obtain: $$\begin{aligned}
\langle \hat{J}_{\pm}(t)\rangle &=& \sum_{m=-N/2}^{N/2-1} e^{\mp i (\Delta
t - \pi/2)} {\cal D}_{1}(\theta,m) \nonumber \\
&& \times \sin \left [ 2\Omega t - q \left ( m + \frac{1}{2} \right ) t + \Delta t + \phi \right ] \nonumber \\
&& + \sum_{m=-N/2}^{N/2} e^{ \mp i \Delta t} {\cal D}_{0}(\theta,m) \left
( \frac{m}{j-m} \right ). \label{Jpt2}\end{aligned}$$ from which $\langle \hat{J}_{x}(t) \rangle$ and $\langle \hat{J}_{y}(t)
\rangle$ may be deduced: $\hat{J}_{x} = \frac{1}{2}( \hat{J}_{+} + \hat{J}_{-})$ and $\hat{J}_{y} = \frac{1}{2i}( \hat{J}_{+} - \hat{J}_{-})$.
![$\langle \hat{J}_{z}(t)\rangle/N$, $\langle \hat{J}_{x}(t)\rangle/N$, and $\langle \hat{J}_{y}(t)\rangle/N$ for $q = 0$. Solid line: initial Dicke state, $|\theta = 0, \phi = 0 \rangle$ Dashed line: initial phase state $|\theta = \pi/2, \phi = 0 \rangle$. Left column: $\Delta = q_{j}$. Right column: $\Delta = 10q_{j}$, where $q_{j}/\Omega = 3/N$[]{data-label="Jq0"}](figureJq0.ps){height="7cm"}
We first present in Figs. \[Jq0\] and \[Jqsmall\], the expectation values $\langle \hat{J}_{\alpha}(t)\rangle$, $\alpha = x,y,z$ scaled to the total number of atoms $N$ for $q = 0$, and $q = q_{j}$ respectively for the detunings $\Delta = q_{j}$ and $\Delta = 10q_{j}$. It is noted that no significant qualitative difference is observed in these scaled amplitudes when one uses different number of atoms. In all the figures in this paper, the solid line represents the result for the initial CSS $|\theta, \phi=0 \rangle$ with $\theta = 0$ i.e. the Dicke state, while the dashed line corresponds to the initial phase state, $\theta = \pi/2$. In Fig. \[Jq0\], it is seen that, for the initial state $|\theta = 0, \phi = 0 \rangle$, the $y$ and $z$ components of spin, $\langle \hat{J}_{y} \rangle$ and $\langle \hat{J}_{z} \rangle$, undergo oscillatory evolution with a $\pi/2$ phase shift. This implies that the macroscopic spin vector may be visualized as if undergoing a circular motion in the $y$-$z$ plane. Superposed on this motion is the gradual increase of the oscillation in the $x$ component, $\langle \hat{J}_{x} \rangle$ with the amplitude of the $y$ component being proportionally reduced with a $\pi/2$ phase shift. This implies that the spin vector initially undergoing a circular motion in the $y$-$z$ plane rotates around the $z$-axis with increasing amplitude. This motion gets more pronounced for the larger $\Delta$ in that the spin vector undergoing circular motion in the $y$-$z$ plane rotates around the $z$-axis relatively quickly into a circular motion in the $x$-$z$ plane. This continues on to return to the $y$-$z$ plane, repeating this pattern over time. For the initial state $|\theta = \pi/2, \phi = 0 \rangle$, there is no oscillation in the $y$-$z$ plane, and only a slow rotation through the $x$-$y$ plane is observed. This state may be identified as a “self-trapping” state, as there is no transfer of populations during the evolution. With higher $\Delta$, the frequency of this rotation is clearly increased.
![$\langle \hat{J}_{z}(t)\rangle/N$, $\langle \hat{J}_{x}(t)\rangle/N$, and $\langle \hat{J}_{y}(t)\rangle/N$ for $q = q_{j}$. Identical details as in Fig. \[Jq0\].[]{data-label="Jqsmall"}](figureJqsmall.ps){height="7cm"}
For the case of $q = q_{j}$, Fig. \[Jqsmall\] shows the dephasing or “collapse” of oscillations for the initial state $|\theta = 0, \phi = 0 \rangle$. The three spin components collapse to give average spin of almost zero. As will be seen below, this has a significant impact on the variance, and consequently on squeezing and entanglement in a TBEC. For the self-trapping initial state $|\theta = \pi/2, \phi = 0 \rangle$, there is no change in the behavior with $q = q_{j}$. As to be expected, higher values of nonlinearity $q$ results in shorter time scales for the collapse and with higher number of atoms, no significant qualitative difference is observed except that the collapse happens somewhat faster. The long time simulation is provided in Fig. \[Jqsmalllong\], demonstrating clear revivals.
![$\langle \hat{J}_{z}(t)\rangle/N$, $\langle \hat{J}_{x}(t)\rangle/N$, and $\langle \hat{J}_{y}(t)\rangle/N$ for $q = q_{j}$, simulated for a longer time to observe revivals. Identical details as in Fig. \[Jq0\].[]{data-label="Jqsmalllong"}](figureJqsmalllong.ps){height="7cm"}
The zero mean spin due to collapse indicates that the quantum state of a TBEC evolves from a CSS which is similar to the usual coherent state in quantum optics to a state which is similar to the number eigenstate with an equal number of atoms in each mode. The state that the CSS evolves into is, however, not exactly a self-trapping state as it clearly revives back into a CSS; it may be viewed as a quantum state with a small spread around the mean self-trapping state. For lack of better terminology we shall refer to this collapsed quantum state as a “[*quasi*]{}-self-trapping state” in this paper.
Quantum squeezing and fluctuations
----------------------------------
In the language of spin squeezing, any reduction of the quantum noise for $q = 0$ corresponds to an effective squeezing due to the rotation of the coordinate axes, while squeezing observed for $q = q_{j}$ comes from a complex combination of coordinate rotation on top of the interatomic collision effect of the one-axis twisting term proportional to $\hat{J}^{2}_{z}$[@ueda]. In order to calculate the amount of squeezing in the system, we first calculate the variances in the three components of the macroscopic spin vector, i.e. $%
[\Delta \hat{J}_{\alpha}(t)]^2 = \langle \hat{J}^{2}_{\alpha}(t) \rangle -
\langle \hat{J}_{\alpha}(t) \rangle^{2}$, $\alpha = x,y,z$ which, as mentioned above are experimentally accessible quantities. The unitary property of $\hat{R}$ i.e. $\hat{%
R}^{\dagger}\hat{R} = \openone$ is used to evaluate terms of the form $\hat{R}\hat{J}%
_{\alpha} \hat{J}_{\beta} \hat{R}^{\dagger} \equiv \hat{R} \hat{J}_{\alpha}
\hat{R}^{\dagger} \hat{R} \hat{J}_{\beta} \hat{R}^{\dagger}$ along with the identities Eqs. (\[RJR\]-\[RJmR\]).
It can be shown that $\langle \hat{J}_{z}^{2}(t)\rangle$ is given by $$\begin{aligned}
\langle \hat{J}_{z}^{2}(t)\rangle &=& \frac{1}{2} \sum_{m=-N/2}^{N/2-2}
{\cal D}_{2}(\theta,m) \left( \frac{N}{2} - m - 1 \right) \nonumber \\
& & \times \cos [4\Omega t - 2q(m+1) t + 2\Delta t + 2\phi] \nonumber \\
& & + \frac{1}{4} \sum_{m=-N/2}^{N/2-1} {\cal D}_{2}(\theta,m) \left( \frac{N}{2} - m \right ) \nonumber \\
&& + {\cal D}_{0}(\theta,m) \left( \frac{N}{2} + m + 1
\right), \label{Jzsqt2}\end{aligned}$$ so that the variance in the $z$ direction, $\lbrack \Delta \hat{J}_{z}(t)]^{2}$, is simply given by subtracting the square of Eq. (\[Jzt2\]) from Eq. (\[Jzsqt2\]). On the other hand, $\langle \hat{J}^{2}_{x} \rangle = \frac{1}{4} [ \langle \hat{J}^{2}_{+} \rangle + \langle
\hat{J}^{2}_{-} \rangle + \langle \hat{J}_{+} \hat{J}_{-} \rangle + \langle
\hat{J}_{-} \hat{J}_{+} \rangle ]$ and $\langle \hat{J}^{2}_{y} \rangle =
\frac{1}{4} [\langle \hat{J}_{+} \hat{J}_{-} \rangle + \langle \hat{J}_{-} \hat{%
J}_{+} \rangle - \langle \hat{J}^{2}_{-} \rangle - \langle \hat{J}^{2}_{+}
\rangle]$. The required expectation values of $\langle \hat{J}^{2}_{\pm} \rangle$ and $\langle \hat{J}_{\pm}\hat{J}_{\mp} \rangle$ are given by:
$$\begin{aligned}
\langle \hat{J}_{+}^{2}(t)\rangle &=& \sum_{m=-N/2}^{N/2-2}
\frac{e^{-2i \Delta t}}{2} {\cal D}_{2}(\theta,m) \left( \frac{N}{2} - m - 1 \right
) \cos [ 4\Omega t - 2q(m+1) t + 2\Delta t + 2\phi ] \nonumber \\
&& - \sum_{m=-N/2}^{N/2-1} \frac{e^{-2i \Delta t}}{4} \left \{
{\cal D}_{2}(\theta,m) \left( \frac{N}{2} - m \right ) + {\cal D}_{0}(\theta,m) \left( \frac{N}{2} + m + 1 \right ) \right. \nonumber \\
& & + \left. (2m+1)i \sin \left [ 2\Omega t - q \left (m + \frac{1}{2} \right ) t + \Delta t + \phi \right ] \right \} + \sum_{m=-N/2}^{N/2} e^{-2i \Delta t} {\cal D}_{0}(\theta,m) \left (\frac{m^2}{j-m} \right ), \label{Jpsqt2}\end{aligned}$$
$$\begin{aligned}
\langle \hat{J}_{+}(t)\hat{J}_{-}(t)\rangle &=& -
\sum_{m=-N/2}^{N/2-2} \frac{1}{2} {\cal D}_{2}(\theta,m) \left( \frac{N}{2} - m - 1
\right ) \cos[4\Omega t - 2q(m+1) t + 2 \Delta t + 2 \phi ] \nonumber \\
&& + \sum_{m=-N/2}^{N/2-1} \frac{1}{4} \left \{ {\cal D}_{2}(\theta,m) \left( \frac{N}{2} - m \right ) + {\cal D}_{0}(\theta,m) \left ( \frac{N}{2} + m + 1 \right ) \right. \nonumber \\
&& - \left. 4{\cal D}_{1}(\theta,m) \cos \left [ 2\Omega t - q\left ( m + \frac{1}{2} \right ) t + \Delta t + \phi \right ] \right \} + \sum_{m=-N/2}^{N/2} {\cal D}_{0}(\theta,m) \left ( \frac{m^2}{j-m} \right ), \label{Jpmsqt2}\end{aligned}$$
with $\langle \hat{J}_{-}^{2}(t)\rangle \equiv \langle \hat{J}_{+}^{2}(t)\rangle^{*}$ where asterisk (\*) denotes complex conjugate, and $\langle \hat{J}_{-}(t)\hat{J}_{+}(t)\rangle = \langle \hat{J}_{+}(t)\hat{J}_{-}(t)\rangle - 2 \langle \hat{J}_{z}(t) \rangle$, as to be expected from the commutation relation $[ \hat{J}_{+}, \hat{J}_{-}] = 2\hat{J}_{z}$. From these expressions and those for $\langle \hat{J}_{x}(t) \rangle$ and $\langle \hat{J}_{y}(t) \rangle$ obtained above, the variances $[\Delta \hat{J}_{x}(t)]^{2}$ and $\lbrack \Delta \hat{J}_{y}(t)]^{2}$ may be calculated.
We find that, for $q=0$, the variances $[\Delta \hat{J}%
_{\alpha}]^2$, $\alpha = x,y,z$, show an oscillatory behavior which is bounded above by $N/4$ \[Fig. \[Varq0\]\]. This is an expected result for a CSS, which is known to have variances in the standard quantum limit (SQL) of $J/2$. One key observation regarding Fig. \[Varq0\] is that the variances do go below the SQL. This kind of squeezing which is due to the rotation of the coordinate axis[@ueda] naturally occurs since a spin vector is an eigenstate of spin in one direction with zero variance in that direction. As the spin vector traverses the phase space due to dynamical evolution, the error ellipsoid follows the path of the spin vector in such a way that the minor axis of the ellipsoid periodically lines up with the $x$, $y$ or $z$ axis, resulting in the reduction of quantum fluctuations in that direction.
![Variances $[\Delta \hat{J}_{z}(t)]^2$, $[\Delta \hat{J}_{x}(t)]^2$, and $[\Delta \hat{J}_{y}(t)]^2$ for $q = 0$, scaled to SQL of $N/4$ for clarity. Scaled variance of 1 in this figure corresponds to the SQL. Identical details as in Fig. \[Jq0\].[]{data-label="Varq0"}](figureVarq0.ps){height="7cm"}
For small $\Delta$, the variance in the $z$ direction show oscillatory reduction with $\pi/2$ phase shift with that of the $y$ component. The variance in the $x$ direction is seen as more or less maintaining its value near the SQL. For larger $\Delta$, the uncertainty in the $x$ direction is found oscillating with increasing amplitude, which is consistent with the error ellipsoid following the motion of the rotating spin vector discussed above.
For the $q = q_{j}$ case, the variances in the $x$, $y$, and $z$ directions give significantly different behavior, as shown in Fig. \[Varqsmall\]. The unexpected feature is the large variance equivalent to the maximum relative uncertainty (standard deviation) of order $\pm 35 \%$ per measurement. Mathematically, this can be understood as the corollary of the collapsing mean spin components; the collapsing spin vector implies that the the variance $[\Delta \hat{J}_{\alpha}]^2 = \langle \hat{J}_{\alpha}^2 \rangle - \langle \hat{J}_{\alpha} \rangle^2$ is dominated by the $\langle \hat{J}_{\alpha}^2 \rangle$ term of the order $N^2$ as it cannot be cancelled by the $\langle \hat{J}_{\alpha} \rangle^2$ term. Careful analysis of the variance for the $q=0$ case discussed above reveals that the terms of the form $\langle \hat{J}_{\alpha}^2 \rangle$ are almost exactly cancelled by the $\langle \hat{J}_{\alpha} \rangle^2$ which are also of the order $N^2$ to maintain the variances of the order of the SQL, $N/4$. This type of cancellation cannot happen as the spin vector collapses with $q = q_{j}$. Physically, the increased uncertainty can be attributed to the quantum state evolving away from the minimum uncertainty state of CSS into a quasi-self-trapping state with much higher quantum fluctuations. In order to display periodic behavior in the variances, we plot in Fig. \[Varqsmalllong\] the variances over the identical time period as in Fig. \[Jqsmalllong\]. The result is in agreement with the purely numerical result of Tonel [*et al.*]{}[@tonel]. It is found that, as the revivals occur, the variances do go below the SQL periodically. Although it is not clearly depicted due to scaling, the results for the initial phase state $| \theta = \pi/2, \phi = 0 \rangle$ are identical to that for the $q = 0$ case i.e. the variance stays at 1 for the atom number difference, and oscillates between $0$ and the SQL for the relative phase.
![Variances $[\Delta \hat{J}_{z}(t)]^2$, $[\Delta \hat{J}_{x}(t)]^2$, and $[\Delta \hat{J}_{y}(t)]^2$ for $q = q_{j}$, again scaled to the SQL, $N/4$. Identical details as in Fig. \[Jq0\].[]{data-label="Varqsmall"}](figureVarqsmall.ps){height="7cm"}
![Variances $[\Delta \hat{J}_{z}(t)]^2$, $[\Delta \hat{J}_{x}(t)]^2$, and $[\Delta \hat{J}_{y}(t)]^2$ for $q = q_{j}$ scaled to the SQL, $N/4$, simulated for a longer time to observe revivals. Identical details as in Fig. \[Jq0\].[]{data-label="Varqsmalllong"}](figureVarqsmalllong.ps){height="7cm"}
Dynamical evolution of quantum entanglement in a TBEC
-----------------------------------------------------
As mentioned in Ref. [@Milburn], the amount of entanglement depends on the way a system is partitioned into subsystems. In this subsection we calculate the dynamical evolution of quantum entanglement in two different subsystem decomposition of the TBEC: entanglement between the particles and that between the two modes.
### Entanglement between the particles
The atom-atom entanglement is parameterized by Eq. (\[Sxi\]). Equation (\[Sxi\]) is useful since, regardless of the actual spin direction at a given time, one need only to identify three unit orthogonal directions ${\bf n}_{i}$ to check whether the system has become entangled due to system dynamics. If one’s goal is to identify maximum entanglement possible between the atoms, it is necessary to scan through all possible ${\bf n}_{i}$’s within the unit sphere at each point in time. However, it should be noted that keeping to a fixed direction may help in constructing an actual experimental scheme to measure $\xi_{\alpha}^{2}(t)$. In particular, we consider in this paper experimentally meaningful variances and amplitudes of the macroscopic spin vector in the $x$, $y$, and $z$ directions. We therefore consider $$\xi^{2}_{\alpha}(t) =\frac{N[\Delta \hat{J}_{\alpha}(t)]^{2}}{%
\langle \hat{J}_{\beta}(t)\rangle ^{2}+ \langle \hat{J}_{\gamma}(t)\rangle ^{2}}$$ where $\alpha, \beta, \gamma$ cycle through $x, y, z$.
For $q=0$, it was found that $\xi^{2}_{\alpha}(t) \equiv 1$, $\alpha = x,y,z$, at all times for both initial states i.e. not entangled despite the fact that, as already seen in Fig. \[Varq0\], squeezing due to the rotation of the coordinate axis does occur. This shows that the observed squeezing is not accompanied by an increase in quantum correlations between the particles. On the other hand, for $q=q_{j}$, it is found that $\xi^{2}_{\alpha}$ is maintained at 1 for the initial phase state as to be expected, while for the initial Dicke state, the very quickly increasing quantum fluctuations and the collapsing spin vector due to dephasing $\xi_{\alpha}^{2}(t)$ large. Physically, this is due to the fact that the CSS is evolving into a quasi-self-trapping quantum state with low atom-atom correlations. A correct balance between the rapidly increasing variance and the collapsing spin vector is clearly needed in order to maintain entanglement with $\xi_{\alpha}^2 (t) <1$. From Fig. \[Varqsmall\], it is seen that the variance in $\hat{J}_{x}$ increases most slowly amongst the variances, while Fig. \[Jqsmall\] indicates the spin vector $\langle \hat{J}_{y} \rangle + \langle \hat{J}_{z} \rangle$ oscillating with a non-zero amplitude, showing the most promise in finding entanglement in this spin direction. In Fig. \[Xiqsmall\] we show $\xi^2_{x} (t)$ for $q = q_{j}$, for $\Delta = q_j$ and $\Delta = 10q_j$. Differently from other results, it was found that with $\Delta = 0$, $\xi^2_{x}(t) \geq 1$ for all times, indicating the crucial role $\Delta$ plays in this system. For the initial Dicke state the parameter quickly becomes large, clearly indicating that quantum correlations between the atoms are destroyed rapidly. However, it is seen that the parameter $\xi_{x}^2 (t)$ does dip below 1 for a brief time period demonstrating atom-atom entanglement.
![Entanglement parameter $\xi_{x}^2$ for $q = q_{j}$. Identical details as in Fig. \[Jq0\].[]{data-label="Xiqsmall"}](figureXiqsmall.ps){height="7cm"}
![Entanglement parameter $\xi_{x}^2$ for reduced nonlinearity, $q = q_{j}/10$. Identical details as in Fig. \[Jq0\].[]{data-label="Xiqsmaller"}](figureXiqsmaller.ps){height="7cm"}
Somewhat counter-intuitively, one may extend the duration over which the atoms are entangled by [*reducing*]{} the nonlinearity constant $q$; it is clear that with a smaller $q$, the variance in $\hat{J}_{x}$ increases more gently while the collapse of the spin vector occurs over a longer time scale, extending the duration over which $\xi^2_{x} < 1$. This is presented in Fig. \[Xiqsmaller\] for $q = q_{j}/10$. As before, the increase in $\Delta$ results in the increase in oscillation frequency. This indicates a potential for the quantum control of entanglement properties of a TBEC via externally adjustable parameters.
### Entanglement between the two modes
Writing the quantum state of a TBEC in the form $$\begin{aligned}
|\Psi (t) \rangle & = & \sum_{m = 0}^{N} c_{m}(t) |N\rangle |N-m \rangle \\
& = & \sum_{m = -j}^{j} c_{m}(t) |j, m \rangle\end{aligned}$$ the well-known von Neumann entropy that shows degree of entanglement can be written[@Milburn]: $$E(t) = -\frac{1}{\log_2(N+1)} \sum_{m = -j}^{j} |c_m(t)|^2 \log_2 |c_m(t)|^2 \label{scaledE}$$ where the normalization factor $\log_2(N+1)^{-1}$ was included so that $0< E < 1$. The expansion coefficients are given by $$\begin{aligned}
c_m(t) & = & \langle m | \hat{U}(t) | \psi(0) \rangle \\
& = & \langle m | \hat{R}^{\dagger}(t) e^{-i H' t} \hat{R}(0) | \theta, \phi \rangle , \end{aligned}$$ where the rotation operator $\hat{R}$, reduced Hamiltonian $H'$ and the CSS $|\theta, \phi \rangle$ are as defined above. Using Eq. (\[Rphi\]), and inserting the completeness relation $\sum_{m'} |m' \rangle\langle m'| = \openone$ one can write $$\begin{aligned}
c_m(t) & = & \sum_{m'} {\cal F}^{(\lambda)}_{m,m'}(t) {\cal R}_{m'}^{j}(\theta - \pi/2,\phi -\Delta t) \nonumber \\
&& \times e^{-i[2m'\Omega t - qm'^2 t/2]}, \label{cmt}\end{aligned}$$ where $$\begin{aligned}
{\cal F}^{(\lambda)}_{m,m'}(t) & = & \langle m | \hat{R}^{\dagger}(t) |m' \rangle \end{aligned}$$ represents the matrix element of the rotation operator. This can be evaluated by first applying the disentangling theorem on the rotation operator $\hat{R}$[@bloch]: $$\begin{aligned}
\hat{R}(\lambda, \Delta t) & = & \exp \left [ \frac{\lambda}{2} ( \hat{J}_{-}e^{-i\Delta t} - \hat{J}_{+}e^{i\Delta t} ) \right ] \\
& = & e^{\tau \hat{J}_{+}}e^{\ln (1 + |\tau|^2 ) \hat{J}_{z}}e^{-\tau^{*} \hat{J}_{-}} ,\end{aligned}$$ where $\tau = e^{i\Delta t}\tan (-\frac{\lambda}{2})$. Taking into account the unitary nature of the rotation operator$\hat{R}^{\dagger}(\theta, \phi) = \hat{R}(-\theta, \phi)$ and expanding the exponential operators containing $\hat{J}_{+}$ and $\hat{J}_{-}$ one can write $$\begin{aligned}
{\cal F}^{(\lambda)}_{m,m'}(t) & = & \sum_{n, n'} \frac{(-\tau)^{n}}{n!} \frac{\tau^{* n'}}{n' !} (1+ |\tau|^2)^{m' - n' } \nonumber \\
&& \times \langle m | \hat{J}_{+}^{n} \hat{J}_{-}^{n' }|m' \rangle .\end{aligned}$$ The final expression for the matrix element, which was obtained by using the ladder operator nature of $\hat{J}_{\pm}$ and found to be identical to the irreducible representation of full rotation group [@bloch] is: $$\begin{aligned}
{\cal F}^{(\lambda)}_{m,m'}(t) & = & \sum_{n = 0}^{2j} \frac{(-1)^{m'-m+n}}{(n + m' - m)! n!} \frac{(j - m + n)!}{(j + m - n)!} \nonumber \\
&& \times \left [ \frac{(j + m)!(j+m')!}{(j - m)!(j-m')!} \right ]^{1/2} e^{i(m-m')\Delta t}
\nonumber \\
&& \times \sin^{2n + m' - m} \left ( \frac{\lambda}{2} \right )\cos^{-m'-m} \left (\frac{\lambda}{2} \right ). \label{Fmatrix}\end{aligned}$$ The normalized entanglement parameter $E(t)$ is plotted in Fig. \[E\], for the four combinations of $q$ and $\Delta$: $q = 0$, $q = q_{j}$ and $\Delta = q_{j}$, $\Delta = 10q_{j}$. It is seen that the nonlinearity is crucial for achieving high degree of entanglement. The effect of $\Delta$ was, as expected, to increase the frequency of oscillations. To generate Fig. \[E\], we used Eqs. (\[scaledE\]), (\[cmt\]), and (\[Fmatrix\]) with $N=40$ due to the computational limitations in numerically calculating factorials of large numbers. However, it was found that, with non-zero $q$, as long as the ratio of the various parameters such as $q/N$ is kept the same, there are no discernible changes in the plot as a function of $N$. Our result agrees with the numerical result obtained by Tonel [*et al.*]{}[@tonel] with higher number of atoms.
It was found that, for the initial phase state $|\theta = \pi/2, \phi = 0 \rangle$, the entanglement parameter does depend on the number of atoms, as the normalized coefficients $|c_m(t)|^2$ are given in this case by the binomial distribution, namely, $|c_m(t)|^2 \equiv C^{2j}_{j+m}$. It is also notable that the coefficients are no longer time dependent. As the number of particles $N$ and hence $j$ increases, the broad binomial distribution approximates narrower and narrower Gaussian as a function of $m$, and hence the state becomes less and less entangled. For example, for $N = 400$, it is found that the scaled $E = 0.62$. In Fig. \[Ephase\], we plot the entanglement parameter as a function of the number of atoms $N$ for the initial phase state. It is seen that the value tends towards 0.6 as the number is increased up to 1000 atoms. Such consistent behavior exhibited for the bipartite entanglement suggests a possibility to use this property to create a macroscopic matter-wave state with known entanglement.
![Scaled von Neumann entanglement parameters $E$ for $N = 40$. Top row: $q = 0$, Bottom row: $q = q_{j}$. Left column: $\Delta = q_{j}$, Right column: $\Delta = 10q_{j}$[]{data-label="E"}](figureE.ps){height="7cm"}
![Scaled von Neumann entanglement parameters $E$ as a function of the number of atoms $N$ for the initial phase state.[]{data-label="Ephase"}](figureEphase.ps){height="5cm"}
Conclusion
==========
We have studied the dynamical evolution of a TBEC in the presence of an adiabatically varying, off-resonant atom-light coupling. The macroscopic, fictitious spin vector was found to undergo a rather complex motion, and further complications in the form of collapses and revivals were found with nonzero nonlinearity. As the main result of this paper, we used an exact solution to the time-dependent Schrödinger equation to calculate analytically the amount of quantum mechanical squeezing and entanglement in a TBEC under various conditions. In particular, we considered entanglement generated between the atoms and that between the modes. For the case with the nonlinearity turned off, although the variance was found to be reduced below the SQL (i.e. squeezed), the system never demonstrated atom-atom entanglement, and the entanglement between the two modes remained well below its maximal value. With the nonlinearity turned on, it was found that the atom-atom entanglement can be generated initially, and then become unentangled rapidly. On the other hand, the entanglement between the two modes was found to more or less maintain maximal values throughout, except for some fluctuations. The dynamics of entanglement was found to be controllable via various parameters present in this system, namely the nonlinearity $q$ and the detuning $\Delta$.
Potentially useful results of this work are the squeezing below the SQL in the atom number difference and relative phase which may help reduce the projection noise in spectroscopic and interferometric applications as discussed by Wineland [*et al.*]{}[@wineland], and the consistently high degree of entanglement between the two modes with the nonlinearity turned on, which could be useful in the context of quantum information science. In addition, the dynamically stable features of the self-trapping state and the time-independent entanglement between the two modes which is also relatively insensitive to the changes in the atom number (for large $N$) could prove useful in designing a robust, macroscopic bipartite quantum state with known entanglement. Future work could involve studying quantum control methods for maintaining optimal squeezing and entanglement in a TBEC, for such applications as precision matter-wave interferometry using entangled BECs[@interferometry].
[99]{}
J. A. Dunningham, M. J. Collett, and D. F. Walls, Phys. Lett. A [**245**]{} 49 (1998)
J. Rogel-Salazar, S. Choi, G. H. C. New, and K. Burnett, Phys. Lett. A [**299**]{} 476 (2002)
J. Rogel-Salazar, G. H. C. New, S. Choi, and K. Burnett, Phys. Rev. A [**65**]{} 023601 (2002)
V. Chernyak, S. Choi, and S. Mukamel, Phys. Rev. A [**67**]{}, 053604 (2003)
S. Choi, and N. P. Bigelow, J. Mod. Opt. [**52**]{}, 1081 (2005)
D. F. Walls, Nature [**306**]{}, 141 (1983); D. F. Walls and G. J. Milburn, [*Quantum Optics*]{} (Springer-Verlag, Berlin, 1994)
D. F. Walls, and P. Zoller, Phys. Rev. Lett. [**47**]{}, 709 (1981)
M. Kitagawa, and M. Ueda, Phys. Rev. A [**47**]{}, 5138 (1993)
D. J. Wineland, J. J. Bollinger, W. M. Itano, and D. J. Heinzen, Phys. Rev. A [**50**]{}, 67 (1994)
A. S[o]{}rensen, L.-M. Duan, J. I. Cirac, and P. Zoller Nature (London) [**409**]{}, 63 (2001)
A. Micheli, D. Jaksch, J. I. Cirac, and P. Zoller, Phys. Rev. A, [**67**]{}, 013607 (2003)
S. D. Jenkins, and T. A. B. Kennedy Phys. Rev. A [**66**]{}, 043621 (2002)
L.-M. Duan, A. S[o]{}rensen, J. I. Cirac, and P. Zoller Phys. Rev. Lett. [**85**]{}, 3991 (2001)
M. G. Moore and P. Meystre Phys. Rev. A [**59**]{}, R1754 (1999)
A. P. Hines, R. H. McKenzie, and G. J. Milburn Phys. Rev. A [**67**]{}, 013609 (2003)
C. J. Myatt, E. A. Burt, R. W. Ghrist, E. A. Cornell, and C. E. Wieman, Phys. Rev. Lett [**78**]{}, 586 (1997)
J. Stenger, S. Inouye, D.M. Stamper-Kurn, H.-J. Miesner, A.P. Chikkatur, and W. Ketterle, Nature (London) [**396**]{}, 345 (1998); H.-J. Meisner, D.M. Stamper-Kurn, J. Stenger, S. Inouye, A.P. Chikkatur, and W. Ketterle, Phys. Rev. Lett. [**82**]{}, 2228 (1999)
F. T. Arrechi, E. Courtens, R. Gilmore, and H. Thomas, Phys. Rev. A [**6**]{}, 2211 (1972)
S. M. Barnett and P. M. Radmore, [*Methods in Theoretical Quantum Optics*]{} (Oxford University Press, Oxford, 1997)
I. Fuentes-Guridi, J. Pachos, S. Bose, V. Vedral, and S. Choi Phys. Rev. A [**66**]{}, 022102 (2002)
Z.-D. Chen, J.-Q. Liang, S.-Q. Shen, and W.-F. Xie Phys. Rev. A [**69**]{}, 023611 (2004)
S. Inouye, M.R. Andrews, J. Stenger, H.-J. Miesner, D.M. Stamper-Kurn, and W. Ketterle, Nature (London) [**392**]{}, 151 (1998)
J.-Q. Liang, and H. J. W. Müller-Kirsten, Ann. Phys. (N.Y.) [**219**]{}, 42 (1992); Y. Z. Lai, J.-Q. Liang, H. J. W. Müller-Kirsten, and J.-G. Zhou, Phys. Rev. A [**53**]{}, 3691 (1996)
L.-M. Kuang, and Z.-W. Ouyang, Phys. Rev. A [**61**]{}, 023604 (2000); W.-D. Li, X. J. Zhou, Y. Q. Wang, J. Q. Liang, and W. M. Liu, Phys. Rev. A [**64**]{}, 015602 (2001)
D. Gordon, and C. M. Savage, Phys. Rev. A [**61**]{}, 023604 (1999)
Y. Castin and J. Dalibard, Phys. Rev. A [**55**]{}, 4330 (1997)
A. J. Leggett, Rev. Mod. Phys. [**73**]{}, 307 (2001)
A. P. Tonel, J. Links, and A. Foerster, J. Phys. A: Math. Gen. [**38**]{}, 1235 (2005)
J. A. Dunningham, K. Burnett, and S. M. Barnett, Phys. Rev. Lett. [**89**]{}, 150401 (2002)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In-plane Nernst signal and resistivity have been measured for three $La_{2-x}Sr_xCuO_4$ single crystals (x=0.09, 0.11 and 0.145) with the magnetic field parallel to c-axis. A quadratic temperature dependence of resistivity, i.e., $\rho=\rho_0+aT^2$ is observed below a certain temperature $T_R$. It is found that the upper boundary of the Nernst signal $T_n$ coincides with $T_R$, which points to an inherent relation between the anomalous Nernst signal and the nodal quasiparticle transport in the pseudogap region. Finally a phase diagram together with the pseudogap temperature $T^*$ is presented, which suggests a second energy scale in the pseudogap region.'
author:
- 'Hai-Hu Wen, Hong Gao, Hao Jin, Lei Shan, Fang Zhou, Jiwu Xiong, Wenxin Ti'
title: Inherent Relation between Nernst Signal and Nodal Quasiparticle Transport in Pseudogap Region of Underdoped High Temperature Superconductors
---
One of the core issues in high temperature superconductors is the origin of a pseudogap above $T_c$ in underdoped region. In order to understand the physics behind the pseudogap, many models have been proposed, such as resonating valence bond (RVB)[@RVB] theory, spin fluctuation[@SpinFluc], preformed Cooper pairs[@Emery], charge stripes[@EmeryKivelson], d-density wave (DDW)[@Chakravarty; @Affleck], etc. Among many of them, the pseudogap state has been considered as a precursor to the superconducting state. In this precursor state, Cooper pairs have already formed before the long range phase coherence (or Bose-Einstein like condensation) is established. Measurements on the high frequency complex conductivity illustrate that a short-life phase coherence can persist up to about 30 K above $T_c$ and these data can be described by the dynamics of thermally generated topological defects (free vortices)[@Corson]. The Princeton group found that a significant in-plane Nernst signal[@XuZA] appears in the pseudogap region with H$\|$ c. This may be understood by the phase-slip due to the thermal drifting of vortex-like excitations. However, it remains unknown how does this strong Nernst signal relate to the nodal quasiparticles which are supposed to dominate the in-plane transport properties in low temperature region. In this Letter we present the evidence for an inherent relation between the Nernst signal and nodal quasiparticle transport: The upper boundary temperature $T_n$ of Nernst signal is found to coincide very well with a crossover temperature $T_R$ below which a quadratic temperature dependence of resistivity is observed. Possible reasons are given to explain this coincidence.
The single crystals measured in this work were prepared by travelling solvent floating-zone technique. Samples with three different doping concentrations p=0.09($T_c$=24.4K, as grown, x=0.09), 0.11($T_c$=29.3K, as grown, x=0.11), 0.145($T_c$=36.1K, nominal x=0.15) have been investigated. The quality of our samples has been characterized by x-ray diffraction, and $R(T)$ data showing a narrow transition $\Delta T_c \leq $ 2 K. For some samples, the full width at the half maximum (FWHM) of the rocking curve of (008) peak is only 0.10$^\circ$[@ZhouFang]. The samples have also been checked by AC and DC magnetization and resistive measurements. Inset(a) of Fig.1 shows a typical curve of the diamagnetic transition of sample $p=0.11$ measured at $H=20
Oe$.
![Temperature dependence of resistivity for sample $p=0.11$ (shown here as a typical example). An enhancement of resistivity is observed below a temperature $T_R$ (here about 164 K). Below $T_R$=164$\pm$20 K a quadratic temperature dependence of resistivity has been observed (see the linear part of Fig.1(b)). In Fig.1(c) we show the derivative of resistivity versus $T$. It is clear that the derivative $dR/dT$ starts to drop down at about 164 K. Similar behavior is found for other two samples with $p=0.09$ and $p=0.145$ although the $T_R$ are different.[]{data-label="fig1"}](fig1){width="8cm"}
The resistivity was measured by using standard four-point technique. For measuring the Nernst effect we adopted the one-heater-two-thermometer technique. A heating power of 2 mW is applied to one end of the single crystal and two tiny thermometers with distance of 1.5 mm are attached to the sample for measuring the temperatures along the heat flow (longitudinal) direction. The Nernst voltage is measured through two contacts on two opposite planes at the symmetric positions. Very small contacting resistance ($\leq 0.1 \Omega$) has been achieved by using silver paste. All samples are shaped into a bar structure with dimensions of $4-5 mm$ (length) $\times$ 1 mm (width) $\times$ 0.5 mm (thickness). All measurements are based on an Oxford cryogenic system ( Maglab-12 ) with temperature fluctuation less than 0.04%$T$ and magnetic fields up to 12 T. During the measurement for Nernst signal the magnetic field is applied parallel to $c-axis$ and swept between 7 to -7 T and the Nernst signal $V_N$ is obtained by subtracting the positive field value with the negative one in order to remove the Faraday signal during sweeping the field and the possible thermal electric power due to asymmetric contacts. The Nernst voltage is measured by a Keithley 182-Nanovoltmeter with a resolution of about 5 nV in present case. In this paper we show the Nernst signal $e_y=E_N/|\bigtriangledown
T|$, where $E_N = V_N/d$ with $d$ the distance between the two contacts for Nernst voltage, $\bigtriangledown T$ is the temperature gradient along the heat flow direction.
Fig.1 shows the temperature dependence of resistivity for sample $p=0.11$ (as a typical example). One can see that a quasi-linear behavior is observed above a characteristic temperature $T_R$ at about $164 K$. It is necessary to note that this quasi-linear part (as marked by the dotted line) is different from the linear part above $T^*$ which is about $400 K$ for $p=0.11$. Below $T_R$ the resistivity deviates from the linear behavior. In Fig.1(b) we present $\rho$ vs. $T^2$. One can see that below about $164\pm20
K$, the curve posses a linear behavior which indicates a relation of $\rho=\rho_0+aT^2$. In order to determine $T_R$ with higher accuracy, we present in Fig.1(c) the derivative of resistivity $dR/dT$ vs. $T$, it is evident that at about $T_R=164\pm20 K$ the derivative starts to drop down showing a crossover from a quasi-linear behavior (above $T_R$) to a quadratic behavior below $T_R$. Similar features are observed for other two samples ($p=0.09$ and $p=0.145$). Note that Ando et al.[@AndoT2] recently reported that the relation $\rho = \rho_0 + aT^2$ in low temperature region for underdoped La-214 and Y-123 systems, and proposed that this behavior is characteristic of the nodal quasiparticles on the so-called Fermi arcs. It is found also by these authors that $T_R$ increases towards more underdoping.
Fig.2 presents the Nernst signal of sample $p=0.11$ with thermal stream along \[110\] direction (the sample was cut with the longitudinal direction along \[110\]) and at temperatures from $5 K$ to $220 K$. In order to show the reliability of the measurement, we repeated the measurements at $20$, $25$ and $220 K$ after a 2-day delay. One can see that the two sets of data for each temperature coincide very well showing a high reproducibility. In low temperature region, the Nernst signal is dominated by the motion of Abrikorsov vortices. One can easily see that the background when the vortices are freezed (see data at $T=5, 10, 15
K$ in low field region) is actually precisely zero. This is contrasted by the slight negative background at high temperatures (above $150 K$). As first discovered by the Princeton group[@XuZA] and later checked by us[@WenHHNernst], a strong in-plane Nernst signal can be measurable far above $T_c$ (here $T_c$ = 29.3 K for $p=0.11$). Here it shows the same case. When $T$ is above $80 K$, the signal becomes negative and gradually it approaches a background with a negative slope. And when $T$ is above about $150 K$ the Nernst signal does not change anymore with $T$, therefore it is reasonable to define a upper boundary $T_n$ for the Nernst signal which locates in the region of $150-180 K$ for sample $p=0.11$ with thermal flow direction along \[110\].
![ Nernst signal of sample $p=0.11$ at temperatures from (a) 5 K to 45 K and (b) 45 k to 220 K. The measurement at 20, 25 and 220 K were repeated and data overlap each other for each temperatures. At temperatures of 5 K, 10 K, 15 K, the vortices are freezed showing a background with precisely zero resistivity. At temperatures above about 150 K all data overlap to a background with a negative slope.[]{data-label="fig2"}](fig2){width="8cm"}
![ Net Nernst signal $\Delta e_y$ (subtracted with the background at 220 K) at 6 T shown in (a) semi-logarithmic and (b) linear scale for samples ($p=0.11$) with thermal stream along \[110\] (open triangle) and \[100\] (open square). From both figures one can see that the Nernst signal vanishes in the region of 150 K to 180 K as marked by a rectangular in Fig.3(b). The solid line shows an empirical relation $\Delta e_y=0.135/T^{3.3}$ which describes the data above $T_c$ very well but the reason remains unknown. Inset of Fig.3(a) shows the same data in double logarithmic scale.[]{data-label="fig3"}](fig3){width="8cm"}
In Fig.3 the Nernst signal from both samples ($p=0.11$ along \[110\] and \[100\]) at 6 T is shown vs. T. The red (solid) curve is an empirical relation $\Delta e_N=e_N^0/T^{3.3}$ which describes the data very well up to about $120 K$ and a deviation from this behavior with a faster dropping rate occurs above about $120 K$. It is thus safe to conclude that the Nernst signal vanishes in the region of $150 K$ to $180 K$ as marked by a rectangular in Fig.3(b). Surprisingly this upper boundary temperature $T_n$ of Nernst signal coincides rather well with the crossover temperature $T_R=164\pm 20 K$ as mentioned above. For sample ($p=0.11$) we have measured on two specimens, one with the heat stream along \[100\] and another one along \[110\]. The results on these two samples turn out to be almost identical (as shown by open squares and triangles in Fig.3). Then we checked for other two samples with doping $p$=0.09,0.145 and found that this coincidence of $T_R$ and $T_n$ is true also for them. From the recent published data of Ando et al.[@AndoT2] the $T_R$ has also been determined and shown together with our data in Fig.4. It is clear that the data from both groups overlap each other. In Fig.4 we also present the doping dependence of $T_n$ given by the Princeton group. In the very underdoped region, according to Ando et al.[@AndoT2], the crossover of resistivity occurs at higher temperatures. For example, at $p=0.02$, $T_R$ is about $230 K$. However the upper boundary of Nernst signal is obscured by both the small Nernst signal and strong thermal power in very underdoped region. Therefore we don’t know whether this coincidence holds still for very underdoped samples. If taking a fixed value which is determined by the resolution of the voltmeter as a criterion for the Nernst signal, according to the recent data of the Princeton group[@WangYYPRB], the Nernst upper boundary temperature $T_n$ will drop down in very underdoped region. Despite the uncertainty in determining the value of $T_n$, in the doping region of our samples ($p=0.09,0.11,0.145$), the data are not far from that of the Princeton group.
![ Phase diagram of $La_{2-x}Sr_xCuO_4$ single crystals. The bottom dotted line represents an empirical curve of $T_c/38=1-82.6(p-0.16)^2$. The filled symbols represent the temperatures $T^*$ (quoted from[@Tstar]) below which the Fermi surface is partially gapped. The $T^*-\rho$ was determined from the resistivity and $T^*-mag$ was determined as the point where the Knight shift starts to decrease. The open symbols are determined from our measurement for $T_n$ and $T_R$, $T_R$ from Ando et al.[@AndoT2] and $T_n$ from Wang et al.[@WangYYPRB]. The upper boundary of Nernst signal ($T_n$) and the crossover temperature of resistivity $T_R$ coincide rather well in the doping regime of our present samples. []{data-label="fig4"}](fig4){width="8cm"}
In the following we will try to understand the coincidence of $T_n$ and $T_R$ in the intermediate doping region based on several possible pictures. We first discuss the scenario of vortices. It is known that the in-plane resistivity normally reflects the dynamics of the nodal quasiparticles, while the Nernst signal is induced by the thermal drifting of vortices as suggested by the Princeton group, therefore it seems difficult to relate one and another. However one way to understand this point is that the nodal quasiparticles partially form Cooper pairs and induce a precursor superconducting state which has strong phase fluctuation with vortices as one of the important excitations. Therefore below this specific temperature the Nernst signal which senses the vortex motion starts to appear and the temperature dependence of resistivity changes behavior (due to probably the reduction of the effective charge carriers). As a partial support to this picture, it was shown that this strong in-plane Nernst signal may be explained by the superconducting fluctuation which has the Kosterlitz-Thouless vortex-anti-vortex pairs as the topological excitations[@Ussishkin]. The second picture close to this vortex scenario is based on the bosonic RVB[@WengZY] which predicts the co-existence of spinon vortices and quasiparticles below $T_n$ and the Nernst signal is induced by the motion of these spinon vortices. This picture naturally predicts a second energy scale ($T_n$ or $T_R$) and different type of vortices below and above $T_c$ (still lacking of experimental evidence for the spinon vortices). However both pictures are facing a common difficulty: No any trace of vortex (or spinon vortex) flow resistivity has been reported far above $T_c$. One explanation to this point would be that the resistivity induced by quasiparticle scattering is overwhelmingly larger than the flux flow resistivity. Meanwhile it remains to be understood why the temperature dependence of resistivity should be precisely quadratic. Two close relatives of the picture based on vortices are the one with a mixture of composite charge carriers[@Kim; @Levin] or the pair density wave (PDW)[@ZhangSCPDW] (coexistence of Cooper pairs and free electrons below $T_n$). The paired electrons tend to localize and the rest free electrons contribute to the electrical conductivity. According to Ando et al.[@AndoT2], the relation $\rho=\rho_0+aT^2$ may not be simply due to the scattering rate $1/\tau \propto T^2$ of the nodal quasipartcles since the effective charge carrier density changes with temperature even below $T_R$. A naive explanation for $\rho=\rho_0+aT^2$ in this region would be that the nodal quasiparticle density (roughly proportional to the length of Fermi arc[@Norman] or the electronic specific heat coefficient $\gamma$[@Loram]) drops down linearly with $T$, but the scattering rate $1/\tau$ of nodal quasiparticles with features close to Dirac fermions is proportional to $T^3$. One more possibility for the coincidence of $T_R$ and $T_n$ relies on a exotic picture that some kind of unconventional charge density wave (UCDW)[@Maki] occurs at $T_n$. Below $T_n$ or $T_R$ the quasiparticle spectrum on top of this UCDW will contribute a strong Nernst signal, while the nodal quasiparticles are responsible for the in-plane resistivity, thus both the in-plane Nernst signal and resistivity will certainly correlate each other. This picture needs both theoretical justification and experimental evidence, especially a recent calculation seems ruling out the D-density wave as one of the possible causes[@DDWUssishkin]. Since the Nernst signal is detectable only below $T_R$, it may have no direct relation with the “flat band” near $(\pi,0)$ as observed by ARPES since it influences the transport properties only at high temperatures[@Flatband]. The coincidence of $T_R$ and $T_n$ found in our experiment strongly suggests an inherent relation between the Nernst signal and the nodal quasiparticle transport. This may also imply a second energy scale in the pseudogap region.
In conclusion, the in-plane Nernst signal and a quadratic temperature dependence of resistivity are found to occur in the same temperature region above $T_c$ in the intermediate underdoped regime. This suggests a close and inherent relation between Nernst signal and the nodal quasiparticle transport. Possible reasons are given to explain this coincidence. This coincidence may be obscured by other effects in other doping regimes. Our observation may imply a second energy scale or temperature within the pseudogap region.
This work is supported by the National Science Foundation of China , the Ministry of Science and Technology of China, the Chinese Academy of Sciences. We thank Dr. Yoichi Ando and his group (CRIEP, Komae, Tokyo, Japan) for providing us the sample with $p$=0.145. We are grateful to Z. X. Zhao, Z. Y. Weng, N. Nagaosa, T. Xiang and Z. A. Xu for fruitful discussions.
Correspondence should be addressed to [email protected]
[00]{} P. W. Anderson et al., Phys. Rev. Lett. [**58**]{}, 2790 (1987); P. W. Anderson, Science [**235**]{}, 1196 (1987). D. Pines, Physica C [**235**]{}, 113 (1994); D. Scalapino, Phys. Rep. [**250**]{}, 329 (1995). V. J. Emery, S. A. Kievelson, Nature [**374**]{}, 434 (1995); A. J. Millis, Nature [**398**]{}, 193(1999). V. J. Emery, S. A. Kievelson, and O. Zachar, Phys. Rev. [**B56**]{}, 6120 (1997). S. Chakravarty et al., Phys. Rev. [**B63**]{}, 94503 (2001). I. K. Affleck and J. B. Marston, Phys. Rev. [**B 37**]{}, 3774(1988). J. Corson, et al. Nature [**398**]{}, 221-223(1999). Z. A. Xu, et al., Nature [**406**]{}, 486(2000). F. Zhou, et al., Supercond. Sci. Tech. [**16**]{}, L7(2003). Y. Ando, Y. Kurita, S. Komiya, S. Ono, K. Segawa, Phys. Rev. Lett.[**92**]{}, 197001(2004). H. H. Wen, et al., Europhys. Lett.[**63**]{},583(2003). Y. Y. Wang et al., Phys. Rev. Lett. [**88**]{}, 257003(2002); Y. Y. Wang et al., Phys. Rev. [**B 64**]{}, 224519(2001); N. P. Ong, et al., Annalen der Physik [**13**]{}, 9(2004). B. Batlogg, et al., Physica [**C235-240**]{}, 130(1994). A figure with sumarized data of $T^*$ see T. Timusk, B. Statt, Rep. Prog. Phys.[**62**]{}, 61(1999). I. Ussishkin, S. L. Sondhi, D. A. Huse, Phys. Rev. Lett. [**89**]{}, 287001(2002). Z. Y. Weng and V. N. Mathakumar, Phys. Rev. [**B 66**]{}, 094510(2002). Y.H. Kim et al., J. Phys.: Condens. Matter [**15**]{}, 8485(2003). S. N. Tan, K. Levin, Phys. Rev. [**B 69**]{}, 064510(2004). H. D. Chen, et al., Phys. Rev. Lett. [**89**]{}, 137004(2002); Phys. Rev. [**B 70**]{}, 024516(2004); Cond-mat/0402323. M. R. Norman, et al., Nature [**392**]{}, 157(1998). P. Coleman, Nature [**392**]{}, 134(1998). A. Damascelli, Z. Hussain, Z. X. Shen, Rev. Mod. Phys. [**75**]{}, 473(2003). J. W. Loram, et al., [**71**]{}, 1740(1993). B. Dora et al., Phys. Rev. [**B 68**]{}, 241102(2003). V. Oganesyan, I. Ussishkin, Phys. Rev.[**70**]{}, 054503(2004). A. Ino, et al., Phys. Rev. [**B 65**]{}, 094504(2002); T. Yoshida, et al., Phys. Rev. Lett. [**91**]{}, 027001(2003).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'L. J. Hadfield'
- 'Paul A. Crowther'
- 'H. Schild'
- 'W. Schmutz'
title: 'A Spectroscopic Search for the non-nuclear Wolf-Rayet Population of the metal-rich spiral galaxy M83 [^1]'
---
Introduction
============
Massive stars play a major role in the ecology of galaxies via radiative, mechanical and chemical feedback [@smith05]. Wolf-Rayet (WR) stars in particular, albeit rare and short-lived, make a significant contribution to their environment via the mechanical return of nuclear processed material to the interstellar medium (ISM) through their exceptionally powerful stellar winds.
Metallicity, $Z$, is a key factor in determining the number and subtype distribution of a WR population. Although the metallicity dependence of WR wind properties still remains unclear, mass-loss prior to this phase has been established to depend on metallicity, with the latest models predicting M $\propto Z^{\sim 0.8}$ for O stars [@vink01]. Evolutionary models for single stars predict that the minimum mass cut-off required for WR formation should decrease as metal content increases. It is anticipated that the minimum mass required for progression to the WR phase decreases from $\sim 32 \mbox{M}_{\,\odot}$ for a SMC-like metallicity to $\sim 21
\mbox{M}_{\,\odot}$ for super-Solar metallicity [@meynet04]. Single star predictions are broadly consistent with the initial masses of WR stars in the Milky Way, LMC and SMC inferred from cluster membership [@massey00; @massey01]. The fractional distribution of carbon-rich (WC) to nitrogen-rich (WN) stars is also known to increase with metallicity, such that empirically one would expect to observe a large WC population in a metal-rich environment [@massey98]. The formation of WR stars at low metallicity is anticipated primarily via close binary Roche Lobe Overflow (RLOF), yet the observed WR binary fraction in the SMC does not differ from that of the Milky Way [@foellmi03].
With the aim of substantiating such predictions, surveys of WR stars in Local Group galaxies (typically 0.2–1Z$_{\odot}$) have been carried over the last two decades. At sub-Solar metallicities, the LMC and SMC have been well sampled [@breysacher99; @massey03], as has M33 [@massey98; @abbott04]. In contrast, M31 is the only Local Group member with super-Solar metallicity, but its unfavourable inclination and large spatial extent makes surveying the complete WR population very challenging. In order to increase the variety of galaxies sampled, our group has begun to look beyond the Local Group [e.g. NGC300, @schild03].
Galaxies hosting substantive WR populations are known as ‘WR galaxies’ [@kunth81; @scp99], where the number of WR stars ranges from $\sim$35 in NGC1569-A [@gonzalez97] to 2$\times 10^{4}$ in Mrk 309 [@schaerer00]. Within specifically metal-rich environments, previous studies of WR populations have generally been restricted to integrated spectra from bright star forming knots [e.g. @schaerer99] or H[ii]{} regions [e.g. @pindao02]. Here we present the results of a deep imaging and spectroscopic survey of the disk WR population within the metal-rich galaxy M83, in which WR signatures have previously been identified by [@rosa87] and [@bresolin02].
M83 (NGC 5236) is a massive, grand-design southern spiral (SBc(s)II) with on-going star formation in its spiral arms plus an active nuclear starburst [@elme98; @harris01]. M83 is the principal member of a small galaxy group ($\sim
11$ members) within the Centaurus A complex [@kara02]. Located at a distance of 4.5$\pm$0.3Mpc [@thim03], its favourable inclination and apparently high metal abundance of log(O/H)+12=9.2 [@bresolin02] makes M83 an ideal candidate for studies of massive stellar populations at high metallicity.
More recently, oxygen abundances in metal-rich galaxies have been revised downward [@pilyugin04; @bresolin04], such that M83 may have a metal abundance closer to log(O/H)+12=9.0 (Bresolin, 2004, priv. comm.), i.e. approximately twice the Solar oxygen abundance of log(O/H)+12=8.66 recently derived by @asplund04.
We present the results of an imaging and spectroscopic survey of the WR content of M83 using the ESO Very Large Telescope (VLT). The present paper complements the initial findings of this study reported in @pac04 [hereafter PaperI]. In Sect.\[obs:red\] we briefly describe the observations and data reduction techniques employed. Section \[wr:con\] discusses the method followed to obtain a global WR population of M83. Sect.\[discussion\] discusses the properties of metal rich WR stars with those of Local Group galaxies and evolutionary models. Finally, conclusions are drawn in Sect.\[conc\].
Observations and data reduction {#obs:red}
===============================
We have observed M83 with the ESO Very Large Telescope UT4 (Yepun) and Focal Reduced/Low Dispersion Spectrograph \#2 (FORS2). The detector consists of a mosaic of two $2048 \times 1024$ MIT/LL CCDs which in conjunction with the standard collimator provides a field-of-view $6.8'\times6.8'$ and an image scale of 0.126$''$/pixel. Photometric observations of M83 were made between May–June 2002 with follow-up spectroscopic data being acquired during April–June 2003.
Imaging
-------
M83 subtends 12.9$'$ by 11.5$'$ on the sky, preventing it being imaged by a single FORS2 frame. In order to obtain complete coverage, the galaxy was divided into four overlapping regions, covering the NE (Field A), NW (B), SE (C) and SW (D) as indicated in Fig.\[fields\]. Occulting bars were positioned in Field C to prevent detector saturation by bright foreground stars. The central 15$''$ appears saturated on all images obtained, and as a result the WR population of the nucleus can not be discussed further.
FORS2 was used on 2 June 2002 to obtain narrow-band images with central wavelengths 4684Å, 4781Å and band widths of 66Å and 68Å respectively. These were obtained consecutively for each Field in seeing conditions between 0.6 – 0.8$''$[^2] with individual exposures of 1800s. The $\lambda$4684 filter is coincident with the strong WR emission features which incorporates the N[iii]{} ($\lambda 4640$Å), [C]{}($\lambda
4650$Å) and [[ii]{} ]{}($\lambda 4686$Å) emission lines, whereas the latter samples a wavelength region relatively free from emission, providing a measure of the continuum level. In addition to these, 2 exposures (60s and 600s) were taken using narrow-band on- and off-[$\mbox{H}\alpha \ $]{}filters $(\lambda 6563,\,6665\,\mbox{\AA}, \,
\mbox{FWHM}=61,\,65\,\mbox{\AA}$) on 16 May 2002. Finally, in order to supplement the primary dataset, 2 exposures (60s and 120s) were also acquired using a Bessell B filter on 21 May 2002.
![Four combined [$\mbox{H}\alpha \ $]{}VLT FORS2 frames of M83. The overall size of the image is $\sim 12\,'
\times 12\,'$. The fields A, B, C and D used to image M83 are marked. North is up and east is to the left of the image.[]{data-label="fields"}](2262fig1.ps){width="9cm"}
Photometry
----------
Images were prepared following standard procedures i.e. debiased, flat field corrected and cosmic ray cleaned. Photometry of individual sources within M83 was performed using the package [daophot]{}, a point-spread function (PSF) fitting routine within [iraf]{}. Absolute photometry in the [broad-band B ]{}filter was achieved with the aid of photometric standard fields Ru 152 and PG 1528+062 (containing a total of 10 photometric standards, 11.9 $\leq$ B $\leq$16.3). For the narrow-band images such standards are not available and photometric zero-points have been obtained by observing spectrophotometric standards LTT7987 (B = 12.2) and G138-31 (B = 16.5).
The majority of our sources appear point-like on the ground-based images. However, a number of bright sources are surrounded by a faint, extended halo, which was not accounted for in the PSF photometry and as a result only a lower limit to the magnitude is given, based on PSF photometry. A further subset of the bright sources are spatially extended, indicating that PSF photometry is inappropriate, as indicated in Table A1 in the appendix.
Typical formal photometric errors range between 0.02mag ($\sim$18mag), 0.05mag ($\sim$20mag) and 0.08mag ($\sim$22mag). Significantly higher errors, of up to 0.15mag, are obtained for regions of the galaxy where the background levels are high, or they are located in spatially crowded regions.
As a consistency check we have compared results obtained for the two Bessell B exposures (for which the PSF model was based on different template stars) and also derived magnitudes for objects which appear in multiple fields. Excellent agreement was observed in both cases, with results agreeing to within the formal errors. In a minority of cases this was not achieved due to severe crowding.
Candidate Selection
-------------------
WR candidates were identified by searching for He[ii]{} / [C]{} excess emission (at $\lambda$4684) relative to the continuum ($\lambda$4781), i.e. a negative value of $\Delta m$ = [$m_{\lambda 4684} \ $]{}- $m_{\lambda 4781}$. The optimal method of identifying suitable candidates was found to be via ‘blinking’ individual $\lambda$4781 and $\lambda$4684 frames together with the difference image obtained by subtracting the $\lambda$4781 image from the $\lambda$4684 frame. In total, 283 candidate $\lambda$4684 emission sources were identified.
For 75% of our candidates we have obtained a magnitude in at least the $\lambda$4684 filter. For cases where we did not obtain photometry, the object was either too faint or was located in a spatially crowded region. In addition, for a significant fraction of the fainter sources it was not possible to measure a $\lambda$4871 magnitude.
Candidates were grouped according to continuum brightness, $\Delta m$, and association with underlying H[ii]{} regions. To ensure we spectroscopically observed a representative sample, a selection from each group was chosen for spectroscopic follow up. In Fig. \[photom:excess\] we show $\Delta m$ as a function of continuum magnitude for the sources in which WR signatures were either spectroscopically confirmed, rejected or no spectroscopy was obtained, i.e. the remaining candidates. The majority of confirmed sources have a $\lambda$4684 excess between –1.5 $\leq \Delta m \leq$ –0.4 mag, although a few do exhibit rather smaller values of $\Delta
m$. In contrast, all rejected regions have $\Delta m \geq$ –0.2 mag, suggesting that remaining candidates which display a moderate $\lambda$4684 excess should represent regions that genuinely host WR stars, together with a subset of those for with $\Delta m \sim$ 0.0 mag.
![Comparison between [$m_{\lambda 4781} \ $]{} magnitude and [$m_{\lambda 4684} \ $]{}excess of the WR candidates located in M83. Regions which have been spectroscopically observed and subsequently eliminated or classified as WR regions are presented in the key. Regions which still await spectroscopic observations are also marked. Sources for which PSF photometry was unavailable are not marked.[]{data-label="photom:excess"}](2262fig2.ps){width="7.8cm"}
Spectroscopy
------------
Spectroscopic data was obtained using FORS2 with the Multi Object Spectroscopy (MOS) mode. MOS datasets of individual WR candidates were obtained during seeing conditions of $\sim$0.5–1.0$''$, using a slit width of $0.8 ''$. The CCD was binned by a factor of 2 in the dispersion direction, resulting in a dispersion of 3.3Åpixel$^{-1}$ with the 300V grism and a spectral resolution of $\sim 7$Å, as measured from comparison arc lines. The wavelength range of individual targets depended on their position within the MOS mask but typical wavelength coverage was $\sim$3700Å to $\sim$7500Å .
MOS allows the spectra of up to 19 candidates to be recorded simultaneously. However, due to positional limitations this was generally restricted to $\sim$15, supplemented where possible by H[ii]{} regions. In total, 198 candidates have been spectroscopically observed using 17 different MOS masks. To maximise continuum S/N, sources were grouped according to brightness, with total on-source integration times ranging from 720s for the brightest objects to 4800s for the faintest. Details of the spectroscopic log can be found in Table\[Tab:spec\], and includes DIMM seeing measurements. The MOS masks were labelled according to the region of M83 in which they were observed, i.e. Field A was observed using 5 different masks labelled A1 to A5.
\[Tab:spec\]
------------ ------- ----------------- -------------
Date MOS Exposure DIMM Seeing
Mask ($''$)
2003-04-06 D2 $3 \times 900$ 0.6
2003-04-07 D1 $1 \times 2400$ 0.5
D4 $3 \times 240$ 0.5
2003-04-13 D1 $1 \times 2400$ 0.7
D3 $3 \times 500$ 0.7
D5 $3 \times 600$ 0.5
2003-05-20 C1 $2 \times 2400$ 1.0
2003-05-21 C2 $3 \times 900$ 0.9
C3 $3 \times 240$ 0.9
2003-05-24 A1 $ 1 \times2400$ 0.6
A3 $ 3 \times 500$ 0.5
2003-05-26 A2,B2 $3 \times 900$ 0.7, 0.8
A5 $ 3 \times 600$ 0.8
A4 $ 3 \times 240$ 0.6
2003-06-17 A1 $1 \times 2400$ 0.6
B1 $2 \times 2400$ 0.6, 0.5
B3 $3 \times 250$ 0.7
B4 $3 \times 240$ 0.5
------------ ------- ----------------- -------------
: FORS2 Multi Object Spectroscopy (MOS) observing log for M83.[]{data-label="Tab:spec"}
Datasets were prepared and processed using standard [iraf]{} and [figaro]{} packages i.e. the data were bias subtracted, flat field corrected, spectra were traced, extracted, and subsequently wavelength and flux calibrated. For very faint sources, where no continuum was evident, identification was made solely on the basis of strong emission lines, a neighbouring continuum source was used as the trace.
Spectroscopic magnitudes (hereafter m$_{\rm spec}$) were obtained by convolving the individual spectra with the transmission curves of the imaging filters. A comparison between the spectroscopic and the PSF photometry at $\lambda 4684$ permitted absolute flux calibration. For 160 sources brighter than 24 mag the average slit correction factor was found to be 3.1 ($\sigma=1.9$). This large factor is due in part to the often crowded nature of sources, such that the full profile extent was not extracted. For 20 spectroscopically observed regions, where photometry was unavailable, we corrected the spectroscopy by a factor of 3$\pm$1.5.
The blue ($\sim$4500Å) continuum S/N ranged from $\sim$80 for the brightest sources, to $\leq$1 for the faintest sources. Nevertheless, lines were typically detected even in the faintest sources at the 5–10$\sigma$ level, and a source was only confirmed if WR emission lines were detected at a $\geq 3\sigma$ level.
The Wolf-Rayet Population of M83 {#wr:con}
================================
The WR content of the disk of M83 has been determined by visually inspecting the extracted spectra for the characteristic WR emission signatures, i.e. [N[iii]{}]{}$\lambda$4634–41, [N[v]{}]{}$\lambda$4603–20, [C]{}[$\lambda 4650$ ]{}– [[ii]{} ]{}[$\lambda 4686$ ]{}blend (hereafter blue WR features) and/or C[iii]{} $\lambda$5696, [C[iv]{} ]{}[$\lambda 5812$ ]{}features (hereafter yellow WR features).
Of the 198 sources spectroscopically observed, 132 contain WR emission features at a 3$\sigma$ level. These are presented in a catalogue in Table A1, in the Appendix, which includes coordinates, PSF (or spectroscopic) photometry, interstellar reddening, line measurements and WR populations. Deprojected distances are included using M83 properties presented in Table1 of @lundgren04. We present finding charts for all confirmed sources in the (electronic only) Appendix, Figures B1–B17.
![ Observed, velocity corrected ($v_r = 513 \mbox{km\,s}^{-1}$) blue (top) and yellow (bottom) spectral regions of Source \#74. Gaussian line profiles are overplotted for the WR features: [N[iii]{}]{}($\lambda 4634-41$), [C]{} ($\lambda 4647-50$) and [[ii]{} ]{}($\lambda 4686$) in the blue, and [C]{}($\lambda 5696$) and [C[iv]{} ]{}($\lambda 5801-12$) in the yellow.[]{data-label="fig:fits"}](2262fig3.ps){width="6cm"}
Of the remaining 66 sources, 40 displayed an early-type spectrum with no WR emission present, 16 resembled that of a late-type carbon star, while 10 sources revealed WR features below the 3$\sigma$ level, or lacked the blue region necessary for WN identification. The latter two groups, along with the 79 regions which were not spectroscopically observed, are given in our candidate list (TableA2 in the Appendix).
Interstellar reddening
----------------------
Estimates of the interstellar reddening for our confirmed WR sources have generally been derived using measurements of the nebular H$\alpha$ (accounting for nearby \[N[ii]{}\] emission) and H$\beta$ features present in the extracted spectra.
Assuming Case B recombination theory for typical electron densities of $10^{2}\,\mbox{cm}^{-3}$ and a temperature of $10^{4} \mbox{K}$ [@hummer87], we obtain 0.2$\leq$E(B-V)=[*c*]{}(H$\beta$)/1.46 $\leq$ 0.8 mag for the majority of the sources, with a few outliers, and typical formal uncertainty of $\pm$0.02 mag. Where Balmer emission was observed, typical H$\beta$ equivalent widths lay in the range $\mbox{and} \sim 20 \ \mbox{to} \sim 150$Å. Consequently, the underlying stellar absorption components ($\leq$1Å at H$\beta$) are neglected. In 41 sources no nebular lines were observed. For those with a well defined continuum, E(B–V) was estimated by assuming an intrinsic optical flux distribution equivalent to a late O-type star, with typical uncertainty of $\pm$0.05–0.1 mag. In 15 cases, the continuum S/N was insufficient for this comparison and an average reddening of E(B–V)= 0.5$\pm$0.3 was adopted. Correction for reddening adopt a standard @seaton79 extinction law with R=3.1=$A_{\rm V}$/E(B-V).
Spectral classification
-----------------------
In order to classify and quantify the WR population within each region, we have fit Gaussian line profiles to the blue and yellow WR features, revealing line fluxes, equivalent widths and FWHM. An example of the fits to the blue and yellow WR features is presented in Fig.\[fig:fits\], where a source (\#74) hosting a mixed WN and WC population is presented.
In general, it was straightforward to distinguish between WN (strong [[ii]{} ]{}$\lambda 4686$) and WC subtypes (strong [C]{}$\lambda$4650 and [C]{}$\lambda$5696 and/or [C[iv]{} ]{}$\lambda$ 5801-12). The following classification scheme was applied for further subdivision. In a minority of cases it was not possible to separate the $\lambda$4650 – $\lambda$4686 features into individual components, and as a result an overall blend was measured. Since WC subtypes were assigned on the basis of [$\lambda 5696$ ]{}and [$\lambda 5812$ ]{}features, this did not prevent accurate classification.
Late and early WN subtypes were assigned if [[ii]{} ]{}$\lambda 4686$ emission was accompanied by [N [iii]{}]{}$\lambda 4634-41$ or [N[v]{}]{}$\lambda 4603-20$ emission, respectively. If nitrogen lines were undetected, we assigned a WNE subtype if FWHM ([[ii]{} ]{}$\lambda 4686$) $>$ 20Å, and WNL otherwise. For WC stars, WC4 – 6 was assigned if [C[iv]{} ]{}$\lambda 5801-12$ was present along with either weak or absent [C]{}$\lambda 5696$. For $0.25
\leq F_{\lambda}$ ([C]{}$\lambda$5696/[C[iv]{} ]{}$\lambda$5801-12) $\leq 0.8$ sources were classified WC7, and WC8–9 if [C]{}$\lambda 5696$ was present, with [C[iv]{} ]{}$\lambda 5801-12$ weak or absent.
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![ De-reddened spectral comparison between WC members in M83 with Galactic WC stars scaled to the distance of M83 (Galactic distances from @derhucht01). To avoid confusion, WCE and WC7 sources are offset by 2$\times 10^{-17}$ ergs$^{-1}$cm$^{-2}$Å$^{-1}$ whereas WC8 and WC9 spectra are offset by 1$\times 10^{-17}$ ergs$^{-1}$cm$^{-2}$Å$^{-1}$.[]{data-label="spectra"}](2262fig4a.ps "fig:"){width="7cm"}
![ De-reddened spectral comparison between WC members in M83 with Galactic WC stars scaled to the distance of M83 (Galactic distances from @derhucht01). To avoid confusion, WCE and WC7 sources are offset by 2$\times 10^{-17}$ ergs$^{-1}$cm$^{-2}$Å$^{-1}$ whereas WC8 and WC9 spectra are offset by 1$\times 10^{-17}$ ergs$^{-1}$cm$^{-2}$Å$^{-1}$.[]{data-label="spectra"}](2262fig4b.ps "fig:"){width="7cm"}
![ De-reddened spectral comparison between WC members in M83 with Galactic WC stars scaled to the distance of M83 (Galactic distances from @derhucht01). To avoid confusion, WCE and WC7 sources are offset by 2$\times 10^{-17}$ ergs$^{-1}$cm$^{-2}$Å$^{-1}$ whereas WC8 and WC9 spectra are offset by 1$\times 10^{-17}$ ergs$^{-1}$cm$^{-2}$Å$^{-1}$.[]{data-label="spectra"}](2262fig4c.ps "fig:"){width="7cm"}
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![De-reddened spectral comparison between WN complexes in M83 with individual Milky Way WN stars scaled to the distance of M83 (Galactic distances from @derhucht01). To avoid confusion, individual sources are successively offset by 1$\times 10^{-17}$ ergs$^{-1}$cm$^{-2}$Å$^{-1}$.[]{data-label="WN_sp"}](2262fig5.ps){width="7cm"}
To ensure consistency with previous studies [e.g. @schaerer99; @bresolin02; @chandar04] we have derived WR populations based on individual line fluxes adapted from @schaerer. As discussed in PaperI, we adopt He[ii]{} $\lambda$4686 lines fluxes of 5.2$\times 10^{35}$ ergs$^{-1}$ and 1.6$\times 10^{36}$ ergs$^{-1}$ for WN2-5 and WN6-10 stars, respectively. For WC stars, we adopt C[iv]{} $\lambda$5801 line fluxes of 1.6$\times 10^{36}$ ergs$^{-1}$ and 1.4$\times 10^{36}$ ergs$^{-1}$ for WC4–6 and WC7 stars, respectively, and a C[iii]{} $\lambda$5696 line flux of 7.1$\times 10^{35}$ ergs$^{-1}$ for WC8–9 stars. WR contents of individual sources then follow, with populations rounded to the nearest integer ($\geq$1). In one source (\#117), we were unable to reliably extract the spectrum since it was located at the very edge of the slit, and so a measure of the reddening/line flux was not possible. Nevertheless, broad He[ii]{} $\lambda$4686 is clearly present, with no WC signature, such that we indicate a population of $\geq 1$ early-type WN star.
In Fig.\[spectra\] we compare sources containing representative late, mid and early WC stars from M83 with extinction corrected Milky Way counterparts, scaled to the distance of M83. Large line widths amongst M83 members hosting late WC stars are apparent, particularly for \#32 versus HD192103 (WC8) and \#81 versus HD164270 (WC9). In contrast, sources containing WC4–7 stars indicate similar line widths to individual Galactic counterparts. Comparisons between sources containing WNL and WNE stars in M83 and two Galactic counterparts are shown in Fig.\[WN\_sp\], revealing similar spectral morphologies. Other examples of sources hosting WN and WC populations are presented in Fig.1 of PaperI.
The M83 WR population – individual stars, binaries, complexes or clusters?
--------------------------------------------------------------------------
What is the nature of the 132 sources in M83 that are known to host WR stars? In Fig. \[syn:excess\](a) we compare the spectroscopic continuum magnitude to the spectroscopic excess, $\Delta m_{\rm spec} = $[$m_{\lambda 4684} \ $]{}– $m_{\lambda 4781}$, for all sources. This is more complete than Fig.\[photom:excess\], since it was generally possible to estimate the spectral [$m_{\lambda 4781} \ $]{} magnitude for the fainter sources, where PSF-photometry was not available.
----------------------------------------------------------------------------- --
{width="7.8cm"} {width="7.8cm"}
----------------------------------------------------------------------------- --
The brightest confirmed WR sources in our sample ([$m_{\lambda 4781} \ $]{}$\, \sim$20 mag) exhibit –0.3 $\leq \Delta m_{\rm spec} \leq $0.0 mag. Such values are consistent with luminous complexes, greatly diluting the WR emission signature. In contrast, the faintest confirmed sources ([$m_{\lambda 4781} \ $]{}$\, \sim$25 mag) possess large spectroscopic excesses of –2 $\leq
\Delta m_{\rm spec} \leq$ –0.5 mag, consistent with isolated, single or binary WR systems. Intermediate brightness sources span the full range in excess, corresponding to less luminous regions hosting a few WR stars to those containing large WR populations.
Fig. \[syn:excess\](b) compares the spectroscopic $\lambda 4686$ excess to the [C]{}$\lambda$4650/[[ii]{} ]{}$\lambda$4686 equivalent width, confirming the expected tight correlation between line strength and $\Delta m_{\rm spec}$, where the scatter indicates the observational accuracy. Typical excesses of –0.2 mag equate to small line equivalent widths of $\sim$10Å, whilst an excess of –1.0 mag corresponds to $\sim$100Å, and the largest excesses equate to $\sim$500Å. For comparison, single Galactic and LMC WR stars possess [[C]{}$\lambda$4650/[[ii]{} ]{}$\lambda$4686]{} equivalent widths of 10–500Å (WN subtypes) or 150–2000Å(WC subtypes).
The global disk WR population of M83
------------------------------------
We identify 1035$\pm$300 WR stars, comprising 564$\pm$170 WC and 471$\pm$130 WN stars, within our 132 spectroscopically observed regions, where errors quoted here were obtained from simply adding individual uncertainties for all regions.
The most important discovery of our spectroscopic survey is the dominant late-type WC population of M83. Over half of the spectroscopically identified WR stars in M83 fall into the WC8–9 subtype, with few WC4–7 stars identified. For comparison, no WC8–9 stars are observed in the SMC, LMC or M33 and the total number of such stars in the Milky Way and M31 is less than 50 [@derhucht01; @moffat87]. The distribution among late- and early-type WN stars is more even, with WNL/WNE $\sim 1$. This value is much greater than that observed in the SMC ($\sim$0) and LMC ($\sim$0.25), but comparable to that of $\sim 1.3$ determined for the Milky Way [@derhucht01].
How robust is this derived WR population for M83? For each source, we have propagated uncertainties in the distance, reddening, photometry and line flux measurements. Together, these translate to a typical uncertainty of $\sim 20 - 30$%, or somewhat higher for regions in which an interstellar reddening or a slit loss correction factor have been adopted.
One of the main limitations in estimating the content of an unresolved WR population is the conversion from WR line flux to WR content. Given the large late WC population identified in M83, we have reconsidered the line flux of individual WC8–9 stars determined by @schaerer. From unpublished data for 5 Galactic, and 2 M31 WC8–9 stars, each with well derived distances, we find a mean $\lambda 5696$ flux of $5.1 \times 10^{35} \mbox{ergs}^{-1}$ and $4.7
\times 10^{35} \mbox{ergs}^{-1}$ respectively. This is $\sim$30% lower than @schaerer, and suggests that, if anything, we may be underestimating the true WC population of M83.
We have also estimated the WC population using the alternative [C]{}$\lambda 4650$ line. Based on individual WR $\lambda 4650$ line fluxes of $3.4 \times 10^{36},\, 4.5 \times 10^{36}\, \mbox{and}\, 1.0
\times 10^{35}\,\mbox{ergs}^{-1} \, $ for individual WC4–6, WC7 and WC8–9 subtypes, respectively [@schaerer], populations of individual sources were found to agree to within a factor of 2, relative to the yellow features. The total WC population was calculated to be 594 using [C]{}$\lambda$4650, in excellent agreement with that of 564 obtained from [C]{}$\lambda 5696$ and C[iv]{} $\lambda$5808.
Turning to the candidates for which spectroscopy was not obtained, all regions in Fig.\[photom:excess\] with $\Delta{\rm{m}} \leq$ –0.3 mag correspond to spectroscopically confirmed WR complexes. Therefore, we would expect that at least 25 out of the 49 candidates, for which $\lambda$4684 and $\lambda$4781 photometry is available, also possess WR stars. Adopting the same fraction for regions where PSF photometry is not available, we expect $\geq$50 of the remaining 89 candidate regions to contain WR stars. Indeed, \#159 has already been observed by @bresolin02. Designated M83-5 in their study, WR emission is spectroscopically confirmed and a population of 2 WCL and 6 WNL stars (scaled to a distance of 4.5Mpc) is inferred from its line luminosity. On average, our confirmed sources host $\sim$5 WR stars, such that we expect $\sim$250 WR stars await identification in M83, bringing the total disk population to $\sim$1300.
The inferred WR population of M83 is greater than that known in the entire Local Group, to date [@massey98]. As anticipated from Figs.\[photom:excess\] and \[syn:excess\], some sources host a single WR star, whilst others contain larger WR populations ($\sim 10$). Regions which contain an exceptionally large WR population will be discussed in more detail in the next section.
Complexes hosting large WR populations {#clusters}
--------------------------------------
In the Milky Way, the most massive open clusters (e.g. Arches, Westerlund 1) host at most 10–20 WR stars [@blum01; @negueruela05]. Similar numbers are observed in the largest H[ii]{} regions of M33, and 30 Doradus in the LMC. We identify 10 regions in M83 with large ($\geq$20), or mixed, WR populations. Mixed WN and WC populations are observed in a total of 5 complexes, \#66 (8$\pm$2 WNL, 4$\pm$1 WC7), \#38 (7$\pm$2 WNL, 21$\pm$6 WCL), \#41 (14$\pm$4 WNL, 13$\pm$6 WC7), \#86 (9$\pm$4 WNL, 24$\pm$10 WCL) and \#74 which will be discussed separately.
Are the sources that host WR stars in M83 compact clusters (e.g. Arches) or extended, giant H[ii]{} regions (30 Doradus)? Massive compact clusters are generally rare in normal disk galaxies, although M83 is known to host many examples, from HST imaging [@larsen04]. Of the 60 bright H regions in M83 identified by @devau83, between 28–38 host WR populations. Indeed, the 3 complexes hosting the largest WR populations are all associated with H regions identified by @devau83. Optical spectroscopy of these were presented in PaperI, together with an estimate of their O star population.
![5$'' \times 5''$ images of the WR cluster M83 \#74 from VLT/FORS2 and HST/ACS. Top Panel: $\lambda$4684 filter, middle panel: continuum subtracted H$\alpha$ filter, lower panel: F475W filter (WFC). North is up and east is to the left. It is apparent that the brightest H$\alpha$ source lies $\sim2''$ to the S-W from the continuum (and WR) source.[]{data-label="74"}](2262fig7.ps){width="7cm"}
### Source \#74
From our sample \#74 is exceptional, with 230$\pm$50 late-type WN and WC stars inferred from the de-reddened line fluxes (recall Fig.\[fig:fits\]). This source has the highest interstellar reddening of our entire sample with E(B-V)=1.0$\pm$0.03, although it is closest to the nucleus. However, the H$\alpha$/H$\beta$ nebular value is supported from fitting its stellar continuum to a young ($\sim$4 Myr) instantaneous burst model at Z=0.04 from Starburst99 [@leitherer99]. In PaperI, we estimated a Lyman continuum flux of 8$\times 10^{51}$ s$^{-1}$ from the de-reddened H$\alpha$ flux, such that \#74 has an ionizing flux equivalent to the giant H[ii]{} region 30 Doradus. However, it possesses a WR content which is a factor of ten times larger, i.e. N(WR)/N(O)$\sim$0.25 versus 0.02 in 30 Doradus.
We have inspected archival HST/Advanced Camera for Surveys (ACS) Wide Field Camera (WFC) F475W datasets of M83 (Proposal 9299, P.I. H. Ford). This revealed that \#74 is very compact, with a FWHM of $\sim$0.2 arcsec or $\sim$4.5pc (for a distance of 4.5Mpc). For H[ii]{} regions with solar or super-solar metallicities, WR signatures are expected to be present in bursts of age 3–6Myr. We have compared the absolute F475W magnitude of \#74 with evolutionary synthesis models for an instantaneous burst of age 3–5Myr [@leitherer99], from which we estimate a mass of 1.4–2$\times 10^{5} M_{\odot}$. Therefore, its mass [*and*]{} size indicate that it is a young massive compact cluster, or Super Star Cluster [@whitmore03].
In Fig.\[74\] we present $5\times5$ arcsec ($\sim110 \times 110$ pc) images of \#74 obtained with FORS2 and ACS. It is apparent that the peak H$\alpha$ source, i.e. H[ii]{} region \#35 from @devau83, lies $\sim$2 arcsec to the S-W of the brightest continuum source (the WR cluster). The spectrum presented in Fig.2 of PaperI is that of the WR cluster, whilst the H$\alpha$ flux, and corresponding O7V star content of $\sim$810 represents the integrated total from both regions. The WR cluster provides approximately 1/3 of the total H[ii]{} luminosity, such that the WR/O ratio of this region approaches unity, comparable to the WR cluster NGC3125-1 [@chandar04].
### Other Clusters in M83
@larsen04 has identified $\sim$80 young massive clusters in M83 based on HST/WFPC2 images. Three such regions are in common with our catalogue of sources containing WR stars, namely n5236-607 (\#61), -617 (\#73) and -277 (\#79), although none host more than a few WR stars. Larsen (priv. comm.) has compared the UBVI colours of these clusters with Solar metallicity @Bruzual93 models, suggesting age estimates of $\log (\tau) = 6.20 \pm$0.51, 6.90$\pm$0.54 and 9.89$\pm$1.87, respectively. The first two are fully consistent with a young cluster which contains WR stars, while the third suggests a dominant old population.
Five additional clusters from @larsen04 are also in common with our remaining candidates, namely n5236-169 (\#193), -805 (\#179), -818 (\#163), -1011 (\#157) and -1027 (\#173). Of course, such candidates have the potential to also host a large WR population - indeed three of these clusters appear young ($\sim$1.5–6Myr) from UBVI photometry (Larsen, priv. comm.), i.e. \#193, \#179 and \#157. Note that \#179 is one of two clusters for which dynamical mass estimates has been made by @larsen04b. Follow-up spectroscopic observations would be required for the identification of additional WR rich clusters.
Comparisons with previous studies
---------------------------------
To date, there have only been two previous studies relating to WR stars within M83. [@rosa87] and [@bresolin02] have both studied stellar populations within M83 and identify six H [ii]{} regions which exhibit WR characteristics. Four of these have been re-examined in this study. @rosa87 obtain optical spectra with very poor signal-to-noise preventing a quantitative discussion, consequently we shall restrict any comparisons solely to results obtained by [@bresolin02].
Both studies followed a similar methodology in estimating the WR population, except that @bresolin02 adopted a distance of 3.2Mpc to M83 (versus 4.5Mpc adopted here). This introduces a factor of 2 between intrinsic line luminosities observed in this study and that by @bresolin02.
- [ *\#40* (M83-2) –]{} The derived WR population for this region is estimated to be 6$\pm$2 WC8–9, contrasting that of 1–2 WNL obtained by @bresolin02. We achieve a 3$\sigma$ detection for the [$\lambda 5696$ ]{} and [$\lambda 5812$ ]{} carbon features, suggesting that poor signal-to-noise in the original investigation prevented positive WC identification.
- [ *\#41* (M83-3) –]{} We confirm the detection of 14$\pm$4 WNL stars identified in region M83-3. In addition we estimate the presence of 13$\pm$6 WC7 stars. @bresolin02 state that [C]{} may be present, but not at a significant level (versus $5 \sigma$ here).
- [ *\#74* (M83-8) –]{} @bresolin02 failed to detect any WR emission in this H[ii]{} region. However, we find the largest individual WR population of any source, namely 230 stars. As stated in Sect.\[clusters\] the WR emission is offset by several arc-secs to the N-E of the peak [$\mbox{H}\alpha \ $]{}emission. Since the @bresolin02 concentrated on bright H regions, their slit was probably centred on the peak [$\mbox{H}\alpha \ $]{}emission, such that the WR signature was missed.
- [ *\#103* (M83-9) –]{} Both investigations infer a late WN population. The present study obtains a population of 29$\pm$9 WNL stars, in agreement with that estimated by @bresolin02 after allowing for differences in the assumed distance.
![Distribution among WCL subtypes as determined using $\mbox{W}_{\lambda}$ (C[iv]{} 5808) / $\mbox{W}_{\lambda}$ (C[iii]{} 5696) versus FWHM (C[iii]{} $\lambda$5696) in Å. For comparison, Galactic (unpublished WHT, AAT and 2.3m ANU data), M31 (unpublished WHT / ISIS data), M33 [@abbott04] and IC10 [@pac03] WCL stars have been included. The subtype divisions marked are those derived by @pac98.[]{data-label="WCL"}](2262fig8.ps){width="7.8cm"}
![Equivalent width ($\mbox{W}_\lambda$ in Å) vs. FWHM (Å) of the [C]{} $\lambda 5696$ for WCL stars in M83 and Local Group galaxies (identical dataset as presented in Fig.\[WCL\]).[]{data-label="WCL:EW"}](2262fig9.ps){width="7.8cm"}
Discussion
==========
We have identified up to $\sim$200 regions in the disk of M83 that host WR stars. We now compare the properties of WR stars at the high metallicity of M83 with those of Local Group galaxies, attempt to explain the dominant late subtypes amongst WC stars, and make comparisons with current evolutionary models.
Properties of WR stars at high metallicity
------------------------------------------
How do the line strengths and widths of sources containing WR stars in M83 compare with those of other galaxies? In Fig.\[WCL\], we show the classification ratio $\mbox{W}_\lambda$ (C[iv]{} 5808) / $\mbox{W}_\lambda$ (C[iii]{} 5696) versus FWHM ([C]{} 5696). Data for WC7–9 stars in four Local Group galaxies are included, along with subtype boundaries as derived by @pac98. This figure highlights the dominance of WC8 and WC9 subtypes, which comprise 95% of the total WC content of M83, by number.
The Galactic WC9 population is very homogeneous, centred on a FWHM of $\sim$25Å and $\mbox{W}_\lambda$ (C[iv]{} 5808) / $\mbox{W}_\lambda$ (C[iii]{} 5696) $\sim$0.3. In contrast, the WC9 population of M83 is very heterogeneous, spanning a wide range of both FWHM and $\mbox{W}_\lambda$ (C[iv]{} 5808) / $\mbox{W}_\lambda$ (C[iii]{} 5696), with the maximum FWHM reaching $\sim 60$Å, 2.5 times greater than the typical Galactic WC9 star. Line widths of sources hosting WC8 stars are much greater than typical Galactic WC8 stars, whilst the few sources containing WC7 subtypes are more indicative of Milky Way counterparts.
In Fig.\[WCL:EW\] we present [C]{}$\lambda$5696 line width versus line strength for WCL stars observed in M83, along with data for single or binary WCL stars in Local Group galaxies. The majority of WR complexes observed in M83 display evidence for line dilution from underlying stellar continua, since line strengths fall well below those observed in Local Group counterparts. Some WR complexes in M83 do display similar line strengths to those in the Milky Way or M31/M33, suggesting little evidence of line dilution in those cases.
In Fig.\[WCE\] we compare $\mbox{W}_{\lambda}$ ([C[iv]{} ]{}$\lambda
5808$) to FWHM ([C[iv]{} ]{}$\lambda 5808$) for early-type WC stars in M83, with (mostly single) Galactic and LMC counterparts. Again, the observed line strengths of sources containing early WC stars in M83 fall below those of Galactic and LMC stars. This is again attributed to line dilution by the underlying continua from early-type stars.
From Fig.\[WCE\], the observed WCE line widths of M 83 members are generally comparable to, or lower than, those of other Local Group members, in contrast with WC8, and especially, WC9 subtypes. Indeed, there are no cases for which FWHM (C[iv]{} $\lambda$5808) $\geq$ 100Å, corresponding to WO subtypes in the Milky Way/LMC [@kingsburgh95; @drew04], except possibly \#6 for which no evidence of O[vi]{} $\lambda$3811-34 is observed.
Finally, in Fig.\[WN\] we compare the equivalent width and FWHM of [[ii]{} ]{}$\lambda$4686 for all WN sources identified in M83. Again, we have included data for single/binary Galactic and LMC WN stars. Aside from the effect of line dilution, some late-type WN stars in M83 possess unusually large line widths. In some M83 complexes hosting multiple WN stars, line widths are a factor of two greater than Galaxy or LMC counterparts. These are reminiscent of observations of broad, strong [N[iii]{}]{}$\lambda$4640 in unresolved WR galaxies [@schmutz99].
![Equivalent width ($\mbox{W}_\lambda$ in Å) and FWHM (Å) of the [C[iv]{} ]{} $\lambda 5808$ line for early WC stars. Galactic, LMC, M31, M33 and IC10 WCE (unpublished WHT, AAT and 2.3m ANU data) and WO [@drew04; @kingsburgh95] stars are also indicated.[]{data-label="WCE"}](2262fig10.ps){width="7.8cm"}
![A comparison between the equivalent width ($\mbox{W}_\lambda$ in Å) of the [[ii]{} ]{}$\lambda 4686$ line and its FWHM (Å). WN stars in M83 and Local Group galaxies are marked, populations have been divided into WNE and WNL subtypes. Data for Galactic and LMC WNL stars are taken from @pac97, WNE information can be found in @conti89.[]{data-label="WN"}](2262fig11.ps){width="7.8cm"}
Origin of late WC stars at high metallicity?
--------------------------------------------
In Fig.\[WCL:WCE\] we compare the fractional distribution of WC8–9 to WC4–7 stars in galaxies with a wide range of metallicity. This clearly illustrates the extreme WCL population hosted by M83, indicating that WCL stars are uniquely associated with metal-rich environments. In M83 the relative number of late to early WC stars is found to be $\sim 9$, much greater than 0.9 and $\sim 0.2$ observed for the Milky Way and M31 respectively. C[iii]{} $\lambda$5696 has been observed in a small number of metal-rich WR galaxies [@phillips92; @pindao02], but as these represent integrated populations, the ‘average’ WC subtype is generally WC7–8.
It has long been recognised that Milky Way WC9 stars are universally observed towards the Galactic Centre. @smith91 argued that the apparent trend towards later subtypes was due to heavy mass-loss, revealing WC subtypes at an earlier evolutionary phase, [*assuming*]{} the surface (C+O)/He ratio decreases from early to late WC subtypes. However, subsequent spectral analysis failed to confirm any systematic trend in C/He with subtype [@koesterke95], arguing against late WC stars being exposed earlier due to prior mass-loss.
Instead, @pac02 claimed that WC subtypes resulted from primarily metallicity-dependent wind strengths. They suggested that the strength of C[iii]{} $\lambda$5696 scales very sensitively with wind density. If wind strengths increase with increasing (heavy element) metallicity, as already established for OB stars, stars which are otherwise identical will only reveal strong C[iii]{} $\lambda$5696 emission at high metallicities, with a corresponding late subtype.
![The fractional distribution of late to early WC stars in M83 and Local Group galaxies versus metallicity from @massey98, @schild03, @pac03 and the present study. The oxygen content of M83 is taken from @bresolin02, although more recent results suggest a reduction by $\sim$0.2 dex [@pilyugin04; @bresolin04][]{data-label="WCL:WCE"}](2262fig12.ps){width="6.5cm"}
Indeed, late-type WC stars are observed across the disk of M83 at a lower average galactocentric distance of 2.9 $\pm$ 0.9 arcmin ($\rho$ = 0.4 $\pm$ 0.1 $\rho_{0}$) versus 3.6 $\pm$ 0.9 arcmin ($\rho$ = 0.5 $\pm$ 0.1 $\rho_{0}$) for early-type WC stars. This can be explained by the weak metallicity gradient observed in M83 [@pilyugin04] since stars at smaller galactocentric distances will be more metal-rich than those at larger galactocentric distances.
As discussed in the introduction, the lower mass limit for stars that ultimately become WR stars decreases with increasing metal content, i.e. the lifetimes of low (initial) mass WR stars are greatly enhanced relative to those at lower metallicity. Could late WC stars be the descendants of such (low initial mass) stars, such that they greatly outnumber the (initially more massive) early WC stars?
In this scenario, one would expect late WC stars to be observed in Milky Way clusters with low mass turn-off’s. @schild86 identified the Galactic WC stars WR77 (WC8) and WR95 (WC9) in open clusters with progenitor masses as low as 35$M_{\odot}$, whilst early WC stars appear to originate from more massive progenitors ($\geq$ 60$M_{\odot}$). More recently, @massey01 identified the late WC WR93 (WC7) with a very high progenitor mass of $\geq 120
M_{\odot}$, comparable to or higher than early WC subtypes. Unfortunately, WC8–9 stars were not included in their study. In addition, the late WC component of WR11 ($\gamma$ Vel, WC8+O7.5) originates from a mass somewhat in excess of 30$M_{\odot}$, the current mass of the O star companion [@demarco99]. Overall, there is limited evidence for a distinction between the progenitor masses of early and late WC stars. Since nebular H$\alpha$ emission scales inversely with age, the complexes hosting early- and late-type WC stars appear to be located in both young and old clusters, such that they do indeed originate from the same parent population.
Consequently, the observed WC population in M83 can most naturally be explained if WC winds are metallicity dependent. Increased mass-loss would elevate the strength of the classification line C[iii]{} $\lambda$5696 resulting in predominantly late WC subtypes. Increased mass-loss rates would, of course, have implications for the life-times of WC stars. For an initial mass of 40 M$_{\odot}$ and $Z=0.04$, a WR star would spend 50% less time in the WC phase with a metal dependent stellar wind [See @meynet04]. At the current stage, however, it is not possible to firmly exclude different progenitor masses for early and late WC populations.
Finally, a similar metallicity effect was earlier proposed by @pac98b, i.e. WO subtypes would be restricted to environments with weak winds (i.e. low metallicities) due to the inverse sensitivity of O[vi]{} $\lambda$3811-34 with increased mass-loss. The absence of WO stars in M83 is also consistent with the inverse sensitivity of the classification line O[vi]{} $\lambda$3811-34 to increasing wind strength due to higher metallicity.
Galaxy log(O/H) N(WN) N(WC) N(WR) WC/WN
--------------------- ---------- ------------- ------------- -------------- -------
LMC$\,^{(1)}$ 8.4 109 24 134 0.2
– 30Dor$\,^{(1)}$ 15 3 18
Milky Way$\,^{(2)}$ 8.7 132 92 237 0.7
– Arches$\,^{(3)}$ 15 0 15
– Wd1$\,^{(4)}$ $\geq$ 12 $\geq$ 7 $\geq$ 19
M83 9.2 471$\pm$130 564$\pm$170 1035$\pm$300 1.2
– \#74 52$\pm$12 179$\pm$42 231$\pm$50
: The WR population of the LMC, Milky Way and M83. The cluster/association hosting the largest WR population is also indicated.[]{data-label="large:clus"}
$\,^{(1)}$@breysacher99, $\,^{(2)}$ @derhucht01, $\,^{(3)}$ @blum01, $\,^{(4)}$ @negueruela05.
Comparison with evolutionary predictions
----------------------------------------
Surveys for WR stars in Local Group galaxies over the past three decades have revealed a strong correlation between the relative number of WC to WN stars and oxygen content of the host galaxy [@massey96]. Extrapolating from previous observations, one would expect N(WC)/N(WN) $\geq 1$ for a galaxy forming stars continuously with $\sim$twice the Solar oxygen content [@massey98].
A summary of results for the disk population of M83 are presented in Table \[large:clus\] together with Local Group members, such that the observed N(WC)/N(WN) ratio is presented in Fig.\[WC:WN\]. Indeed, M83 continues the observed trend rather well with N(WC)/N(WN)$\sim$1.2. Undoubtedly, completeness should obviously be kept in mind given that WC stars are more readily identified in external galaxies than WN stars due to their intrinsically stronger lines. Nevertheless, our approach is optimised for net emission at $\lambda$4686, such that we achieve cases of 4$\sigma$ spectroscopic WNL detections with $W_{\lambda}$(He[ ii]{} $\lambda$4686) $\sim$ 1Å.
Recent bursts of star formation may cause strong deviations from the general trend via a strong enhancement of the WC population at an age of $\sim$5Myr [@pindao02]. Consequently, galaxies in which there is a significant recent starburst episode may strongly deviate from this correlation. IC10 strongly deviates from the overall trend in Fig.\[WC:WN\] due to an apparent ‘galaxy-wide’ starburst [@pac03], although the WN population may be significantly incomplete [@massey02].
![The WC/WN Ratio: Relative number of WC to WN stars in M83 is compared with those for nearby galaxies as a function of oxygen abundance, as determined by @massey98, @schild03, @pac03. We include recent evolutionary predictions for rotating massive stars from @meynet04 (dotted line).[]{data-label="WC:WN"}](2262fig13.ps){width="6cm"}
In the case of M83, since we have no information on the nuclear starburst, the above statistics should be reasonable for the quiescent star forming regions, except that the presence of a recent starburst in \#74 represents a non-negligible fraction of the total disk WR population. If we were to exclude \#74 from our statistics, we would obtain N(WC)/N(WN)=1 (equivalent to the Milky Way) and N(WC8–9)/N(WC4–7)=6 (vs 0.9 in the Milky Way) for the quiescent disk. The subtype ratios remain far from current evolutionary predictions at high metallicity [@meynet04].
Recently, @meynet04 have constructed a set of evolutionary models for rotating (initially $v_{\sin i}=300$ kms$^{-1}$) massive stars from $Z$=0.004 (SMC) to 0.04 ($\sim$M83). For low metallicity Local Group galaxies, predictions from rotating models are in good agreement with observed WC to WN ratios (see Fig.\[WC:WN\]). However, at higher metallicities, even allowing for metallicity dependent WR winds, evolutionary models fail and dramatically underestimate the number of WC stars, i.e. WC/WN=0.36 at $Z$=0.04. In fact, non-rotating models provide a better match to our observations of M83, although such models are unsuccessful in predicting the correct fraction of Type Ib/c to Type II Supernovae at high metallicity. Consequently, there remains a significant discrepancy between the observed and predicted WR populations above Solar metallicities.
Evolutionary models distinguish between late and early WN subtypes via the presence or absence of hydrogen, whilst spectroscopic definitions relate to the observed ionization of nitrogen lines. Determinations of hydrogen content are possible for Local Group WN stars, but the strong nebulosity and potential multiplicity for sources at the distance of M83 prevent such measurements. For a twice Solar metallicity ($Z$=0.04), @meynet04 predicts WNL/WNE $\sim$ 4 by number, allowing for metallicity dependent WR wind strengths. This should provide a reasonable analogue to the observational statistics for M83, on the basis of a good correspondence between late WN stars (with hydrogen) and early WN stars (without hydrogen) in the Milky Way. Consequently, as with the WC to WN ratio, the predicted distribution amongst WN subtypes ($\sim$4) differs from observations ($\sim$1) by a significant factor.
Nitrogen in WN stars is of course partially processed from carbon and oxygen, such that the CNO equilibrium abundance linearly scales with metallicity. As a consequence, the abundance of nitrogen in M83 WN stars will exceed that in the Milky Way and other Local Group galaxies. @pac00 demonstrated that for otherwise identical parameters, N[iii]{} $\lambda$4634–41 reacts more sensitively than N[iv]{} $\lambda$4058 to increased nitrogen content, i.e. a later WN subtype results. If WN winds are also metallicity dependent, the effect will be magnified such that one will expect a predominantly late WN population at high metallicities.
Indeed, given the weak metallicity gradient of M83 [@pilyugin04], late-type WN stars are located within the inner regions of M83, at an average galactocentric distance of 2.2$\pm$0.9 arcmin ($\rho$ = 0.3$\pm$0.1 $\rho_{0}$) versus 3.5$\pm$1.3 arcmin ($\rho$ = 0.5$\pm$0.2 $\rho_{0}$) for early-type WN stars.
Summary {#conc}
=======
Our analysis of the VLT/FORS2 imaging and spectroscopic data indicates that the disk of M83 hosts a large WR population. Using narrow-band optical images we have identified 283 candidate WR regions within M83, of which 198 have been spectroscopically observed. Of these we find that 132 regions contain WR stars. Absolute WR populations have been derived using line flux conversions adapted from @schaerer. We estimate a total WR population of 1035$\pm$300, consisting of 564$\pm$170 WC and 471$\pm$130 WN stars from this population, i.e. a quiescent N(WC)/N(WN) ratio of $\sim$1.2, or $\sim$1.0 excluding the starburst cluster \#74. This differs greatly from current evolutionary predictions at high metallicity, which suggest N(WC)/N(WN)$\sim$0.36, even allowing for metallicity dependent WR mass-loss rates [@meynet04].
The observed statistics exclude both the potentially large WR population in the central starburst, plus the population of perhaps $\sim$250 WR stars resulting from remaining candidates. @pellerin04 has carried out spectral synthesis of $FUSE$ observations of the nucleus of M83(30$''\times 30''$) suggesting a mass of 1.5$\times 10^{6} M_{\odot}$ and age 3.5 Myr, with an inferred WR population of 1700. This appears plausible given that the nuclear starburst has a star formation rate approaching that of the disk [@harris01; @bell01]. Consequently, the total WR population of M83 may exceed 3000.
Using the WR population derived in this study, the global surface density of M83 is found to be $\sim$3 WR/kpc$^{2}$, typical of that observed in the Solar Neighbourhood and M33 [@massey98]. If the nucleus and remaining candidates are accounted for this likely to increase this by a factor of $\sim$3.
The WC population of M83 is dominated by late-type stars. The relative number of WC8–9 to WC4–7 stars is found to be $\sim 9$ (or $\sim$6 excluding cluster \#74), outnumbering that of any other Local Group galaxy tenfold, as illustrated in Fig. \[WCL:WCE\]. WO stars are not observed in M83 suggesting that there is a genuine trend to later subtype at higher metallicities. Observed line widths in early WC subtypes appear to be comparable to those observed in the Milky Way and LMC. This population is most readily explained by a metallicity dependent wind strength amongst WC subtypes.
M83 has a substantial WN population, evenly split between early and late subtypes. The high WNL population likely results from the sensitivity of nitrogen diagnostics to the high global metallicity, whilst evidence in favour of metallicity dependent winds amongst WN subtypes is less clear. Evolutionary models, in contrast, predict a far higher late WN population at high metallicities, with WNL/WNE$\sim$4–5 [@meynet04].
From the present study we infer 9 complexes in M83 which contain a large WR population ($> 20$), with \#74 hosting over 200 late-type WN and WC stars, outnumbering 30 Doradus tenfold. HST/ACS images indicate that \#74 is a compact cluster, whilst several other massive compact clusters from @larsen04 host more modest WR populations.
Within galaxies located at up to $\sim$10Mpc, how unusual are the sources in M83 with regard to WR content? @schaerer99 identified $\sim$40 WN and WC stars in clusters A and B of NGC5253, similar to \#41 in M83, whilst He2–10 dwarfs even \#74, with 1100 WN and $>$250 WC stars. Recently, @chandar04 use HST/STIS UV spectroscopy to claim NGC3125-1 hosts 5000 WNL stars, a factor of ten times greater than optically derived by @schaerer99, and comparable to the global WR population in M83. Consequently, the WR population of M83, and source \#74 in particular, is extreme only with respect to Local Group galaxies. Where our results for M83 stand out from previous studies is the ability to resolve the disk WR population into $\sim$200 regions.
Given such a large WR population in the optically visible disk of M83, it is likely that one such massive star will undergo a core-collapse each century, given typical WR lifetimes of $\sim$10$^{5}$ yr. Indeed, SN1983N [@porter87], one of the six optically visible SN that have been reported in M83 since 1923, was a type Ib SN, for which WR stars are considered as the most probable precursors.
In order to directly witness a WR explode as a supernova on a shorter timeframe of a decade or so, a survey of the WR population in 10–20 nearby ($<$10Mpc) massive star forming galaxies would be required. The present observational program will continue towards this goal, complementing existing broad-band pre-Supernova surveys [@smartt03; @vandyk03].
LH acknowledges financial support from PPARC and a Royal Society summer studentship award, PAC acknowledges financial support from the Royal Society. We thank Soeren Larsen for providing ages estimates of clusters from broad-band photometry. Some of the data presented in this paper were obtained from the Multimission Archive at the Space Telescope Science Institute (MAST). STScI is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS5-26555. Support for MAST for non-HST data is provided by the NASA Office of Space Science via grant NAG5-7584 and by other grants and contracts.
Catalogue
=========
Tables A1 and A2 list properties of confirmed and candidate regions containing WR stars in M83.
Finding charts (electronic only)
================================
Fig. B1 provides a master key to individual finding charts B2–B17, obtained from FORS2 $\lambda$4684 images.
[64]{} natexlab\#1[\#1]{}
Abbott, J. B., Crowther, P. A., Drissen, L., Dessart, L., Martin, P., & Boivin, G. 2004, MNRAS, 350, 552
Asplund, M., Grevesse, N., Sauval, A., Allende Prieto, C., & Kiselman, D. 2004, A&A, 417, 751
Bell, E. & Kennicutt, R. 2001, ApJ, 548, 681
Blum, R. D., Schaerer, D., Pasquali, A., Heydari-Malayeri, M., Conti, P. S., & Schmutz, W. 2001, AJ, 122, 1875
Bresolin, F., Garnett, D. G., & Kennicutt, R. C. 2004, ApJ, 615, 228
Bresolin, F. & Kennicutt, R. C. 2002, ApJ, 572, 838
Breysacher, J., Azzopardi, M., & Testor, G. 1999, A&AS, 137, 117
Bruzual, A. & Charlot, S. 1993, ApJ, 405, 538
Chandar, R., Leitherer, C., & Tremonti, C. A. 2004, ApJ, 604, 153
Conti, P. S. & Massey, P. 1989, ApJ, 337, 251
Crowther, P. A. 1998, in Wolf-Rayet Phenomena in Massive Stars and Starburst Galaxies, Proc IAU Symp 193, ed. K. van der Hucht, G. Koenigsberger, & P. Eenens (San Francisco: ASP), 116
Crowther, P. A. 2000, A&A, 356, 191
Crowther, P. A., De [M]{}arco, O., & Barlow, M. J. 1998, MNRAS, 296, 367
Crowther, P. A., Dessart, L., Hillier, D., Abbott, J., & Fullerton, A. W. 2002, A&A, 392, 653
Crowther, P. A., Drissen, L., Abbott, J. B., Royer, P., & Smartt, S. J. 2003, A&A, 404, 483
Crowther, P. A., Hadfield, L. J., Schild, H., & Schmutz, W. 2004, A&A, 419, L17
Crowther, P. A. & Smith, L. J. 1997, A&A, 320, 500
De Marco, O. & Schmutz, W. 1999, A&A, 345, 163
de Vaucouleurs, D., Pence, W., & Davoust, E. 1983, ApJS, 53, 455
Drew, J. E., Barlow, M. J., Unruh, Y. C., Parker, Q. A., Wesson, R., Pierce, M. J., Masheder, M. R. W., & Phillipps, S. 2004, MNRAS, 351, 206
Elmegreen, D. M., Chromey, F. R., & Warren, A. R. 1998, AJ, 116, 2834
Foellmi, C., Moffat, A., & Guerrero, M. 2003, MNRAS, 338, 360
Gonzalez Delgado, R. M., Leitherer, C., Heckman, T., & Cervino, M. 1997, ApJ, 483, 705
Harris, J., Calzetti, D., Gallagher III, J. S., Conselice, C. J., & Smith, A. D. 2001, AJ, 122, 3046
Hummer, D. & Storey, P. J. 1987, MNRAS, 224, 801
Karachentsev, I. D., Sharina, M. E., Dolphin, A. E., Grebel, E. K., Geisler, D., Guhathakurta, P., Hodge, P., Karachentseva, V. E., Sarajedini, A., & Seitzer, P. 2002, A&A, 385, 21
Kingsburgh, R., Barlow, M. J., & Storey, P. 1995, A&A, 295, 75
Kosterke, L. & Hamann, W.-R. 1995, A&A, 299, 503
Kunth, D. & Sargent, W. 1981, A&A, 101, L5
Larsen, S. 2004, A&A, 416, 537
Larsen, S. & Richtler, T. 2004, A&A, 427, 495
Leitherer, C., Schaerer, D., Goldader, J. D., Gonzalez Delgado, R. M., & Carmelle, R. 1999, ApJS, 123, 3
Lundgren, A. A., Wikland, T., Olofsson, H., & Rydbeck, G. 2004, A&A, 413, 505
Massey, P. 1996, in Wolf-Rayet stars in the Framework of Stellar Evolution, ed. J.-. Vreux, C. W. H. et al. De Loore, & A. Willis (Univerity of Liege: Liege), 361
Massey, P. & Holmes, S. 2002, ApJ, 580, L35
Massey, P. & Johnson, O. 1998, ApJ, 505, 793
Massey, P., Olsen, K. A. G., & Parker, J. W. 2003, PASP, 115, 1265
Massey, P., Waterhouse, E., & DeGioia-Eastwood, K. 2000, AJ, 119, 2214
—. 2001, AJ, 121, 1050
Meynet, G. & Maeder, A. 2004, A&A, 429, 581
Moffat, A. F. J. & Shara, M. M. 1987, ApJ, 320, 266
Negueruela, I. & Clark, J. S. 2005, A&A, astro-ph/0503303
Pellerin, A. 2004, PhD thesis, Université Laval, Quebec
Phillips, A. & Conti, P. 1992, ApJ, 385, L91
Pilyugin, L. S., Contini, T., & Vilchez, J. M. 2004, A&A, 423, 427
Pindao, M., Schaerer, D., Gonzalez Delgado, R. M., & G., S. 2002, A&A, 394, 443
Porter, A. C. & Filippenko, A. V. 1987, AJ, 93, 1372
Rosa, M. & Richter, O. G. 1988, A&A, 192, 57
Schaerer, D., Contini, T., & Kunth, D. 1999, A&A, 341, 399
Schaerer, D., Contini, T., & Pindao, M. 1999, A&AS, 136, 35
Schaerer, D., Guseva, N., Izotov, Y., & Thuan, T. 2000, A&A, 362, 53
Schaerer, D. & Vacca, W. D. 1998, ApJ, 497, 618
Schild, H., Crowther, P. A., Abbott, J. B., & Shmutz, W. 2003, A&A, 397, 859
Schild, H. & Maeder, A. 1986, A&A, 136, 237
Schmutz, W. & Vacca, W. D. 1999, New Ast. Rev., 4, 197
Seaton, M. J. 1979, MNRAS, 187, 73[p]{}
Smartt, S. J., Maund, J. R., Gilmore, G. F., Tout, C., Kilkenny, D., & Benetti, S. 2003, MNRAS, 343, 735
Smith, L. F. & Maeder, A. 1991, A&A, 241, 77
Smith, L. J. 2005, in Formation and Evolution of Young Massive Clusters, ed. H. J. G. L. M. Lamers, A. Nota, & L. J. Smith, Vol. 233 (San Francisco: ASP), 159
Thim, F., Tammann, G. A., Saha, A., Dolphin, A., Sandage, A., Tolstoy, E., & Labhardt, L. 2003, ApJ, 590, 256
van der Hucht, K. A. 2001, New Ast. Rev., 45, 135
Van Dyk, S., Li, W., & Filippenko, A. 2003, PASP, 115, 1
Vink, J. S., de [K]{}oter, A., & Lamers, H. J. G. L. M. 2001, A&A, 369, 574
Whitmore, B. 2003, in A Decade of HST Science, Proc. STScI Symp Series Vol 14, ed. M. Livio & et al. (Cambridge: CUP), 153
[**[Table A1.]{}**]{} Catalogue of sources with confirmed WR spectroscopic signatures in M83. Errors are indicated on the second row for each source. Note $\Delta m$ and M$_{\rm B}$ shown in parentheses correspond to spectroscopic photometry. De-projected distances are expressed as a fraction of the Holmberg radius $\rho_{0} = 7.3 ' =
9.56\,\mbox{kpc}$, equivalent widths $\mbox{W}_\lambda$ and FWHM are in Åand observed line fluxes, $F_{\lambda}$, are expressed in $\mbox{erg s}^{-1} \mbox{cm}^{-2}$. The number of, and error in, WR stars in each source are estimated from the line luminosities indicated in Sect 3.2 and a distance of 4.5$\pm$0.3Mpc [@thim03].\
[c@c@c@c@c@c@l@c@c@r @c @c @r @c @c @r @c @c @c @c@c@c@c@c @ ]{} Cat & RA & Dec& $\rho/\rho_0$ & $m_{4686}$ & $\Delta m$ & $E_{\rm B-V}$ & $M_{\rm B}$ & DeV & & & & & & MOS\
\# & & & mag & mag & mag & mag & & $W_{\lambda}$ & FWHM & $F_{\lambda}$ & $W_{\lambda}$ & FWHM & $F_{\lambda}$ & $W_{\lambda}$ & FWHM & $F_{\lambda}$ & 2–5 & 6–10 & 4–6 & 7 & 8–9 & Mask\
1[$^{a}$]{}&13:36:40.28&-29:51:21.1&0.65&&0.50&&1&3&15& 3.8$\times 10^{-16}$ && &&& & & &3& & & &D5,B5\
&&&& & & 0.01& & &&& 1.5$\times 10^{-16}$&&& &&& & & 1& & & &\
2&13:36:41.04&-29:51:55.9&0.63& 18.2& -0.2& 0.54& -11.8& &6&15& 6.1$\times 10^{-16}$&1&15& 2.5$\times 10^{-16}$&2&35& 2.9$\times 10^{-16}$& & & & 2& &D4\
&&&& 0.0& 0.1& 0.01& 0.2& &&& 2.9$\times 10^{-16}$&&& 7.0$\times 10^{-17}$&&& 9.1$\times 10^{-17}$& & & & 1& &\
3[$^{a}$]{}&13:36:41.14&-29:51:48.8&0.62&(24.6)&(-1.9)& 0.41&(-5.3)&&$>$400&20& 1.8$\times 10^{-16}$&$>$400&25& 2.4$\times 10^{-16}$&$>$300&20& 5.6$\times 10^{-17}$& & & & & 2&D2\
&&&& & & 0.02& & &&& 6.8$\times 10^{-17}$&&& 7.8$\times 10^{-17}$&&& 3.3$\times 10^{-17}$& & & & & 1&\
4&13:36:41.16&-29:52:21.7&0.63& 21.1& -0.3& 0.70& -9.8& &35&35& 3.0$\times 10^{-16}$&&& &&& & 16& & & & &D2\
&&&& 0.1& 0.1& 0.05& 0.3& &&& 6.7$\times 10^{-17}$&&& &&& & 5& & & & &\
5&13:36:41.28&-29:52:24.4&0.63& 19.6& -0.2& 0.39& -9.9& &20&35& 1.2$\times 10^{-15}$&15&50& 7.8$\times 10^{-16}$&4&30& 4.0$\times 10^{-16}$& & & & & 8&B2\
&&&& 0.1& 0.1& 0.03& 0.2& &&& 1.1$\times 10^{-16}$&&& 1.1$\times 10^{-16}$&&& 1.1$\times 10^{-16}$& & & & & 1&\
6&13:36:41.64&-29:48:24.6&0.75& 23.0& (-0.5)& 0.50[$^{b}$]{}&(-5.0)&&&& &&& &$>$300&100& 6.0$\times 10^{-16}$& & & 3& & &B4\
&&&& 0.1& & 0.30 & & &&& &&& &&& 1.5$\times 10^{-16}$& & & 3& & &\
7&13:36:43.40&-29:52:24.1&0.56& 17.7& -0.1& 0.56& -12.5& 3&1&10& 4.3$\times 10^{-16}$&&& &&& & & 5& & & &B3\
&&&& 0.1& 0.1& 0.01& 0.2& &&& 1.7$\times 10^{-16}$&&& &&& & & 2& & & &\
8[$^{a}$]{}&13:36:43.46&-29:52:08.3&0.55&(23.6)&(-0.6)& 0.20&(-6.2)&&90&30& 3.4$\times 10^{-16}$&50&20& 1.2$\times 10^{-16}$&15&10& 2.5$\times 10^{-17}$& & & & & 1&D1\
&&&& & & 0.05& & &&& 3.0$\times 10^{-17}$&&& 1.8$\times 10^{-17}$&&& 1.4$\times 10^{-17}$& & & & & 0&\
9&13:36:43.86&-29:52:36.4&0.55& 20.8& -0.7& 0.55& -8.8& &10&40& 2.5$\times 10^{-16}$&&& &&& & & & 1& & &D3\
&&&& 0.1& 0.1& 0.05& 0.3& &&& 7.3$\times 10^{-17}$&&& &&& & & & 0& & &\
10&13:36:44.17&-29:52:26.8&0.54& 22.4& (-0.6)& 0.50[$^{b}$]{}&(-7.1)&&70&25& 2.8$\times 10^{-16}$&60&60& 1.5$\times 10^{-16}$&30&35& 6.8$\times 10^{-17}$& & & & & 2&D1\
&&&& 0.1& & 0.30& & &&& 3.5$\times 10^{-17}$&&& 3.9$\times 10^{-17}$&&& 2.4$\times 10^{-17}$& & & & & 2&\
11& 13:36:45.21& -29:51:05.5& 0.50& 22.6& (-1.8)& 0.28& (-5.6)& & 500& 45& 3.4$\times 10^{-16}$& & & & & & & 4& & & & & D1\
& & & & 0.1& & 0.02& & & & & 1.6$\times 10^{-17}$& & & & & & & 2& & & & &\
12& 13:36:45.85& -29:53:35.0& 0.55& 23.2& (-0.9)& 0.33& (-5.6)& & 100& 35& 1.1$\times 10^{-16}$& & & & & & & 2& & & & & D1\
& & & & 0.2& & 0.01& & & & & 1.1$\times 10^{-17}$& & & & & & & 0& & & & &\
13& 13:36:45.87& -29:52:37.8& 0.49& $<$19.2[$^{c}$]{}&-0.6& 0.38& $<$-10.1& & 20& 75& 1.9$\times 10^{-15}$& & & & & & & & & 3& & & D3\
& & & & 0.1& 0.1& 0.01& 0.2& & & & 3.2$\times 10^{-16}$& & & & & & & & & 1& & &\
14& 13:36:45.98& -29:49:17.1& 0.58& 22.5& (-1.5)& 0.50& -6.9& & $>$350& 35& 4.3$\times 10^{-16}$& $>$120& 60& 2.4$\times 10^{-16}$& $>$60& 29& 1.2$\times 10^{-16}$& & & & & 4& B1-2\
& & & & 0.1& &0.30 & 1.2& & & & 5.3$\times 10^{-17}$& & & 5.4$\times 10^{-17}$& & & 3.3$\times 10^{-17}$& & & & & 3&\
15& 13:36:46.06& -29:53:40.3& 0.55& 20.4& -0.4& 0.22& -8.5& & 20& 55& 7.2$\times 10^{-16}$& & & & 20& 30& 2.9$\times 10^{-16}$& & & 1& & & D3,B4\
& & & & 0.1& 0.1& 0.02& 0.2& & & & 2.0$\times 10^{-16}$& & & & & & 4.9$\times 10^{-17}$& & & 0& & &\
16& 13:36:47.53& -29:53:47.3& 0.52& 22.0& (-0.8)& 0.56& -7.4& & 150& 40& 5.6$\times 10^{-16}$& & & & 90& 30& 3.0$\times 10^{-16}$& & & 2& & & D2\
& & & & 0.1& & 0.01& 0.2& & & & 6.0$\times 10^{-17}$& & & & & & 3.2$\times 10^{-17}$& & & 0& & &\
17& 13:36:47.78& -29:51:25.9& 0.41& $<$18.0[$^{c}$]{}& -0.3& 0.25& $<$-9.7& & 5& 20& 1.9$\times 10^{-15}$& 2& 40& 9.5$\times 10^{-16}$& 3& 40& 6.1$\times 10^{-16}$& & & & & 6& B4\
& & & & 0.1& 0.2& 0.05& 0.3& & & & 4.4$\times 10^{-16}$& & & 3.4$\times 10^{-16}$& & & 2.3$\times 10^{-16}$& & & & & 2&\
18& 13:36:47.92& -29:51:30.3& 0.41& 21.3& (-1.9)& 0.45& (-7.1)& & 500& 40& 1.6$\times 10^{-15}$& 245& 60& 5.1$\times 10^{-16}$& 230& 40& 4.8$\times 10^{-16}$& & & & & 6& B1\
& & & & 0.1& & 0.05& & & & & 3.9$\times 10^{-16}$& & & 5.6$\times 10^{-17}$& & & 4.7$\times 10^{-17}$& & & & & 1&\
19& 13:36:48.08& -29:51:32.4& 0.40& 22.0& -0.8& 0.30& -6.9& & 40& 25& 2.4$\times 10^{-16}$& 10& 30& 5.8$\times 10^{-17}$& 15& 25& 5.1$\times 10^{-17}$& & & & & 1& D1\
& & & & 0.1& 0.1& 0.03& 0.2& & & & 1.3$\times 10^{-17}$& & & 1.2$\times 10^{-17}$& & & 9.5$\times 10^{-18}$& & & & & 1 &\
20& 13:36:48.50& -29:51:52.4& 0.39& 22.0& (-1.1)& 0.35& -7.0& & 150& 20& 4.8$\times 10^{-16}$& 75& 25& 1.7$\times 10^{-16}$& 60& 30& 1.2$\times 10^{-16}$& & & & & 2& D1\
& & & & 0.0& & 0.05& 0.3& & & & 1.4$\times 10^{-15}$& & & 1.3$\times 10^{-17}$& & & 1.4$\times 10^{-17}$& & & & & 0&\
[$^{a}$]{}) [[ii]{} ]{}PSF or spectroscopic photometry was unavailable and the average slit loss correction factor has been assumed, [$^{b}$]{}) the average E(B–V) value has been assumed, [$^{c}$]{}) photometry corresponds to a lower limit as the region is extended.
[**[Table A1.]{}**]{} (continued)\
[c @c@c@c@c@c@l@c@c@r @c @c @r @c @c @r @c @c @c @c@c@c@c@c@ ]{} Cat & RA & Dec & $\rho/\rho_0$ & $m_{4686}$ & $\Delta m$ & $E_{\rm B-V}$ & $M_{\rm B}$ & DeV & & & & & & MOS\
\# & & & mag & mag & mag & mag & & $W_{\lambda}$ & FWHM & $F_{\lambda}$ & $W_{\lambda}$ & FWHM & $F_{\lambda}$ & $W_{\lambda}$ & FWHM & $F_{\lambda}$ & 2–5 & 6–10 & 4–6 & 7 & 8–9 & Mask\
21& 13:36:48.54& -29:51:11.4& 0.39& 21.9& -0.6& 0.20& -6.7& & 70& 30& 4.7$\times 10^{-16}$& 45& 45& 2.0$\times 10^{-16}$& 45& 25& 1.6$\times 10^{-16}$& & & & & 1& D2\
& & & & 0.0& 0.1& 0.05& 0.3& & & & 8.7$\times 10^{-17}$& & & 1.0$\times 10^{-16}$& & & 4.1$\times 10^{-17}$& & & & & 1&\
22& 13:36:48.90& -29:51:44.2& 0.37& 22.2& (-1.5)& 0.19& -5.7& & $>$350& 45& 4.8$\times 10^{-16}$& & & & & & & 4& & & & & B1\
& & & & 0.1& & 0.04& 0.2& & & & 4.7$\times 10^{-17}$& & & & & & & 1& & & & &\
23& 13:36:49.22& -29:50:56.4& 0.38& 18.8& -0.1& 0.31& -10.4& & 10& 65& 1.1$\times 10^{-15}$& & & & 4& 30& 4.1$\times 10^{-16}$& & & 1& & & B3\
& & & & 0.1& 0.1& 0.02& 0.2& & & & 2.0$\times 10^{-16}$& & & & & & 1.2$\times 10^{-16}$& & & 0& & &\
24& 13:36:49.28& -29:50:52.8& 0.38& 21.7& (-0.6)& 0.50[$^{b}$]{}& -7.9& & 100& 30& 6.3$\times 10^{-16}$& 70& 45& 3.0$\times 10^{-16}$& 25& 30& 1.0$\times 10^{-16}$& & & & & 4& B4\
& & & & 0.0& & 0.30& 0.3& & & & 6.2$\times 10^{-17}$& & & 3.8$\times 10^{-17}$& & & 2.7$\times 10^{-17}$& & & & & 4&\
25& 13:36:49.77& -29:51:19.8& 0.35& 19.7& -0.4& 0.43& -10.2& 4& 10& 20& 6.1$\times 10^{-16}$& & & & & & & & 4& & & & D3\
& & & & 0.0& 0.0& 0.02& 0.2& & & & 7.2$\times 10^{-17}$& & & & & & & & 1& & & &\
26& 13:36:49.86& -29:52:10.0& 0.35& 19.7& -0.1& 0.48& -10.3& & 3& 25& 2.4$\times 10^{-16}$& & & & & & & & 2& & & & B2\
& & & & 0.1& 0.1& 0.01& 0.2& & & & 3.4$\times 10^{-17}$& & & & & & & & 0& & & &\
27& 13:36:50.34& -29:52:13.6& 0.33& 21.2& (-1.0)& 0.50[$^{b}$]{}& (-8.0)& & 200& 30& 1.3$\times 10^{-15}$& 55& 55& 4.4$\times 10^{-16}$& 65& 40& 4.8$\times 10^{-16}$& & & & & 6& B1\
& & & & 0.1& & 0.30& & & & & 3.5$\times 10^{-16}$& & & 7.5$\times 10^{-17}$& & & 5.7$\times 10^{-17}$& & & & & 8&\
28& 13:36:50.41& -29:52:19.1& 0.34& $<$19.8[$^{c}$]{}& 0.1& 0.50[$^{b}$]{}& $<$-11.1& & 10& 25& 5.0$\times 10^{-16}$& 10& 30& 4.5$\times 10^{-16}$& & & & & & & & 8& D3\
& & & & 0.1& 0.1& 0.30& 0.3& & & & 7.3$\times 10^{-17}$& & & 8.6$\times 10^{-17}$& & & & & & & & 6&\
29& 13:36:50.57& -29:52:58.6& 0.37& 20.8& (-0.1)& 0.60& -9.4& & 15& 40& 3.2$\times 10^{-16}$& 3& 15& 6.0$\times 10^{-17}$& 10& 45& 1.2$\times 10^{-16}$& & & & 1& & D1\
& & & & 0.1& & 0.01& 0.2& & & & 3.8$\times 10^{-17}$& & & 1.4$\times 10^{-16}$& & & 2.6$\times 10^{-17}$& & & & 0& &\
30& 13:36:50.78& -29:54:46.4& 0.54& (24.0)& & 0.50[$^{b}$]{}& & & 2000& 30& 4.4$\times 10^{-16}$& 500& 50& 1.3$\times 10^{-16}$& 1000& 45& 2.6$\times 10^{-16}$& & & & 1& & D1\
& & & & & & 0.30& & & & & 6.3$\times 10^{-17}$& & & 2.7$\times 10^{-16}$& & & 2.4$\times 10^{-17}$& & & & 1& &\
31& 13:36:50.87& -29:50:33.1& 0.36& $<$17.7[$^{c}$]{}& -0.2& 0.60& $<$-12.4& 6& 10& 40& 3.2$\times 10^{-15}$& 4& 50& 2.4$\times 10^{-15}$& 2& 30& 7.1$\times 10^{-16}$& & & & & 42& B4\
& & & & 0.1& 0.1& 0.05& 0.2& & & & 1.7$\times 10^{-17}$& & & 1.3$\times 10^{-17}$& & & 3.8$\times 10^{-18}$& & & & & 8&\
32& 13:36:50.88& -29:50:07.4& 0.39& 20.9& (-0.8)& 0.54& -9.0& & 120& 40& 1.5$\times 10^{-15}$& 110& 70& 8.2$\times 10^{-16}$& 90& 45& 6.2$\times 10^{-16}$& & & & & 12& B1\
& & & & 0.1& & 0.01& 0.2& & & & 9.6$\times 10^{-17}$& & & 1.2$\times 10^{-16}$& & & 8.6$\times 10^{-17}$& & & & & 2&\
33& 13:36:51.05& -29:52:37.0& 0.33& 20.4& -0.2& 0.55& -9.7& & 2& 10& 8.1$\times 10^{-17}$& & & & & & & & 1& & & & B1\
& & & & 0.1& 0.1& 0.01& 0.2& & & & 1.5$\times 10^{-17}$& & & & & & & & 0& & & &\
34& 13:36:51.19& -29:50:11.2& 0.38& 20.2& -0.2& 0.60& -10.4& 7& 5& 35& 2.4$\times 10^{-16}$& 7& 40& 1.8$\times 10^{-16}$& 7& 45& 1.5$\times 10^{-16}$& & & & & 3& B2\
& & & & 0.1& 0.2& 0.05& 0.3& & & & 5.2$\times 10^{-17}$& & & 4.4$\times 10^{-17}$& & & 6.1$\times 10^{-17}$& & & & & 1&\
35& 13:36:51.45& -29:53:12.9& 0.36& 20.6& (-0.1)& 0.55& -9.5& & 15& 55& 4.3$\times 10^{-16}$& & & & & & & 14& & & & & D1\
& & & & 0.1& & 0.01& 0.2& & & & 7.4$\times 10^{-17}$& & & & & & & 3& & & & &\
36& 13:36:51.51& -29:53:13.3& 0.36& 20.6& 0.2& 0.60& -9.7& & 20& 30& 5.1$\times 10^{-16}$& 15& 60& 3.1$\times 10^{-16}$& & & & & & & & 5& D2\
& & & & 0.1& 0.1& 0.01& 0.2& & & & 1.3$\times 10^{-16}$& & & 6.1$\times 10^{-17}$& & & & & & & & 1&\
37& 13:36:52.20& -29:51:22.1& 0.28& $<$19.2[$^{c}$]{}& -0.1& 0.45& $<$-10.4& & 10& 30& 1.0$\times 10^{-15}$& 10& 25& 3.5$\times 10^{-16}$& & & & & & & & 4& D5\
& & & & 0.1& 0.1& 0.05& 0.3& & & & 1.6$\times 10^{-16}$& & & 6.6$\times 10^{-17}$& & & & & & & & 1&\
38& 13:36:52.35& -29:53:15.2& 0.34& $<$16.9[$^{c}$]{}& -0.8& 0.78& $<$-13.5& & 5& 15& 3.0$\times 10^{-16}$& 2& 25& 7.6$\times 10^{-16}$& 2& & 8.2$\times 10^{-16}$& & 7& & & 21& D4\
& & & & 0.1& 0.1& 0.01& 0.2& & & & 8.2$\times 10^{-17}$& & & 1.9$\times 10^{-16}$& & & 2.1$\times 10^{-16}$& & 2& & & 6&\
39&13:36:52.59&-29:51:08.5&0.28&20.0&(-0.1)&0.47&(-9.8)&&15&35&1.6$\times 10^{-16}$&&&&10&35&3.2$\times 10^{-16}$&&&2&&&B2\
& & & & 0.1& & 0.02& & & & & 2.5$\times 10^{-16}$& & & & & & 3.2$\times 10^{-16}$& & & 1& & &\
40& 13:36:52.61& -29:53:03.1& 0.32& 19.2& -0.2& 0.42& -10.0& 10& 30&75 & 2.5$\times 10^{-15}$& 8& 50& 5.6$\times 10^{-16}$& 15& 60& 6.8$\times 10^{-16}$& & & & & 6& D3\
& & & & 0.1& 0.1& 0.02& 0.2& & & & 2.4$\times 10^{-16}$& & & 1.5$\times 10^{-16}$& & & 1.1$\times 10^{-16}$& & & & & 2&\
[$^{a}$]{}) [[ii]{} ]{}PSF or spectrscopic photometry was unavailable and the average slit loss correction factor has been assumed, [$^{b}$]{}) the average E(B–V) value has been assumed, [$^{c}$]{}) photometry corresponds to a lower limit as the region is extended.
[**[Table A1.]{}**]{} (continued)\
[c @c@c@c@c@c@l@c@c@r @c @c @r @c @c @r @c @c @c @c@c@c@c@c@ ]{} Cat & RA & Dec & $\rho/\rho_0$ & $m_{4686}$ & $\Delta m$ & $E_{\rm B-V}$ & $M_{\rm B}$ & DeV & & & & & & MOS\
\# & & & mag & mag & mag & mag & & $W_{\lambda}$ & FWHM & $F_{\lambda}$ & $W_{\lambda}$ & FWHM & $F_{\lambda}$ & $W_{\lambda}$ & FWHM & $F_{\lambda}$ & 2–5 & 6–10 & 4–6 & 7 & 8–9 & Mask\
41[$^{a}$]{}& 13:36:52.64& -29:51:47.5& 0.26& & 0.52& & 12& 5& 20& 1.5$\times 10^{-15}$& 2& 30& 1.4$\times 10^{-15}$& 2& 45& 1.9$\times 10^{-15}$& & 14& & 13& & B3\
& & & & & & 0.02& & & & & 3.9$\times 10^{-16}$& & & 3.5$\times 10^{-16}$& & & 9.0$\times 10^{-16}$& & 4& & 6& &\
42[$^{a}$]{}& 13:36:52.86& -29:52:50.0& 0.29& & 0.51& & 13& 1& 25& 5.9$\times 10^{-16}$& & & & & & & & 5& & & & B3\
& & & & & & 0.01& & & & & 2.5$\times 10^{-16}$& & & & & & & & 2& & & &\
43& 13:36:52.89& -29:51:15.2& 0.26& 20.8& -0.1& 0.50[$^{b}$]{}& -9.1& & 15& 15& 2.9$\times 10^{-16}$& & & & & & & & 3& & & & B2\
& & & & 0.1& 0.1& 0.30& 0.3& & & & 3.8$\times 10^{-17}$& & & & & & & & 3& & & &\
44& 13:36:53.34& -29:51:30.0& 0.24& 18.4& -0.3& 0.31& -10.9& 16& 5& 10& 1.2$\times 10^{-15}$& & & & & & & & 5& & & & D4\
& & & & 0.0& 0.0& 0.01& 0.2& & & & 8.9$\times 10^{-17}$& & & & & & & & 0& & & &\
45& 13:36:53.36& -29:51:52.9& 0.23& 22.4& -1.0& 0.56& -7.7& & 10& 20& 3.6$\times 10^{-17}$& & & & & & & & 1& & & & B2\
& & & & 0.2& 0.5& 0.02& 0.2& & & & 8.6$\times 10^{-18}$& & & & & & & & 0& & & &\
46& 13:36:53.60& -29:52:52.8& 0.28& 22.1& (-0.7)& 0.41& -7.6& & $>$100& 35& 2.4$\times 10^{-16}$& $>$350& 75& 2.7$\times 10^{-16}$& $>$200& 35& 9.1$\times 10^{-17}$& & & & & 3& B1\
& & & & 0.3& & 0.03& 0.2& & & & 6.3$\times 10^{-17}$& & & 8.4$\times 10^{-17}$& & & 2.5$\times 10^{-17}$& & & & & 1&\
47& 13:36:53.72& -29:51:31.2& 0.23& 21.6& (-0.4)& 0.75& -8.4& & 50& 45& 4.1$\times 10^{-16}$& & & & 20& 40& 1.1$\times 10^{-16}$& & & 1& & & D2\
& & & & 0.1& & 0.02& 0.2& & & & 8.4$\times 10^{-17}$& & & & & & 3.4$\times 10^{-17}$& & & 0& & &\
48& 13:36:53.87& -29:51:03.0& 0.24& 20.2& (-1.3)& 0.31& -8.6& & 210& 65& 3.2$\times 10^{-15}$& & & & & & & 44& & & & & B1\
& & & & 0.1& & 0.03& 0.2& & & & 2.9$\times 10^{-16}$& & & & & & & 6& & & & &\
49& 13:36:53.92& -29:48:49.6& 0.48& 18.2& -0.4& 0.40& -11.5& 19& 10& 30& 2.7$\times 10^{-15}$& 7& 35& 1.3$\times 10^{-15}$& 5& 30& 9.2$\times 10^{-16}$& & & & & 13& B3\
& & & & 0.1& 0.1& 0.03& 0.2& & & & 1.4$\times 10^{-15}$& & & & & & & & & & & 1&\
50& 13:36:53.98& -29:51:05.3& 0.24& 19.2& -0.2& 0.61& -11.1& & 14&35 & 6.5$\times 10^{-16}$& & & & & & & & 8& & & & D5\
& & & & 0.1& 0.1& 0.01& 0.2& & & & 2.0$\times 10^{-16}$& & & & & & & & 3& & & &\
51& 13:36:54.48& -29:55:55.0& 0.62& 21.8& -0.9& 0.45& -8.2& & 30& 20& 1.6$\times 10^{-16}$& & & & & & & & 1& & & & C2-3\
& & & & 0.1& 0.1& 0.05& 0.3& & & & 3.6$\times 10^{-17}$& & & & & & & & 0& & & &\
52& 13:36:54.93& -29:55:35.1& 0.57& 21.5& (-0.2)& 0.81& -10.0& & 20& 65& 2.2$\times 10^{-16}$& & & & & & & 18& & & & & D2\
& & & & 0.1& & 0.01& 0.2& & & & 4.6$\times 10^{-17}$& & & & & & & 4& & & & &\
53& 13:36:54.94& -29:55:54.7& 0.62& 22.0& (-1.6)& 0.24& (-6.6)& &350& 25& 8.5$\times 10^{-17}$& 220& 50& 3.6$\times 10^{-16}$&190& 35& 2.9$\times 10^{-16}$& & & & & 2& D1\
& & & & 0.0& &0.03 & & & & & 1.8$\times 10^{-17}$& & & 7.3$\times 10^{-17}$& & & 5.9$\times 10^{-17}$& & & & & 1&\
54& 13:36:55.32& -29:49:47.8& 0.34& 21.8& (-1.2)& 0.52& (-7.4)& & $>$100& 30& 6.9$\times 10^{-16}$& 135& 45& 4.1$\times 10^{-16}$& 100& 30& 1.7$\times 10^{-16}$& & & & & 6& B1,A5\
& & & & 0.1& & 0.03& & & & & 1.5$\times 10^{-16}$& & & 5.9$\times 10^{-17}$& & & 4.0$\times 10^{-17}$& & & & & 1&\
55& 13:36:55.43& -29:48:05.7& 0.56& 21.9& -0.9& 0.76& -8.4& & 50& 55& 3.0$\times 10^{-16}$& & & & & & & 20& & & & & B1\
& & & & 0.0& 0.1& 0.02& 0.2& & & & 6.0$\times 10^{-17}$& & & & & & & 4& & & & &\
56& 13:36:55.81& -29:50:17.8& 0.27& 21.9& -1.6& 0.59& -7.9& & 30& 20& 3.9$\times 10^{-16}$& & & & & & & 15& & & & & A3\
& & & & 0.1& 0.6& 0.01& 0.3& & & & 4.6$\times 10^{-17}$& & & & & & & 2& & & & &\
57& 13:36:56.00& -29:49:53.8& 0.32& $<$20.0[$^{c}$]{}& (-0.5)& 0.43& $<$-9.7& 24& 40& 30& 1.3$\times 10^{-15}$& & & & & & & & 9& & & & A3,B2\
& & & & 0.1& & 0.01& 0.2& & & & 1.1$\times 10^{-16}$& & & & & & & & 2& & & &\
58& 13:36:56.20& -29:49:33.3& 0.36& 21.7& (-1.0)& 0.50& (-7.4)& & 185& 35& 7.4$\times 10^{-16}$& 65& 65& 4.0$\times 10^{-16}$& 70& 45& 4.0$\times 10^{-16}$& & & & & 5& B1-2,A2\
& & & & 0.1& & 0.01& & & & & 1.4$\times 10^{-16}$& & & 7.6$\times 10^{-17}$& & & 5.6$\times 10^{-17}$& & & & & 1&\
59& 13:36:56.95& -29:52:48.0& 0.18& 20.0& -0.2& 1.08& -12.5& 25& 4& 10& 2.1$\times 10^{-16}$& & & & & & & & 14& & & & B2,A3\
& & & & 0.1& 0.1& 0.01& 0.2& & & & 2.7$\times 10^{-17}$& & & & & & & & 2& & & &\
60[$^{a}$]{}& 13:36:57.06& -29:50:19.2& 0.25& (22.1)& (-0.2)& 0.32& (-8.4)& & 20& 45& 4.0$\times 10^{-16}$& & & & & & & 6& & & & & A2\
& & & & & & 0.01& & & & & 1.8$\times 10^{-16}$& & & & & & & 3& & & & &\
[$^{a}$]{}) [[ii]{} ]{}PSF or spectroscopic photometry was unavailable and the average slit loss correction factor has been assumed, [$^{b}$]{}) the average E(B–V) value has been assumed, [$^{c}$]{}) photometry corresponds to a lower limit as the region is extended.
[**[Table A1.]{}**]{} (continued).\
[c @c@c@c@c@c@l@c@c@r @c @c @r @c @c @r @c @c @c @c@c@c@c@c@ ]{} Cat & RA & Dec & $\rho/\rho_0$ & $m_{4686}$ & $\Delta m$ & $E_{\rm B-V}$ & $M_{\rm B}$ & DeV & & & & & & MOS\
\# & & & mag & mag & mag & mag & & $W_{\lambda}$ & FWHM & $F_{\lambda}$ & $W_{\lambda}$ & FWHM & $F_{\lambda}$ & $W_{\lambda}$ & FWHM & $F_{\lambda}$ & 2–5 & 6–10 & 4–6 & 7 & 8–9 & Mask\
61& 13:36:57.95& -29:52:55.0& 0.17& $<$18.6[$^{c}$]{}& 0.1& 0.40& $<$-11.7& & 4& 30& 8.3$\times 10^{-16}$& 2& 30& 3.0$\times 10^{-16}$& 2& 50& 4.1$\times 10^{-16}$& & & & 2& & D4\
& & & & 0.1& 0.1& 0.05& 0.3& & & & 1.8$\times 10^{-16}$& & & 6.0$\times 10^{-17}$& & & 7.9$\times 10^{-17}$& & & & 1& &\
62& 13:36:58.24& -29:48:16.4& 0.53& 19.8& -0.4& 0.45& -9.9& & 20& 50& 7.6$\times 10^{-16}$& 10& 70& 3.6$\times 10^{-16}$& 15& 50& 3.7$\times 10^{-16}$& & & & & 4& B2\
& & & & 0.1& 0.1& 0.05& 0.3& & & & 1.4$\times 10^{-16}$& & & 7.3$\times 10^{-17}$& & & 7.6$\times 10^{-17}$& & & & & 1&\
63& 13:36:58.51& -29:55:19.9& 0.50& 22.1& (-0.7)& 0.91& -9.1& & 85& 30& 3.9$\times 10^{-16}$& 30& 50& 1.2$\times 10^{-16}$& 35& 40& 1.5$\times 10^{-16}$& & & && 5& D1,C1-3\
& & & & 0.1& & 0.08& 0.4& & & & 4.5$\times 10^{-17}$& & & 1.8$\times 10^{-17}$& & & 1.9$\times 10^{-17}$& & & & & 2&\
64& 13:36:58.72& -29:48:06.1& 0.55& 19.9& -0.3& 0.59& & 27& 15& 40& 8.1$\times 10^{-16}$& & & & & & & & 10& & & & B3,A3-4\
& & & & 0.1& 0.1& 0.01& & & & & 2.3$\times 10^{-16}$& & & & & & & & 3& & & &\
65[$^{a}$]{}& 13:36:58.86& -29:50:11.9& 0.25& & 0.42& & & 45& 30& 7.3$\times 10^{-16}$& 25& 45& 3.5$\times 10^{-16}$& 25& 60& 2.9$\times 10^{-16}$& & & & & 4& A5\
& & & & & & 0.02& & & & & 2.7$\times 10^{-16}$& & & 1.3$\times 10^{-16}$& & & 1.2$\times 10^{-16}$& & & & & 2&\
66& 13:36:59.17& -29:54:27.0& 0.37& 17.3& -0.3& 0.23& -11.4& 29& 5& 30& 2.3$\times 10^{-15}$& 3& 50& 7.2$\times 10^{-16}$& 3& 50& 8.9$\times 10^{-16}$& & 8& & 4& & D4\
& & & & 0.1& 0.1& 0.02& 0.2& & & & 4.6$\times 10^{-16}$& & & 2.0$\times 10^{-15}$& & & 1.9$\times 10^{-16}$& & 2& & 1& &\
67& 13:36:59.41& -29:54:23.1& 0.36& 21.1& (-0.6)& 0.33& & & 100& 30& 9.9$\times 10^{-16}$& 90& 65& 6.2$\times 10^{-16}$& 75& 50& 5.5$\times 10^{-16}$& & & & & 5& D2\
& & & & 0.0& & 0.02& & & & & 1.2$\times 10^{-15}$& & & 1.3$\times 10^{-16}$& & & 1.1$\times 10^{-16}$& & & & & 1&\
68& 13:36:59.95& -29:48:42.6& 0.46& $<$18.7[$^{c}$]{}& -0.2& 0.55& $<$-11.6& 32& 10& 80& 1.7$\times 10^{-15}$& 2& 35& 2.5$\times 10^{-16}$& 3& 30& 4.0$\times 10^{-16}$& & & & 3& & B2\
& & & & 0.1& 0.1& 0.01& 0.2& & & & 1.7$\times 10^{-16}$& & & 5.4$\times 10^{-17}$& & & 6.7$\times 10^{-17}$& & & & 1& &\
69& 13:37:00.18& -29:55:39.9& 0.54& 22.5& (-0.6)& 0.30& -6.0& & 70& 20& 2.4$\times 10^{-16}$& 145& 25& 2.7$\times 10^{-16}$& 50& 40& 9.3$\times 10^{-17}$& & & & & 2& D1\
& & & & 0.0& & 0.05& 0.3& & & & 4.4$\times 10^{-17}$& & & 3.4$\times 10^{-17}$& & & 4.4$\times 10^{-17}$& & & & & 0&\
70& 13:37:00.27& -29:48:04.0& 0.56& 20.4& -0.6& 0.24& -8.4& & 30& 35& 1.5$\times 10^{-15}$& 9& 40& 3.3$\times 10^{-16}$& 10& 40& 3.4$\times 10^{-16}$& & & & & 2& A2,B2\
& & & & 0.1& 0.1& 0.01& 0.2& & & & 9.4$\times 10^{-17}$& & & 6.1$\times 10^{-17}$& & & 7.0$\times 10^{-17}$& & & & & 0&\
71[$^{a}$]{}& 13:37:00.40& -29:54:13.2& 0.33& 22.1& (0.0)& 0.50[$^{b}$]{}& & & 20& 20& 2.7$\times 10^{-16}$& & 40& 1.9$\times 10^{-16}$& & 15& 7.9$\times 10^{-17}$& & & & & 3& D1\
& & & & 0.1& & 0.30& & & & & 8.6$\times 10^{-16}$& & & 6.6$\times 10^{-17}$& & & 3.5$\times 10^{-17}$& & & & & 3&\
72& 13:37:00.41& -29:52:54.1& 0.14& 19.9& -0.2& 0.02& -8.1& & 10& 20& 5.6$\times 10^{-16}$& 10& 20& 2.4$\times 10^{-16}$& & & & & & & & 1& A1\
& & & & 0.1& 0.2& 0.02& 0.2& & & & 7.5$\times 10^{-17}$& & & 3.6$\times 10^{-17}$& & & & & & & & &\
73& 13:37:01.18& -29:52:53.9& 0.14& 19.2& -0.3& 0.21& -9.6& & 5& 20& 5.8$\times 10^{-16}$& & & & & & & & 2& & & & D5\
& & & & 0.1& 0.1& 0.02& 0.2& & & & 9.7$\times 10^{-17}$& & & & & & & & 0& & & &\
74& 13:37:01.42& -29:51:25.8& 0.08& $<$17.5[$^{c}$]{}& -0.1& 1.00& $<$-14.2& 35& 5& 20& 1.0$\times 10^{-15}$& 6& 40& 3.5$\times 10^{-15}$& 2& 40& 1.4$\times 10^{-15}$& & 52& & & 179& B4\
& & & & 0.1& 0.1& 0.03& 0.2& & & & 2.1$\times 10^{-16}$& & & 7.1$\times 10^{-16}$& & & 2.9$\times 10^{-16}$& & 12& & & 42&\
75& 13:37:02.49& -29:48:27.4& 0.51& 22.3& (-1.3)& 0.25& (-5.7)& & 375& 60& 6.1$\times 10^{-16}$& $>$100& $>$30& 7.2$\times 10^{-17}$& $>$250& 60& 2.6$\times 10^{-16}$& & & & 1& & A2,B1\
& & & & 0.1& & 0.05& & & & & 8.1$\times 10^{-17}$& & & 2.1$\times 10^{-17}$& & & 2.8$\times 10^{-17}$& & & & 0& &\
76& 13:37:02.56& -29:50:37.4& 0.20& 22.1& (-1.1)& 0.20& (-5.8)& & 100& 30& 2.9$\times 10^{-16}$& & & & & & & & 1& & & & A2,B2\
& & & & 0.0& & 0.04& & & & & 2.9$\times 10^{-17}$& & & & & & & & 0& & & &\
77& 13:37:02.84& -29:49:15.3& 0.40& 22.1& (-0.4)& 0.60& (-8.1)& & 45& 20& 2.3$\times 10^{-16}$& 25& 45& 9.2$\times 10^{-17}$& 25& 35& 7.9$\times 10^{-17}$& & & & & 2& A5\
& & & & 0.1& & 0.05& & & & & 2.1$\times 10^{-17}$& & & 2.4$\times 10^{-18}$& & & 2.1$\times 10^{-18}$& & & & & 0&\
78& 13:37:03.06& -29:50:49.5& 0.18& $<$17.9[$^{c}$]{}& -0.3& 0.35& $<$-11.4& 38&3 &15 & 1.1$\times 10^{-15}$& 6& 35& 1.2$\times 10^{-15}$& 2& 35& 5.0$\times 10^{-16}$& & 11& & & & B4\
& & & & 0.1& 0.1& 0.05& 0.1& & & & 2.1$\times 10^{-16}$& & & 1.3$\times 10^{-16}$& & & 1.7$\times 10^{-16}$& & 2& & & &\
79[$^{a}$]{}& 13:37:03.10& -29:54:26.2& 0.36& (22.8)& (-0.7)& 0.58& (-8.3)& & 100& 35& 6.2$\times 10^{-16}$& 70& 50& 3.2$\times 10^{-16}$& 50& 30& 1.7$\times 10^{-16}$& & & & & 5& C1\
& & & & 0.1& & 0.02& & & & & 3.3$\times 10^{-16}$& & & 1.4$\times 10^{-16}$& & & 8.5$\times 10^{-17}$& & & & & 2&\
80& 13:37:03.83& -29:55:29.8& 0.51& 21.7& -0.6& 0.39& & & 0& 0& 2.1$\times 10 ^{-16}$& & & & & & & & 1& & & & C2-3\
& & & & 0.0& 0.1& 0.01& & & & & 4.1$\times 10 ^{-17}$& & & & & & & & 0& & & &\
[$^{a}$]{}) [[ii]{} ]{}PSF or spectroscopic photometry was unavailable and the average slit loss correction factor has been assumed, [$^{b}$]{}) the average E(B–V) value has been assumed, [$^{c}$]{}) photometry corresponds to a lower limit as the region is extended.
[**[Table A1.]{}**]{} (continued)\
[c @c@c@c@c@c@l@c@c@r @c @c @r @c @c @r @c @c @r @r@r@r@r@c@ ]{} Cat & RA & Dec& $\rho/\rho_0$ & $m_{4686}$ & $\Delta m$ & $E_{\rm B-V}$ & $M_{\rm B}$ & DeV & & & & & & MOS\
\# & & & mag & mag & mag & mag & & $W_{\lambda}$ & FWHM & $F_{\lambda}$ & $W_{\lambda}$ & FWHM & $F_{\lambda}$ & $W_{\lambda}$ & FWHM & $F_{\lambda}$ & 2–5 & 6–10 & 4–6 & 7 & 8–9 & Mask\
81& 13:37:03.88& -29:54:31.1& 0.37& 21.6& -1.2& 0.40& -7.7& & & No Data&& 130& 40& 3.4$\times 10 ^{-16}$& 35& 30& 9.1$\times 10 ^{-17}$& & & & & 3& D1\
& & & & 0.0& 0.1& 0.05& 0.3& & & & & & & 1.4$\times 10 ^{-17}$& & & 1.2$\times 10 ^{-17}$& & & & & 1&\
82& 13:37:03.97& -29:52:47.5& 0.15& 20.2& -0.8& 0.24& -8.6& & 45& 30& 1.2$\times 10 ^{-15}$& 30& 40& 6.0$\times 10 ^{-16}$& 20& 30& 3.2$\times 10 ^{-16}$& & & & & 4& A2,C3\
& & & & 0.0& 0.0& 0.04& 0.2& & & & 8.3$\times 10 ^{-17}$& & & 3.3$\times 10 ^{-17}$& & & 2.6$\times 10 ^{-17}$& & & & & 1&\
83& 13:37:04.39& -29:48:26.9& 0.52& 21.0& -0.6& 0.40& -8.6& & 25& 30& 4.1$\times 10 ^{-16}$& 20& 45& 1.8$\times 10 ^{-16}$& 5& 15& 5.7$\times 10 ^{-17}$& & & & & 2& A5\
& & & & 0.1& 0.1& 0.01& 0.2& & & & 3.1$\times 10 ^{-15}$& & & 4.6$\times 10 ^{-17}$& & & 1.3$\times 10 ^{-17}$& & & & & 1&\
84& 13:37:04.56& -29:49:21.8& 0.40& 20.8& (-0.8)& 0.40& (-8.2)& 43& 475& 45& 1.5$\times 10 ^{-15}$& $>$70& $>$20& 8.6$\times 10 ^{-16}$& $>$30& & 5.7$\times 10 ^{-16}$& & & & & 9& A3\
& & & & 0.1& & 0.05& & & & & 5.3$\times 10 ^{-17}$& & & 2.5$\times 10 ^{-16}$& & & 1.5$\times 10 ^{-16}$& & & & & 3&\
85& 13:37:04.58& -29:54:17.1& 0.35& 21.9& (-1.2)& 0.20& (-6.0)& & $>$270& 20& 6.7$\times 10 ^{-16}$& 110& 30& 6.8$\times 10 ^{-16}$& 50& 35& 4.5$\times 10 ^{-16}$& & & & & 4& C1\
& & & & 0.1& & 0.05& & & & & 1.3$\times 10 ^{-16}$& & & 1.4$\times 10 ^{-16}$& & & 8.6$\times 10 ^{-17}$& & & & & 1&\
86[$^{a}$]{}& 13:37:04.65& -29:50:58.4& 0.19& $<$16.2[$^{c}$]{}& (-0.0)& 0.32& $<$-13.1& 42& 2& 20& 2.0$\times 10 ^{-15}$& 2& 30& 2.9$\times 10 ^{-15}$& & & & & 9& & & 24& B4\
& & & & & & 0.04& 0.2& & & & 8.5$\times 10 ^{-16}$& & & 1.1$\times 10 ^{-15}$& & & & & 4& & & 10&\
87& 13:37:05.34& -29:48:20.4& 0.55& 20.0& -0.2& 0.36& -9.5& & 25& 80& 1.2$\times 10 ^{-15}$& & & & 20& 80& 7.3$\times 10 ^{-16}$& & & 3& & & A4\
& & & & 0.1& 0.1& 0.01& 0.2& & & & 2.0$\times 10 ^{-16}$& & & & & & 2.1$\times 10 ^{-16}$& & & 1& & &\
88& 13:37:05.53& -29:48:19.9& 0.55& 21.9& (-0.6)& 0.22& (-6.8)& & 60& 20& 3.2$\times 10 ^{-16}$& 60& 30& 1.3$\times 10 ^{-16}$& 60& 25& 1.0$\times 10 ^{-16}$& & & & & 1& A5\
& & & & 0.2& & 0.03& & & & & 3.2$\times 10 ^{-17}$& & & 2.6$\times 10 ^{-17}$& & & 2.4$\times 10 ^{-17}$& & & & & 0&\
89& 13:37:05.74& -29:53:41.2& 0.28& 21.3& (-0.4)& 0.50[$^{b}$]{}& (-8.4)& 44& 55& 35& 6.5$\times 10 ^{-16}$& & & & 20 &30& 1.7$\times 10 ^{-16}$& & & & & 3& C1\
& & & & 0.1& & 0.30& & & & & 1.6$\times 10 ^{-16}$& & & 6.1$\times 10 ^{-17}$& & & 4.1$\times 10 ^{-17}$& & & & & 4&\
90& 13:37:06.10& -29:53:43.5& 0.29& 20.4& (-0.1)& 0.32& & 44& 10& 55& 3.4$\times 10 ^{-16}$& 8& 65& 1.8$\times 10 ^{-16}$& 4& 45& 1.4$\times 10 ^{-16}$& & & & & 2& C3\
& & & & 0.2& & 0.02& & & & & 5.1$\times 10 ^{-17}$& & & 3.8$\times 10 ^{-17}$& & & 9.8$\times 10 ^{-18}$& & & & & 0&\
91& 13:37:06.36& -29:49:03.7& 0.46& 20.9& -0.5& 0.43& -8.4& & 15& 40& 4.9$\times 10 ^{-16}$& 15& 40& 2.6$\times 10 ^{-16}$& & & & & & & & 3& A3\
& & & & 0.1& 0.1& 0.01& 0.2& & & & 1.2$\times 10 ^{-16}$& & & 6.5$\times 10 ^{-17}$& & & & & & & & 1&\
92& 13:37:06.60& -29:54:51.6& 0.44& 21.8& (-1.7)& 0.31& (-6.4)& & $>$600& 45& 1.0$\times 10 ^{-15}$& $>$250& 70& 3.5$\times 10 ^{-16}$& $>$150& 40& 3.8$\times 10 ^{-16}$& & & & & 3& C1\
& & & & 0.1& & 0.01& & & & & 3.5$\times 10 ^{-16}$& & & 7.4$\times 10 ^{-17}$& & & 7.2$\times 10 ^{-17}$& & & & & 1&\
93[$^{a}$]{}& 13:37:06.62& -29:56:48.2& 0.71& (23.8)& (-1.9)& 0.50[$^{b}$]{}& (-5.8)& & $>$450& 30& 3.4$\times 10 ^{-16}$& & & & & & & 10& & & & & C1\
& & & & & & 0.30& & & & & 1.3$\times 10 ^{-16}$& & & & & & & 8& & & & &\
94[$^{a}$]{}& 13:37:07.10& -29:50:25.2& 0.31& (19.2)& (-0.1)& 0.36& (-11.7)& & 5& 25& 1.4$\times 10 ^{-15}$& & & & & & & 23& & & & & A1\
& & & & & & 0.01& & & & & 7.0$\times 10 ^{-16}$& & & & & & & 12& & & & &\
95& 13:37:07.35& -29:49:39.2& 0.40& 21.9& (-0.4)& 0.40& (-10.4)& & 15& 15& 1.2$\times 10 ^{-16}$& & & & & & & & 1& & & & A1\
& & & & 0.1& & 0.02& & & & & 1.3$\times 10 ^{-17}$& & & & & & & & 0& & & &\
96& 13:37:07.54& -29:52:53.6& 0.24& 19.3& 0.0& 0.27& -9.9& & 4& 30& 3.9$\times 10 ^{-16}$& & & & & & & & 2& & & & C2\
& & & & 0.2& 0.2& 0.02& 0.2& & & & 9.9$\times 10 ^{-17}$& & & & & & & & 1& & & &\
97[$^{a}$]{}& 13:37:07.55& -29:51:06.3& 0.25& (20.1)& (-0.0)& 0.67& & 46& 2& 15& 3.3$\times 10 ^{-16}$& & & & & & & & 5& & & & A5\
& & & & & & 0.01& & & & & 1.5$\times 10 ^{-16}$& & & & & & & & 2& & & &\
98& 13:37:07.58& -29:54:11.2& 0.37& 19.9& -0.3& 0.30& -9.4& & 20& 30& 9.7$\times 10 ^{-16}$& 15& 45& 5.4$\times 10 ^{-16}$& 15& 35& 3.2$\times 10 ^{-16}$& & & & & 4& C2\
& & & & 0.0& 0.1& 0.05& 0.3& & & & 1.4$\times 10 ^{-16}$& & & 4.9$\times 10 ^{-17}$& & & 3.7$\times 10 ^{-17}$& & & & & 1&\
99& 13:37:07.96& -29:52:07.9& 0.22& 20.9& (-0.2)& 0.28& (-8.0)& & 20& 30& 2.8$\times 10 ^{-16}$& & & & & & & & 1& & & & A3,C2\
& & & & 0.1& & 0.03& & & & & 9.7$\times 10 ^{-17}$& & & & & & & & 0& & & &\
100& 13:37:08.25& -29:51:13.6& 0.27& 20.6& -0.7& 0.63& -10.2& & 15& 25& 3.6$\times 10 ^{-16}$& & & & & & & & 5& & & & C2\
& & & & 0.1& 0.1& 0.02& 0.2& & & & 9.7$\times 10 ^{-17}$& & & & & & & & 1& & & &\
[$^{a}$]{}) [[ii]{} ]{}PSF or spectroscopic photometry was unavailable and the average slit loss correction factor has been assumed, [$^{b}$]{}) the average E(B–V) value has been assumed, [$^{c}$]{}) photometry corresponds to a lower limit as the region is extended.
[**[Table A1.]{}**]{} (continued).\
[c @c@c@c@c@c@l@c@c@r @c @c @r @c @c @r @c @c @r @r@r@r@r@c@ ]{} Cat & RA & Dec& $\rho/\rho_0$ & $m_{4686}$ & $\Delta m$ & $E_{\rm B-V}$ & $M_{\rm B}$ & DeV & & & & & & MOS\
\# & & & mag & mag & mag & mag & & $W_{\lambda}$ & FWHM & $F_{\lambda}$ & $W_{\lambda}$ & FWHM & $F_{\lambda}$ & $W_{\lambda}$ & FWHM & $F_{\lambda}$ & 2–5 & 6–10 & 4–6 & 7 & 8–9 & Mask\
101[$^{a}$]{}& 13:37:08.25& -29:53:30.5& 0.31& (22.5)& (-0.3)& 0.55& (-8.7)& & 30& 40& 3.3$\times 10 ^{-16}$& & & & & & & 11& & & & & C1\
& & & & & & 0.05& & & & & 1.0$\times 10 ^{-16}$& & & & & & & 5& & & & &\
102& 13:37:08.40& -29:52:54.9& 0.27& 18.1& -1.0& 0.34& -11.1& 48& 6& 40& 1.8$\times 10 ^{-15}$& & & & & & & & 9& & & & A4,C3\
& & & & 0.1& 0.2& 0.01& 0.2& & & & 3.1$\times 10 ^{-16}$& & & & & & & & 2& & & &\
103& 13:37:08.53& -29:52:12.0& 0.24& 18.7& -0.3& 0.78& -13.2& 49& 6& 40& 1.2$\times 10 ^{-15}$& & & & & & & & 29& & & & A4,C3\
& & & & 0.0& 0.1& 0.02& 0.2& & & & 3.3$\times 10 ^{-16}$& & & & & & & & 9& & & &\
104& 13:37:08.55& -29:49:06.5& 0.49& 19.7& -0.6& 0.36& -9.4& & 30& 40& 1.8$\times 10 ^{-15}$& 10& 40& 3.6$\times 10 ^{-16}$& & & & & & & & 3& A4\
& & & & 0.1& 0.1& 0.01& 0.2& & & & 5.0$\times 10 ^{-16}$& & & 8.3$\times 10 ^{-17}$& & & & & & & & 1&\
105& 13:37:08.70& -29:52:28.9& 0.25& 17.5& -0.4& 0.39& -12.1& 50& 6& 45& 1.4$\times 10 ^{-15}$& & & & & & & & 8& & & & A4\
& & & & 0.1& 0.1& 0.01& 0.2& & & & 4.6$\times 10 ^{-16}$& & & & & & & & 3& & & &\
106& 13:37:08.91& -29:52:05.7& 0.26& 20.7& (-0.1)& 0.50[$^{b}$]{}& (-9.3)& & 20& 40& 5.2$\times 10 ^{-16}$& 20& 50& 3.5$\times 10 ^{-16}$& 15& 50& 2.9$\times 10 ^{-16}$& & & & & 6& C1\
& & & & 0.1& & 0.30& & & & & 8.0$\times 10 ^{-17}$& & & 8.0$\times 10 ^{-17}$& & & 6.7$\times 10 ^{-17}$& & & & & 5&\
107[$^{a}$]{}& 13:37:09.02& -29:52:45.2& 0.28& 21.5& & 0.50[$^{b}$]{}& & & 0& 40& 4.6$\times 10 ^{-16}$& & & & & & & 15& & & & & C1\
& & & & 0.1& & 0.30& & & & & 1.5$\times 10 ^{-16}$& & & & & & & 10& & & & &\
108& 13:37:09.80& -29:52:36.2& 0.29& 20.3& -0.5& 0.35& -9.3& & 10& 20& 3.2$\times 10 ^{-16}$& & & & & & & & 2& & & & A3\
& & & & 0.1& 0.1& 0.01& 0.2& & & & 3.4$\times 10 ^{-17}$& & & & & & & & 1& & & &\
109& 13:37:10.42& -29:51:28.0& 0.32& 19.9& -0.7& 0.00& -8.0& & 105& 30& 3.3$\times 10 ^{-15}$& 110& 50& 2.1$\times 10 ^{-15}$& 40& 35& 7.9$\times 10 ^{-16}$& & & & & 7& C2\
& & & & 0.1& 0.2& 0.03& 0.2& & & & 5.6$\times 10 ^{-16}$& & & 2.5$\times 10 ^{-16}$& & & 1.5$\times 10 ^{-16}$& & & & & 1&\
110[$^{a}$]{}& 13:37:10.77& -29:50:56.6& 0.36& (22.7)& (-0.4)& 0.31& (-7.6)& & 45& 30& 4.1$\times 10 ^{-16}$& 65& 30& 4.2$\times 10 ^{-16}$& 20& 25& 1.3$\times 10 ^{-16}$& & & & & 3& A2\
& & & & & & 0.05& & & & & 1.5$\times 10 ^{-16}$& & & 1.3$\times 10 ^{-16}$& & & 6.9$\times 10^{-17}$& & & & & 1&\
111& 13:37:10.87& -29:54:22.2& 0.45& 20.1& -0.3& 0.30& -9.3& & 10& 30& 4.8$\times 10 ^{-16}$& & & & & & & & & & & 3& C2-3\
& & & & 0.1& 0.1& 0.02& 0.2& & & & 7.6$\times 10 ^{-17}$& & & & & & & & & & & 1&\
112& 13:37:11.03& -29:49:49.1& 0.46& 20.8& (-0.1)& 0.12& & & 45& 30& 1.1$\times 10 ^{-15}$& 15& 45& 3.5$\times 10 ^{-16}$& 20& 40& 3.6$\times 10 ^{-16}$& & & & & 2& A2\
& & & & 0.1& & 0.02& & & & & 1.8$\times 10 ^{-16}$& & & 1.2$\times 10 ^{-16}$& & & 9.4$\times 10 ^{-17}$& & & & & 1&\
113& 13:37:11.03& -29:55:43.7& 0.61& 21.9& (-0.8)& 0.50[$^{b}$]{}& -6.2& & $>$100& 40& 4.9$\times 10 ^{-16}$& 70& <60& 2.7$\times 10 ^{-16}$& & & & & & & & 5& C1\
& & & & 0.1& & 0.30& 1.2& & & & 5.3$\times 10 ^{-17}$& & & 5.6$\times 10 ^{-17}$& & & & & & & & 6&\
114& 13:37:11.11& -29:48:24.1& 0.63& 22.4& (-1.4)& 0.30& (-8.6)& & 170& 0& 3.0$\times 10 ^{-16}$& & & & & & & 4& & & & & A1\
& & & & 0.0& & 0.10& & & & & 2.4$\times 10 ^{-17}$& & & & & & & 2& & & & &\
115& 13:37:11.24& -29:49:42.0& 0.48& 20.9& -0.1& 0.50[$^{b}$]{}& -9.0& & 20& 40& 4.0$\times 10 ^{-16}$& 20& 35& 2.6$\times 10 ^{-16}$& 10& 30& 1.3$\times 10 ^{-16}$& & & & & 4& A5\
& & & & 0.1& 0.1& 0.30& 1.2& & & & 4.4$\times 10 ^{-17}$& & & 4.9$\times 10 ^{-17}$& & & 3.9$\times 10 ^{-17}$& & & & & 1&\
116& 13:37:11.46& -29:49:52.9& 0.47& 17.1& -0.5& 0.39& -12.4& 57& 3& 0& 2.0$\times 10 ^{-15}$& & & & & & & & 12& & & & A4\
& & & & 0.1& 0.1& 0.01& 0.2& & & & 3.3$\times 10 ^{-16}$& & & & & & & & 2& & & &\
117& 13:37:11.67& -29:49:18.1& 0.53& 22.5& (-0.1)& & & 58& & & No Data& & & & & & & 1& & & & & A2\
& & & & 0.1& & & & & & & & & & & & & & & & & & &\
118& 13:37:12.37& -29:50:21.6& 0.45& $<$19.2[$^{c}$]{}& (-0.1)& 0.38& $<$-10.2& & 10& 35& 9.3$\times 10 ^{-16}$& 10& 40& 6.3$\times 10 ^{-16}$& & & & & & & & 6& A1\
& & & & 0.1& & 0.01& 0.2& & & & 1.3$\times 10 ^{-16}$& & & 1.5$\times 10 ^{-16}$& & & & & & & & 1&\
119& 13:37:12.45& -29:52:03.7& 0.37& 21.2& -0.6& 0.35& -8.3& & 25& 2& 3.8$\times 10 ^{-16}$& 30& 25& 2.7$\times 10 ^{-16}$& & & & & & & & 2& A1\
& & & & 0.0& 0.0& 0.01& 0.2& & & & 5.5$\times 10 ^{-17}$& & & 1.8$\times 10 ^{-17}$& & & & & & & & 0&\
120& 13:37:12.58& -29:51:07.0& 0.40& 21.8& (-1.2)& 0.17& -5.9& & 230& 65& 7.5$\times 10 ^{-16}$& & & & & & & 6& & & & & A3,C1\
& & & & 0.0& & 0.02& 0.2& & & & 4.7$\times 10 ^{-17}$& & & & & & & 0& & & & &\
[$^{a}$]{}) [[ii]{} ]{}PSF of spectroscopic photometry was unavailable and the average slit loss correction factor has been assumed, [$^{b}$]{}) the average E(B–V) value has been assumed, [$^{c}$]{}) photometry corresponds to a lower limit as the region is extended.
[**[Table A1.]{}**]{} (continued)\
[c @c@c@c@c@c@l@c@c@r @c @c @r @c @c @r @c @c @c @c@c@c@c@c@ ]{} Cat & RA & Dec& $\rho/\rho_0$ & $m_{4686}$ & $\Delta m$ & $E_{\rm B-V}$ & $M_{\rm B}$ & DeV & & & & & & MOS\
\# & & & mag & mag & mag & mag & & $W_{\lambda}$ & FWHM & $F_{\lambda}$ & $W_{\lambda}$ & FWHM & $F_{\lambda}$ & $W_{\lambda}$ & FWHM & $F_{\lambda}$ & 2–5 & 6–10 & 4–6 & 7 & 8–9 & Mask\
121& 13:37:12.64& -29:49:50.1& 0.50& 21.1& (-1.1) & 0.46& -8.2& & & & 7.1$\times 10 ^{-16}$& & & & & & & 17& & & & & A1\
& & & & 0.1& & 0.02& 0.2& & & & 7.9$\times 10 ^{-17}$& & & & & & & 2& & & & &\
122& 13:37:12.68& -29:54:47.4& 0.53& 20.8& -1.0& 0.40& -8.3& & 100& 57& 1.4$\times 10 ^{-15}$& & & & 50& 60& 7.0$\times 10 ^{-16}$& & & 3& & & C2-3\
& & & & 0.0& 0.0& 0.01& 0.2& & & & 1.1$\times 10 ^{-16}$& & & & & & 1.4$\times 10 ^{-16}$& & & 1& & &\
123[$^{a}$]{}& 13:37:13.05& -29:51:37.7& 0.39& (21.5)& (0.0)& 0.40& (-9.5)& & 15& 27& 5.7$\times 10 ^{-16}$& 6& 25& 1.5$\times 10 ^{-16}$& & & & & & & & 2& A4\
& & & & & & 0.01& & & & & 1.8$\times 10 ^{-16}$& & & 6.9$\times 10 ^{-17}$& & & & & & & & 2&\
124[$^{a}$]{}& 13:37:15.50& -29:52:26.7& 0.46& 21.2& -0.6& 0.41& -8.4& & & No Data&& 30& 50& 3.2$\times 10 ^{-16}$& 16& 35& 1.3$\times 10 ^{-16}$& & & & & 3& A5,C2\
& & & & 0.1& 0.1& 0.01& 0.2& & & & & & & 1.2$\times 10 ^{-16}$& & & 5.9$\times 10 ^{-17}$& & & & & 1&\
125& 13:37:16.29& -29:53:09.8& 0.51& 21.4& (-0.7)& 0.39& (-7.6)& & 115& 27& 8.5$\times 10 ^{-16}$& 60& 45& 2.9$\times 10 ^{-16}$& 35& 45& 3.0$\times 10 ^{-16}$& & & & & 3& C1\
& & & & 0.1& & 0.03& & & & & 7.7$\times 10 ^{-17}$& & & 5.7$\times 10 ^{-17}$& & & 5.4$\times 10 ^{-17}$& & & & & 1&\
126[$^{a}$]{}& 13:37:16.33& -29:50:56.4& 0.52& 20.7& (-0.4)& 0.20& (-10.3)& & 45& 40& 8.4$\times 10 ^{-16}$& & & & 20& 45& 3.8$\times 10 ^{-16}$& & & 1& & & A1\
& & & & 0.1& & 0.05& & & & & 2.4$\times 10 ^{-16}$& & & & & & 1.4$\times 10 ^{-16}$& & & 0& & &\
127[$^{a}$]{}& 13:37:16.93& -29:50:43.5& 0.55& (20.7)& (-0.9)& 0.00& (-8.0)& & 130& 28& 4.9$\times 10 ^{-15}$& & & & 135& 35& 1.8$\times 10 ^{-15}$& & & 3& & & A1,A3\
& & & & & & 0.05& & & & & 1.4$\times 10 ^{-15}$& & & & & & 6.7$\times 10 ^{-16}$& & & 1& & &\
128& 13:37:17.18& -29:51:28.0& 0.53& $<$21.1[$^{c}$]{}& (-0.3)& 0.30& $<$-7.9& & 40& 25& 4.5$\times 10 ^{-16}$& 30& 40& 2.3$\times 10 ^{-16}$& 20& 30& 1.3$\times 10 ^{-16}$& & & & & 2& A1-2\
& & & & 0.1& & 0.05& 0.3& & & & 2.6$\times 10 ^{-17}$& & & 2.1$\times 10 ^{-17}$& & & 1.7$\times 10 ^{-17}$& & & & & 0&\
129& 13:37:17.52& -29:51:25.2& 0.54& 19.7& -0.3& 0.70& -11.0& & 15& 0& 1.0$\times 10 ^{-15}$& & & & 8& 45& 3.7$\times 10 ^{-16}$& & & 4& & & A3\
& & & & 0.0& 0.1& 0.05& 0.3& & & & 3.8$\times 10 ^{-16}$& & & & & & 9.2$\times 10 ^{-19}$& & & 1& & &\
130& 13:37:18.08& -29:48:04.5& 0.81& 23.0& (-1.5)& 0.15& -4.9& & 215& 25& 1.8$\times 10 ^{-16}$& & & & & & & 1& & & & & A1\
& & & & 0.0& & 0.02& 0.2& & & & 1.8$\times 10 ^{-17}$& & & & & & & 0& & & & &\
131& 13:37:18.63& -29:51:41.3& 0.57& 22.1& (-1.9)& 0.00& & & 530& 25& 6.2$\times 10 ^{-16}$& 525& 50& 2.8$\times 10 ^{-16}$& 440& 55& 2.3$\times 10 ^{-16}$& & & & & 1& A1\
& & & & 0.0& & 0.05& & & & & 4.3$\times 10 ^{-17}$& & & 3.5$\times 10 ^{-17}$& & & 3.5$\times 10 ^{-17}$& & & & & 0&\
132& 13:37:23.66& -29:48:53.8& 0.88& 22.2& (-0.4)& 0.56& (-8.1)& & 30& 30& 1.5$\times 10 ^{-17}$& & & & & & & 1& & & & & A2\
& & & & 0.1& & 0.04& & & & & 4.1$\times 10 ^{-18}$& & & & & & & 0& & & & &\
[$^{a}$]{}) [[ii]{} ]{}PSF or spectrscopic photometry was unavailable and the average slit loss correction factor has been assumed, [$^{b}$]{}) the average E(B–V) value has been assumed, [$^{c}$]{}) photometry corresponds to a lower limit as the region is extended.
[**[Table A2.]{}**]{} Catalogue of sources with candidate WR signatures in M83, either from narrow-band imaging or for which MOS spectroscopy was ambiguous.\
[l@l@l@ c@ c@c@ c@ l@ l@ l @c @c@ c@ c @]{}Cat&RA&Dec&$\rho / \rho_{0}$&$m_{4686}$& $m_{4686 - 4781}$&Dev& Cat&RA&Dec&$\rho / \rho_{0}$&$m_{4686}$& $m_{4686-4781}$&Dev\
&J2000&J2000&&&&&&J2000&J2000&&&&\
133&13:36:44.01&–29:51:47.1&0.53&–&–&&178&13:37:01.25&–29:52:52.8&0.13&19.2&–0.2&\
134&13:36:44.02&–29:52:26.5&0.54&21.7&–&&179&13:37:01.91&–29:52:13.2&0.05&–&–&\
135&13:36:44.48&–29:52:10.5&0.52&–&–&&180$^l$&13:37:02.04&–29:55:30.2&0.51&16.9&–0.3&36\
136&13:36:44.51&–29:53:22.1&0.57&19.2&–0.2&&181&13:37:02.19&–29:57:20.2&0.77&20.6&–0.2&\
137&13:36:44.55&–29:52:10.0&0.52&–&–&&182&13:37:02.24&–29:55:33.8&0.52&19.9&–0.3&\
138&13:36:46.05&–29:52:49.7&0.49&19.9&–0.2&&183&13:37:02.64&–29:49:40.8&0.34&18.0&–0.4&39\
139&13:36:48.07&–29:53:11.3&0.45&19.2&–0.1&&184&13:37:02.84&–29:54:05.5&0.31&21.2&–&\
140&13:36:48.29&–29:52:44.9&0.42&19.9&0.1&&185&13:37:02.90&–29:50:50.8&0.18&18.0&–0.4&38\
141&13:36:49.35&–29:55:28.4&0.65&21.2&–0.2&&186$^p$&13:37:03.00&–29:49:44.4&0.33&–&–&39\
142&13:36:49.63&–29:51:19.5&0.36&21.4&–0.5&5&187&13:37:03.01&–29:48:38.8&0.48&20.1&–0.5&\
143&13:36:49.74&–29:50:15.9&0.41&–&–&&188$^l$&13:37:03.20&–29:55:23.5&0.49&19.3&–0.1&\
144&13:36:50.05&–29:52:14.6&0.35&17.7&–0.2&&189$^l$&13:37:03.38&–29:51:00.3&0.16&18.9&–0.3&\
145$^p$&13:36:50.72&–29:54:05.1&0.46&17.8&–0.5&&190&13:37:03.44&–29:51:02.9&0.16&$<$17.3[$^{c}$]{}&–0.3&41\
146&13:36:51.06&–29:50:39.4&0.35&–&–&&191&13:37:03.79&–29:54:14.3&0.33&–&–&\
147&13:36:51.34&–29:50:10.2&0.38&$<$18.1[$^{c}$]{}&–0.6&&192&13:37:04.08&–29:53:54.7&0.29&19.1&–0.1&\
148$^p$&13:36:52.11&–29:52:35.2&0.30&–&–&9&193$^l$&13:37:04.20&–29:55:29.1&0.51&19.6&0.0&\
149&13:36:52.14&–29:51:20.4&0.28&–&–&&194$^l$&13:37:04.81&–29:49:24.8&0.40&17.6&0.3&43\
150&13:36:52.46&–29:52:48.7&0.30&19.6&–&&195&13:37:05.07&–29:50:59.0&0.20&–&–&\
151&13:36:52.53&–29:52:43.9&0.29&–&–&&196&13:37:06.62&–29:50:51.4&0.25&19.5&–0.3&\
152&13:36:52.54&–29:51:09.5&0.28&19.4&–0.2&14&197&13:37:07.19&–29:51:28.1&0.22&18.1&–0.1&\
153&13:36:52.82&–29:51:15.8&0.26&–&–&&198&13:37:07.62&–29:52:51.4&0.24&17.1&–0.6&\
154&13:36:53.02&–29:52:28.6&0.26&20.0&–0.2&&199&13:37:07.68&–29:49:13.9&0.46&19.3&–0.2&\
155&13:36:53.08&–29:55:48.8&0.63&–&–&&200$^p$&13:37:07.71&–29:51:14.9&0.25&19.7&–0.3&\
156&13:36:53.21&–29:52:58.5&0.29&17.3&–0.3&&201&13:37:07.76&–29:49:18.1&0.46&22.1&–&\
157&13:36:53.21&–29:51:30.7&0.24&18.9&–0.2&16&202&13:37:08.26&–29:54:01.7&0.36&–&–&\
158&13:36:53.31&–29:49:55.9&0.36&22.1&–&&203&13:37:09.01&–29:50:07.6&0.38&18.7&–0.2&\
159&13:36:53.34&–29:53:25.8&0.33&–&–&&204&13:37:09.15&–29:49:20.6&0.48&–&–&\
160&13:36:54.90&–29:53:09.4&0.27&–&–&22&205&13:37:09.18&–29:51:10.1&0.30&–&–&\
161&13:36:55.06&–29:48:49.0&0.47&22.1&–&&206&13:37:09.31&–29:51:50.6&0.27&17.8&–0.2&\
162&13:36:55.15&–29:49:50.5&0.34&23.2&–&&207&13:37:09.38&–29:52:20.0&0.27&19.2&0.1&\
163&13:36:55.20&–29:52:13.6&0.18&–&–&&208&13:37:09.48&–29:52:22.2&0.28&20.6&–0.5&\
164&13:36:55.31&–29:48:10.0&0.55&22.7&–0.6&&209$^l$&13:37:09.48&–29:52:38.1&0.28&23.3&–&\
165$^p$&13:36:55.61&–29:54:15.5&0.38&17.6&–0.4&&210&13:37:09.63&–29:49:56.4&0.42&–&–&\
166&13:36:55.71&–29:55:40.3&0.57&–&–&&211&13:37:09.67&–29:49:01.7&0.52&–&–&\
167&13:36:55.85&–29:47:47.5&0.60&–&–&&212&13:37:09.75&–29:49:08.0&0.51&–&–&\
168&13:36:56.06&–29:52:52.5&0.21&–&–&&213&13:37:09.87&–29:50:10.5&0.40&20.0&–0.3&\
169&13:36:56.32&–29:54:55.6&0.46&–&–&&214&13:37:10.02&–29:49:13.4&0.51&18.9&–0.4&\
170&13:36:56.75&–29:52:12.7&0.14&–&–&&215&13:37:10.88&–29:49:54.7&0.45&–&–&\
171&13:36:56.87&–29:52:48.7&0.18&20.0&–0.3&&216&13:37:12.01&–29:51:48.3&0.36&–&–&\
172$^p$&13:36:57.14&–29:54:49.2&0.44&19.1&–0.2&&217&13:37:12.60&–29:55:37.9&0.62&23.1&–&\
173&13:36:57.23&–29:51:26.3&0.13&19.3&–0.3&&218&13:37:16.07&–29:50:56.4&0.52&–&–&\
174$^p$&13:36:58.55&–29:48:12.0&0.54&21.6&–0.6&&219&13:37:16.37&–29:50:59.2&0.52&20.2&–0.3&\
175&13:36:58.95&–29:51:26.5&0.09&$<$19.3[$^{c}$]{}&–0.3&28&220&13:37:16.69&–29:50:54.5&0.54&20.9&–0.2&\
176&13:37:00.64&–29:54:23.0&0.35&20.9&–&&221&13:37:24.08&–29:48:52.6&0.89&23.2&–&\
177&13:37:01.19&–29:52:52.3&0.13&19.6&–0.2&&&&&\
${^p)}$ Insufficient S/N for classification, ${^l)}$ insufficient spectral coverage for WN identification, [$^{c}$]{}) photometry corresponds to a lower limit as the region is extended.\
{width="18cm"} [**Fig. B1:**]{} Master finding chart, indicating location of confirmed WR sources in M83 (+) and corresponding finding chart. North is up and east to the left on this $\lambda$4684 narrow-band FORS2 image of M83 (12$\times$12 arcmin). \[fc1\]
{width="18cm"} [**Fig. B2:**]{} Finding chart 1. North is up and east to the left on this $\lambda$4684 narrow-band FORS2 image, with the scale indicated. \[fc2\]
{width="18cm"} [**Fig. B3:**]{} Finding chart 2. North is up and east to the left on this $\lambda$4684 narrow-band FORS2 image, with the scale indicated. The FORS2 CCD detector gap runs along the bottom of the image. \[fc3\]
{width="18cm"} [**Fig. B4:**]{} Finding chart 3. North is up and east to the left on this $\lambda$4684 narrow-band FORS2 image, with the scale indicated. \[fc4\]
{width="18cm"} [**Fig. B5:**]{} Finding chart 4. North is up and east to the left on this $\lambda$4684 narrow-band FORS2 image, with the scale indicated. \[fc5\]
{height="25cm"} [**Fig. B6:**]{} Finding chart 5. North is up and east to the left on this $\lambda$4684 narrow-band FORS2 image, with the scale indicated. \[fc6\]
{height="25cm"} [**Fig. B7:**]{} Finding chart 6. North is up and east to the left on this $\lambda$4684 narrow-band FORS2 image, with the scale indicated. \[fc7\]
{width="12cm"} [**Fig. B8:**]{} Finding chart 7. North is up and east to the left on this $\lambda$4684 narrow-band FORS2 image, with the scale indicated. \[fc8\]
{width="18cm"} [**Fig. B9:**]{} Finding chart 8. North is up and east to the left on this $\lambda$4684 narrow-band FORS2 image, with the scale indicated. \[fc9\]
{width="18cm"} [**Fig. B10:**]{} Finding chart 9. North is up and east to the left on this $\lambda$4684 narrow-band FORS2 image, with the scale indicated. \[fc10\]
{height="25cm"} [**Fig. B11:**]{} Finding chart 10. North is up and east to the left on this $\lambda$4684 narrow-band FORS2 image, with the scale indicated. \[fc11\]
{height="25cm"} [**Fig. B12:**]{} Finding chart 11. North is up and east to the left on this $\lambda$4684 narrow-band FORS2 image, with the scale indicated. \[fc12\]
{width="18cm"} [**Fig. B13:**]{} Finding chart 12. North is up and east to the left on this $\lambda$4684 narrow-band FORS2 image, with the scale indicated. The feature on the west side of the image is a chip defect. \[fc13\]
{height="25cm"} [**Fig. B14:**]{} Finding chart 13. North is up and east to the left on this $\lambda$4684 narrow-band FORS2 image, with the scale indicated. The FORS2 CCD detector gap runs along the centre of the image. \[fc14\]
{width="18cm"} [**Fig. B15:**]{} Finding chart 14. North is up and east to the left on this $\lambda$4684 narrow-band FORS2 image, with the scale indicated. \[fc15\]
{width="18cm"} [**Fig. B16:**]{} Finding chart 15. North is up and east to the left on this $\lambda$4684 narrow-band FORS2 image, with the scale indicated. The FORS2 CCD detector gap runs along the top of the image. \[fc16\]
{width="9cm"} [**Fig. B17:**]{} Finding chart 16. North is up and east to the left on this $\lambda$4684 narrow-band FORS2 image, with the scale indicated. The bright feature in the north west corner of the image is a chip defect. \[fc17\]
[^1]: Based on observations made with ESO Telescopes at the Paranal Observatory under programme ID 69.B-0125
[^2]: 0.8$''$ corresponds to a linear scale of 18pc at the distance of M83 [@thim03]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Three events in a probability space form a conjunctive fork if they satisfy specific constraints on conditional independence and covariances. Patterns of conjunctive forks within collections of events are characterized by means of systems of linear equations that have positive solutions. This characterization allows patterns of conjunctive forks to be recognized in polynomial time. Relations to previous work on causal betweenness and on patterns of conditional independence among random variables are discussed.'
address:
- |
Department of Computer Science and Software Engineering\
Concordia University, Montreal, Quebec H3G 1M8\
Canada
- |
Institute of Information Theory and Automation\
Academy of Sciences of the Czech Republic\
Pod vod' arenskou věží 4, 182 08 Prague 8\
Czech Republic
- 'Google DeepMind, London, United Kingdom'
author:
- 'Vašek Chvátal, František Mat'' uš and Yori Zwólš'
title: |
Patterns of conjunctive forks\
---
[^1]
[^2]
Motivation
==========
Hans Reichenbach [@Rei56 Chapter 19] defined a *conjunctive fork* as an ordered triple $(A,B,C)$ of events $A$, $B$ and $C$ in a probability space $({\ensuremath{\mathit\Omega}\xspace},\mathcal F,{\ensuremath{P}\xspace})$ that satisfies $$\begin{aligned}
{\ensuremath{P}\xspace}(AC|B) &= {\ensuremath{P}\xspace}(A|B) {\ensuremath{P}\xspace}(C|B)\,,\label{reich1}\\
{\ensuremath{P}\xspace}(AC|{\overline{B}}) &= {\ensuremath{P}\xspace}(A|{\overline{B}}) {\ensuremath{P}\xspace}(C|{\overline{B}})\,,\label{reich2}\\
{\ensuremath{P}\xspace}(A|B) &> {\ensuremath{P}\xspace}(A|{\overline{B}})\,,\label{reich3}\\
{\ensuremath{P}\xspace}(C|B) &> {\ensuremath{P}\xspace}(C|{\overline{B}})\,\label{reich4}
\end{aligned}$$ where, as usual, $AC$ is a shorthand for $A\cap C$ and ${\overline{B}}$ denotes the complementary event ${\ensuremath{\mathit\Omega}\xspace}{\setminus}B$. (Readers comparing this definition with Reichenbach’s original beware: his notation is modified here by switching the role of $B$ and $C$. To denote the middle event in the fork, he used $C$, perhaps as mnemonic for ‘common cause’.) Implicit in this definition is the asumption $$0<{\ensuremath{P}\xspace}(B)<1\, ,\label{reich5}$$ which is needed to define the conditional probabilities in –.
A similar notion was introduced earlier in the context of sociology [@KenLaz50 Part I, Section 2], but the context of Reichenbach’s discourse was philosophy of science: conjunctive forks play a central role in his causal theory of time. In this role, they have attracted considerable attention: over one hundred publications, such as [@Bre77; @Sal80; @Sal84; @EllEri86; @CarJon91; @Dow92; @Spo94; @HofRedSza99; @Kor99], refer to them. Yet for all this interest, no one seems to have asked a fundamental question:
The purpose of our paper is to answer this question.
An additional stimulus to our work was a previous answer [@ChvWu12] to a similar question,
Here, *causal betweenness* is another ternary relation on sets of events in probability spaces, also introduced by Reichenbach[@Rei56 p. 190] in the context of his causal theory of time. (This relation is reviewed in Section \[S:btw\].) From this perspective, our paper may be seen as a companion to [@ChvWu12].
The main result
===============
Let us write $(A,B,C)_{{\ensuremath{P}\xspace}}$ to signify that $(A,B,C)$ is a conjunctive fork in a probability space $({\ensuremath{\mathit\Omega}\xspace},\mathcal F,{\ensuremath{P}\xspace})$ and let us say that events $A_i$ in this space, indexed by elements $i$ of a set $N$, *fork-represent* a ternary relation ${\ensuremath{\mathfrak r}\xspace}$ on $N$ if and only if $${\ensuremath{\mathfrak r}\xspace}=\{(i,j,k)\in N^3\colon (A_i,A_j,A_k)_{{\ensuremath{P}\xspace}}\,\}\,.$$ In order to characterize ternary relations on a finite ground set that are fork representable, we need a few definitions.
To begin, call a ternary relation [$\mathfrak r$]{} on a ground set $N$ a *forkness* if and only if it satisfies $$\begin{aligned}
(i,j,i)\in{\ensuremath{\mathfrak r}\xspace}\;\;&\Rightarrow\;\;(j,i,j)\in{\ensuremath{\mathfrak r}\xspace}\label{sym}\\
(i,j,i), (j,k,j)\in{\ensuremath{\mathfrak r}\xspace}\;\;&\Rightarrow\;\;(i,k,i)\in{\ensuremath{\mathfrak r}\xspace}\label{E:trans}\\
(i,k,j)\in{\ensuremath{\mathfrak r}\xspace}\;\;&\Rightarrow\;\;(j,k,i)\in{\ensuremath{\mathfrak r}\xspace}\label{flip}\\
(i,j,k)\in{\ensuremath{\mathfrak r}\xspace}\;\;&\Rightarrow\;\;(i,j,j), (j,k,k),(k,i,i)\in {\ensuremath{\mathfrak r}\xspace}\label{lower}\\
(i,k,j), (i,j,k)\in{\ensuremath{\mathfrak r}\xspace}\;\;&\Rightarrow\;\; (j,k,j)\in{\ensuremath{\mathfrak r}\xspace}\label{btw}
\end{aligned}$$ for all choices $i,j,k$ in $N$.
Given a forkness ${\ensuremath{\mathfrak r}\xspace}$, write $$V_{\ensuremath{\mathfrak r}\xspace}=\{i\in N\colon(i,i,i)\in{\ensuremath{\mathfrak r}\xspace}\}$$ and let [$\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{\mathfrak r}\xspace}}{\thicksim}}}$]{}denote the binary relation defined on $V_{\ensuremath{\mathfrak r}\xspace}$ by $$i{\ensuremath{\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{\mathfrak r}\xspace}}{\thicksim}}}}\xspace}j\;\;\Leftrightarrow\;\;(i,j,i)\in{\ensuremath{\mathfrak r}\xspace}\,.$$ This binary relation is reflexive by definition of $V_{\ensuremath{\mathfrak r}\xspace}$, it is symmetric by , and it is transitive by . In short, [$\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{\mathfrak r}\xspace}}{\thicksim}}}$]{}is an equivalence relation. Call a forkness [$\mathfrak r$]{}*regular* if, and only if, $$\label{E:regular}
(i,j,k)\in{\ensuremath{\mathfrak r}\xspace}, \; i{\ensuremath{\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{\mathfrak r}\xspace}}{\thicksim}}}}\xspace}i', \; j{\ensuremath{\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{\mathfrak r}\xspace}}{\thicksim}}}}\xspace}j', \; k{\ensuremath{\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{\mathfrak r}\xspace}}{\thicksim}}}}\xspace}k'
\;\Rightarrow\; (i',j',k')\in {\ensuremath{\mathfrak r}\xspace}\,.$$ The *quotient* of a a regular forkness [$\mathfrak r$]{}is the the ternary relation whose ground set is the set of equivalence classes of [$\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{\mathfrak r}\xspace}}{\thicksim}}}$]{} and which consists of all triples $(I,J,K)$ such that $(i,j,k)\in{\ensuremath{\mathfrak r}\xspace}$ for at least one $i$ in $I$, at least one $j$ in $J$, and at least one $k$ in $K$. (Equivalently, since [$\mathfrak r$]{}is regular, $(I,J,K)$ belongs to its quotient if and only if $(i,j,k)\in{\ensuremath{\mathfrak r}\xspace}$ for all $i$ in $I$, for all $j$ in $J$, and for all $k$ in $K$.)
Call a ternary relation [${\mathfrak q}$]{}*solvable* if and only if the linear system $$\begin{gathered}
\label{system}
x_{\{I,K\}}\!=\!x_{\{I,J\}}+x_{\{J,K\}}\\
\text{ for all $(I,J,K)$ in ${\ensuremath{{\mathfrak q}}\xspace}$ with pairwise distinct $I,J,K$}
\end{gathered}$$ has a solution with each $x_{\{I,J\}}$ positive.
\[T:ChMaZw\] A ternary relation on a finite ground set is fork representable if and only if it is a regular forkness and its quotient is solvable.
Theorem \[T:ChMaZw\] implies that fork-representability of a ternary relation ${\ensuremath{\mathfrak r}\xspace}$ on a finite ground set $N$ can be tested in time polynomial in ${\lvertN\rvert}$. More precisely, polynomial time suffices to test [$\mathfrak r$]{}for being a forkness, for testing its regularity, and for the construction of its quotient [${\mathfrak q}$]{}. Solvability of [${\mathfrak q}$]{}means solvability of a system of linear equations and linear inequalities, which can be tested in polynomial time by the breakthrough result of [@Kha79].
We prove the easier ‘only if’ part of Theorem \[T:ChMaZw\] in in Section \[S:onlyif\] and we prove the ‘if’ part in Section \[S:mainproof\]. In Section \[S:btw\], we comment on causal betweenness and its relationship to causal forks. In the final Section \[S:disc\], we discuss connections to previous work on patterns of conditional independence.
Proof of the ‘only if’ part {#S:onlyif}
===========================
Reichenbach’s definition restated
---------------------------------
Reichenbach’s definition of a conjunctive fork has a neat paraphrase in terms of random variables. To present it, let us first review a few standard definitions.
The *indicator function* ${\ensuremath{\mathds{1}}}_E$ of an event $E$ in a probability space is the random variable defined by ${\ensuremath{\mathds{1}}}_E(\omega)=1$ if $\omega\in E$ and ${\ensuremath{\mathds{1}}}_E(\omega)=0$ if $\omega\in {\overline{E}}$. Indicator functions ${\ensuremath{\mathds{1}}}_A$ and ${\ensuremath{\mathds{1}}}_C$ are said to be *conditionally independent given ${\ensuremath{\mathds{1}}}_B$,* in symbols ${\ensuremath{\mathds{1}}}_A{\ensuremath{\!\perp\!\!\!\!\perp\!}}{\ensuremath{\mathds{1}}}_C|{\ensuremath{\mathds{1}}}_B$, if and only if events $A$, $B$, $C$ satisfy and . The *covariance* of ${\ensuremath{\mathds{1}}}_A$ and ${\ensuremath{\mathds{1}}}_B$, denoted here as ${\ensuremath{\mathrm{cov}({A},{B})}\xspace}$, is defined by $${\ensuremath{\mathrm{cov}({A},{B})}\xspace}={\ensuremath{P}\xspace}(AB)-{\ensuremath{P}\xspace}(A){\ensuremath{P}\xspace}(B).$$ Since means that ${\ensuremath{\mathrm{cov}({A},{B})}\xspace}>0$ and means that ${\ensuremath{\mathrm{cov}({B},{C})}\xspace}>0$, we conclude that $$\begin{gathered}
(A,B,C)_{\ensuremath{P}\xspace}\Leftrightarrow
{\ensuremath{\mathds{1}}}_A{\ensuremath{\!\perp\!\!\!\!\perp\!}}{\ensuremath{\mathds{1}}}_C|{\ensuremath{\mathds{1}}}_B \;\&\; {\ensuremath{\mathrm{cov}({A},{B})}\xspace}\!\!>\!\!0 \;\&\; {\ensuremath{\mathrm{cov}({B},{C})}\xspace}\!\!>\!\!0. \label{fork}\end{gathered}$$
A couple of Reichenbach’s results
---------------------------------
Reichenbach [@Rei56 p. 160, equation (12)] noted that $$\begin{gathered}
\label{rr1}
{\ensuremath{\mathds{1}}}_A{\ensuremath{\!\perp\!\!\!\!\perp\!}}{\ensuremath{\mathds{1}}}_C|{\ensuremath{\mathds{1}}}_B \;\Rightarrow\;\\
{\ensuremath{\mathrm{cov}({A},{C})}\xspace}={\ensuremath{\mathrm{cov}({B},{B})}\xspace}\cdot({\ensuremath{P}\xspace}(A|B)-{\ensuremath{P}\xspace}(A|{\overline{B}}))\cdot({\ensuremath{P}\xspace}(C|B)-{\ensuremath{P}\xspace}(C|{\overline{B}}))\end{gathered}$$ and so [@Rei56 p. 158, inequality (1)] $$\label{covac}
(A,B,C)_{\ensuremath{P}\xspace}\Rightarrow {\ensuremath{\mathrm{cov}({A},{C})}\xspace}>0.$$ Implication was his reason for calling the fork ‘conjunctive’: it is “a fork which makes the conjunction of the two events more frequent than it would be for independent events” [@Rei56 p. 159].
Since $$\begin{gathered}
{\ensuremath{\mathrm{cov}({B},{B})}\xspace}({\ensuremath{P}\xspace}(A|B)-{\ensuremath{P}\xspace}(A|{\overline{B}}))=\\ {\ensuremath{P}\xspace}(AB)(1-{\ensuremath{P}\xspace}(B))-{\ensuremath{P}\xspace}(A{\overline{B}}){\ensuremath{P}\xspace}(B)={\ensuremath{\mathrm{cov}({A},{B})}\xspace}\end{gathered}$$ and, similarly, ${\ensuremath{\mathrm{cov}({B},{B})}\xspace}({\ensuremath{P}\xspace}(C|B)-{\ensuremath{P}\xspace}(C|{\overline{B}}))={\ensuremath{\mathrm{cov}({C},{B})}\xspace}$, Reichenbach’s implication can be stated as $$\label{fero}
{\ensuremath{\mathds{1}}}_A{\ensuremath{\!\perp\!\!\!\!\perp\!}}{\ensuremath{\mathds{1}}}_C|{\ensuremath{\mathds{1}}}_B \;\Rightarrow\;{\ensuremath{\mathrm{cov}({A},{C})}\xspace}\cdot{\ensuremath{\mathrm{cov}({B},{B})}\xspace}={\ensuremath{\mathrm{cov}({A},{B})}\xspace}\cdot{\ensuremath{\mathrm{cov}({B},{C})}\xspace}.$$
The idea of the proof
---------------------
An event $E$ is called *[$P$]{}-nontrivial* if and only if $0<{\ensuremath{P}\xspace}(E)<1$, which is equivalent to ${\ensuremath{\mathrm{cov}({E},{E})}\xspace}>0$. When $E,F$ are [$P$]{}-nontrivial events, the *correlation* of their indicator functions, denoted here as ${\ensuremath{\mathrm{corr}}\xspace}(E,F)$, is defined by $${\ensuremath{\mathrm{corr}}\xspace}(E,F)\;=\; \frac {{\ensuremath{\mathrm{cov}({E},{F})}\xspace}}{{\ensuremath{\mathrm{cov}({E},{E})}\xspace}^{1/2}{\ensuremath{\mathrm{cov}({F},{F})}\xspace}^{1/2}}\, .$$ In these terms, reads $$\label{corident}
{\ensuremath{\mathds{1}}}_A{\ensuremath{\!\perp\!\!\!\!\perp\!}}{\ensuremath{\mathds{1}}}_C|{\ensuremath{\mathds{1}}}_B \;\Rightarrow\;{\ensuremath{\mathrm{corr}}\xspace}(A,C)={\ensuremath{\mathrm{corr}}\xspace}(A,B)\cdot {\ensuremath{\mathrm{corr}}\xspace}(B,C).$$
The strict inequalities , , imply that $$\begin{gathered}
\label{nontriv}
\text{in every conjunctive fork $(A,B,C)$,}\\ \text{all three events $A$, $B$, $C$ are {\ensuremath{P}\xspace}-nontrivial,}\end{gathered}$$ and so ${\ensuremath{\mathrm{corr}}\xspace}(A,B)$, ${\ensuremath{\mathrm{corr}}\xspace}(A,C)$, ${\ensuremath{\mathrm{corr}}\xspace}(B,C)$ are well defined; guarantees that ${\ensuremath{\mathrm{corr}}\xspace}(A,B)>0$, ${\ensuremath{\mathrm{corr}}\xspace}(B, C)>0$ and guarantees that ${\ensuremath{\mathrm{corr}}\xspace}(A,C)>0$.
Fact guarantees that the system $$x_{\{A,C\}}\!=\!x_{\{A,B\}}+x_{\{B,C\}}
\text{ for all conjunctive forks $(A,B,C)$}$$ can be solved by setting $x_{\{E,F\}} = -\ln {\ensuremath{\mathrm{corr}}\xspace}(E,F)$. This observation goes a long way toward proving the ‘only if’ part of Theorem \[T:ChMaZw\], but it does not quite get there: For instance, if $(A,B,C)$ is a conjunctive fork and $A{\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{P}\xspace}}{=}}}}B\,$, then ${\ensuremath{\mathrm{corr}}\xspace}(A,B)=1$, and so $x_{\{A,B\}}=0$, but the ‘only if’ part of the theorem requires $x_{\{A,B\}}>0$. To get around such obstacles, we deal with the quotient of the ternary relation made from conjunctive forks.
Other preliminaries
-------------------
Events $E$ and $F$ are said to be [$P$]{}-equal, in symbols $E{\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{P}\xspace}}{=}}}}F$, if and only if ${\ensuremath{P}\xspace}(E{\vartriangle}F)=0$. We claim that $$\begin{aligned}
& (A,B,A)_{{\ensuremath{P}\xspace}} \text{ if and only if $A$ is {\ensuremath{P}\xspace}-nontrivial and $A{\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{P}\xspace}}{=}}}}B\,$,}\label{iji} \\
& \text{if $(A,B,C)_{{\ensuremath{P}\xspace}}$ and $(A,C,B)_{{\ensuremath{P}\xspace}}$, then $B{\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{P}\xspace}}{=}}}}C$.}\label{fbtw}\end{aligned}$$ To justify claim , note that $(A,B,A)_{{\ensuremath{P}\xspace}}$ means the conjunction of ${\ensuremath{\mathrm{cov}({A},{B})}\xspace}>0$ and at least one of ${\ensuremath{P}\xspace}(A)=0$, ${\ensuremath{P}\xspace}(A)=1$, $A{\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{P}\xspace}}{=}}}}{\overline{B}}$, $A{\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{P}\xspace}}{=}}}}B$; of the four equalities, only the last one is compatible with ${\ensuremath{\mathrm{cov}({A},{B})}\xspace}>0$. To justify claim , note that ${\ensuremath{\mathrm{cov}({B},{B})}\xspace}-{\ensuremath{\mathrm{cov}({B},{C})}\xspace}={\ensuremath{P}\xspace}(B){\ensuremath{P}\xspace}({\overline{B}}C) +{\ensuremath{P}\xspace}({\overline{B}}){\ensuremath{P}\xspace}(B{\overline{C}})$; when $B$ is ${\ensuremath{P}\xspace}$-nontrivial, the right-hand side vanishes if and only if ${\ensuremath{P}\xspace}(B{\vartriangle}C)=0$; it follows that $$\begin{aligned}
& \text{if $(A,B,C)_{{\ensuremath{P}\xspace}}$, then ${\ensuremath{\mathrm{cov}({B},{B})}\xspace}{\geqslant}{\ensuremath{\mathrm{cov}({B},{C})}\xspace}$}\\
& \text{with equality if and only if $B{\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{P}\xspace}}{=}}}}C\,$;}\end{aligned}$$ by , this implies that $$\begin{aligned}
& \text{if $(A,B,C)_{{\ensuremath{P}\xspace}}$, then ${\ensuremath{\mathrm{cov}({A},{B})}\xspace}{\geqslant}{\ensuremath{\mathrm{cov}({A},{C})}\xspace}$}\\
& \text{with equality if and only if $B{\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{P}\xspace}}{=}}}}C\,$,}\end{aligned}$$ which in turn implies .
The proof
---------
\[V1\] A ternary relation is fork representable only if it is a regular forkness.
Consider a fork-representable ternary relation [$\mathfrak r$]{}on a ground set $N$. Proving the lemma means verifying that [$\mathfrak r$]{} has properties , , , , , . Since [$\mathfrak r$]{}is fork representable, there are events $A_i$ in in some probability space $({\ensuremath{\mathit\Omega}\xspace},\mathcal F,{\ensuremath{P}\xspace})$, with $i$ ranging over $N$, such that $$(i,j,k)\in {\ensuremath{\mathfrak r}\xspace}\;\Leftrightarrow\;
(A_i,A_j,A_k)_{{\ensuremath{P}\xspace}}\, .$$ Properties and , $$\begin{aligned}
& (i,j,i)\in{\ensuremath{\mathfrak r}\xspace}\;\;\Rightarrow\;\;(j,i,j)\in{\ensuremath{\mathfrak r}\xspace}\, ,\\
& (i,j,i), (j,k,j)\in{\ensuremath{\mathfrak r}\xspace}\;\;\Rightarrow\;\;(i,k,i)\in{\ensuremath{\mathfrak r}\xspace}\, , \end{aligned}$$ follow from . Property , $$\begin{aligned}
& (i,k,j)\in{\ensuremath{\mathfrak r}\xspace}\;\;\Rightarrow\;\;(j,k,i)\in{\ensuremath{\mathfrak r}\xspace}\, ,\end{aligned}$$ is implicit in the definition of $(A_i,A_j,A_k)_{{\ensuremath{P}\xspace}}$. Property , $$\begin{aligned}
& (i,j,k)\in{\ensuremath{\mathfrak r}\xspace}\;\;\Rightarrow\;\;(i,j,j), (j,k,k),(k,i,i)\in {\ensuremath{\mathfrak r}\xspace}\, , \end{aligned}$$ follows from . Property , $$\begin{aligned}
& (i,j,k)\in{\ensuremath{\mathfrak r}\xspace}\;\;\Rightarrow\;\;(i,j,j), (j,k,k),(k,i,i)\in {\ensuremath{\mathfrak r}\xspace}\, , \end{aligned}$$ follows from and . Property , $$(i,j,k)\in{\ensuremath{\mathfrak r}\xspace}, \; i{\ensuremath{\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{\mathfrak r}\xspace}}{\thicksim}}}}\xspace}i', \; j{\ensuremath{\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{\mathfrak r}\xspace}}{\thicksim}}}}\xspace}j', \; k{\ensuremath{\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{\mathfrak r}\xspace}}{\thicksim}}}}\xspace}k'
\;\Rightarrow\; (i',j',k')\in {\ensuremath{\mathfrak r}\xspace}\, ,$$ follows from alone.
\[V2\] A regular forkness is fork representable only if its quotient is solvable.
Consider a fork-representable regular forkness [$\mathfrak r$]{}on a ground set $N$. Since [$\mathfrak r$]{}is fork representable, there are events $A_i$ in in some probability space $({\ensuremath{\mathit\Omega}\xspace},\mathcal F,{\ensuremath{P}\xspace})$, with $i$ ranging over $N$, such that $$(i,j,k)\in {\ensuremath{\mathfrak r}\xspace}\;\Leftrightarrow\;
(A_i,A_j,A_k)_{{\ensuremath{P}\xspace}}\, .$$ With [${\mathfrak q}$]{}standing for the quotient of [$\mathfrak r$]{}, proving the lemma means finding positive numbers $x_{\{I,J\}}$ such that $$\begin{gathered}
\label{solution}
x_{\{I,K\}}\!=\!x_{\{I,J\}}+x_{\{J,K\}}\\
\text{ for all $(I,J,K)$ in ${\ensuremath{{\mathfrak q}}\xspace}$ with pairwise distinct $I,J,K$.}\end{gathered}$$ We claim that this can be done by first choosing an element $r(I)$ from each equivalence class $I$ of [$\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{\mathfrak r}\xspace}}{\thicksim}}}$]{}and then setting $$x_{\{I,J\}} = -\ln {\ensuremath{\mathrm{corr}}\xspace}({A_{r(I)}},{A_{r(J)}})$$ for every pair $I,J$ of distinct equivalence classes that appear together in a triple in [${\mathfrak q}$]{}. (By , the right hand side depends only on $I$ and $J$ rather than the choice of $r(I)$ and $r(J)$.)
To justify this claim, note first that is satisfied by virtue of . A special case of the [*covariance inequality*]{} guarantees that every pair $A,B$ of events satisfies $$({\ensuremath{\mathrm{cov}({A},{B})}\xspace})^2 {\leqslant}{\ensuremath{\mathrm{cov}({A},{A})}\xspace}\cdot {\ensuremath{\mathrm{cov}({B},{B})}\xspace}$$ and that the two sides are equal if and only if $A{\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{P}\xspace}}{=}}}}B$ or $A{\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{P}\xspace}}{=}}}}{\overline{B}}$ or ${\ensuremath{P}\xspace}(A)=0$ or ${\ensuremath{P}\xspace}(A)=1$ or ${\ensuremath{P}\xspace}(B)=0$ or ${\ensuremath{P}\xspace}(B)=1$. If equivalence classes $I,J$ of [$\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{\mathfrak r}\xspace}}{\thicksim}}}$]{}are distinct, then $A_{r(I)}\not{\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{P}\xspace}}{=}}}}A_{r(J)}$; if, in addition, $I$ and $J$ appear together in a triple in [${\mathfrak q}$]{}, then $A_{r(I)}\not{\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{P}\xspace}}{=}}}}{\overline{A}}_{r(J)}$ (since ${\ensuremath{\mathrm{corr}}\xspace}({A_{r(I)}},{A_{r(J)}})>0$ by and ) and $0<{\ensuremath{P}\xspace}(A_{r(I)})<1$, $0<{\ensuremath{P}\xspace}(A_{r(J)})<1$ (by ). In this case, $${\ensuremath{\mathrm{cov}({A_{r(I)}},{A_{r(J)}})}\xspace}^2<{\ensuremath{\mathrm{cov}({A_{r(I)}},{A_{r(I)}})}\xspace}\cdot{\ensuremath{\mathrm{cov}({A_{r(J)}},{A_{r(J)}})}\xspace},$$ and so $x_{\{I,J\}}>0$.
Proof of the ‘if’ part {#S:mainproof}
======================
\[V3\] If the quotient of a regular forkness on a finite ground set is solvable, then it is fork representable.
Given the quotient [${\mathfrak q}$]{}of a regular forkness on a finite ground set along with positive numbers $x_{\{I,J\}}$ such that $$\begin{gathered}
x_{\{I,K\}}\!=\!x_{\{I,J\}}+x_{\{J,K\}}\\
\text{ for all $(I,J,K)$ in ${\ensuremath{{\mathfrak q}}\xspace}$ with pairwise distinct $I,J,K$,}\label{solution++}\end{gathered}$$ we have to construct a probability space $({\ensuremath{\mathit\Omega}\xspace},\mathcal F,{\ensuremath{P}\xspace})$ and events $A_I$ in this space, indexed by elements $I$ of the ground set $C$ of [${\mathfrak q}$]{}, such that $$\label{qeq}
{\ensuremath{{\mathfrak q}}\xspace}=\{(I,J,K)\in C^3\colon (A_I,A_J,A_K)_{{\ensuremath{P}\xspace}}\,\}\,.$$
For this purpose, we let ${\ensuremath{\mathit\Omega}\xspace}$ be the power set $2^C$ of $C$, we let $\mathcal F$ be the power set $2^{{\ensuremath{\mathit\Omega}\xspace}}$ of ${\ensuremath{\mathit\Omega}\xspace}$, and we set $$A_I=\{\omega\in{\ensuremath{\mathit\Omega}\xspace}: I\in\omega\}.$$ Now let us construct the probability measure ${\ensuremath{P}\xspace}$. Set $n={\lvertC\rvert}$. Given a subset $L$ of $C$, consider the function $\chi_L\colon {\ensuremath{\mathit\Omega}\xspace}\rightarrow \{+1, -1\}$ defined by $$\chi_L(\omega)= (-1)^{{\lvert\,\omega\,\cap\, L\,\rvert}}\,.$$ Let $E_{\ensuremath{{\mathfrak q}}\xspace}$ stand for the family of all two-element subsets $\{I,J\}$ of $C$ such that $I$ and $J$ appear together in a triple in [${\mathfrak q}$]{}. Let $M_{\ensuremath{{\mathfrak q}}\xspace}$ stand for the family of all three-element subsets $\{I,J,K\}$ of $C$ such that $\{I,J\}$, $\{J,K\}$, and $\{K,I\}$ belong to $E_{\ensuremath{{\mathfrak q}}\xspace}$ and no triple in ${\ensuremath{{\mathfrak q}}\xspace}$ is formed by all three $I,J,K$. Finally, for positive numbers $\gamma$ and ${\ensuremath{\varepsilon}\xspace}$ that are sufficiently small in a sense to be specified shortly, define ${\ensuremath{P}\xspace}\colon {\ensuremath{\mathit\Omega}\xspace}\rightarrow {\bf R}$ by $${\ensuremath{P}\xspace}(\omega)=2^{-n}\Big[1+
\textstyle{\sum_{\{I,J\}\in E_{\ensuremath{{\mathfrak q}}\xspace}}\;\chi_{\{I,J\}}(\omega)\,\gamma^{x\{I,J\}}
+{\ensuremath{\varepsilon}\xspace}\sum_{\{I,J,K\}\in M_{\ensuremath{{\mathfrak q}}\xspace}}}\;\chi_{\{I,J,K\}}(\omega)\Big].$$ (In exponents, we write $x\{I,J\}$ in place of $x_{\{I,J\}}$.) Here, readers with background in harmonic analysis will have recognized the Fourier-Stieltjes transform on the group $\Z_2$ and the characters $\chi_L$.
When $L$ is nonempty, $\chi_L$ takes each of the values $\pm1$ on the same number its arguments $\omega$, and so $\sum_{\omega\in{\mathit\Omega}}\chi_L(\omega)=0$, which implies that $\sum_{\omega\in{\mathit\Omega}}P(\omega)=1$. If $\gamma<n^{-2/x\{I,J\}}$ for all $\{I,J\}\in E_{\ensuremath{{\mathfrak q}}\xspace}$ and ${\ensuremath{\varepsilon}\xspace}<n^{-3}$, then $$2^{n}{\ensuremath{P}\xspace}(\omega){\geqslant}1-{\lvertE_{\ensuremath{{\mathfrak q}}\xspace}\rvert}\,n^{-2}-{\lvertM_{\ensuremath{{\mathfrak q}}\xspace}\rvert}\,n^{-3}>0\,,$$ so that [$P$]{}is a probability measure, positive on the elementary events.
The bulk of the proof consists of verifying that this construction satisfies . We may assume that $C\ne\emptyset$ (else ${\ensuremath{\mathit\Omega}\xspace}$ is a singleton, and so holds). Given a subset $S$ of $C$, write $$A_S \;=\; \{\omega\in{\ensuremath{\mathit\Omega}\xspace}: S\subseteq\omega\}.$$ Since $$\textstyle{\sum_{\omega\in A_S}\chi_L(\omega)}=
\begin{cases}
\;(-1)^{|L|}\,2^{n-{\lvertS\rvert}}& \quad\text{if $L\subseteq S$},\\
\;0 & \quad\text{otherwise},
\end{cases}$$ we have $$\begin{gathered}
\label{master}
2^{{\lvertS\rvert}}{\ensuremath{P}\xspace}(A_S)= \\
1+ \textstyle{\sum_{\{I,J\}\in E_{\ensuremath{{\mathfrak q}}\xspace},\:I,J\in S}\,\gamma^{x\{I,J\}}}
-{\ensuremath{\varepsilon}\xspace}\,\big|\{\{I,J,K\}\in M_{\ensuremath{{\mathfrak q}}\xspace}\colon I,J,K\in S\}\big|\,.
\end{gathered}$$ In particular, formula yields for all $I$ in $C$ $${\ensuremath{P}\xspace}(A_I)=\frac12$$ and it yields for all choices of distinct $I,J$ in $C$ $$\label{f2}
{\ensuremath{P}\xspace}(A_{\{I,J\}})= \frac14+
\begin{cases}
\;\frac14\gamma^{\,x\{I,J\}} & \quad\text{if $\{I,J\}\in E_{\ensuremath{{\mathfrak q}}\xspace}$},\\
\;0\, & \quad\text{otherwise.}
\end{cases}$$ It follows that $$\label{cvrgamma}
{\ensuremath{\mathrm{cov}({A_I},{A_J})}\xspace}=
\begin{cases}
\;\frac14\;\gamma^{\,x(\{I,J\})} & \text{if } \{I,J\}\in E_{\ensuremath{{\mathfrak q}}\xspace},\\
\;0 & \text{otherwise.}
\end{cases}$$ For future reference, note also that, by definition, $$\begin{aligned}
\label{qfork}
&\text{{\ensuremath{{\mathfrak q}}\xspace}is a forkness such that $(I,I,I)\in{\ensuremath{{\mathfrak q}}\xspace}$ for all $I$ in $C$}\\
&\text{and such that $(I,J,I)\not\in{\ensuremath{{\mathfrak q}}\xspace}$ whenever $I\ne J$.}\nonumber\end{aligned}$$ and that $$\begin{aligned}
\label{ijj}
& \{I,J\}\in E_{\ensuremath{{\mathfrak q}}\xspace}\;\Leftrightarrow\; (I,J,J) \in{\ensuremath{{\mathfrak q}}\xspace}.\end{aligned}$$ (here, implication $\Rightarrow$ follows from properties and of forkness and implication $\Leftarrow$ follows straight from the definition of $E_{\ensuremath{{\mathfrak q}}\xspace}$).
Now we are ready to verify . Given a triple $(I,J,K)$ in $C^3$, we have to show that $$\label{verq}
(A_I,A_J,A_K)_{{\ensuremath{P}\xspace}} \Leftrightarrow (I,J,K)\in{\ensuremath{{\mathfrak q}}\xspace}\, .$$
[Case 1:]{} $I=J=K$.\
Since $C\ne\emptyset$, all $I$ in $C$ are ${\ensuremath{P}\xspace}$-nontrivial, and so we have $(A_I,A_I,A_I)_{{\ensuremath{P}\xspace}}$. By , we have $(I,I,I)\in{\ensuremath{{\mathfrak q}}\xspace}$.
[Case 2:]{} $I\ne J$, $K=I$.\
Here, $A_I\not{\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{P}\xspace}}{=}}}}A_J$, and so implies that $(A_I,A_J,A_I)$ is not a conjunctive fork. By , we have $(I,J,I)\not\in{\ensuremath{{\mathfrak q}}\xspace}$.
[Case 3:]{} $I\ne J$, $K=J$.\
If $\{I,J\}\in E_{\ensuremath{{\mathfrak q}}\xspace}$, then guarantees ${\ensuremath{\mathrm{cov}({A_I},{A_J})}\xspace}>0$, which implies $(A_I,A_J,A_J)_{{\ensuremath{P}\xspace}}$. By , we have $(I,J,J)\in{\ensuremath{{\mathfrak q}}\xspace}$.
If $\{I,J\}\not\in E_{\ensuremath{{\mathfrak q}}\xspace}$, then guarantees ${\ensuremath{\mathrm{cov}({A_I},{A_J})}\xspace}=0$, and so $(A_I,A_J,A_J)$ is not a conjunctive fork. By , we have $(I,J,J)\not\in{\ensuremath{{\mathfrak q}}\xspace}$.
[Case 4:]{} $I= J$, $K\ne J$.\
This case is reduced to [Case 3]{} by the flip $I\leftrightarrow K$, which preserves both sides of .
[Case 5:]{} $I,J,K$ [*are pairwise distinct and at least one of $\{I,J\}$, $\{J,K\}$, $\{K,I\}$ does not belong to $E_{\ensuremath{{\mathfrak q}}\xspace}$.*]{}\
By , at least one of the covariances ${\ensuremath{\mathrm{cov}({A_I},{A_J})}\xspace}$, ${\ensuremath{\mathrm{cov}({A_J},{A_K})}\xspace}$, ${\ensuremath{\mathrm{cov}({A_K},{A_I})}\xspace}$ vanishes, and so , guarantee that $(A_I,A_J,A_K)$ is not a conjunctive fork. By definition of $E_{\ensuremath{{\mathfrak q}}\xspace}$, we have $(I,J,K)\not\in{\ensuremath{{\mathfrak q}}\xspace}$.
[Case 6:]{} $I,J,K$ [*are pairwise distinct and all of $\{I,J\}$, $\{J,K\}$, $\{K,I\}$ belong to $E_{\ensuremath{{\mathfrak q}}\xspace}$.*]{}\
By , all of ${\ensuremath{\mathrm{cov}({A_I},{A_J})}\xspace}$, ${\ensuremath{\mathrm{cov}({A_J},{A_K})}\xspace}$, ${\ensuremath{\mathrm{cov}({A_K},{A_I})}\xspace}$ are positive. Now implies that $(A_I,A_J,A_K)_{{\ensuremath{P}\xspace}}$ is equivalent to ${\ensuremath{\mathds{1}}}_{A_I}{\ensuremath{\!\perp\!\!\!\!\perp\!}}{\ensuremath{\mathds{1}}}_{A_K}|{\ensuremath{\mathds{1}}}_{A_J}$, which means the conjunction of $$\begin{aligned}
\tfrac12{\ensuremath{P}\xspace}(A_{\{I,J,K\}})&={\ensuremath{P}\xspace}(A_{\{I,J\}}){\ensuremath{P}\xspace}(A_{\{J,K\}})\,,\\
\tfrac12\big[{\ensuremath{P}\xspace}(A_{\{I,K\}})-{\ensuremath{P}\xspace}(A_{\{I,J,K\}})\big]&=
\big[\tfrac12-{\ensuremath{P}\xspace}(A_{\{I,J\}})\big]\cdot
\big[\tfrac12-{\ensuremath{P}\xspace}(A_{\{J,K\}})\big]\,.
\end{aligned}$$ Substitution from converts these two equalities to $$\begin{aligned}
\label{E:vind}
8{\ensuremath{P}\xspace}(A_{\{I,J,K\}})=&
1\!+\!\gamma^{\,x\{I,J\}}\!+\!\gamma^{\,x\{J,K\}}
\!+\!\gamma^{\,x\{I,J\}+x\{J,K\}}\\
8{\ensuremath{P}\xspace}(A_{\{I,J,K\}})=&
1\!+\!\gamma^{\,x\{I,J\}}\!+\!\gamma^{\,x\{J,K\}}
\!-\!\gamma^{\,x\{I,J\}+x\{J,K\}}\!+\!2\gamma^{\,x\{I,K\}}.
\label{E:vind2}
\end{aligned}$$ Conjunction of and is equivalent to the conjunction of and $$\label{E:vind3}
x_{\{I,J\}}+x_{\{J,K\}}=x_{\{I,K\}}\, .$$ To summarize, $(A_I,A_J,A_K)_{{\ensuremath{P}\xspace}}$ is equivalent to the conjunction of and .
[Subcase 6.1:]{} $\{I,J,K\}\in M_{\ensuremath{{\mathfrak q}}\xspace}$.\
In this subcase, formula yields $$8\,{\ensuremath{P}\xspace}(A_{\{I,J,K\}})=1+ \gamma^{\,x\{I,J\}}+ \gamma^{\,x\{J,K\}}+ \gamma^{\,x\{I,K\}}-{\ensuremath{\varepsilon}\xspace}\, ,$$ which reduces to $\gamma^{\,x\{I,K\}}-{\ensuremath{\varepsilon}\xspace}=\gamma^{\,x\{I,J\}+x\{J,K\}}$. This is is inconsistent with , and so $(A_I,A_J,A_K)$ is not a conjunctive fork. By definition of $M_{\ensuremath{{\mathfrak q}}\xspace}$, we have $(I,J,K)\not\in~{\ensuremath{{\mathfrak q}}\xspace}$.
[Subcase 6.2:]{} $\{I,J,K\}\not\in M_{\ensuremath{{\mathfrak q}}\xspace}$.\
In this subcase, formula yields $$8\,{\ensuremath{P}\xspace}(A_{\{I,J,K\}})=1+ \gamma^{\,x\{I,J\}}+ \gamma^{\,x\{J,K\}}+ \gamma^{\,x\{I,K\}} ,$$ which reduces to , and so $(A_I,A_J,A_K)_{{\ensuremath{P}\xspace}}$ is equivalent to alone. Now completing the proof means verifying that $$x_{\{I,J\}}+x_{\{J,K\}}=x_{\{I,K\}} \;\;\Leftrightarrow\;\; (I,J,K)\in{\ensuremath{{\mathfrak q}}\xspace}.$$ Implication $\Leftarrow$ is . To prove the reverse implication, note first that by along with $x_{\{I,J\}}>0$ and $x_{\{J,K\}}>0$, we have $$\begin{aligned}
x_{\{I,J\}}+x_{\{J,K\}}=x_{\{I,K\}} \;\;\Rightarrow\;\; x_{\{J,K\}}\!<\!x_{\{I,K\}} \;\;\Rightarrow\;\; (J,I,K)\not\in {\ensuremath{{\mathfrak q}}\xspace}\, ,\label{jik}\\
x_{\{I,J\}}+x_{\{J,K\}}=x_{\{I,K\}} \;\;\Rightarrow\;\; x_{\{I,J\}}\!<\!x_{\{I,K\}} \;\;\Rightarrow\;\; (I,K,J)\not\in {\ensuremath{{\mathfrak q}}\xspace}\, .\label{ikj}\end{aligned}$$ By assumptions of this case and subcase, some triple in ${\ensuremath{{\mathfrak q}}\xspace}$ is formed by all three $I,J,K$ and so, since [${\mathfrak q}$]{}is a forkness, with [${\mathfrak q}$]{}in place of [$\mathfrak r$]{}guarantees that at least one of $(J,I,K)$, $(I,K,J)$, $(I,J,K)$ belongs to [${\mathfrak q}$]{}. If $x_{\{I,J\}}+x_{\{J,K\}}=x_{\{I,K\}}$, then and exclude the first two options, and so we have $(I,J,K)\in{\ensuremath{{\mathfrak q}}\xspace}$.
\[V4\] If a regular forkness has a fork-representable quotient, then it is fork representable.
Given a regular forkness [$\mathfrak r$]{}, a probability space $({\ensuremath{\mathit\Omega}\xspace}^0,\mathcal F^0,{\ensuremath{P}\xspace}^0)$, and events $A_I^0$ in this space, indexed by elements $I$ of the ground set $C$ of the quotient [${\mathfrak q}$]{}of [$\mathfrak r$]{}, such that $${\ensuremath{{\mathfrak q}}\xspace}=\{(I,J,K)\in C^3\colon (A_I^0,A_J^0,A_K^0)_{{\ensuremath{P}\xspace}^0}\,\},$$ we have to construct a probability space $({\ensuremath{\mathit\Omega}\xspace},\mathcal F,{\ensuremath{P}\xspace})$ and events $A_i$ in this space, indexed by elements $i$ of the ground set $N$ of [$\mathfrak r$]{}, such that $$\label{req}
{\ensuremath{\mathfrak r}\xspace}=\{(i,j,k)\in N^3\colon (A_i,A_j,A_k)_{{\ensuremath{P}\xspace}}\,\}\,.$$ For this purpose, we let ${\ensuremath{\mathit\Omega}\xspace}$ be the power set $2^N$ of $N$, we let $\mathcal F$ be the power set $2^{\ensuremath{\mathit\Omega}\xspace}$ of ${\ensuremath{\mathit\Omega}\xspace}$, and we set $$A_i=\{\omega\in{\ensuremath{\mathit\Omega}\xspace}: i\in\omega\}.$$ For each element $\omega$ of ${\ensuremath{\mathit\Omega}\xspace}$ such that every equivalence class $I$ of [$\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{\mathfrak r}\xspace}}{\thicksim}}}$]{}satisfies $I\subseteq\omega$ or $I\cap\omega=\emptyset$, define ${\ensuremath{P}\xspace}(\omega)={\ensuremath{P}\xspace}^0(\omega^0)$, where $\omega^0$ is the set of equivalence classes of [$\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{\mathfrak r}\xspace}}{\thicksim}}}$]{}contained in $\omega$. For all other elements $\omega$ of ${\ensuremath{\mathit\Omega}\xspace}$, define ${\ensuremath{P}\xspace}(\omega)=0$. Now verifying is a routine matter.
Causal betweenness {#S:btw}
==================
Reichenbach [@Rei56 p. 190] defined an event $B$ to be *causally between* events $A$ and $C$ if $$\begin{aligned}
1 > {\ensuremath{P}\xspace}(A|B) > {\ensuremath{P}\xspace}(A|C)> {\ensuremath{P}\xspace}(A) > 0\,,\\
1 > {\ensuremath{P}\xspace}(C|B) > {\ensuremath{P}\xspace}(C|A)> {\ensuremath{P}\xspace}(C) > 0\,,\\
{\ensuremath{P}\xspace}(C|AB) = {\ensuremath{P}\xspace}(C|B)\,.\qquad\qquad
\end{aligned}$$ Implicit in this definition is the assumption ${\ensuremath{P}\xspace}(B)>0$ that makes ${\ensuremath{P}\xspace}(A|B)$ and ${\ensuremath{P}\xspace}(C|B)$ meaningful. In turn, ${\ensuremath{P}\xspace}(A|B)>0$ means ${\ensuremath{P}\xspace}(AB)>0$, which makes ${\ensuremath{P}\xspace}(C|AB)$ meaningful. If $B$ is causally between $A$ and $C$, then all three events are ${\ensuremath{P}\xspace}$-nontrivial and no two of them ${\ensuremath{P}\xspace}$-equal.
If $(A,B,C)$ is a conjunctive fork, then (contrary to the claim in [@Bre77 p. 179]) $B$ need not be causally between $A$ and $C$ even if no two of $A,B,C$ are ${\ensuremath{P}\xspace}$-equal: for example, if $$\begin{array}{llll}
{\ensuremath{P}\xspace}(A BC)=1/5, &{\ensuremath{P}\xspace}(A B{\overline{C}})=1/5,\\
{\ensuremath{P}\xspace}({\overline{A}}BC)=1/5, &{\ensuremath{P}\xspace}({\overline{A}}B{\overline{C}})=1/5,\\
{\ensuremath{P}\xspace}(A{\overline{B}}C)=0, &{\ensuremath{P}\xspace}(A{\overline{B}}{\overline{C}})=0,\\
{\ensuremath{P}\xspace}({\overline{A}}{\overline{B}}C)=0, &{\ensuremath{P}\xspace}({\overline{A}}{\overline{B}}{\overline{C}})=1/5,
\end{array}$$ then $(A,B,C)$ is a conjunctive fork and ${\ensuremath{P}\xspace}(A|B) = {\ensuremath{P}\xspace}(A|C)$.
If an event $B$ is causally between $A$ and $C$, then $(A,B,C)$ need not be a conjunctive fork: for example, if $$\begin{array}{llll}
{\ensuremath{P}\xspace}(A BC)=1/20, &{\ensuremath{P}\xspace}(A B{\overline{C}})=2/20,\\
{\ensuremath{P}\xspace}({\overline{A}}BC)=2/20, &{\ensuremath{P}\xspace}({\overline{A}}B{\overline{C}})=4/20,\\
{\ensuremath{P}\xspace}(A{\overline{B}}C)=0, &{\ensuremath{P}\xspace}(A{\overline{B}}{\overline{C}})=1/20,\\
{\ensuremath{P}\xspace}({\overline{A}}{\overline{B}}C)=1/20, &{\ensuremath{P}\xspace}({\overline{A}}{\overline{B}}{\overline{C}})=9/20,
\end{array}$$ then $B$ is causally between $A$ and $C$ and ${\ensuremath{P}\xspace}(AC|{\overline{B}}) \ne {\ensuremath{P}\xspace}(A|{\overline{B}}) {\ensuremath{P}\xspace}(C|{\overline{B}})$.
Following [@ChvWu12], we call a ternary relation ${\ensuremath{\mathfrak b}\xspace}$ on a finite ground set $N$ an [*abstract causal betweenness*]{} if, and only if, there are events $A_i$ with $i$ ranging over $N$ such that $$\label{cbtw}
{\ensuremath{\mathfrak b}\xspace}=\{(i,j,k)\in N^3\colon\;\text{$A_j$ is causally between $A_i$ and $A_k$}\}\,.$$ A natural question is which ternary relations ${\ensuremath{\mathfrak b}\xspace}$ form an abstract causal betweenness. This question was answered in [@ChvWu12 Theorem 1] in terms of the directed graph $G({\ensuremath{\mathfrak b}\xspace})$ whose vertices are all two-element subsets of $N$ and whose edges are all ordered pairs $(\{i,j\},\{i,k\})$ such that $(i,j,k)\in{\ensuremath{\mathfrak b}\xspace}$ with $i,j,k$ pairwise distinct: $$\begin{aligned}
&\text{a ternary relation {\ensuremath{\mathfrak b}\xspace}on a finite ground set} \label{ChvWu}\\
&\text{is an abstract causal betweenness if and only if}\nonumber \\
&\quad \bullet\; (i,j,k)\in{\ensuremath{\mathfrak b}\xspace}\;\Rightarrow\; i,j,k \text{ are pairwise distinct,}\nonumber\\
&\quad \bullet\; (i,j,k)\in{\ensuremath{\mathfrak b}\xspace}\;\Rightarrow\; (k,j,i)\in{\ensuremath{\mathfrak b}\xspace},\nonumber \\
&\quad \bullet\; \text{$G({\ensuremath{\mathfrak b}\xspace})$ contains no directed~cycle.}\nonumber\end{aligned}$$ (The third requirement implies that $(i,j,k)\in{\ensuremath{\mathfrak b}\xspace}\;\Rightarrow\; (i,k,j)\not\in{\ensuremath{\mathfrak b}\xspace}$: else $G({\ensuremath{\mathfrak b}\xspace})$ would contain the directed cycle $\{i,j\}\rightarrow\{i,k\}\rightarrow\{i,j\}$.)
An essential difference between abstract causal betweenness and fork-representable relations is that, on the one hand, every triple in an abstract causal betweenness consists of pairwise distinct elements and, on the other hand, a forkness includes with every triple $(i,j,k)$ most of triples formed by at most two of $i,j,k$. This difference notwithstanding, the two can be compared. The trick is to introduce, for every ternary relation [$\mathfrak r$]{}, the ternary relation ${\ensuremath{\mathfrak r}\xspace}^\sharp$ consisting of all triples in [$\mathfrak r$]{}that have pairwise distinct elements.
We claim that $$\begin{aligned}
\label{comp}
&\text{if {\ensuremath{\mathfrak r}\xspace}is a fork-representable relation on a finite ground set}\\
&\text{such that $(i,j,i)\not\in{\ensuremath{\mathfrak r}\xspace}$ whenever $i\ne j$,}\nonumber\\
&\text{then ${\ensuremath{\mathfrak r}\xspace}^\sharp$ is an abstract causal betweenness.}\nonumber\end{aligned}$$ To justify this claim, consider a fork-representable relation [$\mathfrak r$]{}on a finite set $N$ such that $(i,j,i)\not\in{\ensuremath{\mathfrak r}\xspace}$ whenever $i\ne j$. By Lemma \[V1\], [$\mathfrak r$]{}is a forkness; assumption $i\ne j\Rightarrow (i,j,i)\not\in{\ensuremath{\mathfrak r}\xspace}$ implies that [$\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{\mathfrak r}\xspace}}{\thicksim}}}$]{}is the identity relation, and so the quotient of [$\mathfrak r$]{}is isomorphic to [$\mathfrak r$]{}. Now Lemma \[V2\] guarantees that [$\mathfrak r$]{}is solvable: there are positive numbers $x_{\{i,j\}}$ such that $x_{\{i,k\}}=x_{\{i,j\}}+x_{\{j,k\}}$ for all $(i,j,k)$ in ${\ensuremath{\mathfrak r}\xspace}$ with pairwise distinct $i,j,k$. Since $x_{\{i,k\}}>x_{\{i,j\}}$ for every edge $(\{i,j\},\{i,k\})$ of $G({\ensuremath{\mathfrak r}\xspace}^\sharp)$, this directed graph is acyclic, and so guarantees that ${\ensuremath{\mathfrak r}\xspace}^\sharp$ is an abstract causal betweenness.
Assumption $i\ne j\Rightarrow (i,j,i)\not\in{\ensuremath{\mathfrak r}\xspace}$ cannot be dropped from : consider ${\ensuremath{\mathfrak r}\xspace}= N^3$. This [$\mathfrak r$]{}is fork representable (for instance, by ${\ensuremath{\mathit\Omega}\xspace}=\{x,y\}$, ${\ensuremath{P}\xspace}(x)={\ensuremath{P}\xspace}(y)=1/2$, and $A_i=\{x\}$ for all $i$ in $N$). Nevertheless, if ${\lvertN\rvert}{\geqslant}3$, then $G({\ensuremath{\mathfrak r}\xspace}^\sharp)$ contains cycles, and so ${\ensuremath{\mathfrak r}\xspace}^\sharp$ is not an abstract causal betweenness.
The converse of , $$\begin{aligned}
&\text{if ${\ensuremath{\mathfrak r}\xspace}^\sharp$ is an abstract causal betweenness}\phantom{xxxxxxxxxxxxxxxxxx}\\
&\text{then {\ensuremath{\mathfrak r}\xspace}is a fork-representable relation}\\
&\text{such that $(i,j,i)\not\in{\ensuremath{\mathfrak r}\xspace}$ whenever $i\ne j$,}\end{aligned}$$ is false. Even its weaker version, $$\begin{aligned}
&\text{if {\ensuremath{\mathfrak r}\xspace}is a regular forkness}\phantom{xxxxxxxxxxxxxxxxxxxxxxxxxxxxi}\\
&\text{such that ${\ensuremath{\mathfrak r}\xspace}^\sharp$ is an abstract causal betweenness,}\\
&\text{then {\ensuremath{\mathfrak r}\xspace}is a fork-representable relation,}\end{aligned}$$ is false: consider the smallest forkness [$\mathfrak r$]{}on $\{1,2,3,4\}$ that contains the relation $$\begin{aligned}
\{&(1,3,2), (2,3,4), (3,1,4), (1,4,2), \label{example}\\
&(2,3,1), (4,3,2), (4,1,3), (2,4,1)\}. \nonumber\end{aligned}$$ Minimality of [$\mathfrak r$]{}implies that $(i,j,i)\not\in{\ensuremath{\mathfrak r}\xspace}$ whenever $i\ne j$; it follows that [$\mathrel{\smash{\overset{\scriptscriptstyle{\ensuremath{\mathfrak r}\xspace}}{\thicksim}}}$]{}is the identity relation, and so [$\mathfrak r$]{} is a regular forkness. Graph $G({\ensuremath{\mathfrak r}\xspace}^\sharp)$ is acyclic
(40,29)(0,7) (10,30)[(0,-1)[6]{}]{}(10,30)[(0,-1)[10]{}]{} (20,30)[(-1,-1)[7]{}]{} (20,30)[(-1,-1)[10]{}]{} (10,30)[(1,-1)[14]{}]{}(10,30)[(1,-1)[20]{}]{} (30,30)[(-1,-1)[14]{}]{} (30,30)[(-1,-1)[20]{}]{} (10,20)[(0,-1)[6]{}]{}(10,20)[(0,-1)[10]{}]{} (20,30)[(1,-2)[6]{}]{} (20,30)[(1,-2)[10]{}]{} (10,10)[(1,0)[11]{}]{}(10,10)[(1,-0)[20]{}]{} (30,30)[(0,-1)[12]{}]{} (30,30)[(0,-1)[20]{}]{} (10,20)[(0,0)]{} (10,10)[(0,0)]{} (30,10)[(0,0)]{} (10,30)[(0,0)]{} (20,30)[(0,0)]{} (30,30)[(0,0)]{}
and so guarantees that ${\ensuremath{\mathfrak r}\xspace}^\sharp$ is an abstract causal betweenness. By Lemma \[V2\], [$\mathfrak r$]{}is not fork representable: here, system is isomorphic to $$\begin{aligned}
x_{\{1,2\}}=x_{\{1,3\}}+x_{\{2,3\}} \\
x_{\{2,4\}}=x_{\{2,3\}}+x_{\{3,4\}} \\
x_{\{3,4\}}=x_{\{1,3\}}+x_{\{1,4\}} \\
x_{\{1,2\}}=x_{\{1,4\}}+x_{\{2,4\}}
\end{aligned}$$ and this system has no solution with $x_{\{1,4\}}>0$ as the linear combination of its four equations with multipliers $-1$, $+1$, $+1$, $+1$ reads $0=2x_{\{1,4\}}$.
Concluding remarks {#S:disc}
==================
1\. The patterns studied in this work are based on combinations of conditional independence and covariance constraints for events. In recent decades, patterns of conditional independence among random variables have been studied in statistics and in probability theory since they provide insight to decompositions of multidimensional distributions, so sought for in applications. A framework for this activity was developed in the graphical models community[@Lauri].
A general formulation of the problem considers random variables $\xi_i$ indexed by $i$ in $N$ and patterns consisting of the conditional independences $\xi_i{\ensuremath{\!\perp\!\!\!\!\perp\!}}\xi_j|\xi_K$ where $\xi_K=(\xi_k)_{k\in K}$, $i,j\in N$, and $i,j\not\in K$. The case $i=j$ means functional dependence of $\xi_i$ on $\xi_K$, a.s. The problem is highly nontrivial even for four variables [@M.4var.III].
First treatments go back to [@Pearl; @Spo80]. The variant of the problem excluding the functional dependence is most frequent [@Studeny]. Restrictions to Gaussian [@M.Lnenicka; @sul] or binary variables, positivity of the distribution of $\xi_N$, etc., have been studied as well [@Dawid]. The idea to employ the Fourier-Stieltjes transform, as in Section \[S:mainproof\], appeared in [@M.indep], characterizing patterns of unconditional independence.
2\. For patterns of conditional independence, the role of forkness is played by graphoids [@Pearl], semigraphoids [@semigr], imsets [@Studeny], semimatroids [@M.4var.III], etc. Notable are connections to matroid representations theory, see [@M.matroid].
3\. All possible patterns of conjunctive forks on events $A_i$ indexed by $i$ in $N$ arise by varying a probability measure on [$\mathit\Omega$]{}, the power set of $N$. For a ternary relation [$\mathfrak r$]{}on a finite set $N$, the set $\mathcal P_{\ensuremath{\mathfrak r}\xspace}$ of probability measures ${\ensuremath{P}\xspace}$ on [$\mathit\Omega$]{}that satisfy $(i,j,k)\in {\ensuremath{\mathfrak r}\xspace}\;\Leftrightarrow\;
(A_i,A_j,A_k)_{{\ensuremath{P}\xspace}}$ is described by finitely many constraints that require quadratic polynomials in indeterminates $z_{\omega}$ indexed by $\omega$ in ${\ensuremath{\mathit\Omega}\xspace}$ to be positive or zero. For fork-representability of ${\ensuremath{\mathfrak r}\xspace}$, it matters only whether $\mathcal P_{\ensuremath{\mathfrak r}\xspace}$ is empty or not, which can be found out in polynomial time by the main result of the present paper. The shape of $\mathcal P_{\ensuremath{\mathfrak r}\xspace}$, which is a semialgebraic subset of the probability simplex, might be difficult to understand; to reveal it, finer algebraic techniques are needed, as in algebraic statistics [@DoSS09; @Zwi].
4\. One of the two [*Discrete Mathematics*]{} reviewers asked: “Is there some interesting algebraic/combinatorial structure in admissible forknesses? Can they be partially ordered for a fixed $N$?" We leave these questions open.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work began in March 2011 in the weekly meetings of the seminar ConCoCO (Concordia Computational Combinatorial Optimization). We wish to thank its participants, in particular, Luc Devroye and Peter Sloan, for stimulating and helpful interactions. We also thank the two [*Discrete Mathematics*]{} reviewers for their thoughtful comments that made us improve the presentation considerably.
[99]{} P. von Bretzel (1977) Concerning a probabilistic theory of causation adequate for the causal theory of time. *Synthese* [**35**]{} 173–190. N. Cartwright and M. Jones (1991) How to hunt quantum causes. *Erkenntnis* [**35**]{} 205–231. V. Chv' atal and B. Wu (2012) On Reichenbach’s causal betweenness. *Erkenntnis* [**76**]{} 41–48. A.P. Dawid (1979) Conditional independence in statistical theory (with discussion). *J.R. Statist. Soc. B* [**41**]{} [1]{}–[31]{}. P. Dowe (1992) Process causality and asymmetry. *Erkenntnis* [**37**]{} 179–196. M. Drton, B. Sturmfels and S. Sullivant (2009) *Lectures on Algebraic Statistics.* Oberwolfach Seminars, Springer Science [&]{} Business Media. F.S. Ellett, Jr. and D.P. Ericson (1986) Correlation, partial correlation, and causation. *Synthese* [**67**]{} 157–173. G. Hofer-Szabó, M. Rédei and L.E. Szabó (1999) On Reichenbach’s commoncause principle and Reichenbach’s notion of common cause. *The British Journal for the Philosophy of Science* [**50**]{} 377–399. P.L. Kendall and P.F. Lazarsfeld (1950) Problems of survey analysis. In: *Continuities in Social Research: Studies in the Scope and Method of “The American Soldier”* (R.K. Merton and P.F. Lazarsfeld, eds.) The Free Press, Glencoe, IL, pp. 133–196. L.G. Khachiyan (1979) A polynomial algorithm in linear programming. (in Russian) *Doklady Akademii Nauk SSSR* [**244**]{} 1093–1096. K.B. Korb (1999) Probabilistic causal structure. In: *Causation and Laws of Nature* (H. Sankey, ed.), Kluwer Academic Publishers, Dordrecht, The Netherlands, pp. 265–311. S.L. Lauritzen (1996) *Graphical Models.* Oxford University Press, Oxford.
R. Lněnička and F. Matúš (2007) On Gaussian conditional independence structures. *Kybernetika* [**43**]{} 327–342. F. Matúš (1994) Stochastic independence, algebraic independence and abstract connectedness. *Theoretical Computer Science* [**134**]{} 455–471. F. Matúš (1999) Conditional independences among four random variables III: final conclusion. *Combinatorics, Probability & Computing* [**8**]{} 269–276. F. Matúš (1999) Matroid representations by partitions. *Discrete Mathematics* [**203**]{} 169–194. F. Matúš (2004) Towards classification of semigraphoids. *Discrete Mathematics* [**277**]{} 115–145. J. Pearl (1988) *Probabilistic Reasoning in Intelligent Systems.* Morgan Kaufman: San Mateo, California, 1988. H. Reichenbach (1956) *The direction of time.* University of California Press, Berkeley and Los Angeles. W.C. Salmon (1980) Probabilistic causality. *Pacific Philosophical Quarterly* [**61**]{} 50–74. W.C. Salmon (1984) *Scientific Explanation and the Causal Structure of the World.* Princeton: Princeton University Press. M. Studený (2005) *Probabilistic conditional independence structures.* Springer, New York. W. Spohn (1980) Stochastic independence, causal independence, and shieldability. *Journal of Philosophical Logic* [**9**]{} 73–99. W. Spohn (1994) On Reichenbach’s principle of the common cause. In: *Logic, Language, and the Structure of Scientific Theories* (W. Salmon and G. Wolters, eds.), University of Pittsburgh Press, Pittsburgh, pp. 211–235. S. Sullivant (2009) Gaussian conditional independence has no finite complete axiom system. *Journal of Pure and Applied Algebra* [**213**]{} 1502–1506. P. Zwiernik (2015) *Semialgebraic Statistics and Latent Tree Models.* CRC Press.
[^1]: Research of V. Chvátal and Y. Zwólš was supported by the Canada Research Chairs program and by the Natural Sciences and Engineering Research Council of Canada
[^2]: Research of F. Mat' uš was supported by Grant Agency of the Czech Republic under Grant 13-20012S
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This short note aims at (re)proving that the symmetrically normalized graph Laplacian $L=\operatorname{Id}- D^{-1/2}WD^{-1/2}$ (from a graph defined from a Gaussian weighting kernel on a sampled smooth manifold) converges towards the continuous Manifold Laplacian when the sampling become infinitely dense. The convergence rate with respect to the number of samples $N$ is $O(1/N)$.'
author:
- 'L. Jacques'
title: |
Convergence Rate of the\
Symmetrically Normalized Graph Laplacian\
[TR-LJ-2011.01]{}
---
There exist discrete operators which are the equivalent of the gradient and the divergence operators defined on continuous manifold. They share with them some common properties and they converge also to their continuous counterparts for a sufficiently fine sampling of the underlying manifold.
The first one relies on the definition of edge derivative [@zhou2004rfl; @peyre2008nlr]. For a smooth function $f: V\to
{\mathbb{R}}$, the edge derivative of $f$ on $u\in V$ along the edge $e=(u,v)\in E$ formed by the connected vertices $u$ and $v\in V$ reads $$\nabla_e f\big(u)\ =\ \sqrt{\tfrac{w(u,v)}{2d(v)}}
f(v) - \sqrt{\tfrac{w(u,v)}{2d(u)}} f(u),$$ where $d(u) = \sum_{v\in V} w(u,v)$ is the *degree* of the vertex $u$. From this relation, we have obviously $\nabla_e f\big(u) = -\nabla_e
f\big(v)$.
The gradient of $f$ is then defined globally as the vector field $\nabla f: E\to {\mathbb{R}}$ defined on the edge set $E$ as $\nabla f(e) =
\nabla_{e} f(u)$ for $e=(u,v)\in E$. This gradient can be represented as a linear operator $\nabla\in{\mathbb{R}}^{N\times N}$, so that $$\nabla f(u)\ =\ \{\nabla f(u,v):\ v\in V\}\ \in\ {\mathbb{R}}^N,$$ corresponds to the gradient of $f$ on $u\in V$ seen as a vector of ${\mathbb{R}}^N$. The norm of this object on each $u\in V$ is defined naturally as $${\|\nabla f(u)\|}^2\ =\ \sum_{v\in V} |\nabla f(u,v)|^2\ =\
\sum_{v\,\sim\,u} |\nabla f(u,v)|^2.$$
The scalar products ${\langle f, g \rangle}=\sum_{u\in V} f(u) g(v)$ and ${\langle F, G \rangle}=\sum_{e\in E} F(e) G(e)$ between two real functions $f$ and $g$ in the Hilbert space $\ell^2(V)=\{h: \sum_{u\in V} |h(u)|^2 <
\infty\}$ and two vector fields $F$ and $G$ in $\ell^2(E)=\{H:
\sum_{e\in E}
|H(e)|^2 < \infty\}$, defined then the adjoint of the gradient, i.e. the graph divergence $\nabla^*$. Indeed, applying the relation ${\langle \nabla g, F \rangle}={\langle g, -\nabla^* F \rangle}$ valid in the continuous domain [@chambolle2008], we get $$[\nabla^* F](u) \ =\ \sum_{v\, \sim\, u} \sqrt{\tfrac{w(u,v)}{2d(u)}}
\big(F(u,v) - F(v,u)\big).$$
The graph Laplacian of a function $f$ is then defined as $$\Delta f(u)\ =\ \nabla^*\nabla f(u)\ =\ \sum_{v\,\sim\, u}
\tfrac{w(u,v)}{\sqrt{d(u)d(v)}} f(v) - f(u).$$ Using matrix notations, this operator corresponds actually to the common (symmetric) normalized Laplacian defined on graph, i.e. $$\Delta\ =\ D^{-1/2}WD^{-1/2}\ -\ \operatorname{Id},$$ with $D\in{\mathbb{R}}^{N\times N}$ ($N = \# V$) a diagonal matrix such that $D_{uu} = d(u)$ and $W\in{\mathbb{R}}^{N\times N}$ the weight matrix with $W_{uv} = w(u,v)$.
Interestingly, the graph Laplacian converges to the continuous Laplace-Beltrami operator [@carmo1992rg] on the manifold underlying the graph definition.
If the vertex set $V=\{v_1,\,\cdots,v_N\}\subset\,{\mathcal{M}}$ consists of $N$ points taken uniformly and independently at random on a $m$-dimensional compact manifold ${\mathcal{M}}$ embedded in ${\mathbb{R}}^n$, then, for any smooth function $f:V\to {\mathbb{R}}$ and for a weighting function $w(u,v)=\exp\{ -{\|u-v\|}^2/{2\epsilon}\}$, the normalized graph Laplacian defined by the graph $G=(V,E=V\times
V,w)$ satisfies $${{\textstyle\frac{1}{\epsilon}}}\,\lim_{N\to\infty}\Delta f(u)\ =\
{{\textstyle\frac{1}{2}}}\,\Delta_{{\mathcal{M}}} f(u)\ +\ O(\epsilon^{1/2}),$$ with $\Delta_{{\mathcal{M}}}$ the Laplace-Beltrami operator defined on ${\mathcal{M}}$. \[lem:discr\_lapl\_manifold\_lapl\]
We follow a similar development to the one given in [@singer2006gml]. With the hypothesis of the Lemma, for any function $h\in \ell^2(V)$, we have $$\begin{aligned}
[Wh](u)\ =\ \sum_j w(u,v_j)\,h(v_j)&=\sum_{j\neq i} \exp\{
-{\|u-v_j\|}^2/{2\epsilon}\}\,h(v_j)\ +\ h(u).\end{aligned}$$ Since $Y_j=w(u,v_j)\,h(v_j)$ are iid for $j\neq i$, by the law of large numbers we have $$\begin{aligned}
\sum_{j\neq i} w(u,v_j)\,h(v_j)&\simeq\ (N-1)\,
\mathbb{E}_{\mathcal{M}}\big[\exp\{u-\cdot\}h(\cdot)\big]\\
&=\ \tfrac{N-1}{\operatorname{vol}{\mathcal{M}}}\,\int_{\mathcal{M}}\exp\{
-{\|u-y\|}^2/{2\epsilon}\}\,h(y)\ {\mathrm{d}}_{\mathcal{M}}y,\end{aligned}$$ where the integral is performed on the manifold with the local infinitesimal volume element ${\mathrm{d}}_{\mathcal{M}}y$. The relation (2.9) in [@singer2006gml] (or Eq. (10) in [@smolyanov2000bmm]) explains that $$\tfrac{1}{(2\pi\epsilon)^{m/2}}\,\int_{\mathcal{M}}\exp\{
-{\|u-y\|}^2/{2\epsilon}\}\,h(y)\ {\mathrm{d}}_{\mathcal{M}}y\ =\ h(u) +
\tfrac{\epsilon}{2} [E(u)h(u) + \Delta_{\mathcal{M}}h(u)] + O(\epsilon^{3/2}),$$ where $E(u)={\frac{1}{3}}S(u)$ and $S$ is the scalar curvature of ${\mathcal{M}}$ [@carmo1992rg]. Therefore, $$\begin{aligned}
[Wh](u)&=\ \tfrac{(N-1)(2\pi\epsilon)^{m/2}}{\operatorname{vol}{\mathcal{M}}}\big[h(u)
+ \tfrac{\epsilon}{2} [E(u)h(u) + \Delta_{\mathcal{M}}h(u)] +
O(\epsilon^{3/2}) \big] + h(u)\nonumber\\
&=\ \tfrac{(N-1)(2\pi\epsilon)^{m/2}}{\operatorname{vol}{\mathcal{M}}}\big[h(u)
+ \tfrac{\epsilon}{2} [E(u)h(u) + \Delta_{\mathcal{M}}h(u)] +
O\big(\epsilon^{3/2}, 1/(N\epsilon^{m/2})\big)\big],
\label{eq:Wh-tayl-dev}\end{aligned}$$ where the notation $O(\mu,\nu)$ means $|O(\mu,\nu)| < C\mu + D\nu$ for two positive values $\mu,\nu \ll 1$. Taking $h=1$, we get then $$\label{eq:degree-dev}
d(v_j) = \tfrac{(N-1)(2\pi\epsilon)^{m/2}}{\operatorname{vol}{\mathcal{M}}}\big[1
+ \tfrac{\epsilon}{2}\,E(v_j) + O\big(\epsilon^{3/2}, 1/(N\epsilon^{m/2})\big)\big].$$ Therefore, $$\begin{aligned}
\Delta f(u)&=\ \sum_j \tfrac{w(u,v_j)}{\sqrt{d(u) d(v_j)}} f(v_j)\ -\
f(u)\\
&=\ \tfrac{1}{\sqrt{1 +
\frac{\epsilon}{2}E(u)}}\ \tfrac{\operatorname{vol}{\mathcal{M}}}{(N-1)(2\pi\epsilon)^{m/2}} \sum_{j\neq i}
w(u,v_j)\,\tfrac{f(v_j)}{\sqrt{1 + \frac{\epsilon}{2}E(v_j)}}\ -\
f(u)\ +\ O\big(\epsilon^{3/2}, 1/(N\epsilon^{m/2})\big). \end{aligned}$$ Taking now $h(t)=f(t)g(t)$ for $t\in{\mathcal{M}}$ with $g(t)=\tfrac{1}{\sqrt{1 + \frac{\epsilon}{2}E(t)}}$ in , we get $$\begin{gathered}
\tfrac{\operatorname{vol}{\mathcal{M}}}{(N-1)(2\pi\epsilon)^{m/2}}\ \sum_{j\neq i}
w(u,v_j)\,f(v_j)g(v_j)\\
=\ \big[ g(u)f(u)(1
+ \tfrac{\epsilon}{2} E(u)) + \tfrac{\epsilon}{2}\Delta_{\mathcal{M}}h(u) +
O\big(\epsilon^{3/2}, 1/(N\epsilon^{m/2})\big)\big].\end{gathered}$$ However, $$\Delta_{\mathcal{M}}h(u) = f(u)\Delta_{\mathcal{M}}g(u) + 2\,{\langle \nabla_{\mathcal{M}}f(u), \nabla_{\mathcal{M}}g(u) \rangle}_{T_u{\mathcal{M}}} +
g(u) \Delta_{\mathcal{M}}f(u),$$ where the scalar product ${\langle \cdot, \cdot \rangle}_{T_u{\mathcal{M}}}$ occurs in the tangent plane $T_u{\mathcal{M}}\simeq {\mathbb{R}}^d$ of ${\mathcal{M}}$ on $u$ with the metric of the manifold [@carmo1992rg]. For any function $s$ on ${\mathcal{M}}$, the gradient $\nabla_{\mathcal{M}}s(u) \in T_u{\mathcal{M}}$ is composed of the derivative of $s$ according to a local system of coordinates (or *chart*) isomorphic to ${\mathbb{R}}^d$.
From the definition of $g$, it is clear that any differential operator $D_{\mathcal{M}}$ of $g$ with respect to this local system in $T_u{\mathcal{M}}$ is of order $D_{\mathcal{M}}g(u) = O(\epsilon)$. Consequently, $$\tfrac{\epsilon}{2}\Delta_{\mathcal{M}}h(u) = \tfrac{\epsilon}{2}g(u) \Delta_{\mathcal{M}}f(u) + O(\epsilon^2),$$ that provides the final result, $$\begin{aligned}
\Delta f(u)&=\ g^2(u) f(u)\,\big(1 +
\tfrac{\epsilon}{2}E(u)\big)\ +\
\tfrac{\epsilon}{2}\Delta_{\mathcal{M}}f(u)\ -\ f(u)\ +\ O\big(\epsilon^{3/2}, 1/(N\epsilon^{m/2})\big)\\
&=\ \tfrac{\epsilon}{2}\,\Delta_{\mathcal{M}}f(u)\ +\ O\big(\epsilon^{3/2}, 1/(N\epsilon^{m/2})\big). \end{aligned}$$
[1]{} \[1\][\#1]{} url@samestyle \[2\][\#2]{} \[2\][[l@\#1=l@\#1\#2]{}]{}
D. Zhou and B. Scholkopf, “[A regularization framework for learning from graph data]{},” in *ICML Workshop on Statistical Relational Learning and Its Connections to Other Fields*, 2004, pp. 132–137.
G. Peyr[é]{}, S. Bougleux, and L. Cohen, “[Non-local regularization of inverse problems]{},” in *Proceedings of ECCV*, vol. 2008.1em plus 0.5em minus 0.4emSpringer, 2008.
A. Chambolle, “[An Algorithm for Total Variation Minimization and Applications]{},” *Journal of Mathematical Imaging and Vision*, vol. 20, pp. 89–97, 2004
M. Carmo, *[Riemannian Geometry]{}*.1em plus 0.5em minus 0.4em Birkhauser, 1992.
A. Singer, “[From graph to manifold Laplacian: the convergence rate]{},” *Applied and Computational Harmonic Analysis*, vol. 21, no. 1, pp. 128–134, 2006.
O. Smolyanov, H. von Weizsacker, and O. Wittich, “[Brownian motion on a manifold as limit of stepwise conditioned standard Brownian motions]{},” *Stochastic processes, physics and geometry: new interplays, II*, vol. 29, pp. 589–602, 2000.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Quantum batteries are quantum mechanical systems with many degrees of freedom which can be used to store energy and that display fast charging. The physics behind fast charging is still unclear. Is this just due to the collective behavior of the underlying interacting many-body system or does it have its roots in the quantum mechanical nature of the system itself? In this work we address these questions by studying three examples of quantum-mechanical many-body batteries with rigorous classical analogs. We find that the answer is model dependent and, even within the same model, depends on the value of the coupling constant that controls the interaction between the charger and the battery itself.'
author:
- Gian Marcello Andolina
- Maximilian Keck
- Andrea Mari
- Vittorio Giovannetti
- Marco Polini
title: 'Quantum versus classical many-body batteries'
---
Introduction
============
Recently there has been a great deal of interest in quantum batteries (QBs) [@Alicki13; @Hovhannisyan13; @Binder15; @Campaioli17; @Ferraro17; @Le17; @Andolina18; @Andolina18b; @Campaioli18; @Farina18; @Julia-Farre18; @Zhang18], i.e. quantum mechanical systems that are able to store energy. These works have a key common thread in trying to understand whether quantumness yields a temporal speed-up of the charging process. A first, abstract approach [@Binder15; @Campaioli17] studied the possibility to charge $N$ systems via unitary operations. The authors introduced a parallel charging scheme, in which each of the subsystems is acted upon independently of the others, and a collective charging scheme, where global unitary operations acting on the full Hilbert space of all subsystems are allowed. In these works it was shown that the charging time scales with $N$, decreasing for increasing $N$. In the collective charging case and for large $N$, the power delivered by a QB is much larger than the one delivered by the parallel scheme. This speed-up was dubbed “quantum advantage” [@Binder15; @Campaioli17; @Le17; @Ferraro17]. Furthermore, in Ref. it was shown that entanglement is not required to speed-up the evolution of a QB, since states which are confined in the sphere of separable states share an identical speed-up. However, the authors of Ref. pointed out that such highly mixed states host only a vanishing amount of energy, yielding therefore a highly non-optimal result from the point of view of energy storage and delivery.
In the same spirit, the authors of Refs. studied similar issues but in realistic setups which can be implemented in a laboratory, such as arrays of qubits in cavities [@Ferraro17; @Andolina18; @Andolina18b; @Farina18] and spin chains in external magnetic fields [@Le17]. In Refs. , the battery units are not charged via abstract unitaries but, rather, by other quantum mechanical systems dubbed “chargers”. In this framework, the parallel scheme is the one in which each battery is charged by its own charger, independently of the others—see Fig. \[fig:Sketch\]. On the contrary, the collective scheme is the one in which all batteries are charged by the very same charger. Also in this context, the collective scheme outperforms the parallel one in terms of speed of the charging process. Finally, the authors of Ref. demonstrated that quantum batteries have the potential for faster charging over their classical counterparts. As they noticed, however, the classical counterparts were assumed to be composed of non-interacting units.
[Figs/Sketch]{}(4,31)
In this Article we compare the performance of QBs with that of their appropriate classical versions. Such comparison is clearly of great interest for foundational reasons but has no implications on the development of scalable solid-state systems where energy transfer processes and their time scales can be studied experimentally. Indeed, any solid-state QB device is going to operate on the basis of electrons, photons, spins, etc, which are inherently described by quantum mechanics. We focus on three models. In the first, a single bosonic mode (the charger) is coupled to $N$ harmonic oscillators (the proper battery composed of $N$ subunits). In the second one, $N$ qubits playing the role of charging units are coupled to another set of $N$ qubits playing the role of the proper battery. Finally, the third one is the Dicke QB introduced in Ref. . In the first case, the performance of classical and quantum versions of the model is identical. In the second case, the classical version outperforms the quantum one. In the third case, there is a range of values of the charger-matter coupling parameter $g$ for which the quantum (classical) model performs better than the classical (quantum) one.
Our Article is organized as following. In Sect. \[Comp\] we explain how the classical versus quantum comparison is actually carried out in this Article, briefly reviewing the correspondence between quantum commutators and classical Poisson brackets. In Sect. \[Fom\] we recap the charging protocol first introduced in Refs. and introduce the figures of merit needed to evaluate the performance of classical and quantum many-body batteries. In Sect. \[sect:HO\] we discuss the first model (single bosonic mode coupled to $N$ harmonic oscillators). We demonstrate analytically that in this case classical and quantum versions of the model display fast charging with the same time scale. In Sect. \[sect:spin\] we introduce the second model ($N$ qubits coupled to $N$ qubits) and demonstrate how the classical version of the model outperforms the quantum one. In Sect. \[sect:Dicke\] we compare the Dicke QB model introduced in Ref. with the corresponding classical analogue, showing numerically that the relative performance depends on the charger-matter coupling $g$. Finally, in Sect. \[concl\] we report a summary of our main findings and our conclusions.
Comparison between quantum and classical mechanics {#Comp}
==================================================
In quantum mechanics, the evolution of an operator $\hat{O}$ in time $t$ is described by the Heisenberg equation of motion $\hbar~d\hat{O}(t)/dt=i[{\mathcal{H}},\hat{O}(t)]$, where ${\cal H}$ is the Hamiltonian. Moreover, canonically conjugate variables, such as position $\hat{q}_i$ and momentum $\hat{p}_j$, fulfill the commutation relation $[\hat{q}_i,\hat{p}_j]=i\hbar\delta_{i,j}$. In the case of angular momentum $\hat{\bm J}$, a similar relation holds between different components: $[\hat{J}_i,\hat{J}_j]=\sum_k i\hbar\epsilon_{ijk}\hat{J}_k$, where $\epsilon_{ijk}$ is the Levi-Civita tensor.
In Hamiltonian mechanics, a classical physical system is uniquely described by a set of canonical coordinates $\boldsymbol{x}^{\rm T}=( {\bm p}, {\bm q})$, where the components $q_i, p_i$ are conjugate variables obeying $\{q_i, p_j\}=\delta_{i,j}$. Here, $\{u,v\}\equiv\sum_i(\partial_{q_i} u~ \partial_{p_i}v- \partial_{p_i}u~ \partial_{q_i}v)$ denotes the Poisson brackets.
The time evolution of the system is uniquely defined by Hamilton’s equations: $$\begin{aligned}
\label{eq:HJ}
\frac{dq_i}{dt}&=&\partial_{p_i}\mathcal{H}^{\rm cl}({\bm x})~, \nonumber \\
\frac{dp_i}{dt}&=&-\partial_{q_i}\mathcal{H}^{\rm cl}({\bm x})~.\end{aligned}$$
A proper comparison between quantum and classical systems can be made by following the canonical quantization procedure [@DIRAC]. Once the Hamilton’s function $\mathcal{H}^{\rm cl}({\bm x})$ of a classical system is written in terms of conjugate variables with Poisson brackets $\{q_i, p_j\}=\delta_{i,j}$, quantization is carried out by replacing classical coordinates by operators and enforcing canonical commutation relations instead of canonical Poisson brackets.
While finding the classical analog of a quantum system with degrees of freedom that are position and momentum is straightforward and consists in making the replacements $\hat{q}_i\to{q}_i$ and $\hat{p}_j\to{p}_j$, the classical version of quantum mechanical angular momentum is more subtle. It turns out [@BraunBook; @Carlos18] that the right choice is to replace the components $\hat{J}_i$ of the angular momentum operator $\hat{\bm J}$, with $\hat{\bm J}^2=\hbar^2J(J+1)$, with the classical canonical coordinates $J_z=J\cos(\theta)$ and $\phi=\arctan(J_y/J_x)$, so that $\{J \cos(\theta), \phi \}=1$, i.e. $\hat{J}_z \to J\cos(\theta)$, $\hat{J}_x \to J\sin(\theta)\cos(\phi)$, and $\hat{J}_y \to J\sin(\theta)\sin(\phi)$.
In the remainder of this Article we set $\hbar=1$.
Charging protocol and figures of merit {#Fom}
======================================
We start by reviewing a model for the charging process of a QB [@Ferraro17; @Andolina18; @Andolina18b; @Farina18]. As stated above, the classical and quantum cases are both described by an Hamiltonian formalism. We can therefore introduce the charging protocol in terms of a general Hamiltonian, without specifying [*a priori*]{} whether we treat the classical or quantum case. As such, we will describe the protocol in general, commenting explicitly on the classical and quantum cases only when it is needed.
In our charging protocol [@Ferraro17; @Andolina18; @Andolina18b; @Farina18], a first system $\rm A$ acts as the energy “charger” for a second system $\rm B$, which instead acts as the proper battery. They are characterized by local Hamiltonians $\mathcal{H}_{\rm A}$ and $\mathcal{H}_{\rm B}$, respectively, which, for the sake of convenience, are both chosen to have zero ground-state energy. We also assume ${\rm B}$ to be composed by $N$ non-mutually interacting elements. (Effective interactions between these elements are induced by the charger. In the Dicke QB case, for example, the cavity mode induces effective interactions between all the qubits.) In the quantum case, the system at time $t=0$ is in a pure factorized state $| \psi\rangle_{\rm A} \otimes |0\rangle_{\rm B}$, $|0\rangle_{\rm B}$ being the ground state of $\mathcal{H}_{\rm B}$ and $| \psi\rangle_{\rm A}$ having mean local energy $E^{(N)}_{\rm A}(0)\equiv {_{\rm A} \langle} \psi|\mathcal{H}_{\rm A}|\psi\rangle_{\rm A} >0$, where $N$ is the number of elements which compose the battery. Analogously, in the classical case we impose that the system B at time $t=0$ is in the configuration with the lowest energy and we fix the energy in the charger A to be $E^{(N)}_{\rm A}(0)>0$.
By switching on a coupling Hamiltonian $\mathcal{H}_1$ between A and B, our aim is to provide as much energy as possible to $\rm B$, in some finite amount of time $\tau$, the charging time of the protocol. For this purpose, we write the global Hamiltonian of the AB system as $$\label{eq:protocol}
\mathcal{H}(t) \equiv \mathcal{H}_{\rm A}+\mathcal{H}_{\rm B}+\lambda(t)\mathcal{H}_1~,$$ where $\lambda(t)$ is a time-dependent parameter that represents the external control we exert on the system, and which we assume to be given by a step function equal to $1$ for $t\in[0,\tau]$ and zero elsewhere. Accordingly, in the quantum case, we denote by $|\psi(t) \rangle_{\rm AB}$ the evolved state of the AB system at time $t$, its total energy $E(t) \equiv {_{\rm AB}\langle} \psi(t) |\mathcal{H}(t)|
\psi(t) \rangle_{\rm AB}$ being constant at all times with the exception of the switching points, $t=0$ and $t=\tau$, where some non-zero energy can be transferred to ${\rm AB}$ by the external control. (See Ref. for a detailed analysis of the energy cost of modulating the interaction.)
The same conditions hold in the classical case where we denote by $\boldsymbol{x}^{\rm T}(t)=( \boldsymbol{p}(t), \boldsymbol{q}(t))$ and $E(t)=\mathcal{H}^{\rm cl}\big(\boldsymbol{x}(t)\big)$ the solution of Hamilton’s equations of motion and the total energy at time $t$, respectively. Here, $ \boldsymbol{p}$ and $ \boldsymbol{q}$ are classical conjugate variables. It is also useful to define the vector $\boldsymbol{x}_{\rm B}^{\rm T}(t)=( \boldsymbol{p}_{\rm B}(t), \boldsymbol{q}_{\rm B}(t))$, denoting the position in phase space of B at time $t$.
In the quantum case, we are mainly interested in the mean local energy of the battery at the end of the protocol, i.e. $$\label{stored energy}
E^{(N)}_{\rm B}(\tau)\equiv {\rm tr}[\mathcal{H}_{\rm B} \rho_{\rm B}(\tau)]~,$$ $\rho_{\rm B}(\tau)$ being the reduced density matrix of the battery at time $\tau$. It is worth noticing that while $E^{(N)}_{\rm B}(\tau)$ does not necessarily represent the amount of energy that one can recover from the battery after charging, it has been shown that for large enough $N$ this is not a relevant issue [@Andolina18b]. In the classical case, the corresponding quantity is the energy in B, $E^{(N)}_{\rm B}(\tau)=\mathcal{H}^{\rm cl}_{\rm B}(\boldsymbol{x}_{\rm B}(\tau))$.
The performance of the charger-battery set-up can be studied by analyzing the average storing power $P^{(N)}_{\rm B}(\tau)\equiv E^{(N)}_{\rm B}(\tau)/\tau$. Specifically, we define the maximum average power as $\bar P^{(N)}_{\rm B}\equiv \max_\tau [P^{(N)}_{\rm B}(\tau)]$. Finally, we introduce the optimal charging time $\bar{\tau}$, $\bar P^{(N)}_{\rm B}=P^{(N)}_{\rm B}(\bar{\tau})$, and the energy at the maximum power, $\bar{E}^{(N)}_{\rm B}\equiv{E}^{(N)}_{\rm B}(\bar{\tau})$.
Our aim is to compare the parallel charging scenario against the collective one [@Binder15; @Campaioli17; @Ferraro17]. As mentioned above, we define as a parallel charging, the protocol in which $N$ batteries are independently charged by $N$ chargers. Each charger has an energy $E_{\rm A}^{(1)}(0)$. Conversely, the collective charging case is the one in which all $N$ batteries are charged by the same charger. In order to do a clear comparison, in the collective charging case we impose that the charger has total energy equal to the sum of the energies of all the chargers of the parallel charging scheme, i.e. $E_{\rm A}^{(N)}(0)=NE_{\rm A}^{(1)}(0)$.
Since we are interested in comparing the power of the protocols, we denote by the symbol ${\bar{P}_{\sharp}}$ (${\bar{P}_{\parallel}}$) the maximum power in the collective (parallel) protocol. Following Ref. , we introduce the so-called collective advantage: $$\label{Gamma}
\Gamma \equiv \frac{\bar{P}_{\sharp}}{\bar{P}_{\parallel}}~.$$ We have $\bar{P}_{\sharp}=\bar{P}^{(N)}_{\rm B}$ and $\bar{P}_{\parallel}=N\bar{P}^{(1)}_{\rm B}$. The latter property follows from the fact that the power in the parallel charging scheme is trivially extensive.
The figure of merit in Eq. (\[Gamma\]) quantifies how convenient is to charge a battery in a collective fashion rather than in a parallel way. While in Refs. this quantity is named “quantum advantage”, it is possible to define $\Gamma$ also in the classical case. Since our main goal is to compare quantum and classical batteries, we will denote by $\Gamma_{\rm qu}$ the collective advantage produced by a quantum Hamiltonian and by $\Gamma_{\rm cl}$ the collective advantage produced by the analog classical Hamiltonian. What matters is therefore the ratio $$\label{R}
R\equiv \frac{\Gamma_{\rm qu}}{\Gamma_{\rm cl}}~.$$ If $R=1$, the QB and its classical analog share the same collective boost in the charging process. Conversely, having $R>1$ means that there is a genuine quantum advantage. Finally, $R<1$ means that the collective dynamics in the classical model is more beneficial.
The quantity $R$ will be crucial below in determining if fast charging is due to exquisitely quantum resources or, rather, if it has a collective (i.e. many-body) origin due to effective interactions between the battery subunits, which is present also in the classical case.
Harmonic oscillator batteries {#sect:HO}
=============================
In this Section we study a system composed by $N+1$ harmonic oscillators, one acting as a charger while the remaining $N$ playing the role of the proper battery. This system is described by the following Hamiltonian, $$\begin{aligned}
\label{eq:Hhoho}
\mathcal{H}_{\rm A}&=&\omega_0 a^\dagger a~,\nonumber\\
\mathcal{H}_{\rm B}&=&\omega_0\sum_i b^\dagger_ib_i~,\nonumber\\
\mathcal{H}_{1}&=&g\sum_i \big(a b^\dagger_i+a^\dagger b_i\big)~,\end{aligned}$$ where $a$ ($b_i$) is the destruction bosonic operator acting on A (on the $i$-th unit of the battery B), and ${\omega}_0$ and $g$ are the characteristic frequency of both systems and the charger-battery coupling parameter, respectively. For simplicity, we choose $E_{\rm A}^{(1)}(0)=\omega_0$.
It is useful to introduce the bright mode [@Ciuti05] $B=\sum_i b_i/\sqrt{N}$, which is a bosonic mode fulfilling $[B,B^\dagger]=1$. Expressing the Hamiltonian in terms of the bright mode, we obtain: $$\begin{aligned}
\label{eq:HhohoB}
\mathcal{H}_{\rm B}&=&\omega_0 B^\dagger B~,\nonumber\\
\mathcal{H}_{1}&=&g_N\left( a B^\dagger+a^\dagger B\right)~,\end{aligned}$$ where $$\label{eq:g_N}
g_N \equiv \sqrt{N}g~.$$ Hence, the AB system is equivalent to two harmonic oscillators with a renormalized coupling $g_N$. It is straightforward to obtain the dynamics of the energy of B, which is independent of the initial state $\ket{\psi}_{\rm A}$ in A. In order to calculate the stored energy we find then useful to adopt the Heisenberg representation, writing $E_{\rm B}(\tau) ={\rm tr}[\rho_{\rm AB}(0)\mathcal{H}_{\rm B}(\tau)]$, where $\rho_{\rm AB}(0)$ is the density matrix of the full system at the initial time, with $\mathcal{H}_{\rm B}(\tau)\equiv e^{i \mathcal{H} \tau} \mathcal{H}_{\rm B}e^{-i \mathcal{H} \tau}$. Expressing $a$ and $b$ as functions of the normal operators $\gamma_{\pm}=(a\pm B)\sqrt{2}$ and using that the latter evolve simply as $\gamma_{\pm}(t) = e^{-i\omega_{\pm}t}\gamma_{\pm}$ with $\omega_{\pm}=\omega_0\pm g_N$, we obtain $$\begin{aligned}
\label{evolvedH}
\mathcal{H}_{\rm B}(\tau)&=&\frac{\omega_0}{2}\Bigg\{a^\dagger a+B^\dagger B \\&-&\left[\frac{e^{-i2g_{N}\tau}}{2} (a^\dagger a-B^\dagger B+B^\dagger a-a^\dagger B) +{\rm H.c.}\right] \Bigg\}~, \nonumber\end{aligned}$$ and, finally: $$\begin{aligned}
\label{eq:Ebho}
E^{(N)}_{\rm B}(\tau)=N\omega_0\sin^2(g\sqrt{N}\tau)~.\end{aligned}$$ Defining $Y={\max_x}[\sin^2(x)/x]$, the maximum power reads $\bar{P}^{(N)}_{\rm B}=N\sqrt{N}g\omega_0Y$. Accordingly, we have: $$\begin{aligned}
\label{eq:Gammaho}
\Gamma_{\rm qu}=\sqrt{N}~.\end{aligned}$$ We note that if $\ket{\psi}_{\rm A}$ is a coherent state, the evolved state $|\psi(t) \rangle_{\rm AB}$ remains factorized at all times [@Andolina18; @WallsMilburn2007]. This is an example where the advantage is present, despite the total absence of correlations.
Now we study the classical analog of the quantum model in Eq. (\[eq:Hhoho\]), which can be simply obtained by reversing the quantization procedure and substituting quantum commutators with classical Poisson brackets. The corresponding classical Hamiltonian describes a set of coupled springs: $$\begin{aligned}
\label{eq:HhohoCl}
\mathcal{H}^{\rm cl}_{\rm A}&=&\frac{\omega_0}{2}\left(p_a^2+q_a^2\right),\nonumber~\\
\mathcal{H}^{\rm cl}_{\rm B}&=&\frac{\omega_0}{2}\sum_i\left(p_{b_i}^2+q_{b_i}^2\right)\nonumber~,\\
\mathcal{H}^{\rm cl}_{1}&=&g \sum_i \left( q_a q_{b_i}+p_a p_{b_i}\right)~,\end{aligned}$$ where $(p_a,q_a)$ are conjugate variables of the charger and $(\boldsymbol{p}_{b_{i}},\boldsymbol{q}_{b_{i}})$ are conjugate variables of the $i$-th battery. As earlier, we choose $E_{\rm A}^{(1)}(0)=\omega_0$. We now introduce $Q_b=\sum_i q_{b_i}/\sqrt{N}$ and $P_b=\sum_i p_{b_i}/\sqrt{N}$. The classical Hamiltonian becomes $$\begin{aligned}
\label{eq:HhohoClR}
\mathcal{H}^{\rm cl}_{\rm B}&=&\frac{\omega_0}{2}\left(P_{b}^2+Q_{b}^2\right)\nonumber~,\\
\mathcal{H}^{\rm cl}_{1}&=&g_N\left( q_a Q_{b}+p_a P_{b}\right)~.\end{aligned}$$ We conclude that also in the classical case the model maps into that of two coupled oscillators with a renormalized coupling $g_N$.
Hamilton’s equations of motion follow from Eqs. , , and : $$\begin{aligned}
\label{eq:hohoHJ}
\frac{d{p}_a}{dt}&=&-\omega_0q_a-g_N Q_b~,\nonumber \\
\frac{d{q}_a}{dt}&=&\omega_0p_a+g_N P_b~,\nonumber \\
\frac{d{P}_b}{dt}&=&-\omega_0Q_b-g_Nq_a~,\nonumber \\
\frac{d{Q}_a}{dt}&=&\omega_0P_b+g_Np_a~.\end{aligned}$$ Solving these equations we find that, irrespective of the particular initial condition, the stored energy reads $E^{(N)}_{\rm B}(\tau)=N\omega_0\sin^2(g\sqrt{N}\tau)$. This implies $$\begin{aligned}
\label{eq:GammahoCl}
\Gamma_{\rm cl}=\sqrt{N}~,\end{aligned}$$ and $R=1$. This is the main result of this Section. For the case of harmonic oscillator batteries defined in (\[eq:Hhoho\]), fast charging, i.e. $\Gamma \propto \sqrt{N}$, is solely due to the collective behavior of the underlying many-particle system, and does not have its roots in the quantumness of its Hamiltonian.
Spin batteries {#sect:spin}
==============
In this Section we study a system composed by $N$ qubits, acting as charger, coupled to another set of $N$ qubits, which play the role of the battery. The quantum Hamiltonian is $$\begin{aligned}
\label{eq:HSpins}
\mathcal{H}_{\rm A}&=&\omega_0\left(J^{(a)}_z+\frac{N}{2}\right)~,\nonumber\\
\mathcal{H}_{\rm B}&=&\omega_0\left(J^{(b)}_{z}+\frac{N}{2}\right)~,\nonumber\\
\mathcal{H}_{1}&=&4g \left(J^{(a)}_xJ^{(b)}_x+J^{(a)}_yJ^{(b)}_y\right)~,\end{aligned}$$ where $J^{(a)}_\alpha$ ($J^{(b)}_\alpha$) with $\alpha=x,y,z$ are the components of a collective spin operator of length $J=N/2$ acting on the Hilbert space of the charger A (battery B), while all the other parameters have the same meaning as in Eq. .
[Figs/GammaQuantumSpinLL.pdf]{} (1,70)[(a)]{}
[Figs/GammaClassicalSPINS.pdf]{} (1,70)[(b)]{}
[Figs/RSpins.pdf]{} (1,70)[(c)]{}
Defining $\mathcal{H}_{0}=\mathcal{H}_{\rm A}+\mathcal{H}_{\rm B}$, the propagator in the interaction picture simply reads $\tilde{U}_t=e^{i\mathcal{H}_{0}t}e^{-i\mathcal{H}t}=e^{-i\mathcal{H}_1t}$. Hence, in this model there is no dependence of the dynamics on the energy scale $\omega_0$, and $\tilde{U}_t$ depends only on the product $gt$. As in the case of Eq. (\[eq:Gammaho\]), this scaling implies that the collective advantage $\Gamma_{\rm qu}$ for this model does not depend on the value of $g$ but only on $N$. In Fig. \[fig:3\](a) we report the log-log plot of the collective advantage $\Gamma_{\rm qu}$ as a function of $N$. Fits to the numerical data (not shown) indicate a quasi-linear dependence on $N$ for large $N$ of the form $$\label{eq:GammaSpins}
\Gamma_{\rm qu}\propto {N^{\alpha}}~,$$ with $\alpha \sim 1$ and a proportionality constant $\sim0.25$.
We now move on to analyze the classical case. Following the discussion of Sect. \[Comp\], we model the analog classical Hamiltonian as $$\begin{aligned}
\label{eq:HSpinsCL}
\mathcal{H}^{\rm cl}_{\rm A}&=&N\omega_0\frac{\big[\cos(\theta_a)+1\big]}{2},\nonumber~\\
\mathcal{H}^{\rm cl}_{\rm B}&=&N\omega_0\frac{\big[\cos(\theta_b)+1\big]}{2}\nonumber~,\\
\mathcal{H}^{\rm cl}_{1}&=&g N^2\sin(\theta_a)\sin(\theta_b)\cos(\phi_a-\phi_b)~,\end{aligned}$$ where $(N\cos(\theta_a)/2,\phi_a)$ and $(N\cos(\theta_b)/2,\phi_b)$ are conjugate variables [@BraunBook; @Carlos18].
Hamilton’s equations of motion follow from Eqs. and . We find $$\begin{aligned}
\label{eq:HSpinsH}
&&\frac{d\cos(\theta_a)}{dt}=2gN \sin(\theta_a)\sin(\theta_b)\sin(\phi_a-\phi_b),\nonumber \\
&&\frac{d\phi_{a}}{dt}=\omega_0-2 g N\cot(\theta_a) \sin(\theta_b)\cos(\phi_a-\phi_b)~.\end{aligned}$$ Since the Hamiltonian is invariant under the exchange of variables $a\leftrightarrow b$, the equations of motion for $\cos(\theta_{b}) $ and $\phi_{b}$ can be simply obtained by exchanging $a\leftrightarrow b$.
It is now useful to define $\varphi_a=\phi_a+\omega_0 t$ and $\varphi_b=\phi_b+\omega_0 t$, which allow us to write Eq. (\[eq:HSpinsH\]) as following: $$\begin{aligned}
\label{eq:HSpinsH2}
&&\frac{d\cos(\theta_a)}{dt} =2gN \sin(\theta_a)\sin(\theta_b)\sin(\varphi_a-\varphi_b),\nonumber \\
&& \frac{d\varphi_{a}}{dt}=-2 g N\cot(\theta_a) \sin(\theta_b)\cos(\varphi_a-\varphi_b)~.\end{aligned}$$ These equations show that the only energy scale in the problem is $gN$. On the basis of simple dimensional analysis we therefore expect $\bar{\tau}\propto 1/(gN)$. Accordingly, since the energy of the system is extensive, this will yield $\bar{P}_{\parallel} \propto N$ while $\bar{P}_{\sharp}\propto N^2$ leading to $\Gamma_{\rm cl} \propto N$. This argument is not asymptotic, i.e. does not only apply for $N\gg 1$. In Fig. \[fig:3\](b) we plot the classical collective advantage obtained by solving numerically Hamilton’s equation of motion. Indeed, we clearly see a linear growth in $N$, also for small values of $N$, perfectly consistent with the dimensional argument.
Finally, in Fig. \[fig:3\](c) we show the ratio $R$ defined as in Eq. (\[R\]), for the case of our spin batteries. We conclude that, for this model, quantum mechanical dynamics yields a [*disadvantage*]{} rather than an advantage, as $R<1$ for all $N$. This is the second main result of this Article.
Dicke batteries {#sect:Dicke}
===============
[Figs/GammaQuantumLL.pdf]{} (1,70)[(a)]{}
[Figs/GammaClassicalLL.pdf]{} (1,70)[(b)]{}
[Figs//RDicke.pdf]{} (1,70)[(c)]{}
[Figs/RDickeG1.pdf]{} (1,70)[(d)]{}
In this Section we study the case of Dicke batteries [@Ferraro17; @Andolina18b]. In a Dicke QB, one cavity mode, acting as charger, is coupled to $N$ qubits, which play the role of the battery. The quantum Hamiltonian is [@Ferraro17] (see also Refs. ) $$\begin{aligned}
\label{eq:Dicke}
\mathcal{H}_{\rm A}&=&\omega_0 ~a^\dagger a~,\nonumber\\
\mathcal{H}_{\rm B}&=&\omega_0\left(J_z+\frac{N}{2}\right)~,\nonumber\\
\mathcal{H}_{1}&=&2g \left(a^\dagger+a\right)J_x~,\end{aligned}$$ where $J_\alpha$ with $\alpha=x,y,z$ are the components of a collective spin operator of length $J=N/2$, while all the other parameters have the same meaning as in Eq. . As in the other models introduced in previous Sections, we choose $E_{\rm A}^{(1)}(0)=\omega_0$. Moreover, for the sake of simplicity, we fix $\ket{\psi}_{\rm A}$ to be a Fock state. In Ref. it was shown that the particular choice of the initial state does not change qualitatively the collective advantage. While a detailed analysis of Dicke QBs is reported in Ref. , here we summarize the main findings—Fig. \[fig:2\](a)—and compare them with those obtained for the classical analog of a Dicke QB.
In Fig. \[fig:2\](a) we plot the collective advantage $\Gamma_{\rm qu}$ of a Dicke QB for different choices of the coupling parameter $g$. In agreement with Ref. , fits to the numerical data (not shown) suggest the following power-law scaling in the limit of large $N$ $$\begin{aligned}
\label{eq:GammaDicke}
\Gamma_{\rm qu}\propto \sqrt{N}~.\end{aligned}$$ We now analyze the classical case. In the literature there is a well-established classical analog of the Dicke model [@deAguiar92; @Rodriguez18; @Carlos18], which reads as follow $$\begin{aligned}
\label{eq:DickeCl}
\mathcal{H}^{\rm cl}_{\rm A}&=&\frac{\omega_0}{2}\left(p_a^2+q_a^2\right),\nonumber~\\
\mathcal{H}^{\rm cl}_{\rm B}&=&N\omega_0\frac{\big[\cos(\theta)+1\big]}{2}\nonumber~,\\
\mathcal{H}^{\rm cl}_{1}&=&g\sqrt{2}Nq_a \sin(\theta)\cos(\phi)~,\end{aligned}$$ where $(p_a,q_a)$ and $(N\cos(\theta)/2,\phi)$ are classical conjugate variables [@BraunBook; @Carlos18]. This Hamiltonian describes a spring coupled to a nonlinear pendulum of length $N$.
We would like to stress that the model defined by Eq. (\[eq:DickeCl\]) is not a semi-classical approximation of the quantum Hamiltonian in Eq. (\[eq:Dicke\]), but represents instead an intrinsically classical description of a classical spin coupled to a cavity, directly obtainable from classical Hamiltonian mechanics. Our aim is indeed not to approximate the quantum model, but to understand the differences between the quantum and the classical batteries.
As in all previous cases, we choose $E_{\rm A}^{(1)}(0)=\omega_0$. We still have the freedom to choose initial conditions, since the previous condition imposes only the constraint $p_a^2(0)+q_a^2(0)=2N\omega_0$. For the sake of simplicity, we choose $p_a(0)=q_a(0)$. We have checked that other initial conditions do not alter our main conclusions.
From Eqs. and we find Hamilton’s equations of motion for the classical Dicke battery: $$\begin{aligned}
\label{eq:DickeHJ}
&&\frac{d{p}_a}{dt}=-\omega_0q_a-\sqrt{2}Ngq_a\sin(\theta)\cos(\phi)~,\nonumber \\
&&\frac{d{q}_a}{dt}=\omega_0p_a~,\nonumber \\
&&\frac{d\cos(\theta)}{dt}=2\sqrt{2}gq_a\sin(\theta)\sin(\phi),\nonumber \\
&&\frac{d\phi}{dt}=\omega_0-2\sqrt{2}gq_a\cos(\phi)\cot(\theta)~.\end{aligned}$$ We can rescale these equations in such a way to have $P^{2}_{a}(0)+Q^{2}_a(0)=2$, i.e. $P_a=\sqrt{N}p_a$ and $Q_a=\sqrt{N}q_a$. We obtain: $$\begin{aligned}
\label{eq:DickeHJT}
&&\frac{d{P}_a}{dt}=-\omega_0Q_a-\sqrt{2}g_NQ_a\sin(\theta)\cos(\phi)~,\nonumber \\
&&\frac{d{Q}_a}{dt}=\omega_0P_a~,\nonumber \\
&&\frac{d\cos(\theta)}{dt}=2\sqrt{2}g_NQ_a\sin(\theta)\sin(\phi),\nonumber \\
&&\frac{d\phi}{dt}=\omega_0-2\sqrt{2}g_NQ_a\cos(\phi)\cot(\theta)~,\end{aligned}$$ where $g_{N}$ has been defined in Eq. (\[eq:g\_N\]). We note that, in these equations, the only parameters with physical dimensions (of energy) are $\omega_0$ and $g_N$. Since $\bar{\tau}$ has physical dimensions of inverse energy (in our units), the optimal charging time must have the following form: $$\label{DEFEFFE}
\bar{\tau}=\frac{1}{g_{N}}F(\omega_0/g_N)~,$$ where $F(x)$ is an unknown dimensionless function. From this expression we can conclude that, as long as $F(x)$ does not reach zero for $x=0$, also in the classical scenario the collective advantange parameter will exhibit a $\sqrt{N}$ scaling similar to the one in Eq. (\[eq:GammaDicke\]) observed for the quantum counterpart, i.e. $\Gamma_{\rm cl} \propto \sqrt{N}$. Indeed, assuming $F(0)\neq 0$, from (\[DEFEFFE\]) it follows that for large enough $N$ the charging time can be approximated as $\bar{\tau}\simeq F(0)/g_N$ with a $1/\sqrt{N}$ scaling. Accordingly, since the energy is an extensive quantity, we will have, asymptotically, $\bar{P}^{(N)}_{\rm B}\propto N\sqrt{N}$, which implies $\Gamma_{\rm cl} \propto \sqrt{N}$ as anticipated. To put this observation on a firmer ground, we resort to numerical integration of Eqs. . In Fig. \[fig:2\](b) we plot the collective advantage $\Gamma_{\rm cl}$ as a function of $N$, for different values of $g$. A comparison with the expected $\sqrt{N}$ scaling of $\Gamma_{\rm cl}$ in the large-$N$ limit is also shown. (The expected saturation to the $\sqrt{N}$ scaling law requires $g_{N}/\omega_{0}\gg1$ and is therefore difficult to reach numerically for small values of $g/\omega_{0}$.)
We now proceed with a more quantitive comparison between $\Gamma_{\rm qu}$ and $\Gamma_{\rm cl}$. In Fig. \[fig:2\](c) we report the plot of the quantity $R$ of Eq. (\[R\]) as a function of $N$, for different values of $g$. We clearly see that the ratio $R$ can be smaller or larger than unity depending on the value of $g$. This is emphasized in Fig. \[fig:2\](d), where we show $R$ as a function of $g$ for $N=50$. This is the third main result of this Article. The quantum advantage shown by a Dicke QB in a window of values of $g$ is on the order of $10\%$ and therefore not spectacular but clearly indicates the possibility to engineer more complex quantum Hamiltonians to achieve much better quantum performances. These will be the subject of future work.
Summary and conclusions {#concl}
=======================
In this Article we have compared three quantum battery models against their rigorous classical versions in order to better understand the origin of the fast charging phenomenon discussed in previous literature.
In particular, we have defined a genuine [*quantum advantage*]{} (i.e. $R>1$) via the ratio $R$ in Eq. (\[R\]) between the collective advantages in the quantum and classical cases, $\Gamma_{\rm cl}$ and $\Gamma_{\rm qu}$, respectively.
In the case of harmonic oscillator batteries—see Sect. \[sect:HO\]—$R=1$ for all values of $N$ and $g$. Quantum harmonic oscillator batteries defined as in Eq. (\[eq:Hhoho\]) do not therefore display any quantum advantage. The case of spin batteries, discussed in Sect. \[sect:spin\], is even worse. In this model, indeed, $R<1$ for all values of $N$ and $g$.
We can safely conclude that, in these two cases, fast charging in the quantum case (i.e. the fact that $\Gamma_{\rm qu}$ increases for increasing $N$) is solely due to the collective behavior of the many-body systems described by the quantum Hamiltonians in Eqs. (\[eq:Hhoho\]) and (\[eq:HSpins\]), which is also present in the corresponding classical Hamiltonians.
The case of Dicke batteries, discussed in Sect. \[sect:Dicke\], is far more richer. In this case, the ratio $R$ depends on the charger-battery coupling parameter $g$ and, for each fixed $N$, can be larger than unity in a range of values of $g$. As evident from Figs. \[fig:2\](c) and (d), the quantum advantage displayed by a Dicke quantum battery at optimal coupling is on the order of $10\%$. More work is needed to discover quantum models of batteries with larger values of $R$.
For the sake of completeness, we note that the authors of Ref. have very recently proposed to study the evolution of the battery state in the energy eigenspace of the battery Hamiltonian. Combining this geometric approach with bounds on the power, they are able to distinguish whether the quantum advantage in a charging process stems either from the speed of evolution or the non-local character of the battery state.
Acknowledgments
===============
Part of the numerical work has been performed by using the Python toolbox QuTiP2 [@QuTip]. We wish to thank D. Farina, D. Ferraro, P. Erdman, M. Esposito, and, especially, M. Bera, V. Cavina, S. Juliá-Farrè, and M. Lewenstein for many useful discussions.
[77]{} R. Alicki and M. Fannes, [Phys. Rev. E [**87**]{}, 042123 (2013)](http://dx.doi.org/10.1103/PhysRevE.87.042123). K.V. Hovhannisyan, M. Perarnau-Llobet, M. Huber, and A. Acín, [Phys. Rev. Lett. [**111**]{}, 240201 (2013)](https://doi.org/10.1103/PhysRevLett.111.240401). F.C. Binder, S. Vinjanampathy, K. Modi, and J. Goold, [New J. Phys. [**17**]{}, 075015 (2015)](http://dx.doi.org/10.1088/1367-2630/17/7/075015). F. Campaioli, F.A. Pollock, F.C. Binder, L. Céleri, J. Goold, S. Vinjanampathy, and K. Modi, [Phys. Rev. Lett. [**118**]{}, 150601 (2017)](http://dx.doi.org/10.1103/PhysRevLett.118.150601). T.P. Le, J. Levinsen, K. Modi, M. Parish, and F.A. Pollock, Andolina18 [Phys. Rev. A [**97**]{}, 022106 (2018)](https://dx.doi.org/10.1103/PhysRevA.97.022106). D. Ferraro, M. Campisi, G.M. Andolina, V. Pellegrini, and M. Polini, [Phys. Rev. Lett. [**120**]{}, 117702 (2018)](https://dx.doi.org/110.1103/PhysRevLett.120.117702). G.M. Andolina, D. Farina, A. Mari, V. Pellegrini, V. Giovannetti, and M. Polini, [Phys. Rev. B [**98**]{}, 205423 (2018)](https://doi.org/10.1103/PhysRevB.98.205423). G.M. Andolina, M. Keck, A. Mari, M. Campisi, V. Giovannetti, and M. Polini, [arXiv:1807.08656](https://arxiv.org/abs/1807.08656). D. Farina, G.M. Andolina, A. Mari, M. Polini, and V. Giovannetti, [arXiv:1810.10890](https://arxiv.org/abs/1810.10890). S. Juliá-Farrè, T. Salamon, A. Riera, M.N. Bera, and M. Lewenstein, [arXiv:1811.04005](https://arxiv.org/abs/1811.04005). Y.-Y. Zhang, T.-R. Yang, L. Fu, and X. Wang, [arXiv:1811.04395](https://arxiv.org/abs/1811.04395). For a recent review see e.g. F. Campaioli, F.A. Pollock, and S. Vinjanampathy, [arXiv:1805.05507](https://arxiv.org/abs/1805.05507). P.A.M. Dirac, [*Principles of Quantum Mechanics*]{} (Oxford University Press, 1982). J. Ch' avez-Carlos, B. L' opez-del-Carpio, M.A. Bastarrachea-Magnani, P. Str' ansky, S. Lerma-Hern' andez, L.F. Santos, and J.G. Hirsch, [arXiv:1807.10292](https://arxiv.org/abs/1807.10292). D. Braun, [*Dissipative Quantum Chaos and Decoherence*](https://doi.org/10.1007/3-540-40916-5) (Springer Tracts in Modern Physics, 2001). C. Ciuti, G. Bastard, and I. Carusotto, [Phys. Rev. B [**72**]{} 115303 (2005)](https://doi.org/10.1103/PhysRevB.72.115303). D.F. Walls and G.J. Milburn, [*Quantum Optics*](https://doi.org/10.1007/978-3-540-28574-8) (Springer Science & Business Media, 2007). R.H. Dicke, [Phys. Rev. [**93**]{}, 99 (1954)](https://doi.org/10.1103/PhysRev.93.99). The Hamiltonian in Eq. has been widely used in the literature [@Carlos18; @Julia-Farre18] with a different normalization of the coupling constant, namely with the replacement $g\to g/\sqrt{N}$. This different choice guarantees well-defined results [@Julia-Farre18] if one works in the thermodynamic limit defined by $N\to \infty$, $L\to \infty$ with $n \equiv N/L = {\rm const}$, where $L$ is the length of the cavity. In this limit, the length of the cavity scales with the number $N$ of qubits in order to keep the density $n$ of qubits constant. Whether one uses Eq. or Eq. with the replacement $g\to g/\sqrt{N}$ ultimately depends on the experimental setup. For example, in circuit-QED setups like the one realized in Ref. , the length of the photonic cavity (i.e. the length of the transmission line resonator) does not scale with the number of qubits (i.e. the number of transmons). Indeed, in Ref. the resonator is $\sim 20~{\rm mm}$ long, while a transmon has a linear size which is on the order of $300~{\rm \mu m}$. This implies that the resonator used in the setup of Ref. can host something like $N = 40$-$50$ qubits, without any need to scale its length with $N$ for their accommodation. The authors of Ref. used the same Hamiltonian as in Eq. and Ref. to explain their data. J.M. Fink, R. Bianchetti, M. Baur, M. Göppl, L. Steffen, S. Filipp, P.J. Leek, A. Blais, and A. Wallraff, [Phys. Rev. Lett. [**103**]{}, 083601 (2009)](http://dx.doi.org/10.1103/PhysRevLett.103.083601). M.A.M. de Aguiar, K. Furuya, C.H. Lewenkopf, and M.C. Nemes, [Ann. Phys. [**216**]{}, 291 (1992)](https://doi.org/10.1016/0003-4916(92)90178-O). J.P.J. Rodriguez, S.A. Chilingaryan, and B.M. Rodr' iguez-Lara, [arXiv:1808.03193](https://arxiv.org/pdf/1808.03193.pdf). J.R. Johansson, P.D. Nation, and F. Nori, [Comp. Phys. Comm. [**184**]{}, 1234 (2013)](https://doi.org/10.1016/j.cpc.2012.11.019).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Theory and observations suggest that star formation occurs hierarchically due to the fragmentation of giant molecular clouds. In this case we would expect substructure and enhanced stellar multiplicity in the primordial cluster. This substructure is expected to decay quickly in most environments, however historic stellar encounters might leave imprints in a protoplanetary disc (PPD) population. In a low density environment such as Taurus, tidal tails from violent star-disc or disc-disc encounters might be preserved over time-scales sufficient to be observed. In this work, we investigate the possibility that just such an event occured between HV Tau C (itself a component of a triple system) and DO Tau $\sim 0.1$ Myr ago, as evidenced by an apparent ‘bridge’ structure evident in the $160$ $\mu$m emission. By modelling the encounter using smoothed particle hydrodynamics (SPH) we reproduce the main features of the observed extended structure [(‘V’-shaped emission pointing west of HV Tau and a tail-like structure extending east of DO Tau).]{} We suggest that HV Tau and DO Tau formed together in a quadruple system on a scale of $\sim 5000$ au ($0.025$ pc).'
author:
- |
Andrew J. Winter, Richard A. Booth, Cathie J. Clarke\
Institute of Astronomy, University of Cambridge, Madingley Road, Cambridge CB3 0HA, UK
bibliography:
- 'References/HVDO.bib'
date:
-
- 'Accepted XXX. Received YYY; in original form ZZZ'
title: 'Evidence of a [past]{} disc-disc encounter: HV and DO Tau'
---
\[firstpage\]
accretion, accretion discs – stars: kinematics and dynamics, formation, circumstellar matter – submillimetre: ISM
Introduction
============
Star formation occurs predominantly in clustered environments from giant molecular clouds . Simulations suggest that stars form in hierarchical fragmentation of these molecular clouds, resulting in small subclusters [e.g. @2003MNRAS.343..413B]. Such subclusters interact dynamically, merging or dispersing over a similar time scale to the star formation [@2010MNRAS.407.1098A; @2011MNRAS.415.1967A]. In this scenario, substructure within a cluster is only directly observable over short time-scales. However, enhanced local stellar density in turn increases the chance of a close encounter between young stars [@2013ApJ...769..150C], which can have significant consequences for the evolution of a circumstellar disc [e.g. @1999MNRAS.304..425A].
The Taurus star forming region contains almost exclusively young stars of age $\lesssim 3$ Myr and is considered an archetype of low-mass star formation, with a low stellar density and long dynamical time [@Bal99]. @1995MNRAS.272..213L and @2008ApJ...686L.111K find evidence for hierarchical structure in Taurus on large scales, but not on smaller scales ($\sim0.04$ pc), and it is hypothesised that structure has been erased by dynamical interactions in this regime. Although star-disc encounters are rare in most young cluster environments [e.g. @Win18b], if this substructure in Taurus did indeed exist in the past then enhanced numbers of early close encounters could leave evidence in the form of truncated discs or tidal tails [e.g. RW Aurigae, @Cab06; @Dai15]. The low stellar density in Taurus also means that there are fewer disrupting influences, and any tidal tails produced in historic interactions may be preserved for periods long enough to be observed.
Photometric observations of HV and DO Tau, which have a present day separation of $90.8''$ ($0.06$ pc), by @2013ApJ...776...21H using the Photodetector Array Camera and Spectrometer (PACS) of the *Herschel Space Observatory* were made at $70$ $\mu$m, $100$ $\mu$m and $160$ $\mu$m (Figure \[herschel\]). The extended emission from each component, HV and DO, is directed towards the other, with a common envelope or ‘bridge’ (i.e. emission connecting the two) visible at $160$ $\mu$m. While imaged at low resolution, the structure observed is reminiscent of tidal tail structures found in simulations of close encounters between disc-hosting stars [@1993MNRAS.261..190C; @2015MNRAS.446.2010M].
The following is an investigation of the hypothesis that DO Tau plus the 3 stars comprising HV Tau were originally formed as a bound hierarchical multiple, and that the present morphology of the system can be explained in terms of a close, disc mediated encounter and subsequent ejection of DO Tau from the system. We aim to replicate observations using hydrodynamical modelling in order to understand the nature of such an interaction in terms of the disc geometry and stellar kinematics.
Observational Constraints {#sec:obs}
=========================
Stellar Components
------------------
HV Tau is a young triple system in Taurus. It is comprised of a tight optically bright binary AB, projected separation $10$ au [@Sim96], and a third star HV Tau C at approximately $550$ au separation with common proper motion [@Duc10]. The tight binary has an estimated age $2$ Myr and a combined mass of $\sim0.6M_\odot$ [@Whi01]. The separation of AB could be larger than $10$ au due to orbital eccentricity or deprojection, as suggested by a comparatively long orbital period [@Duc10]. A mass of $0.5-1M_\odot$ is inferred from the CO maps of the edge on disc of HV Tau C [@Duc10]. It is observed to be exceptionally red, with a high accretion rate [@Woi98; @Mon00].
DO Tau is a G star located at a projected distance $1.26 \times 10^{4}$ au ($90.8$” at $140$ pc) west of HV Tau, which has position angle $95.3^{\circ}$ relative to DO. Mass and age estimates range between $0.3M_\odot$, $0.16$ Myr [@Har95] and $0.7M_\odot$, $0.6$ Myr [@Bec90]. The whole system is depicted with the components labelled in Figure \[fig:schem\].
Disc Properties
---------------
![\[herschel\] Produced using the data discussed in @2013ApJ...776...21H. The top two images are the specific intensity in the $100$ $\mu$m, and $160$ $\mu$m overlaid with logarithmic contours. Both stars appear to be associated with extended emission. The edge of the image is close to DO Tau (east), which results in excess noise. The bottom panel is the inferred dust temperature distribution assuming that the cloud is optically thin, likely yielding an overestimate close to the stars. The point spread function (PSF) in the $100$ $\mu$m observations also lead to noise in the temperature determination in these regions.](paper_figure_obs.pdf){width="50.00000%"}
@2015ApJ...808..102K used CARMA observations and models to deduce properties of $6$ protoplanetary discs, including DO Tau. Their models found an outer disc radius of $\sim75$ au and consistent values for mass $M_{\mathrm{disc}} \approx 0.013 \, M_\odot$, inclination $\sim -33^{\circ}$, and position angle $\sim 90^{\circ}$, following the convention as described by . There remains ambiguity as to which side of the disc is closer to the observer as the quoted negative inclination angle can produce two rotation senses with the same aspect ratio.
HV Tau A and B have no associated infrared excess and therefore are not expected to host a substantial disc, while C has an edge on disc of radius $50$ au and mass $\sim 2 \times 10^{-3} \, M_\odot$ [@Woi98; @Sta03]. @Mon00 find that the observed disc radius does not depend on wavelength. This suggests the disc has been truncated, as otherwise the grain size-dependent radial drift of dust particles leads to a wavelength-dependent disc extent. To the contrary they note that the ratio of disc size to projected separation between C and close binary AB is $R_{\mathrm{disc}}/x_{\mathrm{min}} \equiv R_{\mathrm{tidal}} \sim 0.1$, where $R_{\mathrm{disc}}$ ($= 50$ au) is the outer disc radius, and $x_{\mathrm{min}}$ is the closest approach distance. This makes truncation due to tertiary interaction at the current separation unlikely as a ratio of around $R_{\mathrm{tidal}} \approx 0.35$ is expected if the masses of C and combined AB are equal . [It remains possible that the orbit of AB is highly eccentric, and that the periastron distance is sufficiently small to cause tidally induced truncation. Alternatively, an historic encounter may have left the disc truncated.]{}
In modelling the disc around HV Tau C, @Duc10 find an inclination $\theta_i \approx 80^{\circ}$ and PA of approximately $20^{\circ}$, corresponding to an orientation such that the blue shifted side of the disc is pointing east with the northern side closer to us. It is further noted that the coplanarity of the centre of mass of AB and the disc of C is unlikely as the nearly edge on angle would lead to a very large actual separation. @Duc10 also suggest that scattered light images might imply a disc size greater than $50$ au, and gas emission alone suggests a radius up to $100$ au. A model with temperature profile $T\propto R^{-q}$ is found to fit well with $0.4<q<0.6$ and a temperature at $50$ au of $15$-$30$ K.
HST and Herschel/PACS Images
----------------------------
The Herschel/PACS survey observations of HV/DO Tau are discussed by @2013ApJ...776...21H, and we use that data to produce Figure \[herschel\]. At $160$ $\mu$m the extended emission connects HV and DO in a common envelope. Of particular interest is the ‘V-shaped’ emission close to HV Tau and the tail to the North-East of DO Tau (see Figure \[fig:schem\]), seen clearly at $100$ and $160$ $\mu$m, which we aim to reproduce as the result of a disc-disc interaction producing two tidal tails.
It has been shown in numerous studies that two tails, or a ‘bridge’ and an external arc, can be produced as a result of prograde or inclined encounters [@1972ApJ...178..623T; @1993MNRAS.261..190C; @2015MNRAS.446.2010M]. Observed morphology is dependent on viewing angle and interaction parameters. Angular momentum transfer between star and disc, and therefore the quantity of circumstellar material ejected during an encounter, is a strong function of the closest approach distance [@Ost94; @Win18]. As we will discuss in Section \[sec:massest\], we expect a collision between the discs, as opposed to a distant encounter, is required to produce the observed emission.
![\[fig:schem\]Schematic diagram of the $160$ $\mu$m dust emission structure visible in Figure \[herschel\] with positions of the stellar components overlaid. The diagram is simplified to highlight the features which we aim to reproduce in our models. HV Tau is a system of three stars, the tight binary HV-AB shown here as one point has a projected separation of $\sim 10$ au. HV-C has a PA of $\sim 45^\circ$ with respect to HV-AB, and HV has a PA of $95.3^\circ$ with respect to DO.](HVDO_schem.pdf){width="50.00000%"}
Cloud Temperature and Mass {#sec:massest}
--------------------------
![\[fig:tdust\] Distribution of the dust temperature of each pixel in Figure \[herschel\] as a function of separation from HV Tau (red) and DO Tau (blue). The error bars are the $1\sigma$ range in separation and temperature for a given bin of pixels. Close to the star the optical depth and the PSF result in considerable errors in the determination of temperature. ](paper_figure_tprof.pdf){width="50.00000%"}
To compare the mass in the envelope of our model to that of the observations, we reproduce the expected flux at 100 $\mu$m and 160 $\mu$m using the methods outlined by @1983QJRAS..24..267H. The specific intensity of radiation at frequency $\nu$ across the envelope can be written: $$I_\nu = \left( 1- e^{-\tau_\nu} \right)B_\nu (T_\mathrm{dust})$$ where $B_\nu(T_\mathrm{dust})$ is the Planck disribution at a given dust temperature $T_\mathrm{dust}$, and $\tau_\nu$ is the optical depth of the dust. The latter can be rewritten $\tau_\nu = \kappa_\nu \Sigma_\mathrm{dust}$ if we assume that $\kappa_\nu$ is spatially uniform.
We make estimates of dust mass and temperature by assuming that $\Sigma_\mathrm{dust}$ is sufficiently small such that the cloud is optically thin ($1- e^{-\tau_\nu}\approx \kappa_\nu \Sigma_\mathrm{dust}$). While this approximation is useful away from the stars (a posteriori we find $\Sigma_\mathrm{dust} \sim 10^{-4}$ g cm$^{-2}$ in this region) it is likely to break down locally to HV and DO Tau where $\Sigma_\mathrm{dust}$ is large. For this reason, when we come to presenting our models and final mass estimates (Section \[sec:hydromodel\]) we will produce an intensity map from the simulation data for comparison with observations. For the two frequencies $\nu_1 = c/100~\mu$m, $\nu_2 = c/160~\mu$m we use the opacity of spherical dust grains with radius $a$ following a power law distribution $n(a) \propto a^{-3}$ between $a_\mathrm{min}=10$ nm and $a_\mathrm{max}=1.023$ cm as computed by . The models in that work are based on abundances appropriate for a circumstellar disc described by @Pol94.
The measured intensities are integrated over the normalised transmission spectra for PACS $S_{\nu_{1,2}}$: $$I_{\nu_{1,2}} = \frac{\int I_\nu(\nu) S_{\nu_{1,2}}(\nu) \, \mathrm{d} \nu}{\int S_{\nu_{1,2}}(\nu) \, \mathrm{d} \nu}$$ and hence $$\frac{I_{\nu_1}}{I_{\nu_2}} \approx \frac{\int B_\nu(\nu;T_\mathrm{dust}) \kappa_\nu (\nu) S_{\nu_{1}}(\nu) \, \mathrm{d} \nu}{\int B_\nu (\nu; T_\mathrm{dust}) \kappa_\nu (\nu)S_{\nu_{2}}(\nu) \, \mathrm{d} \nu} \cdot \frac{\int S_{\nu_{2}}(\nu) \, \mathrm{d} \nu}{\int S_{\nu_{1}}(\nu) \, \mathrm{d} \nu}.$$ We invert this expression to estimate the temperature at each pixel. The result is shown in the bottom panel of Figure \[herschel\]. The point spread function (PSF) of the $100$ $\mu$m observations combined with the greater optical depth result in considerable errors close to the stars. However, by plotting the pixel temperature against projected distance from the nearest star we find evidence for a temperature gradient within the cloud (as expected, Figure \[fig:tdust\]).
Once we have the temperature in each pixel we can determine the column density of dust that is required to match the observed emission map. This can only be performed on regions that are optically thin, and for those in which we have detections at both $100$ $\mu$m and $160$ $\mu$m. We find a dust mass of $\sim 1 -5\times 10^{-4} \,M_\odot$, depending on assumed values of $T_\mathrm{dust}$. For a dust to gas ratio $\Sigma_\mathrm{dust}/\Sigma_\mathrm{gas} = 10^{-2}$ this yields an estimate of the total cloud mass of $M_\mathrm{cloud} \gtrsim 10^{-2} \, M_\odot$. This is greater than the total present day mass of the disc around DO Tau, and would suggest that a large fraction of the circumstellar material has been ejected into the ISM (or possibly accreted onto the stellar components) during the hypothesised past encounter. [However, if the material originates in discs, the dust to gas ratio could be enhanced]{} and our derived cloud mass would be an overestimate.
[Based on the relative intensity of the $100$ $\mu$m and $160$ $\mu$m emission we further find evidence that the extended structure originated in a circumstellar environment. We repeat our mass estimates with opacities calculated from an ISM dust grain distribution $n(a) \propto a^{-3.5}$]{}, and a maximum grain size $a_\mathrm{max}= 1$ $\mu$m . [Such a calculation yields lower temperatures ($\sim 10$-$20$ K) throughout the cloud and a dust mass of $\gtrsim 5 \times 10^{-3} M_\odot$ (or a total cloud mass of $\gtrsim 0.5 M_\odot$). This total mass is extremely large, and physically unlikely given the emission is associated with the stellar components of similar mass. Further, we estimate the Jean’s mass:]{} $$M_\mathrm{J} \approx 2 M_\odot \left( \frac{c_\mathrm{s}}{0.2 \, \mathrm{ km/s}} \right)^3 \sqrt{ \frac{10^3 \, \mathrm{cm}^{-3}}{n_\mathrm{H}}}$$ [where $n_\mathrm{H}$ is the number density of hydrogen, and the sound speed $c_\mathrm{s} \approx 0.5$ km/s for a gas with $T = 15$ K. If the total mass is $0.5$ $M_\odot$ and the volume is $\sim 10^4 \times 2\cdot 10^3 \times 2\cdot 10^3$ au$^{3}$ this yields $M_\mathrm{J} \sim 0.5 \, M_\odot \sim M_\mathrm{cloud} $. The free-fall time-scale in this case is $t_\mathrm{ff} \sim 0.03$ Myr, which is much smaller than the age of the stars. Such a cloud could be interpreted as residual material from an initial star forming core, however it is unclear whether such material could be supported against gravitational collapse on this time-scale. In addition, this interpretation offers no clear mechanism for the formation of the apparently tidal morphology. We therefore focus on the hypothesis that the material between the two systems originated in the dics around HV-C and DO.]{}
Kinematics {#sec:relvels}
----------
[The proper motions DO Tau and the (unresolved) binary AB in HV Tau are recorded in *Gaia DR2*]{} [@GaiaMis_16; @GaiaDR2_18; @Lin18]. [DO Tau has a velocity in declination $v_{\delta \mathrm{, DO}} = -21.340 \pm 0.091$ mas/yr and in right ascension $v_{\alpha \mathrm{, DO}} = 6.128 \pm 0.126$ mas/yr. HV Tau AB has $v_{\delta \mathrm{, HV}} = -21.783 \pm 0.171$ mas/yr and in right ascension $v_{\alpha \mathrm{, HV}} = 4.888 \pm 0.126$ mas/yr. This yields $\Delta v_\delta = v_{\delta \mathrm{, DO}} - v_{\delta \mathrm{, HV}} = 0.29\pm 0.17$ km/s and $\Delta v_\alpha = v_{\alpha \mathrm{, DO}} - v_{\alpha \mathrm{, HV}} = 0.82\pm 0.24$ km/s. If the velocity vector was anti-parallel to the position vector (i.e. the systems were moving away from each other) we would expect $\Delta v_\delta \gtrsim 0$ and $\Delta v_\alpha < 0$. However, as mentioned the HV-A and -B are unresolved and multiplicity introduces uncertainties into the center of mass velocity of HV, for which an upper bound is set by the relative velocity of the AB pair]{} [$\sim 1.5$ km/s; @Duc10]. [Hence the kinematic constraints are consistent with common proper motion of the two systems. Based on the projected separation, the escape velocity is $\sim 0.4$ km/s, and it is possible that HV and DO Tau are marginally bound or unbound. The one dimensional velocity dispersion in the Taurus region is estimated to be $\sigma_v \sim 2$-$4$ km/s, although the value is uncertain due to difficulty in establishing membership]{} [@Ber06; @Riv15]. The relative proper motion components of HV and DO, which are both considerably less than this, hint at a common origin.
[No radial velocity measurement for either star is present in the *Gaia DR2*. DO Tau is estimated to have a radial velocity of $16.04 \pm 0.17$ km/s by @Ngu12, however no such estimate exists for HV Tau. Therefore we cannot place constraints on the geometry of the system using the radial velocity differential.]{}
Summary of Observational Constraints
------------------------------------
We identify the following key criteria to consider in addressing the possibility of a previous tidal encounter.
- For any given parameters of a proposed fly-by, the time of the interaction should not be older than the age of the stars. Because our hypothesis requires that the stars are coeval, we already assume considerable error in the claimed ages. However, $0.16$ Myr is the lowest age estimate for any of the stellar components, and so any interaction time-scale smaller than this is feasible. Longer time-scales may also be reasonable if this is an underestimate of the age of DO Tau.
- Disc orientations should be approximately consistent with the observations, although we note that modelling the evolution of a violent encounter over a long period of time introduces considerable uncertainty in obtaining present day orientation. To obtain a feasible solution we are motivated to explore solutions for which the disc around HV Tau C is edge on, with the plane of the disc aligned with the extended emission, while the disc around DO Tau is face on.
- Solutions for the stellar kinematics should be consistent with the present size of the disc around HV Tau C, and hence we do not expect to see tight binary HV Tau AB orbiting C post-interaction such that $R_{\mathrm{tidal}} >0.5$, where $R_{\mathrm{tidal}}$ is here the ratio of observed disc size ($\sim50$ au) to closest approach. The closest separation between HV Tau C and DO should not be considerably less than twice the outer radius of the disc around DO Tau - i.e. $150$ au. Although it is possible that the viscous spreading of this disc may have an impact on its present extent.
- When recovering a flux from the surface density distribution in a given model, the dust to gas ratio required to reproduce the same flux as in the $100$ $\mu$m and $160$ $\mu$m and initial total disc mass should be sensible, and consistent between wavelengths.
- The parameters of such an interaction should be capable of producing common envelope surrounding both stars with the structure seen in Figure \[herschel\]. Although it may not be possible to reproduce the structure precisely, especially if the binary HV-AB has a significant effect, the aim of the modelling process is to show that the observations can feasibly result from a disc-disc interaction.
Numerical Method {#sec:method}
================
The complexity of the HV/DO system is approached by dividing the problem into a kinematics study of the stellar components, and hydrodynamical modelling of star-disc and disc-disc interactions. For the hydrodynamics we apply a smoothed particle hydrodynamics (SPH) treatment of the gas particles. Its computationally expensive nature means that we cannot rely on Markov Chain Monte Carlo (MCMC) or similar statistical techniques to constrain the parameters which yield the observed structure. [A large number ($\sim 500$) of low resolution models with $10^4$ particles are explored to find a promising configurations for which ejected material approximately traces the observed structure, allowing variation in disc orientations and surface density profiles]{} (see Section \[sec:discics\]). Subsequently we rerun promising models with a resolution of $10^6$ particles and refining the disc properties and viewing angles to establish a model that yields extended structure closest to observations.
Kinematic Modelling {#sec:kmod}
-------------------
[The first stage in obtaining a model is exploring the kinematic parameter space of a multiple encounter of a three star system (DO, HV-C and HV-AB, the latter we will consider one star - see below) to find solutions which satisfy the dynamical conditions discussed in Section]{} \[sec:obs\]. [As in the case of the hydrodynamics, we cannot use an MCMC exploration of the kinematic parameter space due to the chaotic nature of the three body problem. Instead we search for a (probably non-exhaustive) library of kinematic solutions for further hydrodynamical modelling. We do this by uniformly varying parameters which describe the initial conditions of the three bodies and checking for consistency with observations. Viable solutions are expected to be initially bound, but we do not have further a priori constraints. We apply the following parametrisation of the problem (sampling uniformly over each within the defined range) as it allows us to minimise the size of the exploration space by choosing likely ranges, with the caveat that drawing statistical conclusions from our kinematic library is problematic.]{} We simulate the trajectories of the three star particles by applying the $N$-body 4$^{\mathrm{th}}$ order Hermite integrator [@1992PASJ...44..141M] in the <span style="font-variant:small-caps;">gandalf</span> code [which is also used for the SPH simulations described in Section \[sec:hydromodel\], @Hub18].
![\[fig:icdiag\] [Schematic diagram illustrating the parameters used to define the initial conditions for our 3-body simulations. The blue line traces the HV/DO trajectory, with coordinates centred on the centre of mass of the HV system. The red line traces the HV-AB/C trajectory. The circular markers represent the locations of the components of each orbit at the time of the closest approach between DO Tau and the centre of mass of HV (blue circles). The positions of HV-AB and -C are shown as red circles. The angles as discussed in the text are annotated]{}](paper_figure_paramschem.pdf){width="50.00000%"}
[Our parametrisation is described below, and illustrated in Figure]{} \[fig:icdiag\]. Firstly, we are helped by the small separation of the binary AB, which we hereafter consider as a single star with the combined mass. With this approximation all stellar components now have the same mass within errors, and this is estimated to be $0.7 \, M_{\odot}$. In order to parametrise the interaction of the three remaining stellar components, we consider two distinct orbital equations of the form $$\label{orbit}
x = \frac{h^2}{\mu}\left( \frac 1 {1+ e \cos\left( \phi -\theta\right)}\right)$$ for HV and for HV/DO, where HV is the orbit of HV-C and HV-AB, while HV/DO is the ‘two-body’ system comprised DO and the centre of mass of HV. In Equation \[orbit\], $x$ is the separation between bodies, $\phi$ is phase, $\theta$ is the angle of the periastron in the plane (equivalent to rotation in the $z$-axis), $h$ is the specific angular momentum and $\mu = G(m_1 +m_2)$. For HV/DO we fix $\theta =0^\circ$. In the case of HV, the orbit of C and AB is rotated in the $y$-axis by angle $i$ and in the $x$-axis by angle $\omega$. The final parameter $\Delta \phi$ is defined as the difference in phases as DO reaches periastron, with respective separation $x$ found for the initial conditions by integrating back along the arc of both orbits. This leaves 8 initial values fully parameterising the system: $e^{\mathrm{HV/DO}}_0$, $x_{\mathrm{min}}^{\mathrm{ HV/DO}}$, $e^{\mathrm{HV}}_0$, $x_{\mathrm{min}}^{\mathrm{HV}}$, $\theta^{\mathrm{HV}}$, $i^{\mathrm{HV}}$, $\omega^{\mathrm{HV}}$, $\Delta \phi$.
----------------- -------------------------------------- ------------------------ ----------------------------------- --------------------- ---------------------------------- ----------------------------- ---------------------------------- -------------------------
$x_{\mathrm{min}}^\mathrm{HV/DO}$/au $e^{\mathrm{HV/DO}}_0$ $x_{\mathrm{min}}^\mathrm{HV}$/au $e^{\mathrm{HV}}_0$ $\theta^{\mathrm{HV}}$ /$^\circ$ $i^{\mathrm{HV}}$ /$^\circ$ $\omega^{\mathrm{HV}}$ /$^\circ$ $\Delta \phi$ /$^\circ$
\[0.5ex\] Range $0$-$2000$ $0$-$1$ $100$-$1500$ $0$-$1$ $0$-$360$ $0$-$360$ $0$-$180$ $0$-$360$
\[1ex\]
----------------- -------------------------------------- ------------------------ ----------------------------------- --------------------- ---------------------------------- ----------------------------- ---------------------------------- -------------------------
[The ranges for each parameter over which we search for successful kinematic solutions are summarised in Table]{} \[table:par\_range\]. We focus on the solutions for which DO is initially bound to HV ($e^{\mathrm{HV/D0}}_0<1$) [as they offer the most likely scenarios for a close encounter between stellar components. Further, highly hyperbolic encounters in a low density stellar environment are physically unlikely.]{} We apply one further restriction that configurations for which the energy of the HV initial orbit exceeds the energy of the DO trajectory are discounted. This is both because in this regime our orbital parametrisation does not make physical sense, and because our investigation finds that solutions for which the orbital energies are comparable are also relatively rare. We search uniformly over the remaining parameter space for successful solutions.
Our criteria for a ‘successful’ kinematic solution are as follows. A lower limit of $50$ au is placed on all interactions as this is a conservative constraint, a distance below which either disc would be significantly over-truncated. Additionally an upper limit on the closest approach between HV Tau C and DO Tau is set at $300$ au. This is motivated both by the present day disc outer radii and the study of and our own findings that a close flyby is required to produce the observed extended structure in the tidal tails (see Section \[sec:sph\]). After encounter, DO must either be unbound from the whole system, or reach a maximum separation $>1.2 \times 10^4$ au. HV Tau C and AB must remain bound. Acceptable final maximum separation of HV wide binary is defined to be between $400$ and $1500$ au, in line with observed projected separation of $550$ au. A minimum periastron distance is placed at $125$ au to prevent over-truncation of the disc around HV Tau C.
Hydrodynamics Model {#sec:hydromodel}
-------------------
The SPH code <span style="font-variant:small-caps;">gandalf</span> is used in the simulation of the discs [@Hub18]. It is adapted here to include a locally isothermal equation of state as a function of radial separation from the nearest star. Self-gravity is disregarded, the gravitational potential being dominated by the stellar component.
Artificial viscosity parameters as prescribed by @1997JCoPh.136...41M are applied to minimise the effects of viscous diffusion in the tidal tails. However, inevitably at the required integration times on the order of $0.1$ Myr, the effect of numerically accelerated viscous spreading and magnified inter-particle torques will result in a loss of structure. This is especially the case where there is considerable mass loss from the disc, as during the violent interactions necessary to produce significant external structure.
Disc Interaction Initial Conditions {#sec:discics}
-----------------------------------
@2005ApJ...629..526P showed that for discs in which there is significant mass transfer one cannot analogously extrapolate structure from star-disc interactions, and hence both discs are required simultaneously for all models where closest approach is of order the disc radius. For disc-disc simulations the work of @2015MNRAS.446.2010M offers a starting point in terms of the expected closest approach between HV Tau C and DO Tau, where extremely close interactions with $R_{\mathrm{tidal}}\sim 10.0$ both result in the near-destruction of the original discs and also in significant sapping of orbital energy and stellar capture (although a large disc mass approximately $10$% of the star mass is used in this study). Conversely, encounters with a wide closest approach such that $R_{\mathrm{tidal}} < 0.5$ do not produce significant external structure.
[Due to the uncertainty in the line of sight separation (and therefore the angle of orientation) of the present day system, the appropriate disc orientations are not immediately clear. For the initial conditions of the three star encounter, a snapshot is taken from an appropriate kinematic model at a time before close encounter. In order to ensure that discs are dynamically settled prior to the encounter, this time is chosen to be five orbital periods at the radius of the outer disc before closest approach between any two stellar components. The discs around HV Tau C and DO are added at an orientation which matches the present day orientation if the two stellar systems are in the plane of the sky. The simulation is then continued with SPH discs included to examine the hydrodynamic evolution of the multiple star interaction. Subsequently disc orientations in promising models are modified to better match the extended structure.]{}
The surface density profile of the discs is both important to the structure and quantity of ejected material, and hard to constrain given that it may be significantly altered in a close interaction. It is treated as a power law such that $$\Sigma = \Sigma_0 \left(\frac{R}{R_0}\right)^{-p}$$ where both ‘shallow’ ($p=0$) and ‘steep’ ($p=1$) surface density gradients are tested.
Temperatures in the disc are defined by distance to the nearest star by $$\centering
T =\mathrm{max}\left\{T_0 \left( \frac R R_0\right)^{-q}, 15\mathrm{~K}\right\}$$ with a value of $q=0.6$ and a temperature at $50$ au of $20$ K is adopted for HV Tau C and the same profile assumed for DO Tau. Variations in temperature are expected only to have a modest effect on the observed structure as a result of star-disc interaction [@Dai15]. Our choice of temperature profile for the hydrodynamic simulations is based on the observations by @Duc10 and is lower than the observed temperature through the extended cloud discussed in Section \[sec:massest\]. This discrepancy could be due to heating of the ejected material during the disc-disc encounter, which we do not model here as there are considerable uncertainties in the temperature estimates. The temperature in both the disc and the cloud are both empirically derived and therefore represent reasonable choices.
Outer radii of the discs prior to interaction are not well constrained, as it is unknown the proximity of the closest approach and therefore the extent of truncation by the initial fly-by. Further, the post-interaction relaxation of the disc, including viscous spreading and possible further dynamical binary interactions in the case of HV-C, is not well characterised. To eject sufficient material to produce observed structure, initial tests suggest that $R_{\mathrm{out}}$ such that $R_{\mathrm{tidal}}\equiv R_{\mathrm{disc}}/x_{\mathrm{min}}\approx 0.8$ is reasonable. This is the initial estimate for a given kinematic model, and the outer radii are subsequently tuned to fit observations. The inner radius is defined to be $R_{\mathrm{disc}}/20$. Choosing a conservative inner radius is necessary given that a significant proportion of the discs pass though each other. [The smoothing lengths of the sink particles are chosen to be half of the inner radius of the disc with the smallest extent. ]{}
The final parameter required to define the disc interactions is the relative masses of the two discs (i.e. how many SPH particles each contains). For each configuration we allow the mass ratio to vary.
Modelling Results {#sec:sph}
=================
Before presenting our chosen model, we note that while we will refer to it as the ‘best-fitting model’, this is in the sense that it best matches observations of all the models studied. As discussed, the size of the parameter space involved and the computational expense of the simulations means that the number of models examined is not exhaustive, and that usual statistical parameter space exploration techniques were not practical.
Kinematic Properties
--------------------
-------------------------------------- ---------------------- ----------------------------------- ------------------- ---------------------------------- ----------------------------- ---------------------------------- -------------------------
$x_{\mathrm{min}}^\mathrm{HV/DO}$/au $e^{\mathrm{HV/DO}}$ $x_{\mathrm{min}}^\mathrm{HV}$/au $e^{\mathrm{HV}}$ $\theta^{\mathrm{HV}}$ /$^\circ$ $i^{\mathrm{HV}}$ /$^\circ$ $\omega^{\mathrm{HV}}$ /$^\circ$ $\Delta \phi$ /$^\circ$
\[0.5ex\] $864$ $0.85$ $653$ $0.38$ $28$ $158$ $10$ $94$
\[1ex\]
-------------------------------------- ---------------------- ----------------------------------- ------------------- ---------------------------------- ----------------------------- ---------------------------------- -------------------------
![\[fig:kindist\] [The distribution of the initial semi-major axes of the HV ($a_0^\mathrm{HV}$) and HV/DO ($a_0^\mathrm{HV/DO}$) trajectories for successful solutions of our kinematic parameter space exploration. The solid lines (horizontal blue for $a_0^\mathrm{HV}$ and vertical red for $a_0^\mathrm{HV/DO}$) represent the median of the results. The associated dashed lines indicate the associated $16^\mathrm{th}$ and $84^\mathrm{th}$ percentile values. The green circle represents the location of our chosen ‘best-fit’ solution in reproducing the extended emission between the stellar systems (see Section \[sec:hydrobridge\]). ]{}](paper_figure_kinprops.pdf){width="50.00000%"}
[The distribution of semi-major axes in the initial systems ($a_0^\mathrm{HV/DO}$ and $a_0^\mathrm{HV}$) are shown for successful kinematic solutions is shown in Figure]{} . As discussed in Section \[sec:kmod\], it is not possible to draw statistical conclusions from this distribution. However, we note that most solutions exist for $a_0^\mathrm{HV/DO} \sim 10^4$ au, although the model which best reproduces the extended bridge structure (Section \[sec:hydrobridge\]) has $a_0^\mathrm{HV/DO} \approx 5800$ au. The parameters of this model are presented in Table . [We note that the orientation of the HV/DO angular momentum vector is approximately anti-parallel that of HV-AB/C. This reversal of the orbits appears surprising. However, if the forming stars were initially separated by $\sim 4 \cdot 10^4$ au (initial apastron) it is possible that local velocity fields in the collapsing gas of the primordial system lead to non-aligned orbits.]{}
The important dynamical properties of the chosen kinematic model are summarised in Table \[table:kinprops\]. By integrating backwards, all stellar components in this model are found to remain bound on time-scales $>1$ Myr. Initially HV-AB/C has an orbit with a semi-major axis $a^\mathrm{HV} _0 \approx 10^3$ au, and eccentricity $e_0 \approx 0.37$. The encounter with DO removes angular momentum from the HV system, and results in DO being marginal bound, with a large semi-major axis $a_\mathrm{f}^{\mathrm{HV/DO}} \approx 1.5 \times 10^4$ au, sufficient to reach the observed present day projected separation.
The closest encounter between each stellar component is also consistent with observations. The single encounter between HV Tau C and DO Tau is the closest between any of the components at $285$ au, and is close enough to truncate discs to $\sim 100$ au. No interaction involving AB is close enough such that a $\sim10$ au binary is likely to be disrupted. The minimum distance between HV Tau C and AB is equivalent to the final periastron distance as no closer interaction occurred.
Finally, the time since the closest encounter to reach the projected present day separation for our preferred system orientation is $\sim 0.1$ Myr, which is consistent with even the lowest estimate for the age of any of the stellar components.
------------------- ----------------------- ------------------- -------- --------------------- ------------------
$x_{\mathrm{min}}$/au $a_0$/au $e_0$ $a_{\mathrm{f}}$/au $e_{\mathrm{f}}$
\[0.5ex\] HV-C/DO $285$ - - - -
HV-AB/DO $657$ - - - -
HV-AB/C $445$ $1.05 \cdot 10^3$ $0.37$ $859$ $0.48$
HV/DO - $5.76\cdot 10^3$ $0.85$ $1.48 \cdot 10^4$ $0.95$
\[1ex\]
------------------- ----------------------- ------------------- -------- --------------------- ------------------
: Dynamical properties of the stellar components of the best-fitting model, where $x_{\mathrm{min}}$ is the closest approach $a_0$, $a_{\mathrm{f}}$, $e_0$, $e_{\mathrm{f}}$ are the initial and final semi-major axes and eccentricities of the binaries respectively.[]{data-label="table:kinprops"}
Disc properties
---------------
{width="\textwidth"}
The properties of the circumstellar discs found by tuning to best match the Herschel observations in Figure \[herschel\] are shown in Table \[table:discprops\], and the snapshots of the gas surface density distribution during the encounter are shown in Figure \[fig:ICs\]. The initial radii for HV Tau C and DO Tau discs are $320$ au and $355$ au respectively, which means that the stellar components penetrate the discs at the closest approach distance of $285$ au. We find that both a smaller mass and outer radius are required for the disc around HV-C with respect to DO. The present day observed disc mass ratio is $M_{\mathrm{disc}}^{\mathrm{HV-C}}/M_{\mathrm{disc}}^{\mathrm{DO}} \approx 0.15$, while our chosen model has an initial mass ratio of $0.33$. At the time of our chosen snapshot this ratio in the simulation becomes $\sim 0.13$, with the disc around HV Tau C losing a greater fraction of the initial mass.
In our model the orientation is such that the disc around HV Tau C is approximately edge on with the plane along the direction of the ‘V’-shaped emission, as suggested by observations (see Figure \[fig:ICs\]). The disc around DO Tau is also approximately face-on, and thus the geometry of the system is compatible with the observed extended structure discussed below. These disc orientations lead to a collision in which the discs collide approximately perpendicular in a strongly penetrating encounter. This violent interaction induces significant pressure gradients and justifies the need for hydrodynamic simulations.
---------------- ------------ ----------------------- --------------------- -------------------------- ---------------------------- ----- -- -- -- -- --
$R_{0}$/au $x_{\mathrm{min}}$/au $R_\mathrm{obs}$/au $M_{\mathrm{rel}, \, 0}$ $M_\mathrm{obs}$/$M_\odot$ $p$
\[0.5ex\] HV-C $320$ $285$ $\sim 50-100$ $0.33$ $\sim0.002$ 0
DO $355$ $285$ $\sim 75$ $1.0$ $0.013$ 0
\[1ex\]
---------------- ------------ ----------------------- --------------------- -------------------------- ---------------------------- ----- -- -- -- -- --
: Disc properties of the best-fit model. The quantities are as follows: $R_{ 0}$ is the initial outer radius of the disc, $x_{\mathrm{min}}$ is the closest encounter with any stellar component, $M_{\mathrm{rel},0}$ is the initial relative mass of each disc, $M_{\mathrm{obs}}$ is the observed total disc mass, $p$ is the power law index for the surface density. The subscript $0$ pertains to initial values in the model and ‘obs’ the observed (present-day) values.[]{data-label="table:discprops"}
External Structure {#sec:hydrobridge}
------------------
{width="\textwidth"}
![\[fig:vzgas\] Simulated variation in line of sight gas velocity $\delta v_z = v_z-\langle v_z \rangle $ density contours in the extended gas cloud. The contours are spaced over a factor 5 in surface density in arbitrary units. The blue contours are for SPH particles with $0$ km/s $< \delta v_z < 1$ km/s, while the red contours are for $-1$ km/s $< \delta v_z < 0$ km/s.](paper_figure_vzgas.pdf){width="50.00000%"}
In order to reproduce the extended structure between HV and DO, we have introduced a moderate temperature gradient with respect to the projected distance $d$ from each stellar component: $$T_\mathrm{dust} = 35 \, \mathrm{K} \left( \frac{d}{950 \, \mathrm{au}}\right)^{-0.32}$$ with a maximum temperature of $35$ K, which is consistent with the temperature profile found in Section \[sec:massest\]. The resulting surface brightness of the extended structure at $100$ $\mu$m and $160$ $\mu$m in our model is shown in Figure \[finalfig\]. In order to obtain this flux distribution we have had to assume a large initial total gas mass of $M_{\mathrm{tot},0} = 0.18\, M_\odot$ (with $\Sigma_\mathrm{dust}/\Sigma_\mathrm{gas} = 10^{-2}$). This is on the order of the mass we would expect if the interaction occurred at an early evolutionary stage. Approximately $50 \%$ of the mass is accreted in our simulations at the time of the snapshot, which leaves $0.09 \, M_\odot$ total mass, of which $\sim 0.027 \, M_\odot$ is retained in the disc around DO Tau and $3.5 \times 10^{-3} \, M_\odot$ in that of HV Tau C. [The remaining mass occupies the external structure.]{} These disc masses are a factor $\sim 2$ greater than the present day, and indeed the mass of the total system is expected to be an [overestimate due both observational and numerical factors. First we find resolution-dependent diffusion of SPH particles into the ISM (away from what we consider the ‘bridge’ between HV and DO). As we increase the resolution, for simulations run at a resolution lower than $10^6$ particles, a smaller fraction of SPH particles are lost to the ISM. Therefore we expect that increasing the resolution further would decrease the required total initial mass of the system. Additionally, increasing the initial radii of the discs has a similar effect of increasing the mass of the bridge while preserving the observed structure; however this additionally enhances accretion rates and therefore compounds resolution issues at late times. Alternatively, the dust-to-gas ratio in the original discs may be enhanced [@Ans16], which would mean our gas mass is overestimated. ]{}
We also note that we have chosen a snapshot at a separation between HV and DO of $\sim 5\times 10^3$ au, half of the observed present day separation. This is because, as discussed in Section \[sec:hydromodel\], resolution effects mean that the structure diffuses as the model is integrated in time. Integrating further to the present day results in a numerical loss of structure due to low resolution in the region between the stellar components. Contrary to the diffusive numerical effects described above, this means that additional initial mass would be required to produce sufficient surface density at the present day separation.
Overall, the main features seen in the $100$ $\mu$m and $160$ $\mu$m observations are well produced in our model, [namely the V-shaped emission close to HV Tau and the tidal tail close to DO Tau.]{} The broad envelope shape is less well reflected in our models, however we note that these regions have a low resolution of SPH particles which can result in a loss of structure. Additionally, uncertainties in the temperature profile discussed in Section \[sec:massest\], particularly at the outer edge and centre of the envelope where we only have detections at $160$ $\mu$m, mean that we are unable to accurately map the surface density to an intensity distribution. However, the agreement between our model and the observations is sufficient to suggest that a disc-disc interaction $\sim 0.1$ Myr ago is a viable mechanism by which the extended structure between HV and DO Tau has been produced.
Gas Velocity
------------
In Figure \[fig:vzgas\] [we demonstrate that we expect to find some substructure in the line of sight gas velocities. The standard deviation in line of sight velocity of the SPH particles $v_z$ for the best fit model is $\sigma_{v_z} \approx 1.3$ km/s. We divide the deviation from the mean gas velocity $\delta v_z = v_z - \langle v_z \rangle$ into two bins, red shifted ($-1$ km/s$<\delta v_z <0$ km/s) and blue shifted ($0$ km/s$<\delta v_z <1$ km/s). The results in Figure]{} \[fig:vzgas\] illustrate both the large scale velocity structure of the whole system, and the line of sight motion of the wide binary HV Tau C and AB.
Although, as previously discussed, the present day system is at approximately double the separation of the snapshot, Figure \[fig:vzgas\] [is indicative of the velocity field we would expect to obtain from observations if a past encounter produced the observed extended emission. Future observations of the gas in the region can be compared with our results to establish the likeliness of the scenario we suggest here.]{}
Conclusions {#sec:conc}
===========
We have used hydrodynamic modelling to lend evidence to the conclusion that the three stars making up HV Tau and the apparently unrelated star DO Tau had a past encounter $\sim 0.1$ Myr ago. While it is difficult to make hard conclusions about the nature of the dynamical history of the system and subsequent disc evolution, our modelling suggests the following scenario:
- HV Tau A, B and C initially formed a quadruple system with DO Tau $\gtrsim 0.1$ Myr ago, with a spatial scale of $\sim 5000$ au (and an orbital period of $\sim 0.3$ Myr).
- The highly eccentric orbit of DO Tau led to a close encounter with HV Tau C $0.1$ Myr ago. During this encounter the disc around HV Tau C interacted strongly with the disc around DO Tau, leading to rapid accretion and truncation of the discs. [This was likely the first encounter and therefore we expect the age of the original system to be $\lesssim 0.4$ Myr. ]{}
- Subsequent to this encounter the DO Tau trajectory became either marginally bound or marginally unbound to reach a separation $> 10^4$ au.
- The tidal tails of this event can be observed in the $160$ $\mu$m dust emission to the present day.
In terms of the history of Taurus, this supports the idea that there previously existed substructure down to smaller scales which has now been dynamically erased . Given the improbability of such a close encounter producing tidal tails that can be observed for time-scales $\sim 1$ Myr after the encounter, it is likely that many more such encounters which cannot be inferred have also occurred.
Acknowledgements {#acknowledgements .unnumbered}
================
We are appreciative for the comments of the anonymous referee which helped to significantly strengthen the arguments presented here. We would like to thank Marco Tazzari for useful discussion and for providing the dust opacities used in this work. AJW thanks the Science and Technology Facilities Council (STFC) for their studentship. This work has been supported by the DISCSIM project, grant agreement 341137 funded by the European Research Council under ERC-2013-ADG. It has also used the DIRAC Shared Memory Processing system at the University of Cambridge, operated by the COSMOS Project at the Department of Applied Mathematics and Theoretical Physics on behalf of the STFC DiRAC HPC Facility (www.dirac.ac.uk). This equipment was funded by BIS National E-infrastructure capital grant ST/J005673/1, STFC capital grant ST/H008586/1, and STFC DiRAC Operations grant ST/K00333X/1. DiRAC is part of the National E-Infrastructure. This work has made use of data from the European Space Agency (ESA) mission [*Gaia*]{} (<https://www.cosmos.esa.int/gaia>), processed by the [*Gaia*]{} Data Processing and Analysis Consortium (DPAC, <https://www.cosmos.esa.int/web/gaia/dpac/consortium>). Funding for the DPAC has been provided by national institutions, in particular the institutions participating in the [*Gaia*]{} Multilateral Agreement.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Generalization of Pascal’s triangle and its fascinating properties attract the attention of many researchers from the very beginning. Sir Isaac Newton first observed that the first five rows of Pascal’s triangle can be obtained from the power of 11 and claimed without proof that the subsequent rows of Pascal’s triangle can also be generated by the power of eleven. The allegation later proved by Arnold but the visualization of the rows restricted till $5^{th}$ row due to the limitation of 11. In the concept of 11, Morton showed that dividing each row (considering multi-digit numerals as single place value) by $11$ we get an immediately preceding row, but he didn’t give any formula for getting the full row. In this paper, a formula is derived as an extension of the concept of $11^n$ to generate any row of Pascal’s triangle. We extended the concept of $11^n$ to $1\Theta1^n$. We briefly discussed how our proposed concept works for any number of $n$ by employing an appropriate number of zeros between $1$ and $1$ ($11$) represented by $\Theta$. We generated the formula for getting the value of $\Theta$ stands for the number of zeros between 1 and 1. The evaluation of our proposed concept verified with Pascal’s triangle and matched successfully. Finally, we demonstrate Pascal’s triangle for a large n such as $51^{st}$ row as an example using our proposed formula.'
bibliography:
- 'main.bib'
---
[**Finding any row of Pascal’s triangle\
extending the concept of power of $11$**]{}\
[**Md. Robiul Islam**]{}^$\ast$,1^ and [**Md. Shorif Hossan**]{}^2^\
\
\
\
\
[**Keywords:**]{} Pascal’s triangle, power of $11$, Finding any rows, generalized Pascal’s triangle.
Introduction {#sec:intro}
============
Algebra is a spacious part of the science of mathematics which provides the opportunity to express mathematical ideas more precisely. In algebra, the Binomial expansion and Pascal’s triangle are considered important. Pascal’s triangle is a triangular arrangement of the binomial coefficients and one of the most known integer models. Though it was named after French scientist Blaise Pascal, it was studied in ancient India, Persia, China, Germany, and Italy by different mathematicians afore him. In fact, the definition of the triangle was made centuries ago. It is thought that in 450 BC, Indian mathematician Pingala was included the concept of this triangle in the book of poetry in Sanskrit. At the same time, the commentators of this book acquaint that the diagonal surface of the triangle is the sum of the Fibonacci numbers. It is the same idea among Chinese mathematicians and calls the triangle “Yang Hui’s triangle”. Later, Persian mathematician Al-Karaji and Persian astronomer-poet Omar Khayyam named the triangle as the “Khayyam triangle”. It also has multidimensional shapes, the three-dimensional shape is referred to as Pascal’s pyramid or Pascal’s tetrahedron, while the other general-shaped ones are called Pascal’s simplifications. Mathematicians find the application of this triangle in mathematics and many modern physics subjects. Various studies have been conducted in many different disciplines about Pascal’s triangle. The studies conducted in the last century can be analyzed as follows.\
In [@white], the importance of the Pascal’s triangle in modern mathematics and properties of this triangle with an application are discussed. In [@low], applications on Pascal’s triangle using modular arithmetic are showed. In each application, the first number was increased by one, and correlated the results with the Pascal triangle. Pascal method is narrated in [@sgroi] as “the usual method of selection for middle school or higher level students, which determines the number of a number of subsets”. Here @sgroi mentioned that in the construction of Pascal’s triangle, each line starts with 1 and ends with 1, and this series can be expanded with simple cross-joints. In his study, @jansson [@jansson] developed three geometric forms related to Pascal’s triangle and included examples on each form. In [@hoffman], 17 different properties of Pascal’s triangle and their relations with each other are studied. The relationship between Pascal’s triangle and Binomial expansion are investigated by using permutations [@smith]. In his study, @toschi [@toschi] constructed new types of Pascal’s triangles using different permutations and created generalizations. @duncan [@duncan] discussed about the reconstruction of Pascal’s triangle with the individuals. Here they collected data on the opinions of individuals using qualitative methods, and determined the methods of constructing the Pascal’s triangle in different ways with the attained findings. The relationship between the Pascal triangle and the Fibonacci numbers had been discussed In [@ouellette]. In [@staib], the Pascal pyramid concept created and visualized the Pascal triangle. In his study, @putz [@putz] developed the concept of Pascal Polytope using the concept of permutation and associated it with the Fibonacci concept. In [@houghton], @houghton gave concept about the relationship between successive differential operation of a function and Pascal’s triangle. Here, he tried to integrate the concept of differentiable function into Pascal’s triangle with an application. In [@hinz], relationship between Pascal’s triangle and Tower of Hanoi had been expressed. While forming this relationship, he benefited from the Kummer’s theorem. In his study, @osler [@osler] affirmed that Pascal’s triangle is the oldest and most important tool in mathematics. In addition, he used it in brackets, square brackets and higher forces, and identified each of these expansions with Pascal’s triangle.\
In 1956, @Freund [@Freund] elicited that the generalized Pascal’s triangles of $s^\text{th}$ order can be constructed from the generalized binomial coefficients of order $s$. In [@bankier] @bankier gave the Freud’s alternative proof. @Kall_1 tried to generalize Pascal triangle using power of integers [@Kall_1], different based triangle [@Kall_2] and their connections with prime number [@Farkas]. Kuhlmann tried to generate Pascal’s triangle using the T-triangle concept [@kuhlmann]. Some fascinating properties of Pascal’s triangle are available in [@Bondarenko; @Korec].\
The concept of power of $11$ was first introduced by sir Issac Newton. He observed that first five rows of Pascal’s triangle are generated by power of $11$ and claimed (without proof) for the later rows, that is successive rows can also be generated by power of eleven [@newton1736]. In [@Arnold] @Arnold supported Newton’s assertion and proved it generally. In [@Mueller] @Mueller noted that from the $n^\text{th}$ row of the Pascal’s triangle with positional addition, one can get the $n^\text{th}$ power of $11$. In this study, we try to extend the concept of power of $11$ and proposed a formula to attain any row of Pascal’s triangle.
Methods
=======
The very basic definition to get any element of a row of the Pascal’s triangle is the summation of two adjacent elements of the previous row. Each number in Pascal’s triangle is the sum of two numbers above that number. Usually, the lines of Pascal’s triangle are numbered starting from n = 0 from the top and the numbers in each line are starting from k = 0 from the left. For k=0 their is only one value 1. As the next lines are created, The remaining right most and left most element for new row is taken as 1. $$\begin{array}{ccccccccccccc} {} & {} & {} &{} & {} & {} & {1} & {} & {} & {} &{} & {} & {} \\ {} & {} & {} &{} & {} & {1} & {} & {1} & {} & {} &{} & {} & {} \\ {} & {} & {} &{} & {1} & {} & {2} & {} & {1} & {} &{} & {} & {} \\ {} & {} & {} &{1} & {} & {3} & {} & {3} & {} & {1} &{} & {} & {}\\ {} & {} & {1} &{} & {4} & {} & {6} & {} & {4} & {} &{1} & {} & {}\\ {} & {1} & {} &{5} & {} & {10} & {} & {10} & {} & {5} &{} & {1} & {}\\ {1} & {} & {6} &{} & {15} & {} & {20} & {} & {15} & {} &{6} & {} & {1} \\ {} & {} & {\ldots} &{} & {} & {} & {\ldots} & {} & {} & {} &{\ldots} & {} & {} \end{array}$$ **Figure 1:** Pascal’s triangle\
The concept of power of $11$ leads to us $11^{1}=11$, $1^{\text{st}}$ row of Pascal triangle and so $11^{2}=121$, $11^{3}=1331$ and $11^{4}=14641$ reveal $2^{\text{nd}}$, $3^{\text{rd}}$ and $4^{\text{th}}$ row respectively. Before finding the general rule for subsequent rows we first elaborate the previous concept of 11. The reason behind getting Pascal’s triangle by the power of 11 lies on the general rule of multiplication. What do we get from multiplication of a number by 11? multiplication of 11 leads to us the following figure.\
$$\begin{array}{@{}r@{}}
2^{nd}\text{ row of Pascal's triangle} \rightarrow 121 \\
{}\times 11\\
\hline
1${\color{red}2}$1\\
\text{left shift of all digits by 1 time} \rightarrow 12${\color{red}1}$0\\
\hline
3^{rd}\text{ row of Pascal's triangle} \rightarrow 13{\color{red}{\large \textcircled{\small\color{black} 3}}}1%\\
\vspace{-0.1in}
\end{array}$$
The number given inside the circle same as the summation of the two adjacent numbers of the previous row
**Figure 2:** Results after multiplication by $11$\
Figure 2 shows, multiplication of a number by $11$ gives an output which is similar to the addition of the two adjacent numbers of previous row of Pascal’s triangle.\
\
Patently $11^{5}=161051$ and $11^6=1771561 $ but the $5^{\text{th}}$ and $6^{\text{th}}$ row of Pascal’s triangle are
[1]{} [ ]{} [ ]{} [5]{} [ ]{} [ ]{} [10]{} [ ]{} [ ]{} [10]{} [ ]{} [ ]{} [5]{} [ ]{} [ ]{} [1]{}\
[ ]{} [ ]{} [ ]{} [ ]{} [ ]{} [ ]{} [ ]{} [ ]{} [ $\qquad~~\&$]{} [ ]{} [ ]{} [ ]{}\
[1]{} [ ]{} [ ]{} [6]{} [ ]{} [ ]{} [15]{} [ ]{} [ ]{} [20]{} [ ]{} [ ]{} [15]{} [ ]{} [ ]{} [6]{} [ ]{} [ ]{} [1]{}\
respectively. The above scheme fails for $11^5$ or $11^6$. Why are we not getting the $5^{th}$ row or why does the power of $11$ fail here? The answer is the middle value from the $5^{\text{th}}$ row of Pascal’s triangle are of two decimal places whereas the power of $11$ represents Pascal’s row as a representation of one decimal place. So for finding $5^{\text{th}}$ or any frontal row, we need a formula that can represent the number generated from the power of 11 as two or higher decimal places. Now, we will endeavor to formulate a specific rule that generates the required number of decimal places for the representation of Pascal’s triangle. At first, we attempt to represent the number as two decimal places using the very basic rules of multiplication. Figure 3, displays the impact of multiplication by $101$\
$$\begin{array}{@{}r@{}}
101 \\
{}\times 101\\
\hline
${\color{red}1}$01\\
\text{zeros cause the left shift of all digits by 1 time} \rightarrow $~{\color{red}00}$00\\
\text{left shift of all digits by 2 times} \rightarrow 1${\color{red}01}$00\\
\hline
1{\color{red}{\large \textcircled{\small \color{black} 02}}}01\\
{}\times 101\\
\hline
${\color{blue}1}{\color{green}02}$01\\
${\color{blue}00}{\color{green}00}$00\\
\text{left shift of all digits by 2 times} \rightarrow 1${\color{blue}02}{\color{green}01}$00\\
\hline
1{\color{blue}{\large \textcircled{\small {\color{black}03}}}} {\color{green}{\large \textcircled{\small {\color{black}03}}}}01\\
\end{array}$$
Numbers given inside the circles are same as the summation of two adjacent numbers of the previous row, but multiplication by 101 displays the rows as a representation of two decimal places
**Figure 3:** Results after multiplication by $101$\
Now, $101^{5}=1051010051$, from which we can construct $5^{\text{th}}$ row of Pascal’s triangle by omitting extra zeros and separating the digits.
[1]{} [5]{} [10]{} [10]{} [5]{} [1]{}
Similarly from $101^{6}=10615201560$ and $101^{7}=107213535210701$, we can easily construct $6^{\text{th}}$ and $7^{\text{th}}$ row respectively.\
[1]{} [ ]{} [ ]{} [06]{} [ ]{} [ ]{} [15]{} [ ]{} [ ]{} [20]{} [ ]{} [ ]{} [15]{} [ ]{} [ ]{} [06]{} [ ]{} [ ]{} [01]{}\
[ $\qquad~~\&$]{}\
[1]{} [ ]{} [ ]{} [07]{} [ ]{} [ ]{} [21]{} [ ]{} [ ]{} [35]{} [ ]{} [ ]{}[35]{} [ ]{} [ ]{} [21]{} [ ]{} [ ]{} [07]{} [ ]{} [ ]{} [01]{}
$101^{5}$, $101^{6}$ and $101^{7}$ all are representing $5^{th}$, $6^{th}$ and $7^{th}$ row of Pascal’s triangle respectively as a representation of two decimal places due to the addition of one zero between $1$ and $1~$(11) such that $101$. $11^5$, $11^6$ and $11^7$ could also represent the respective rows according to the Newton’s claim but $101^n$ makes the visualization.\
Can a conclusion be drawn for the generating any row of Pascal’s triangle with the help of extended concept of power of 11 such as $101^n$ ? Let’s have a look for n=9. Plainly, $101^9 =1093685272684360901$. But the $9^{\text{th}}$ row of Pascal’s triangle is $$1~~ 9~~ 36 ~84~~ 126~~ 126 ~~84 ~~36~ 9~~ 1$$ This is due to the three digits in the central element 126. So, we need a formula for the representation of three decimal places. The previous context directed that multiplication of a number by $11$ and $101$ makes the left shift of all digits by one and two times respectively. So the representation of three decimal places requires multiplication by $1001$.
Figure 4, proofs the left shift of all digits by $3$ times when a number is multiplied by $1001$\
$$\begin{array}{@{}r@{}}
1001 \\
{}\times 1001\\
\hline
${\color{red}1}$001\\
${\color{red}00}$000\\
${\color{red}000}$000\\
\text{left shift of all digits by 3 time} \rightarrow 1${\color{red}001}$000\\
\hline
1{\color{red}{\Large \textcircled{\normalsize {\color{black}002}}}}001\\
{}\times 1001\\
\hline
${\color{blue}1}{\color{green}002}$001\\
${\color{blue}00}{\color{green}000}$000\\
${\color{blue}000}{\color{green}000}$000\\
\text{left shift of all digits by 3 time} \rightarrow 1${\color{blue}002}{\color{green}001}$000\\
\hline
1{\color{blue}{\Large \textcircled{\normalsize {\color{black}003}}}}{\color{green}{\Large \textcircled{\normalsize {\color{black}003}}}}001
\end{array}$$
Numbers given inside the circles are summation of the two adjacent numbers of the previous row as a representation of three decimal places
**Figure 4:** Results after multiplication by $1001$\
By continuing the multiplication, we get $$1001^9 =1009036084126126084036009001$$ from which one may form the $9^{th}$ row of Pascal’s triangle with the representation as three decimal places:
$1~~009~~036~~084~~126~~126~~084~~036~~009~~001$
Similarly, $1001^{10}$ represents the $10^{\text{th}}$ row of Pascal triangle with the representation as three decimal places: $$1010045120210252210120045010001\longmapsto 1~~010~~045~~120~~210~~252~~210~~120~~045~~010~~001$$
Results and discussion
======================
From the above study, it can be easily concluded that the representation of three decimal places requires the left shift of all digits by three times, and three times the left shift of all digits requires two zeros between $1$ and $1~$(11), that is $1001$. Why do we require three decimal places representation for $9^{th}$ and $10^{th}$ rows of Pascal’s triangle?. Because the central element of $9^{\text{th}}$ and $10^{\text{th}}$ row is of three decimal places. Similarly, we required two decimal places representation for $5^{th}$ to $8^{th}$ rows since the central element of these rows are numbers of two decimal places. And, the $1^{st}$ four rows satisfied $11^n$ since those are numbers of one decimal place. So for any row, the number of decimal places representation should be equal to the number of digits exist in the central value of that row.\
Now we seek to generate a formula to find the central value of any row of the Pascal’s triangle. For an odd number, say $n=9$ we get $n+1=10$ elements in $9^{th}$ row. So the central value should be $\left(\frac{10}{2}\right)^{th}=5^{th}$ observation of that row, which is $\binom{9}{5-1}=\binom{9}{4}=126$. For an even number, say $n=10$ we get $n+1=11$ elements and the central value should be $\left(\frac{11}{2}\right)=5.5\Rightarrow 6^{th}~(Ceiling~value)$ observation, which is $\binom{10}{6-1}=\binom{10}{5}=252$.\
Subtraction by $1$ each time can be omitted by taking the *floor* value of $\frac{n}{2}$. So the formula for having central value of $n^{th}$ row is $\binom{n}{floor(\frac{n}{2})}$. But we never need for the central value rather get the number of digits to exist in the central value. Let’s make it more facilitate, using Logarithmic function we can directly calculate how many digits (or decimal places) should the central number have?. Applying the property of Logarithmic function the formula becomes $\log_{10} \binom{n}{floor(\frac{n}{2})}$, since $ceil(\log_{10}(X))$ represents the number of digits of $X$. However, if the central value is of $d$ decimal places then we require one less number of zeros between $1~\text{and}~1~$(11) such that $\left(1\colorbox{red!40!green!30!}{\textbf{(d-1)~zeros}}1\right)^{n}$. So, we can get the number of zeros required between $1~\text{and}~1~$(11) by taking the *floor* value of $\log_{10} \binom{n}{floor(\frac{n}{2})}$ that is $floor(\log_{10} \binom{n}{floor(\frac{n}{2})})$.\
Let us consider $\Theta$ represents the number of zeros between $1$ and $1~$(11). Then $\Theta=floor\left(\log_{10} \binom{n}{floor(\frac{n}{2})}\right)$. Let’s verify it for an odd number $n=9$ and an even number $n=10$.\
Currently, $n=9$ gives
&=floor(\_[10]{} )\
&=floor(\_[10]{} )\
&=floor(2.10)=2
And, $n=10$ gives
&=floor(\_[10]{} )\
&=floor(\_[10]{} )\
&=floor(2.401)=2
For both of the numbers we need $2$ zeros between $1$ and $1~$(11). So, to get $9^{th}$ and $10^{th}$ rows we have to calculate $1001^{9}$ and $1001^{10}$ respectively. Both are verified above already.\
It’s time to generate the formula to find any row of Pascal’s triangle. The general formula for generating $n^{th}$ row of Pascal’s triangle is $(1\Theta1)^n$, where $\Theta$ represents the number zeros required to generate the desired row and defined by $$\Theta=floor\left(\log_{10} \binom{n}{floor(\frac{n}{2})}\right)$$ For a random number such as $n=15$ we get $\Theta=3$. So we have to insert 3 zeros and the $15^{th}$ row can be constructed from the following $$10001^{15}=
1001501050455136530035005643564355005300313650455010500150001$$ Same thing goes for an even number such as for $n=16$, $\Theta=4$. So the $16^{th}$ row can be constructed from the following\
$(100001)^{16} =
\seqsplit{%
100016001200056001820043680800811440128701144008008043680182000560001200001600001}
$\
One can verify both of these from the Pascal’s triangle. The above formula can be used for a large $n$. We now exemplify $51^{st}$ row of Pascal’s triangle. Hence $n=51$ gives
&=floor(\_[10]{} )\
&=floor(\_[10]{} )\
&=floor(\_[10]{} )\
&=floor(\_[10]{}(247959266474052))\
&=floor(14.394)=14
So, we have to put $14$ zeros between $1~\text{and}~1~$(11), that is $(1000000000000001)^{51}$.\
Now $(1000000000000001)^{51} =
\seqsplit{%
1
{\color{red}000000000000051}
{\color{blue}000000000001275}
{\color{red}000000000020825}
{\color{blue}000000000249900}
{\color{red}000000002349060}
{\color{blue}000000018009460}
{\color{red}000000115775100}
{\color{blue}000000636763050}
{\color{red}000003042312350}
{\color{blue}000012777711870}
{\color{red}000047626016970}
{\color{blue}000158753389900}
{\color{red}000476260169700}
{\color{blue}001292706174900}
{\color{red}003188675231420}
{\color{blue}007174519270695}
{\color{red}014771069086725}
{\color{blue}027900908274925}
{\color{red}048459472266975}
{\color{blue}077535155627160}
{\color{red}114456658306760}
{\color{blue}156077261327400}
{\color{red}196793068630200}
{\color{blue}229591913401900}
{\color{red}247959266474052}
{\color{blue}247959266474052}
{\color{red}229591913401900}
{\color{blue}196793068630200}
{\color{red}156077261327400}
{\color{blue}114456658306760}
{\color{red}077535155627160}
{\color{blue}048459472266975}
{\color{red}027900908274925}
{\color{blue}014771069086725}
{\color{red}007174519270695}
{\color{blue}003188675231420}
{\color{red}001292706174900}
{\color{blue}000476260169700}
{\color{red}000158753389900}
{\color{blue}000047626016970}
{\color{red}000012777711870}
{\color{blue}000003042312350}
{\color{red}000000636763050}
{\color{blue}000000115775100}
{\color{red}000000018009460}
{\color{blue}000000002349060}
{\color{red}000000000249900}
{\color{blue}000000000020825}
{\color{red}000000000001275}
{\color{blue}000000000000051}
{\color{red}000000000000001}}
$ The desired $51^{\text{st}}$ row can be obtained by separating each $15$ digits (except the first digit $1$) from the above result. For readers convenient, we marked each entry with different colors and showing that the above formula generates a Pascal’s triangle with a representation of $15$ digits.
Conclusion
==========
Pascal’s triangle is a startling mathematical tool that has vastly infliction throughout various mathematical topics. So, forming pascal’s triangle easily and quickly is an expectation of all analysts who are interested in it. Here, w e extended the existing formula from $11^{n}$ to $(1\Theta1)^n$. In view of the above discussion, we may conclude that, as multiplication by $11$ leads us to the addition of the adjacent numbers of the previous row so we can find any row of Pascal’s triangle by inserting proper number of zeros between $1~\text{and}~1$(11).\
The number of zeros yields from: $ \Theta=floor\left(\log_{10} \binom{n}{floor(\frac{n}{2})}\right) $ and the [$n^{th}$ ]{} row is obtained by $(1\Theta1)^n$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We compare various viewpoints on down-sets (simplicial complexes), illustrating how the combinatorial inclusion-exclusion principle may serve as an alternative to more advanced methods of studying their face numbers.'
address: 'Institute of Mathematics, University of Bremen Bibliothekstr. 1, 28359 Bremen, Germany'
author:
- 'Micha[ł]{} Adamaszek'
title: 'Face numbers of down-sets'
---
For any family $\Delta\subseteq 2^{[n]}$ of subsets of some ground set $[n]=\{1,\ldots,n\}$ we define a two-variable generating function $$H_\Delta(x,y)=\sum_{\sigma\in\Delta} x^{|\sigma|}y^{n-|\sigma|}$$ and its two specializations, the $f$- and $K$-polynomial: $$f_\Delta(t)=H_\Delta(t,1) \quad\textrm{and}\quad K_\Delta(t)=H_\Delta(t,1-t).$$
The family $\Delta$ is called a *down-set* if it is closed under taking subsets; that is $\sigma\in\Delta$ and $\tau\subseteq \sigma$ imply $\tau\in\Delta$. For a down-set $\Delta$, let ${\mathcal{F}}=\{F_1,\ldots,F_m\}$ be the family of *maximal elements* in $\Delta$ and let ${\mathcal{M}}=\{M_1,\ldots,M_k\}$ be the family of minimal elements of $2^{[n]}\setminus \Delta$, which will also be called the *blockers* of $\Delta$.
Clearly every object of the triple $\Delta$, ${\mathcal{F}}$, ${\mathcal{M}}$ determines the other two, but the representation via ${\mathcal{F}}$ or ${\mathcal{M}}$ is usually much smaller than the listing of all of $\Delta$. We will start with a brief survey of down-sets in various branches of mathematics. Different points of view highlight different aspects of the relationship between ${\mathcal{F}}$, ${\mathcal{M}}$, $f_\Delta(t)$ and $K_\Delta(t)$. We will then give a formula for $H_\Delta(x,y)$ in terms of the combinatorics of ${\mathcal{F}}$ and ${\mathcal{M}}$ and show its interpretations in the different approaches to down-sets.
The ways of thinking of down-sets we wish to consider are the following.
- In geometry, a down-set is usually called an *abstract simplicial complex*. The elements of $\Delta$ are its *faces*, enumerated by the face polynomial $f_\Delta(t)$ which is a well-studied combinatorial invariant (see [@SS]). The elements of ${\mathcal{F}}$ are the highest-dimensional faces (or *facets*) of $\Delta$. By taking a simplex spanned on $\ell$ vertices for each $\ell$-element face in $\Delta$ and gluing them together, we obtain the geometric realization of $\Delta$, as in the figure.
![The geometric realization of the down-set $\Delta=\{\emptyset,1,2,3,4,12,23,24,34,234\}\subseteq 2^{[4]}$ with maximal elements ${\mathcal{F}}=\{12,234\}$ and blockers ${\mathcal{M}}=\{13,14\}$](fig-1)
- In commutative algebra, every element $M\in{\mathcal{M}}$ can be identified with a square-free monomial $\mathbf{x}_M=\prod_{i\in M} x_i$ in the polynomial ring $R=\mathbb{F}[x_1,\ldots,x_n]$. If $I=(\mathbf{x}_{M_1},\ldots,\mathbf{x}_{M_k})$, then $K_\Delta(t)/(1-t)^n$ is the Hilbert series of the $R$-module $R/I$, with the latter usually referred to as the Stanley-Reisner ring of $\Delta$. As one would expect, the correspondence between simplicial complexes and their Stanley-Reisner rings is the source of a fruitful interaction between combinatorial geometry and commutative algebra.
- In complexity theory, ${\mathcal{M}}$ determines a monotone Boolean function in disjunctive normal form $$\varphi_{\mathcal{M}}(x_1,\ldots,x_n)=\bigvee_{M\in{\mathcal{M}}}\bigwedge_{i\in M} x_i$$ with variables $x_1,\ldots, x_n$. Then $\Delta$ is the set of non-satisfying assignments for $\varphi_{\mathcal{M}}$. The problem of expressing the formula $\overline{\varphi_{\mathcal{M}}}(\overline{x_1},\ldots,\overline{x_n})$ again in disjunctive normal form, known as *monotone dualization* or *hypergraph transversal*, is equivalent to computing ${\mathcal{F}}$ from ${\mathcal{M}}$. The decision version of this problem is “Given ${\mathcal{M}}$ and ${\mathcal{F}}$, do they represent the same $\Delta$?”. The question has received a lot of attention due to its applicability in various areas of theoretical computer science, and it is open whether it can be answered in time polynomial with respect to the size of the combined description of ${\mathcal{M}}$ and ${\mathcal{F}}$. The best known algorithm requires quasi-polynomial time [@D].
The reader is welcome to consult [@SM; @EMG] for a more thorough treatment of the above topics. For example, [@SM] provides an exposition based on the ideas of combinatorial commutative algebra. Here is our main result, illustrating how the elementary inclusion-exclusion principle may serve as an alternative to the homological algebra of monomial ideals.
\[mainthm\] For a down-set $\Delta\subseteq 2^{[n]}$ with maximal elements ${\mathcal{F}}$ and blockers ${\mathcal{M}}$ we have $$\begin{aligned}
H_\Delta(x,y)&=& \sum_{\emptyset\neq S\subseteq{\mathcal{F}}}(-1)^{|S|+1}(x+y)^{|\bigcap S|}y^{n-|\bigcap S|}\\
&=&\sum_{T\subseteq{\mathcal{M}}}(-1)^{|T|}(x+y)^{n-|\bigcup T|}x^{|\bigcup T|}.\end{aligned}$$
The coefficient of $x^\ell y^{n-\ell}$ in $H_\Delta(x,y)$ is the number of sets of cardinality $\ell$ in $\Delta$. It is given by either of the formulae $$\sum_{\emptyset\neq S\subseteq{\mathcal{F}}}(-1)^{|S|+1}{|\bigcap S|\choose \ell}\quad\textrm{or}\quad\sum_{T\subseteq{\mathcal{M}}}(-1)^{|T|}{n-|\bigcup T|\choose \ell-|\bigcup T|}.$$ Both expressions are instances of the inclusion-exclusion principle, except that instead of including-excluding elements, as it is usually done, we include-exclude entire $\ell$-subsets of sets in ${\mathcal{F}}$ or $\ell$-supersets of sets in ${\mathcal{M}}$. From this the formulae for $H_\Delta(x,y)$ immediately follow.
Here are some applications.
- We can express the $K$-polynomial of $\Delta$ in terms of the blockers ${\mathcal{M}}$: $$\label{eq:kk}
K_\Delta(t)=H_\Delta(t,1-t)=\sum_{T\subseteq{\mathcal{M}}}(-1)^{|T|}t^{|\bigcup T|}.$$ This is exactly the main result of [@G], with special cases appearing also in earlier work. In fact the formula $H_\Delta(x,y)=(x+y)^nK_\Delta(x/(x+y))$ implies that is equivalent to Theorem \[mainthm\].
- The reduced Euler characteristic of $\Delta$ is $\widetilde{\chi}(\Delta)=f_\Delta(-1)=H_\Delta(-1,1)$. In a similar fashion we get $$\widetilde{\chi}(\Delta)=H_\Delta(-1,1)=\sum_{T\subseteq{\mathcal{M}}}(-1)^{|T|}0^{n-|\bigcup T|}(-1)^{|\bigcup T|}=(-1)^n\sum_{\substack{T\subseteq{\mathcal{M}}\\\bigcup T=[n]}}(-1)^{|T|},$$ as shown in [@MT Thm.4.2].
- The dual version of the above is $$\begin{aligned}
\widetilde{\chi}(\Delta)=\sum_{\emptyset\neq S\subseteq{\mathcal{F}}}(-1)^{|S|+1}0^{|\bigcap S|}&=&\sum_{\substack{S\subseteq{\mathcal{F}}\\\bigcap S=\emptyset}}(-1)^{|S|+1}
=\sum_{\substack{S\subseteq{\mathcal{F}}\\\bigcap S\neq\emptyset}}(-1)^{|S|},\end{aligned}$$ which follows also from the fact, familiar to topologists, that a simplicial complex $\Delta$ is homotopy equivalent to the nerve of the family ${\mathcal{F}}$, hence their Euler characteristics are equal.
- In the language of the $f$-polynomial Theorem \[mainthm\] reads $$f_\Delta(t)=H_\Delta(t,1)=\sum_{\emptyset\neq S\subseteq{\mathcal{F}}}(-1)^{|S|+1}(t+1)^{|\bigcap S|}.$$
It is worth mentioning that the apparent symmetry of the two formulae in Theorem \[mainthm\] is a manifestation of Alexander duality, which prevails in all examples of down-sets we mentioned earlier. For a down-set $\Delta\subseteq 2^{[n]}$ define $$\Delta^*=\{[n]\setminus\sigma~:~\sigma\in 2^{[n]}\setminus\Delta\}.$$ Then $\Delta^*$ is again a down-set which we call the *Alexander dual* of $\Delta$. The maximal elements of $\Delta^*$ are $F_i^*=[n]\setminus M_i$ for the blockers $M_1,\ldots,M_k$ of $\Delta$, while the blockers of $\Delta^*$ are $M_j^*=[n]\setminus F_j$ for the maximal elements $F_1,\ldots,F_m$ of $\Delta$. We will denote those new families by ${\mathcal{F}}^*=\{F_1^*,\ldots,F_k^*\}$ and ${\mathcal{M}}^*=\{M_1^*,\ldots,M_m^*\}$. Note that $(\Delta^*)^*=\Delta$ and $\overline{\varphi_{\mathcal{M}}}(\overline{x_1},\ldots,\overline{x_n})=\varphi_{{\mathcal{M}}^*}(x_1,\ldots,x_n)$.
\[prop\] For a down-set $\Delta\subseteq 2^{[n]}$ we have $$H_\Delta(x,y)+H_{\Delta^*}(y,x)=(x+y)^n.$$
We will give a proof which uses Theorem \[mainthm\], but the reader is encouraged to find a proof based only on the definition of $H_\Delta(x,y)$ and then check that under Proposition \[prop\] each of the two parts of Theorem \[mainthm\] implies the other. We obtain $$\begin{aligned}
&H_\Delta(x,y)+ H_{\Delta^*}(y,x)=\\
&=\sum_{T\subseteq{\mathcal{M}}}(-1)^{|T|}(x+y)^{n-|\bigcup T|}x^{|\bigcup T|}+\sum_{\emptyset\neq S^*\subseteq{\mathcal{F}}^*}(-1)^{|S^*|+1}(x+y)^{|\bigcap S^*|}x^{n-|\bigcap S^*|}\\
&=\sum_{T\subseteq{\mathcal{M}}}(-1)^{|T|}(x+y)^{n-|\bigcup T|}x^{|\bigcup T|}+\sum_{\emptyset\neq T\subseteq{\mathcal{M}}}(-1)^{|T|+1}(x+y)^{n-|\bigcup T|}x^{|\bigcup T|}.\end{aligned}$$ The only contribution which is not cancelled out is $(x+y)^n$ arising from $T=\emptyset$.
If we only perform the first approximation to the inclusion-exclusion formula, that is take the union bound, after some manipulation we derive for $x,y\geq 0$ the inequality $$H_\Delta(x,y)\leq \sum_{F\in{\mathcal{F}}} (x+y)^{|F|}y^{n-|F|} = \sum_{M^*\in{\mathcal{M}}^*} (x+y)^{n-|M^*|}y^{|M^*|}.$$ Taking $x=y=\frac12$ in Proposition \[prop\] and using the above we obtain $$\label{eq}
\sum_{M^*\in {\mathcal{M}}^*}2^{-|M^*|}+\sum_{M\in {\mathcal{M}}}2^{-|M|}\geq 1,$$ which is well known in the theory of the dualization of disjunctive normal forms [@D Lemma 1] as a tool for estimating the joint size of a formula and its dual. To see how much deviates from equality one takes $x=y=\frac12$ in Proposition \[prop\]: $$\sum_{T^*\subseteq{\mathcal{M}}^*}(-1)^{|T^*|}2^{-|\bigcup T^*|}+\sum_{T\subseteq{\mathcal{M}}}(-1)^{|T|}2^{-|\bigcup T|}=1.$$
Acknowledgment. {#acknowledgment. .unnumbered}
---------------
The author thanks the referees for helpful remarks concerning the presentation. This work was supported by the DFG grant FE 1058/1-1.
[99]{}
T. Eiter, K. Makino, G. Gottlob, Computational aspects of monotone dualization: A brief survey, *Discrete Appl. Math.* [**156**]{} (2008) 2035–2049.
M. L. Fredman, L. Khachiyan, On the Complexity of Dualization of Monotone Disjunctive Normal Forms, *Journal of Algorithms* [**21**]{} (1996) 618–628.
A. Goodarzi, On the Hilbert series of monomial ideals, *J. Combin. Theory Ser. A* [**120**]{} (2013) 315–317.
M. Marietti, D. Testa, Conical and spherical graphs, *European J. Combin.* [**33**]{} (2012) 1609–1618.
E. Miller, B. Sturmfels, *Combinatorial Commutative Algebra*, Graduate Texts in Mathematics, Vol. 227, Springer-Verlag, New York, 2005.
R. Stanley, *Combinatorics and commutative algebra*, second edition. Progress in Mathematics, Vol. 41, Birkäuser Boston, Inc., Boston, MA, 1996.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We analyze mechanisms and regimes of wave packet spreading in nonlinear disordered media. We predict that wave packets can spread in two regimes of strong and weak chaos. We discuss resonance probabilities, nonlinear diffusion equations, and predict a dynamical crossover from strong to weak chaos. The crossover is controlled by the ratio of nonlinear frequency shifts and the average eigenvalue spacing of eigenstates of the linear equations within one localization volume. We consider generalized models in higher lattice dimensions and obtain critical values for the nonlinearity power, the dimension, and norm density, which influence possible dynamical outcomes in a qualitative way.'
address: 'Max-Planck-Institut für Physik komplexer Systeme, Nöthnitzer Str. 38, 01187 Dresden, Germany'
author:
- 'S. Flach'
title: Spreading of waves in nonlinear disordered media
---
disorder ,Anderson localization ,nonlinearity ,chaos ,diffusion
05.45-a ,05.60Cd ,63.20Pw
Introduction {#Introduction}
============
In this paper we will discuss the mechanisms of wave packet spreading in nonlinear disordered systems. More specifically, we will consider cases when i) the corresponding linear wave equations yield Anderson localization, ii) the localization length is bounded from above by a finite value, iii) the nonlinearity is compact in real space and therefore does not induce long range interactions between eigenstates of the linear equations.
There are several reasons to analyze such situations. First, wave propagation in spatially disordered media has been of practical interest since the early times of studies of waves. In particular, it became of much practical interest for the conductance properties of electrons in semiconductor devices more than half a century ago. It was probably these issues which motivated P. W. Anderson to perform his groundbreaking studies on what is now called Anderson localization [@PWA58]. With evolving technology, wave propagation became of importance also in photonic and acoustic devices [@Exp; @Exp2]. Finally, recent advances in the control over ultracold atoms in optical potentials made it possible to observe Anderson localization there as well [@BECEXP]. Peter Hänggi and collaborators studied properties of wave propagation in disordered media by phase space visualization of the underlying dynamical system in high dimensional phase spaces [@hanggi].
Second, in many if not all cases wave-wave interactions are of importance, or can even be controlled experimentally. Screening effects can reduce the long range character of these interactions considerably for electrons. Short range interactions also hold for s-wave scattering of atoms. When many quantum particles interact, mean field approximations often lead to effective nonlinear wave equations. As a result, nonlinear wave equations in disordered media become of practical importance.
Third, there is fundamental interest in understanding, how Anderson localization is modified for nonlinear wave equations. All of the above motivates the choice of corresponding linear wave equations with finite upper bounds on the localization length. Then, the linear equations admit no transport. Analyzing transport properties of nonlinear disordered wave equations allows to observe and characterize the influence of wave-wave interactions on Anderson localization in a straightforward way.
A number of studies was recently devoted to the above subject [@Mol98; @PS08; @Shep08; @kkfa08; @fks08; @skkf09; @hvyksf09; @mm09; @mmkaapds09]. In the present work we will present a detailed analysis of the chaotic dynamics which is at the heart of the observed destruction of Anderson localization. In particular, we will show that an optional intermediate strong chaos regime of subdiffusive spreading is followed by an even slower subdiffusive spreading process in the regime of weak chaos.
Wave equations
==============
We will use the Hamiltonian of the disordered discrete nonlinear Schrödinger equation (DNLS) $$\mathcal{H}_{D}= \sum_{l} \epsilon_{l}
|\psi_{l}|^2+\frac{\beta}{2} |\psi_{l}|^{4}
- (\psi_{l+1}\psi_l^* +\psi_{l+1}^* \psi_l)
\label{RDNLS}$$ with complex variables $\psi_{l}$, lattice site indices $l$ and nonlinearity strength $\beta \geq 0$. The random on-site energies $\epsilon_{l}$ are chosen uniformly from the interval $\left[-\frac{W}{2},\frac{W}{2}\right]$, with $W$ denoting the disorder strength. The equations of motion are generated by $\dot{\psi}_{l} = \partial
\mathcal{H}_{D}/ \partial (i \psi^{\star}_{l})$: $$i\dot{\psi_{l}}= \epsilon_{l} \psi_{l}
+\beta |\psi_{l}|^{2}\psi_{l}
-\psi_{l+1} - \psi_{l-1}\;.
\label{RDNLS-EOM}$$ Eqs. (\[RDNLS-EOM\]) conserve the energy (\[RDNLS\]) and the norm $S
= \sum_{l}|\psi_l|^2$. We note that varying the norm of an initial wave packet is strictly equivalent to varying $\beta$. Eqs. (\[RDNLS\]) and (\[RDNLS-EOM\]) are derived e. g. when describing two-body interactions in ultracold atomic gases on an optical lattice within a mean field approximation [@oberthaler], but also when describing the propagation of light through networks of coupled optical waveguides in Kerr media [@yskgpa03].
Alternatively we also refer to results for the Hamiltonian of the quartic Klein-Gordon lattice (KG) $$\mathcal{H}_{K}= \sum_{l} \frac{p_{l}^2}{2} +
\frac{\tilde{\epsilon}_{l}}{2} u_{l}^2 +
\frac{1}{4} u_{l}^{4}+\frac{1}{2W}(u_{l+1}-u_l)^2,
\label{RQKG}$$ where $u_l$ and $p_l$ are respectively the generalized coordinates and momenta, and $\tilde{\epsilon}_{l}$ are chosen uniformly from the interval $\left[\frac{1}{2},\frac{3}{2}\right]$. The equations of motion are $\ddot{u}_{l} = - \partial \mathcal{H}_{K}
/\partial u_{l}$ and yield $$\ddot{u}_{l} = - \tilde{\epsilon}_{l}u_{l}
-u_{l}^{3} + \frac{1}{W} (u_{l+1}+u_{l-1}-2u_l)\;.
\label{KG-EOM}$$ Equations (\[KG-EOM\]) conserve the energy (\[RQKG\]). They serve e.g. as simple models for the dissipationless dynamics of anharmonic optical lattice vibrations in molecular crystals [@aaovchinnikov]. The energy of an initial state $E \geq 0$ serves as a control parameter of nonlinearity similar to $\beta$ for the DNLS case. For small amplitudes the equations of motion of the KG chain can be approximately mapped onto a DNLS model [@KG-DNLS-mapping]. For the KG model with given parameters $W$ and $E$, the corresponding DNLS model (\[RDNLS\]) with norm $S=1$, has a nonlinearity parameter $\beta\approx 3WE$. The norm density of the DNLS model corresponds to the normalized energy density of the KG model.
The theoretical considerations will be performed within the DNLS framework. It is straightforward to adapt them to the KG case.
Anderson localization
=====================
For $\beta=0$ with $\psi_{l} = A_{l}
\exp(-i\lambda t)$ Eq. (\[RDNLS\]) is reduced to the linear eigenvalue problem $$\lambda A_{l} = \epsilon_{l} A_{l}
- A_{l-1}-A_{l+1}\;.
\label{EVequation}$$ The normalized eigenvectors $A_{\nu,l}$ ($\sum_l A_{\nu,l}^2=1)$ are the NMs, and the eigenvalues $\lambda_{\nu}$ are the frequencies of the NMs. The width of the eigenfrequency spectrum $\lambda_{\nu}$ of (\[EVequation\]) is $\Delta=W+4$ with $\lambda_{\nu} \in \left[ -2 -\frac{W}{2}, 2 + \frac{W}{2}
\right] $.
The asymptotic spatial decay of an eigenvector is given by $A_{\nu,l} \sim
{\rm e}^{-l/\xi(\lambda_{\nu})}$ where $\xi(\lambda_{\nu})$ is the localization length and $\xi(\lambda_{\nu}) \approx
24(4-\lambda_{\nu}^2)/W^2$ for weak disorder $W \leq 4$ [@PWA58; @KRAMER]. The NM participation number $p_{\nu} = 1/\sum_l A_{\nu,l}^4$ is one possible way to quantize the spatial extend (localization volume) of the NM. The localization volume $V$ is on average of the order of $3 \xi(0)$ for weak disorder, and tends to $V=1$ in the limit of strong disorder. The average spacing $d$ of eigenvalues of NMs within the range of a localization volume is therefore of the order of $d \approx \Delta / V$, which becomes $d \approx \Delta W^2 /300 $ for weak disorder. The two scales $ d \leq \Delta $ are expected to determine the packet evolution details in the presence of nonlinearity.
Due to the localized character of the NMs, any localized wave packet with size $L$ which is launched into the system for $\beta=0$ , will stay localized for all times. If $L \ll V$, then the wave packet will expand into the localization volume. This expansion will take a time of the order of $\tau_{lin}=2\pi/d$. If instead $L \geq V$, no substantial expansion will be observed in real space. We remind that Anderson localization is relying on the phase coherence of waves. Wave packets which are trapped due to Anderson localization correspond to trajectories in phase space evolving on tori, i.e. quasiperiodically in time.
Adding nonlinearity
===================
The equations of motion of (\[RDNLS-EOM\]) in normal mode space read $$i \dot{\phi}_{\nu} = \lambda_{\nu} \phi_{\nu} + \beta \sum_{\nu_1,\nu_2,\nu_3}
I_{\nu,\nu_1,\nu_2,\nu_3} \phi^*_{\nu_1} \phi_{\nu_2} \phi_{\nu_3}\;
\label{NMeq}$$ with the overlap integral $$I_{\nu,\nu_1,\nu_2,\nu_3} =
\sum_{l} A_{\nu,l} A_{\nu_1,l}
A_{\nu_2,l} A_{\nu_3,l}\;.
\label{OVERLAP}$$ The variables $\phi_{\nu}$ determine the complex time-dependent amplitudes of the NMs.
The frequency shift of a single site oscillator induced by the nonlinearity is $\delta_l = \beta |\psi_l|^{2}$. If instead a single mode is excited, its frequency shift can be estimated by $\delta_{\nu} = \beta |\phi_{\nu}|^2/
p_{\nu}$.
As it follows from (\[NMeq\]), nonlinearity induces an interaction between NMs. Since all NMs are exponentially localized in space, each normal mode is effectively coupled to a finite number of neighbouring NMs, i.e. the interaction range is finite. However the strength of the coupling is proportional to the norm density $n = |\phi|^2$. Let us assume that a wave packet spreads. In the course of spreading its norm density will become smaller. Therefore the effective coupling strength between NMs decreases as well. At the same time the number of excited NMs grows.
One possible outcome would be: (I) that after some time the coupling will be weak enough to be neglected. If neglected, the nonlinear terms are removed, the problem is reduced to the linear wave equation, and we obtain again Anderson localization. That implies that the trajectory happens to be on a quasiperiodic torus. Then it has to be on that torus from the beginning. Another possibility is: (II) that spreading continues for all times. That would imply that the trajectory evolves not on a quasiperiodic torus, but in some chaotic part of phase space. A third possibility is: (III) that the trajectory was initially strongly chaotic, but manages in the course of spreading to get trapped between denser and denser torus structures in phase space after some spreading, leading again to localization as an asymptotic outcome.
Consider a wave packet with size $L$ and norm density $n$. Replace it by a [*finite*]{} system of size $L$ and norm density $n$. Such a finite system will be in general nonintegrable. Therefore the only possibility to generically obtain a quasiperiodic evolution is to be in the regime where the KAM theorem holds. Then there is a finite fraction of the available phase space volume which is filled with KAM tori. For a given $L$ it is expected that there is a critical density $n_{KAM}(L)$ below which the KAM regime will hold. We do not know this $L$-dependence. Computational studies may not be very conclusive here, since it is hard to distinguish a regime of very weak chaos from a strict quasiperiodic one on finite time scales.
The above first possible outcome (I) (localization) will be realized if the packet is launched in a KAM regime. Whether that is possible at all for an infinite system is an open issue. The second outcome (II) (spreading) implies that we start in a chaotic regime and remain there. Since the packet density is reduced and is proportional to its inverse size $L$ at later times, this option implies that the critical density $n_{KAM}(L)$ decays faster than $1/L$, possibly faster than any power of $1/L$. The third possibility (III) (asymptotic localization) should be observable by some substantial slowing down of the spreading process.
The secular normal form
-----------------------
Let us perform a further transformation $\phi_{\nu} = {\rm e}^{-i \lambda_{\nu} t} \chi_{\nu}$ and insert it into Eq. (\[NMeq\]): $$i \dot{\chi}_{\nu} = \beta \sum_{\nu_1,\nu_2,\nu_3}
I_{\nu,\nu_1,\nu_2,\nu_3} \chi^*_{\nu_1} \chi_{\nu_2} \chi_{\nu_3} {\rm e}^{i(\lambda_{nu} +
\lambda_{\nu_1}-\lambda_{\nu_2}-\lambda_{nu_3})t}
\;.
\label{NMeqchi}$$ The right hand side contains oscillating functions with frequencies $$\lambda_{\nu,\vec{\nu}} \equiv \lambda_{\nu} +
\lambda_{\nu_1}-\lambda_{\nu_2}-\lambda_{\nu_3}\;,\;\vec{\nu} \equiv (\nu_1,\nu_2,\nu_3)\;.
\label{dlambda}$$ For certain values of $\nu,\vec{\nu}$ the value $\lambda_{\nu,\vec{\nu}}$ becomes exactly zero. These secular terms define some slow evolution of (\[NMeqchi\]). Let us perform an averaging over time of all terms in the rhs of (\[NMeqchi\]), leaving therefore only the secular terms. The resulting secular normal form equations (SNFE) take the form $$i \dot{\chi}_{\nu} = \beta \sum_{\nu_1}
I_{\nu,\nu_,\nu_1,\nu_1} |\chi_{\nu_1}|^2 \chi_{\nu}
\;.
\label{NMeqRNF}$$ Note that possible missing factors due to index permutations can be absorbed into the overlap integrals, and are not of importance for what is following. The SNFE can be now solved for any initial condition $\chi_{\nu}(t=0)=\eta_{\nu}$ and yield $$\chi_{\nu}(t) = \eta_{\nu} {\rm e}^{-i \Omega_{\nu} t}\;,\; \Omega_{\nu} = \beta \sum_{\nu_1}
I_{\nu,\nu_,\nu_1,\nu_1} |\eta_{\nu_1}|^2
\;.$$ Since the norm of every NM is preserved in time for the SNFE, it follows that Anderson localization is preserved within the SNFE. The only change one obtains is the renormalization of the eigenfrequencies $\lambda_\nu$ into $\tilde{\lambda}_{\nu} = \lambda_{\nu}+\Omega_{\nu}$. Moreover, the phase coherence of NMs is preserved as well. Any different outcome will be therefore due to the nonsecular terms, neglected within the SNFE.
Measuring properties of wave packets
------------------------------------
We order the NMs in space by increasing value of the center-of-norm coordinate $X_{\nu}=\sum_l l A_{\nu,l}^2$. We analyze normalized distributions $n_{\nu}
\geq 0$ using the second moment $m_2= \sum_{\nu}
(\nu-\bar{\nu})^2 n_{\nu}$, which quantifies the wave packet’s degree of spreading and the participation number $P=1 / \sum_{\nu} n_{\nu}^2$, which measures the number of the strongest excited sites in $n_{\nu}$. Here $\bar{{\nu}} = \sum_{\nu} \nu n_{\nu}$. We follow norm density distributions $n_{\nu}\equiv |\phi_{\nu}|^2/\sum_{\mu} |\phi_{\mu}|^2$. The second moment $m_2$ is sensitive to the distance of the tails of a distribution from the center, while the participation number $P$ is a measure of the inhomogeneity of the distribution, being insensitive to any spatial correlations. Thus, $P$ and $m_2$ can be used to quantify the sparseness of a wave packet through the compactness index $$\zeta=\frac{P^2}{m_2}.
\label{eq:ci}$$ A thermalized wave packet yields $\zeta=3$. Distributions with larger gaps between equally excited isolated sites attain a compactness index $\zeta<3$.
Expected regimes of spreading
-----------------------------
Previous studies suggested the existence of various dynamical regimes of spreading of wave packets [@PS08; @fks08; @skkf09]. Some of these definitions were contradictory. Below we will resolve this.
Consider a wave packet at $t=0$ which has norm density $n$ and size $L$. If $\beta n \geq \Delta$, a substantial part of the wave packet will be selftrapped [@kkfa08; @skkf09]. This is due to the above discussed nonlinear frequency shifts, which will tune the excited sites immediately out of resonance with the nonexcited neighborhood. As a result, discrete breather like structures will be formed, which can persist for immensely long times. While selftrapping and discrete breather formation are interesting localization phenomena at strong nonlinearity, they are very different from Anderson localization since they require the existence of gaps in the spectrum of the linear wave equations [@DB]. If now $\beta n < \Delta$, selftrapping is avoided, and the wave packet can start to spread. For $L < V$ and $\beta=0$, the packet will spread over the localization volume during the time $\tau_{lin}$. After that, the new norm density will drop down to $n(\tau_{lin}) \approx n \frac{L}{V}$. For $L > V$ the norm density will not change appreciably up to $\tau_{lin}$, $n(\tau_{lin}) \approx n$. The nonlinear frequency shift $\beta n(\tau_{lin})$ can be now compared with the average spacing $d$. If $\beta n(\tau_{lin}) > d$, all NMs in the packet are resonantly interacting with each other. This regime will be coined strong chaos. If instead $\beta n(\tau_{lin}) < d$, NMs are weakly interacting with each other. This regime will be coined weak chaos. To summarize: $$\begin{aligned}
\beta n(\tau_{lin}) < d \;:\; {\rm weak\; chaos}
\nonumber
\\
\beta n(\tau_{lin}) > d \;:\; {\rm strong\; chaos}
\nonumber
\\
\beta n > \Delta \;:\; {\rm selftrapping}
\nonumber\end{aligned}$$ In terms of the above wave packet characteristics $n$ , $L$ it follows $$\begin{aligned}
\beta n \tilde{L} < \Delta \;:\; {\rm weak\; chaos}
\nonumber
\\
\beta n \tilde{L} > \Delta \;:\; {\rm strong\; chaos}
\label{regimes}
\\
\beta n > \Delta \;:\; {\rm selftrapping}
\nonumber\end{aligned}$$ where $\tilde{L}=L$ for $L < V$ and $\tilde{L}=V$ for $L > V$. It follows that the regime of strong chaos can be observed only if $L > 1$. For $L=1$ we expect only two regimes - selftrapping and weak chaos. Furthermore, we obtain that the regimes of strong and weak chaos are separated by the quantity $\beta n = d$, i.e. the average spacing $d$ is the only characteristic frequency scale here.
Discussion of numerical results
-------------------------------
Let us discuss the above in the light of published computational experiments. We show results for single site excitations from [@skkf09] in Fig.\[fig1\] with $W=4$, $L=1$ and $n=1$.
![(color online) Single site excitations. $m_2$ and $P$ versus time in log–log plots. Left plots: DNLS with $W=4$, $\beta=0,0.1,1,4.5$ \[(o), orange; (b), blue; (g) green; (r) red\]. Right plots: KG with $W=4$ and initial energy $E=0.05,0.4,1.5$ \[(b) blue; (g) green; (r) red\]. The orange curves (o) correspond to the solution of the linear equations of motion, where the term $u_l^3$ in (\[KG-EOM\]) was absent. The disorder realization is kept unchanged for each of the models. Dashed straight lines guide the eye for exponents 1/3 ($m_2$) and 1/6 ($P$) respectively. Insets: the compactness index $\zeta$ as a function of time in linear–log plots for $\beta=1$ (DNLS) and $E=0.4$ (KG). Adapted from [@skkf09]](fig1.eps){width="0.99\columnwidth"}
. \[fig1\]
For the DNLS model (left plots in Fig.\[fig1\]) with $\beta=4.5$ it follows $\beta n = 4.5$. Already at these values selftrapping of a part of the wave packet is observed. Therefore $P$ does not grow significantly, while the second moment $m_2\sim
t^{\alpha}$ with $\alpha \approx 1/3$ (red curves). A part of the excitation stays highly localized [@kkfa08], while another part delocalizes. For $\beta=1$ selftrapping is avoided since $\beta n < \Delta$. With $V\approx 20$ and $d \approx 0.4$ it follows that $\tau_{lin} \approx 16$ and $\beta n(\tau_{lin}) \approx 0.05 \ll d$. Therefore we expect to observe the regime of weak chaos. It is characterized by subdiffusive spreading with $m_2\sim t^{\alpha}$ and $P \sim t^{\alpha/2}$ (green curves). For $\beta=0.1$ we will remain in the regime of weak chaos, however the time scales for observing spreading grow. Therefore one finds no visible spreading up to some time $\tau_d$ which increases with further decreasing nonlinearity. For $t < \tau_d$ both $m_2$ and $P$ are not changing. However for $t > \tau_d$ the packet shows visible growth with the characteristics of weak chaos (blue curves). The simulation of the equations of motion in the absence of nonlinear terms (orange curves) shows Anderson localization. Since $L=1$ in the above numerical data, strong chaos has not been observed.
Notably, the authors of Ref. [@skkf09] also considered single mode excitations with total norm $S=1$. Using the above terminology, $n\approx 1/V$ and $L=V$ with $W=4$ and therefore again $V\approx 20$. For the case $\beta=6.5$ the authors detected a growth of $m_2$ which was subdiffusive but faster than $t^{1/3}$. We think that these observations are a clear hint towards the realization of strong chaos, which should be observable for $5...10 < \beta < 30...40$ in these cases.
The time evolution of $\zeta$ for excitations in the regime of weak chaos is shown in the insets of Fig. \[fig1\]. As one can see the compactness index oscillates around some constant nonzero value both for the DNLS and the KG models. This means that the wave packet spreads but does not become more sparse. The average value $\overline{\zeta}$ of the compactness index over 20 realizations of single mode excitations at $t=10^8$ for the DNLS model with $W=4$ and $\beta=5$ was found to be $\overline{\zeta}=2.95 \pm 0.39$ [@skkf09].
![Norm density distributions in the NM space at time $t=10^8$ for the initial excitations in the regime of weak chaos of the DNLS model. Left plots: single site excitation for $W=4$ and $\beta=1$. Right plots: single mode excitation for $W=4$ and $\beta=5$. $|\phi_\nu|^2$ is plotted in linear (logarithmic) scale in the upper (lower) plots. The average localization volume $V\approx20$ (shown schematically in the lower plots) is much smaller than the length over which the wave packets have spread. Adapted from [@skkf09]. []{data-label="fig2"}](fig2.eps){width="0.99\columnwidth"}
The norm density distribution for the DNLS model at $t=10^8$ is plotted in Fig. \[fig2\]. The distribution is characterized by a flat plateau of almost ideally thermalized NMs. The width of this plateau is more than an order of magnitude larger than the localization volume of the linear equations. Therefore Anderson localization is destroyed. The plateau is bounded by exponentially decreasing tails, with exponents corresponding to the localization length of the linear equations. With growing time the plateau widens, drops in height, and is pushing the tails to larger distances. Another remarkable feature are the huge fluctuations of norm densities in the tails, reaching 4-6 orders of magnitude. Such fluctuations are observed even in the case $\beta=0$. They are due to the fact, that NMs are ordered in space. Neighbouring NMs in space may have different eigenfrequencies, and therefore different values of their localization length. Tail NMs are excited by the core. The further away they are, the weaker the excitation. But within a small tail volume, NMs with larger localization length will be more strongly excited than those with smaller localization length, hence the large observed fluctuations, which on a logarithmic scale are of the order of the relative variation of the localization length. The remarkable observation is, that these fluctuations in the tails persist for the nonlinear case. Anderson localization is destroyed in the core (plateau) of the wave packet due to mode-mode interactions. The tail NMs are slaved to the core and excited by it. The interaction between neighbouring tail NMs is negligible, and the huge fluctuations persist. Therefore, Anderson localization is preserved in the tails of the distributions over very long times (essentially until the given tail volume becomes a part of the core).
For single site excitations in the regime of weak chaos the exponent $\alpha$ of subdiffusive spreading does not appear to depend on $\beta$ in the case of the DNLS model or on the value of $E$ in the case of KG. We find no visible dependence of the exponent $\alpha$ on $W$. Therefore the subdiffusive spreading is rather universal and the parameters $\beta$ (or $E$) and $W$ are only affecting the prefactor. Excluding selftrapping, any nonzero nonlinearity appears to completely delocalize the wave packet and destroy Anderson localization. Fittings were performed by analyzing 20 runs in the regime of weak chaos with different disorder realizations. For each realization the exponent $\alpha$ was fitted, and then averaged over all computational measurements. We find $\alpha = 0.33 \pm 0.02$ for DNLS, and $\alpha = 0.33 \pm 0.05$ for KG [@fks08; @skkf09]. Therefore, the universal exponent $\alpha=1/3$ [@fks08] appears to explain the data.
Another intriguing test was performed on the same disorder realizations and single site initial conditions, by additionally dephasing the NMs in a random way every hundred time units [@skkf09]. In that case, subdiffusion speeds up, and $m_2$ grows as $t^{1/2}$ implying $\alpha_{deph} = 1/2$. This regime of complete decoherence of NM phases exactly corresponds to the above anticipated one of strong chaos, but here enforced by explicit dephasing.
From strong to weak chaos, from resonances to nonlinear diffusion
=================================================================
We can think of two possible mechanisms of wave packet spreading. A NM with index $\mu$ in a boundary layer of width $V$ in the cold exterior, which borders the packet, is either incoherently [*heated*]{} by the packet, or [*resonantly excited*]{} by some particular NM from a boundary layer with width $V$ inside the packet.
For heating to work, the packet modes $\phi_{\nu}(t)$ should contain a part $\phi_{\nu}^{c}(t)$, having a continuous frequency spectrum (similar to a white noise), in addition to a regular part $\phi_{\nu}^{r}(t)$ of pure point frequency spectrum: $$\phi_{\nu}(t) = \phi_{\nu}^{r}(t) + \phi_{\nu}^{c}(t)\;.
\label{cpp}$$ Therefore at least some NMs of the packet should evolve chaotically in time. The more the packet spreads, the less the mode amplitudes in the packet become. Therefore its dynamics should become more and more regular, implying $
\lim_{t\rightarrow \infty} \phi_{\nu}^{c}(t)/\phi_{\nu}^{r}(t) \rightarrow 0$.
Strong chaos {#sec:dephase}
------------
Let us assume that all NMs in the packet are strongly chaotic, and their phases can be assumed to be random on the time scales of the observed spreading. According to (\[NMeq\]) the heating of the exterior mode should evolve as $i
\dot{\phi}_{\mu} \approx \lambda_{\mu} \phi_{\mu} + \beta n^{3/2} f(t)$ where $\langle f(t) f(t') \rangle = \delta(t-t')$ ensures that $f(t)$ has a continuous frequency spectrum. Then the exterior NM increases its norm according to $|\phi_{\mu}|^2 \sim \beta^2 n^3 t$. The momentary diffusion rate of the packet is given by the inverse time $T$ it needs to heat the exterior mode up to the packet level: $D = 1/T \sim \beta^2 n^2$. The second moment is of the order of $m_2 \sim 1/n^2$, since the packet size is $1/n$. The diffusion equation $m_2 \sim D t$ yields $m_2 \sim \beta t^{1/2}$. This agrees very well with the numerical results for dephasing in NM space. Moreover, we expect it to hold also without explicit dephasing, provided the initial wave packet satisfies the above conditions for strong chaos (\[regimes\]). First numerical tests show that this is correct [@tljb10], but it contradicts the observations of the numerical data in the regime of weak chaos without additional dephasing. Thus, in the regime of weak chaos, not all NMs in the packet are chaotic, and dephasing is at best some partial outcome.
Resonance probability
---------------------
Chaos is a combined result of resonances and nonintegrability. Let us estimate the number of resonant modes in the packet for the DNLS model. Excluding secular interactions, the amplitude of a NM with $|\phi_{\nu}|^2 = n_{\nu}$ is modified by a triplet of other modes $\vec{\mu}\equiv (\mu_1,\mu_2,\mu_3)$ in first order in $\beta$ as (\[NMeq\]) $$|\phi_{\nu}^{(1)}| = \beta \sqrt{n_{\mu_1}n_{\mu_2}n_{\mu_3}}
R_{\nu,\vec{\mu}}^{-1}\;,\; R_{\nu,\vec{\mu}} \sim
\left|\frac{\lambda_{\nu,\vec{\mu}}}{I_{\nu,\mu_1,\mu_2,\mu_3}}\right| \;,
\label{PERT1}$$ where $\lambda_{\nu,\vec{\mu}} =
\lambda_{\nu}+\lambda_{\mu_1}-\lambda_{\mu_2}-\lambda_{\mu_3}$. The perturbation approach breaks down, and resonances set in, when $\sqrt{n_{\nu}}
< |\phi_{\nu}^{(1)}|$. Since all considered NMs belong to the packet, we assume their norms to be equal to $n$ for what follows. Then the resonance condition for a given NM with index $\nu$ is met if there is at least one given triplet of other NMs $\vec{\mu}$ such that $$\beta n < R_{\nu,\vec{\mu}}\;.
\label{resonance_R}$$ If three of the four mode indices are identical, one is left with interacting NM pairs. A statistical analysis of the probability of resonant interaction was performed in Ref. [@fks08]. For small values of $n$ (i.e. when the packet has spread over many NMs) the main contribution to resonances are due to rare multipeak modes [@fks08], with peak distances being larger than the localization volume. However pair resonances are expected not to contribute to the spreading process [@hvyksm10]. When distances between the peaks of multipeak modes are larger than the localization volume, the time scale of excitation transfer from one peak to another will grow exponentially with the distance. Such processes are too slow in order to be observed in numerical experiments [@hvyksm10].
If two or none of the four mode indices are identical, one is left with triplets and quadruplets of interacting NMs respectively. In both cases the resonance condition (\[resonance\_R\]) can be met at arbitrarily small values of $n$ for NMs from one localization volume.
For a given NM $\nu$ we define $ R_{\nu,\vec{\mu}_0} = \min_{\vec{\mu} }
R_{\nu,\vec{\mu}}$. Collecting $R_{\nu,\vec{\mu}_0}$ for many $\nu$ and many disorder realizations, we can obtain the probability density distribution $\mathcal{W}(R_{\nu,\vec{\mu}_0})$. The probability $\mathcal{P}$ for a mode, which is excited to a norm $n$ (the average norm density in the packet), to be resonant with at least one triplet of other modes at a given value of the interaction parameter $\beta$ is therefore given by $$\mathcal{P} = \int_0^{\beta n} \mathcal{W}(x) {\rm d}x\;.
\label{resprob}$$ The main result is that $\mathcal{W}(R_{\nu,\vec{\mu}_0} \rightarrow 0)
\rightarrow C(W) \neq 0$ [@skkf09]. For the cases studied, the constant $C$ drops with increasing disorder strength $W$. This result of nonzero values of $C$ is not contradicting the fact of level repulsion of neighbouring NMs, since triplet and quadruplet combinations of NM frequencies can yield practically zero values of $\lambda_{\nu,\vec{\mu}}$ with finite distances between the eigenfrequencies.
For the case of strong disorder ($W \gg 1$) the localization volume tends to one, and quadruplet resonances are rare. Excluding also pair resonances for the above reasons, we are left with triplet resonances. A given mode may yield a triplet resonance with its two nearest neighbours to the left and right. Replacing the overlap integrals by some characteristic average, and assuming that the three participating modes have essentially uncorrelated eigenfrequencies, it follows that $$\mathcal{W}(R) \approx C \left( 1-\frac{CR}{3} \right) ^2
\;.$$ Due to the nonnegativity of $\mathcal{P}$ it would follow that $\mathcal{P}=0$ for $R \geq 3/C$. In reality we expect an exponential tail for large $R$. As a simple approximation, we may instead use $$\mathcal{W} (R) \approx C {\rm e}^{-CR}
\label{approxp}$$ which in turn can be expected to hold also for the case of weak disorder. It leads to the approximative result $$\mathcal{P} = 1-{\rm e}^{-C\beta n}\;.
\label{approxpp}$$
For $\beta n \rightarrow 0$ it follows $$\mathcal{P} \approx C \beta n\;.
\label{resprobas}$$ Therefore the probability for a mode in the packet to be resonant is proportional to $C \beta n$ in the limit of small $n$ [@fks08; @skkf09]. However, on average the number of resonant modes in the packet is proportional to the product of $\mathcal{P}$ and the total number of modes in the packet. Since the total number is proportional to $1/n$, the the average number of resonant modes in a packet is constant, proportional to $C \beta$, and their fraction within the packet is $\sim C \beta n$ [@fks08; @skkf09]. Since packet mode amplitudes fluctuate in general, averaging is meant both over the packet, and over suitably long time windows. A detailed numerical analysis of the statistical properties of resonances and related issues is in preparation [@dksf10].
Finally we consider the process of resonant excitation of an exterior mode by a mode from the packet. The number of packet modes in a layer of the width of the localization volume at the edge, which are resonant with a cold exterior mode, will be proportional to $\beta n$. After long enough spreading $\beta n
\ll 1$. On average there will be no mode inside the packet, which could efficiently resonate with an exterior mode. Resonant growth can be excluded [@fks08; @skkf09]. Thus, a wave packet is trapped at its edges, and stays localized until the interior of the wave packet decoheres (thermalizes). On these (growing) time scales, the packet will be finally able to incoherently excite the exterior and to extend its size.
A conjecture leading to the correct asymptotics
-----------------------------------------------
We assume, that the continuous frequency part of the dynamics of a packet mode is $\mathcal{P}(\beta n)$. It follows that $\phi_{\nu}^{c}(t) / \phi_{\nu}^{r}(t) \sim \mathcal{P}(\beta n)$. As expected initially, the chaotic part in the dynamics of packet modes becomes weaker the more the packet spreads, and the packet dynamics becomes more and more regular in the limit of large times. Therefore the chaotic component is conjectured to be a small parameter $\phi_{\nu}^{c}(t) \ll \phi_{\nu}^{r}(t)$. Expanding the term $|\phi_{\nu}|^2\phi_{\nu}$ to first order in $\phi_{\nu}^{c}(t)$, the heating of the exterior mode should evolve according to $i \dot{\phi}_{\mu} \approx \lambda_{\mu} \phi_{\mu} + \beta n^{3/2} \mathcal{P}(\beta n)
f(t)$. It follows $|\phi_{\mu}|^2 \sim \beta^2 n^3 (\mathcal{P}(\beta n))^2 t$, and the rate $$D =
1/T \sim \beta^2 n^2 (\mathcal{P}(\beta n))^2\;.$$ With (\[approxp\]),(\[approxpp\]) and $m_2 \sim 1/n^2$ the diffusion equation $m_2 \sim D t$ yields $$\frac{1}{n^2} \sim \beta (1-{\rm e}^{-C\beta n }) t^{1/2}\;.
\label{strong-weak}$$ The solution of this equation yields a crossover from subdiffusive spreading in the regime of strong chaos to subdiffusive spreading in the regime of weak chaos: $$\begin{aligned}
m_2 \sim (\beta^2 t)^{1/2}\;,\;{\rm strong}\;{\rm chaos}\;,\;C\beta n > 1\;,
\nonumber
\\
m_2 \sim C^{2/3} \beta^{4/3} t^{1/3}\;,\;{\rm weak}\;{\rm chaos}\;,\;C\beta n < 1\;,
\nonumber\end{aligned}$$
The crossover from strong chaos to weak chaos
---------------------------------------------
According to (\[resprob\]) the probability of resonance for a packet NM will be practically equal to one, if $\beta n$ is sufficiently larger than $1/C$. Such a situation can be generated for packets with large enough $\beta n$, and should yield spreading, provided one avoids selftrapping $\beta n \leq 4+W$ [@kkfa08; @skkf09]. This spreading will be different from the asymptotic behaviour discussed above over potentially large time scales.
Let us use as an example $W=4$ and $\beta=1$, with the constant $C\approx 6.2$ [@skkf09]. Single site excitations with norm $S=1$ lead after very short times to a spreading of the excitation into the localization volume of the linear wave equations, which is of the order of 10-20. The attained norm density is therefore of the order of $n \leq 0.1$. The observed spreading is the asymptotic one since $\mathcal{P} \sim C \beta n$. However, if we choose a packet size $L$ to be of the order of the localization volume, and the norm density $n$ of the order of $n=1$, initially $\mathcal{P} \approx 1$. Thus every mode in the packet will be resonant, and the condition for strong chaos should hold. At the same time $\beta n=1$ is far below the selftrapping threshold $4+W = 8$. For strong chaos we derived $m_2 \sim t^{1/2}$. With spreading continuing, the norm density in the packet will decrease, and eventually $\beta n \leq 1/C$. Then there will be a crossover from strong chaos to weak chaos, and $m_2\sim t^{1/3}$ for larger times. This crossover happens on logarithmic time scales, and it will be not easy to confirm it numerically [@tljb10].
In Fig.\[fig3\] we show the resulting time dependence of $m_2$ on $t$ from (\[strong-weak\]) in a log-log plot, where we used $\beta=1$, $C=6.2$, $L=20$ and $n(t=10^2)=1$.
With $x=log_{10}(t)$ and $y=log_{10}(m_2)$ it is straightforward to calculate the zero of the third derivative $d^3y/dx^3=0$ to obtain the crossover position: $$C\beta n_c \approx 1.86\;.$$ Therefore the only characteristic frequency scale here is $1/C$. From the above discussion of the different spreading regimes (\[regimes\]) it follows, that this scale is corresponding to the average spacing $d$: $$\frac{1}{C} \approx d\;.$$ Then $$\begin{aligned}
C \approx \frac{100}{W^2}\;,\;W \leq 4 \;,
\\
C \approx \frac{1}{W}\;,\; W \gg 4\;.\end{aligned}$$ Our results can be used to predict the critical value of the norm density $n_c$ at which the crossover should take place. For $W=4$ and $\beta=1$ it follows $n_c\approx 0.3$.
Scaling of nonlinear diffusion equations
----------------------------------------
With the above results on the diffusion coefficient, we may consider a set of nonlinear diffusion equations for the norm density distributions in NM space. For simplicity we replace the discrete NM indices by a continuous variable: $$\frac{\partial n}{\partial t} = \frac{\partial}{\partial \nu} D(n) \frac{\partial n}{\partial \nu}
\;.$$ In particular, we are interested in cases $D(n) \sim n^{\kappa}$. With a single scaling assumption $n(\nu,t/a) = b n(c\nu,t)$, and using the conservation of the total norm and $n(\nu \rightarrow \pm \infty,t)
\rightarrow 0$ we obtain $b=c=a^{1/(\kappa+2)}$. Then the second moment $m_2$ will grow in time according to $$m_2(t) = \left( \frac{t}{t_0}\right) ^{\alpha} m_2(t_0)\;,\;\alpha=\frac{2}{\kappa+2}\;.$$
Notably an explicit self-similar solution was calculated by Tuck in 1976 [@tuck76] which has the following spatial profile: $$n(\nu)=\left( B - \frac{\kappa \nu^2}{2(\kappa+2)}\right)^{1/\kappa}\;.$$ Here $B$ is an integration constant (see also [@nonlineardiffusion]).
For the case of strong chaos $\kappa=2$ and $\alpha=1/2$ in accord with the above results. Also for $\kappa=4$ we obtain $\alpha=1/3$ again in agreement with the above results.
Generalizations
===============
Let us consider $\boldsymbol{D}$-dimensional lattices with nonlinearity order $\sigma > 0$: $$i\dot{\psi_{\boldsymbol{l}}}= \epsilon_{\boldsymbol{l}} \psi_{\boldsymbol{l}}
-\beta |\psi_{\boldsymbol{l}}|^{\sigma}\psi_{\boldsymbol{l}}
-\sum\limits_{\boldsymbol{m}\in
D(\boldsymbol{l})}\psi_{\boldsymbol{m}}\;.
\label{RDNLS-EOMG}$$ Here $\boldsymbol{l}$ denotes an $\boldsymbol{D}$-dimensional lattice vector with integer components, and $\boldsymbol{m}\in
D(\boldsymbol{l})$ defines its set of nearest neighbour lattice sites. We assume that (a) all NMs are spatially localized (which can be obtained for strong enough disorder $W$), (b) the property $\mathcal{W}(x \rightarrow 0) \rightarrow const \neq 0$ holds, and (c) the probability of resonances on the edge surface of a wave packet is tending to zero during the spreading process. A wavepacket with average norm $n$ per excited mode has a second moment $m_2 \sim 1/n^{2/\boldsymbol{D}}$. The nonlinear frequency shift is proportional to $\beta n^{\sigma/2}$. The typical localization volume of a NM is still denoted by $V$, and the average spacing by $d$.
Consider a wave packet with norm density $n$ and volume $L < V$. A straightforward generalization of the expected regimes of spreading leads to the following: $$\begin{aligned}
\beta n^{\sigma/2} \left( \frac{L}{V}\right)^{\sigma/2} V < \Delta \;:\; {\rm weak\; chaos}
\nonumber
\\
\beta n^{\sigma/2} \left( \frac{L}{V}\right)^{\sigma/2} V > \Delta \;:\; {\rm strong\; chaos}
\nonumber
\\
\beta n^{\sigma/2} > \Delta \;:\; {\rm selftrapping}
\nonumber\end{aligned}$$ The regime of strong chaos, which is located between selftrapping and weak chaos, can be observed only if $$L > L_c = V^{1-2/\sigma}\;,\; n > n_c = \frac{V}{L} \left( \frac{d}{\beta}\right)^{2/\sigma}\;.$$ For $\sigma =2$ we need $L>1$, for $\sigma \rightarrow \infty$ we need $L > V$, and for $\sigma < 2$ we need $L \geq 1$. Thus the regime of strong chaos can be observed e.g. in a one-dimensional system with a single site excitation and $\sigma < 2$.
If the wave packet size $L > V$ then the conditions for observing different regimes simplify to $$\begin{aligned}
\beta n^{\sigma/2} < d \;:\; {\rm weak\; chaos}
\nonumber
\\
\beta n^{\sigma/2} > d \;:\; {\rm strong\; chaos}
\nonumber
\\
\beta n^{\sigma/2} > \Delta \;:\; {\rm selftrapping}
\nonumber\end{aligned}$$ The regime of strong chaos can be observed if $$n > n_c = \left( \frac{d}{\beta}\right)^{2/\sigma}\;.$$
Similar to the above we obtain a diffusion coefficient $$D \sim \beta^2 n^{\sigma} (\mathcal{P}(\beta n^{\sigma/2}))^2
\;.
\label{ggeneralizeddiffusion}$$ In both regimes of strong and weak chaos the spreading is subdiffusive [@fks08]: $$\begin{aligned}
m_2 \sim (\beta^2 t)^{\frac{2}{2+\sigma \boldsymbol{D}}}\;,\;{\rm strong}\;{\rm chaos}\;,
\label{sigma_strong}
\\
m_2 \sim (\beta^4 t)^{\frac{1}{1+\sigma \boldsymbol{D}}}\;,\;{\rm weak}\;{\rm chaos}\;.
\label{sigma_weak}\end{aligned}$$
Let us calculate the number of resonances in the wave packet volume ($N_{RV}$) and on its surface ($N_{RS}$) in the regime of weak chaos: $$N_{RV} \sim \beta n^{\sigma/2-1}\;,\; N_{RS} \sim \beta n^{\frac{\boldsymbol{D}(\sigma-2)+2}{2\boldsymbol{D}}}\;.$$ We find that there is a critical value of nonlinearity power $\sigma_c = 2$ such that the number of volume resonances grows for $\sigma < \sigma_c$ with time, drops for $\sigma > \sigma_c$ and stays constant for $\sigma=\sigma_c$. Therefore subdiffusive spreading will be more effective for $\sigma < \sigma_c$.
We also find that the number of surface resonances will grow with time for $$\boldsymbol{D} > \boldsymbol{D_c}=\frac{1}{1-\sigma/2}\;,\; \sigma < 2\;.$$ Therefore, for these cases, the wave packet surface will not stay compact. Instead surface resonances will lead to a resonant leakage of excitations into the exterior. This process will increase the surface area, and therefore lead to even more surface resonances, which again increase the surface area, and so on. The wave packet will fragmentize, perhaps get a fractal-like structure, and lower its compactness index. The spreading of the wave packet will speed up, but will not anymore be due to pure incoherent transfer, instead it will become a complicated mixture of incoherent and coherent transfer processes.
Mulansky computed spreading exponents for single site excitations with $\beta=1$, $W=4$, $L=1$, $\boldsymbol{D}=1$ $n=1$ and $\sigma =1,2,4,6$ [@mm09]. Since for $\sigma=2,4,6$ strong chaos is avoided, the fitting of the dependence $m_s(t)$ with a single power law is reasonable. The corresponding fitted exponents $0.31\pm0.04$ ($\sigma=2$), $0.18 \pm 0.04$ ($\sigma=4$) and $0.14 \pm 0.05$ ($\sigma=6$) agree well with the predicted weak chaos result $1/3,1/5,1/7$ from (\[sigma\_weak\]). For $\sigma=1$ the initial condition is launched in the regime of strong chaos. A single power law fit will therefore not be reasonable. Since the outcome is a mixture of first strong and later possibly weak chaos, the fitted exponent should be a number which is located between the two theoretical values $1/2$ and $2/3$. Indeed, Mulansky reports a number $0.56\pm 0.04$ confirming our prediction. Veksler et al [@hvyksf09] considered short time evolutions of single site excitations (up to $t=10^3$). While the time window may happen to be too short for conclusive results, the observed increase of fitted exponents with increasing $\beta$ for $\sigma < 2$ is possibly also influenced by the crossover from weak to strong chaos. Note that Skokos et al [@sf10] performed a more detailed analysis for the KG lattice, which confirm many the above results.
Discussion and conclusions
==========================
Let us summarize the findings. If the strength of nonlinearity is large enough, a wave packet (or at least an appreciable part of it) is selftrapped due to the finite bounds for the spectrum of the linear equation. If the nonlinearity is weak enough so as to avoid selftrapping, two possible outcomes are predicted, which now depend also on the volume $L$ of the packet. If $L > L_c$ and $n > n_c$, the NMs in the packet will be all resonant, strongly interacting with each other and quickly dephase. That leads to a regime of strong chaos. As time grows, the norm density $n$ drops below $n_c$, and the spreading continues in the regime of weak chaos. If however either $L < L_c$ or $n < n_c$, strong chaos is avoided, and the packet will spread in the regime of weak chaos. Lowering $\beta$ or $n$ further will keep the spreading in the regime of weak chaos, but time scales of subdiffusion will grow, and the process will not be observable on the finite time window currently accessible by computational experiments. The above holds if $\boldsymbol{D} < \boldsymbol{D_c}$ which implies that Anderson localization is preserved in the tails and destroyed in the wave packet core. In other words, the time scales for destroying Anderson localization in the tails are much larger than the time scales which lead to a thermalization of the core and the corresponding subdiffusive spreading. In order to observe the crossover from strong to weak chaos, one has to carefully choose the system parameters. In particular, it is desirable to make the crossover region more narrow.
If $\boldsymbol{D} > \boldsymbol{D_c}$ then the spreading process will be different from the above predictions, because resonant interaction in the surface and the tails of the wave packet will destroy Anderson localization as well. The spreading will presumably stay subdiffusive. But we do not know currently how to estimate and characterize the details of this process.
Our results rely on a conjecture of the dependence of a diffusion coefficient on the probability of resonances. Future investigations may consider the connection between this conjecture and the dependence of Lyapunov coefficients, relaxation times of correlation functions, and detrapping times on the system parameters.\
Acknowledgements\
I thank I. Aleiner, B. Altshuler, S. Aubry, J. Bodyfelt, S. Fishman, D. Krimer, Y. Krivolapov, T. Lapteva, N. Li, Ch. Skokos, and H. Veksler for useful discussions.
[00]{}
P. W. Anderson, Phys. Rev. [**109**]{} 1492 (1958).
T. Schwartz, G. Bartal, S. Fishman and M. Segev, Nature **446** 52 (2007)
Y. Lahini, A. Avidan, F. Pozzi, M. Sorel, R. Morandotti, D. N. Christodoulides and Y. Silberberg, Phys. Rev. Lett. **100** 013906 (2008).
D. Clement A. F. Varon, J. A. Retter, L. Sanchez-Palencia, A. Aspect and P. Bouyer, New J. Phys. **8** 165 (2006); L. Sanches-Palencia D. Clement, P. Lugan, P. Bouyer, G. V. Shlyapnikov and A. Aspect, Phys. Rev. Lett. **98** 210401 (2007); J. Billy, V. Josse, Z. Zuo, A. Bernard, B. Hambrecht, P. Lugan, D. Clement, L. Sanchez-Palencia, P. Bouyer and A. Aspect, Nature [**453**]{}, 891 (2008); G. Roati, C. D’Errico, L. Fallani, M. Fattori, C. Fort, M. Zaccanti, G. Modugno, M. Modugno and M. Inguscio, Nature [**453**]{}, 895 (2008).
D. Weinmann, S. Kohler, G.-L. Ingold and P. Hänggi, Ann. Phys. (Leipzig) **8** 277 (1999); C. Aulbach, A. Wobst, G.-L. Ingold, P. Hänggi and I. Varga, New J. Phys. [**6**]{} 70 (2004).
M. I. Molina, Phys. Rev. B **58** 12547 (1998).
A. S. Pikovsky and D. L. Shepelyansky, Phys. Rev. Lett. **100** 094101 (2008).
I. García-Mata and D. L. Shepelyansky, Phys. Rev. E, **79** 026205 (2009).
G. Kopidakis, S. Komineas, S. Flach and S. Aubry, Phys. Rev. Lett. **100** 084103 (2008).
S. Flach, D. Krimer and Ch. Skokos, Phys. Rev. Lett. **102** 024101 (2009).
Ch. Skokos, D. O. Krimer, S. Komineas and S. Flach, Phys. Rev. E **79** 056211 (2009).
H. Veksler, Y. Krivolapov and S. Fishman, Phys. Rev. E **80** 037201 (2009).
M. Mulansky, [*Localization Properties of Nonlinear Disordered Lattices*]{}, Diplomarbeit Universität Potsdam (2009); http://opus.kobv.de/ubp/volltexte/2009/3146/ .
M. Mulansky, K. Ahnert, A. Pikovsky and D. Shepelyansky, Phys. Rev. E **80** 056212 (2009).
O. Morsch and M. Oberthaler, Rep. Prog. Phys. **78** 176 (2006).
Yu. S. Kivshar and G. P. Agrawal, [*Optical Solitons: From Fibers to Photonic Crystals*]{} (Academic Press, Amsterdam, 2003).
A. A. Ovchinnikov, N. S. Erikhman and K. A. Pronin, Vibrational-Rotational Excitations in Nonlinear Molecular Systems, Kluwer Academic / Plenum Publishers (New York) 2001.
Yu. S. Kivshar and M. Peyrard, Phys. Rev. A [**46**]{}, 3198 (1992); Yu. S. Kivshar, Phys. Lett. A [**173**]{}, 172 (1993); M. Johansson, Physica D [**216**]{}, 62 (2006).
B. Kramer and A. MacKinnon, Rep. Prog. Phys. **56** 1469 (1993).
S. Flach and C. R. Willis, Phys. Rep. **295** 181 (1998); S. Flach and A. V. Gorbach, ibid. **467**, 1 (2008).
J. Bodyfelt and T. Lapteva, private communication.
H. Veksler, Y. Krivolapov and S. Fishman, Phys. Rev. E **81** 017201 (2010).
D. O. Krimer and S. Flach, in preparation.
B. Tuck, J. Phys. D **9** 1559 (1976).
W. F. Ames, [*Non-linear Partial Differential Equations in Engineering*]{} vol. 1 (Academic, New York, 1965); F. A. Cunnell and C. H. Gooch, J. Phys. Chem. Solids **15** 127 (1960); H. Ikezi, Y. Kiwamoto, K. E. Lonngren, C. M. Burde and H. C. S. Hsuan, Plasma Phys. **15** 1141 (1973); M. A. H. Kadhim and B. Tuck, J. Mater. Sci. **7** 68 (1972); K. E. Lonngren, W. F. Ames, A. Hirose and J. Thomas, Phys. Fluids **17** 1919 (1974). H. Skokos and S. Flach, in preparation.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'General expressions for the longitudinal and transverse resistivities of single-crystalline cubic and tetragonal ferromagnets are derived from a series expansion of the resistivity tensor with respect to the magnetization orientation. They are applied to strained (Ga,Mn)As films, grown on (001)- and (113)A-oriented GaAs substrates, where the resistivities are theoretically and experimentally studied for magnetic fields rotated within various planes parallel and perpendicular to the sample surface. We are able to model the measured angular dependences of the resistivities within the framework of a single ferromagnetic domain, calculating the field-dependent orientation of the magnetization by numerically minimizing the free-enthalpy density. Angle-dependent magnetotransport measurements are shown to be a powerful tool for probing both anisotropic magnetoresistance and magnetic anisotropy. The anisotropy parameters of the (Ga,Mn)As films inferred from the magnetotransport measurements agree with those obtained by ferromagnetic resonance measurements within a factor of two.'
author:
- 'W. Limmer'
- 'M. Glunk'
- 'J. Daeubler'
- 'T. Hummel'
- 'W. Schoch'
- 'R. Sauer'
- 'C. Bihler'
- 'H. Huebl'
- 'M. S. Brandt'
- 'S. T. B. Goennenwein'
title: 'Angle-dependent magnetotransport in cubic and tetragonal ferromagnets: Application to (001)- and (113)A-oriented (Ga,Mn)As'
---
Introduction
============
Realization of ferromagnetism in III-V semiconductors by introducing high concentrations of magnetic elements has motivated intense research on the dilute magnetic semiconductor (Ga,Mn)As. This system is considered a potential candidate for spintronic applications due to its compatibility with conventional semiconductor technology.[@Ohn98; @Jun05] In (Ga,Mn)As, magnetic Mn acceptors are predominantly incorporated on cation sites as Mn$^{2+}$ ions having a total spin of $S$ = 5/2. (Ga,Mn)As is paramagnetic at room temperature and undergoes a transition to the ferromagnetic phase at the Curie temperature $T_C$, where maximum values of up to $\sim$170 K have been reported so far.[@Jun05] The ferromagnetism has been successfully explained within the Zener mean-field model by an indirect Mn-Mn exchange interaction mediated by delocalized holes.[@Die01] (Ga,Mn)As is grown by low-temperature molecular-beam epitaxy and, if necessary, subjected to post-growth annealing to reduce the density of compensating defects. Considerable progress has been made in understanding its structural, electronic, and magnetic properties. In particular, anisotropic magnetoresistance (AMR),[@Jun03; @Jun02; @Bax02; @Mat04; @Wan05a; @Goe05] planar Hall effect (PHE),[@Tan03] and magnetic anisotropy (MA),[@Wel03; @Liu03; @Liu05; @Saw05; @Wan05b; @Bih06; @Lim06] have been identified as characteristic features, making (Ga,Mn)As potentially suitable for field-sensitive devices and non-volatile memories. These properties have been shown to be governed by several parameters, such as Mn concentration, hole density, strain, or temperature. Most of the work carried out on AMR in (Ga,Mn)As, however, has been restricted to special cases where the magnetic field was applied parallel or perpendicular to the layer and equations for the angular dependence of the longitudinal and transverse resistivities, describing the AMR and the PHE, respectively, have been given only for in-plane configuration and polycrystalline films. A comprehensive theoretical model describing the resistivities as a function of arbitrary field orientation is still missing. Moreover, (Ga,Mn)As layers are usually grown on GaAs(001) substrates and only little is known about the magnetic properties of films grown on high-index substrates.[@Bih06; @Lim06; @Dae06; @Omi01; @Weg05; @Wan05c]
In this work, the longitudinal and transverse resistivities of (Ga,Mn)As layers, grown on (001)- and (113)A-oriented GaAs substrates, are studied for arbitrarily orientated magnetic fields. The anisotropy of the resistivities and the MA are experimentally probed by rotating the magnetic field $\bm{H}$ at fixed field strengths within different planes parallel and perpendicular to the sample surface. General expressions for the resistivities, holding for single-crystalline cubic and tetragonal ferromagnets, are derived from a series expansion of the resistivity tensor with respect to the direction cosines of the magnetization $\bm{M}$. The measured data are well modeled by applying the expressions to the given experimental configurations, assuming the (Ga,Mn)As films to consist of a single ferromagnetic domain. Analytical expressions, widely used in the literature to describe the angular dependence of AMR and PHE, are shown to be inapproriate to single-crystalline materials. Finally, anisotropy parameters are estimated from the low-field magnetotransport data and compared with those obtained from ferromagnetic resonance (FMR) spectroscopy.
Experimental details
====================
40-nm-thick (001) and (113)A (Ga,Mn)As films with Mn concentrations of $\sim$5% were simultaneously grown by low-temperature molecular-beam epitaxy (MBE) in a RIBER 32 MBE machine on semi-insulating GaAs(001) and GaAs(113)A substrates mounted together on the same Mo holder. A conventional Knudsen cell and a hot-lip effusion cell were used to provide the Ga and Mn fluxes, respectively. A valved arsenic cracker cell was operated in the non-cracking mode to supply As$_4$ with a maximum V/III flux ratio of about 5. First, a 30-nm-thick GaAs buffer layer was grown at a temperature of $T_s$ $\approx$ 580$^\circ$C (conventional growth temperature for GaAs), then the growth was interrupted and $T_s$ was lowered to $\sim$250$^\circ$C. The Mn concentrations in the (Ga,Mn)As films were determined by flux measurements.
![\[coordinates\] Orientation of the (001) and (113)A samples with respect to the crystallographic axes.](fig1.eps)
For the magnetotransport and FMR studies, the (001) and (113)A samples were cleaved into small rectangular pieces with edges along \[110\] and \[$\bar{1}$10\], and along \[33$\bar{2}$\] and \[$\bar{1}$10\], respectively. The \[110\] and \[$\bar{1}$10\] directions of the (001) sample were identified by selective wet chemical etching and the orientation of the \[001\] crystal axis in the (113)A sample was determined by x-ray diffraction. The crystallographic orientations of the samples are shown in Fig. \[coordinates\]. Hall bars with Ti-AuPt-Au contacts were prepared on several pieces of the cleaved (001) and (113)A samples with the current direction along \[110\] and \[33$\bar{2}$\], respectively. The width of the Hall bars was 0.3 mm and the longitudinal voltage probes were separated by 1 mm. Hole densities of 3$\times 10^{20}$ cm$^{-3}$ for the (001) sample and 1.2$\times 10^{20}$ cm$^{-3}$ for the (113)A sample were determined by means of high-field magnetotransport measurements (up to 14.5 T) at 4.2 K using an Oxford SMD 10/15/9 VS liquid helium cryostat with superconducting coils. Least squares fits were performed to separate the contributions of the normal and anomalous Hall effect. Curie temperatures of $T_C$ $\approx$ 65 K and 54 K, respectively, were estimated from the peak positions of the temperature-dependent sheet resistivities at 10 mT. For the angle-dependent magnetotransport measurements carried out at 4.2 K, the Hall bars were mounted on the sample holder of a liquid-He-bath cryostat, which was positioned between the poles of a LakeShore electromagnet system providing a maximum field strength of 0.7 T. The sample holder possesses two perpendicular axes of rotation, allowing for an arbitrary alignment of the Hall bars with respect to the applied magnetic field $\bm{H}$. Using a dc current density of 8$\times 10^{2}$ Acm$^{-2}$, the longitudinal and transverse resistivities $\rho_{\mathrm{long}}$ and $\rho_{\mathrm{trans}}$ were measured at fixed magnitudes $\mu_0H$ = 0.1, 0.25 and 0.7 T of $\bm{H}$ while rotating its orientation. Prior to each angular scan, the magnetization $\bm{M}$ was put into a clearly defined initial state by raising the field to 0.7 T where $\bm{M}$ is nearly saturated and aligned with the external field. The field was then lowered to one of the above mentioned magnitudes and the scan was started.
The FMR measurements were carried out at 5 K in a commercial Bruker ESP 300 electron paramagnetic resonance spectrometer operated at a fixed frequency of $\omega_{HF}/2\pi$ $\approx$ 9.3 GHz (X-band). The spectrometer consists of a microwave bridge for the high-frequency radiation and an electromagnet providing a variable dc magnetic induction up to 1 T. To increase the sensitivity, lock-in techniques were used in which the dc induction is superimposed by a 100 kHz modulation field of 3.2 mT.
Theoretical overview
====================
In our theoretical considerations the total magnetic moment arising from the Mn-ion/hole spin complex is treated within the framework of the Stoner-Wohlfarth model,[@Sto48] i.e., for temperatures below $T_C$ the whole (Ga,Mn)As layer is assumed to consist of a single homogeneous ferromagnetic domain. This simple model has been astoundingly successful in describing a large variety of magnetization-related phenomena in (Ga,Mn)As. Under the given experimental conditions described above, domain nucleation and expansion, which have been shown to accompany in-plane and perpendicular magnetization-reversal processes,[@Liu05; @Wel03] are expected to play only a minor role. Accordingly, we may write the magnetization as a vector $\bm{M}=M\bm{m}$ where $M$ denotes its magnitude and the unit vector $\bm{m}$ its direction. In terms of the polar and azimuth angles $\theta$ and $\varphi$, respectively, which are defined in Fig. \[coordinates\], the components of $\bm{m}$ read as $m_x=\sin\theta\cos\varphi$, $m_y=\sin\theta\sin\varphi$, and $m_z=\cos\theta$. The equations used in the discussion of the angle-dependent magnetotransport data can be written in a concise way by introducing the unit vectors $\bm{j}$, $\bm{n}$, and $\bm{t}$, which specify the current direction, the surface normal, and an in-plane vector defined by $\bm{t}$ = $\bm{n}\times\bm{j}$, respectively. Throughout this work, all vector components refer to the cubic coordinate system with the \[100\], \[010\], and \[001\] directions of the crystal denoted by $x$, $y$, and $z$, respectively.
Longitudinal and transverse resistivities
-----------------------------------------
In standard magnetotransport measurements the longitudinal and transverse voltages, measured along and across the current direction, arise from the components $E_{long} = \bm{j}\cdot\bm{E}$ and $E_{trans} = \bm{t}\cdot\bm{E}$ of the electric field $\bm{E}$, respectively. Starting from Ohm’s law $\bm{E}=\bar{\rho}\cdot \bm{J}$, where $\bar{\rho}$ represents the resistivity tensor and $\bm{J}=J\bm{j}$ the current density, the corresponding longitudinal resistivity $\rho_{\mathrm{long}}$ (sheet resistivity) and transverse resistivity $\rho_{\mathrm{trans}}$ (Hall resistivity) can be written as $$\begin{aligned}
\label{define_resistivities}
\rho_{\mathrm{long}} &=& \frac{E_{long}}{J} = \bm{j} \cdot \bar{\rho} \cdot \bm{j},
\nonumber \\
\rho_{\mathrm{trans}} &=& \frac{E_{trans}}{J} = \bm{t} \cdot \bar{\rho} \cdot \bm{j}.\end{aligned}$$ In (Ga,Mn)As, as in many other ferromagnets, the resistivity tensor sensitively depends on the orientation of $\bm{M}$ with respect to the crystallographic axes.[@Bir64] Thus, in order to quantitatively model the measured resistivities in the general case of an arbitrarily oriented magnetization, a universal mathematical relationship between $\rho_{\mathrm{long}}$ and $\rho_{\mathrm{trans}}$ and the direction cosines $m_i$ of $\bm{M}$ has to be derived. For these purposes, we follow the ansatz of Birss[@Bir64] and Muduli et al.[@Mud05] and write the resistivity tensor $\bar{\rho}$ as a series expansion in powers of $m_i$ using the Einstein summation convention: $$\label{series_expansion}
\rho_{ij} = a_{ij} + a_{kij}m_k + a_{klij}m_km_l + \ldots\;.$$ For cubic symmetry T$_\mathrm{d}$, most of the components $a_{ij}$, $a_{kij}$, ..., of the galvanomagnetic tensors vanish and, considering terms up to the second order, we obtain $$\begin{aligned}
\label{rho_cubic}
\nonumber
\bar{\rho}^{\,\mathrm{cubic}} &=&
A
\left( \begin{array}{ccc}
1 & 0 & 0 \\
0 & 1 & 0 \\
0 & 0 & 1 \\
\end{array} \right) +
B
\left( \begin{array}{ccc}
m_x^2 & 0 & 0 \\
0 & m_y^2 & 0 \\
0 & 0 & m_z^2 \\
\end{array} \right) \nonumber \\
&+&
C
\left( \begin{array}{ccc}
0 & m_xm_y & m_xm_z \\
m_xm_y & 0 & m_ym_z \\
m_xm_z & m_ym_z & 0 \\
\end{array} \right) \nonumber \\
&+&
D
\left( \begin{array}{ccc}
0 & m_z & -m_y \\
-m_z & 0 & m_x \\
m_y & -m_x & 0 \\
\end{array} \right),\end{aligned}$$ with the resistivity parameters $$\begin{aligned}
\label{resistivity_parameters_1}
A &=& a_{11}+a_{1122}\;, \quad B = a_{1111}-a_{1122}\;, \nonumber \\
C &=& a_{2323}\;, \quad D = a_{123}.\end{aligned}$$ Insertion of Eq. (\[rho\_cubic\]) into Eqs. (\[define\_resistivities\]) and elementary vector algebra yields the general expressions $$\begin{aligned}
\label{rho_general_cubic}
\rho_{\mathrm{long}}^{\mathrm{cubic}} &=& A + C (\bm{j}\cdot \bm{m})^2
+ \left( B-C \right) \sum_{i}j_i^2 m_i^2, \nonumber \\
\rho_{\mathrm{trans}}^{\mathrm{cubic}} &=& C (\bm{j}\cdot \bm{m})(\bm{t}\cdot \bm{m})
+ \left( B-C \right) \sum_{i} t_ij_i m_i^2 \nonumber \\
&-& D (\bm{n}\cdot \bm{m}),\end{aligned}$$ which apply to single crystalline ferromagnetic materials of cubic symmetry. The transverse resistivity in Eqs. (\[rho\_general\_cubic\]) includes the contribution of the anomalous Hall effect, which correlates with the perpendicular component of $\bm{M}$, but it does not account for the ordinary Hall effect. For magnetic field strengths $\mu_0H < 1$ T and hole concentrations $p > 10^{20}$ cm$^{-3}$ as in our experiments, however, the maximum contribution of the ordinary Hall effect is $\mu_0H/ep \approx 6\cdot 10^{-6}$ $\Omega$ cm ($e$ denotes the elementary charge), and thus about two orders of magnitude smaller than the measured peak values of $\rho_{\mathrm{trans}}^{\mathrm{cubic}}$ (see Section IV).
In the following, Eqs. (\[rho\_general\_cubic\]) are applied to three different experimental configurations using the relation $$\label{mi_general}
m_i=j_i(\bm{j}\cdot \bm{m})+ t_i(\bm{t}\cdot \bm{m})+
n_i(\bm{n}\cdot \bm{m}).$$ In the simplest case of a sample with (001) surface, where the current flows along the \[100\] or \[010\] direction, Eqs. (\[rho\_general\_cubic\]) reduce to $$\begin{aligned}
\label{rho001_100_cubic}
\rho_{\mathrm{long}}^{\mathrm{cubic}} &=& A+ B(\bm{j}\cdot \bm{m})^2, \nonumber \\
\rho_{\mathrm{trans}}^{\mathrm{cubic}} &=& C (\bm{j}\cdot \bm{m})(\bm{t}\cdot \bm{m})
- D (\bm{n}\cdot \bm{m}).\end{aligned}$$ The magnetotransport studies presented in this work were performed on (001)- and (113)A-oriented samples with the current direction $\bm{j}$ along \[110\] and \[33$\bar{2}$\], respectively. The corresponding resistivities are:\
[*(001) surface and $\bm{j}$ $\parallel$ \[110\]*]{} $$\begin{aligned}
\label{rho001_110_cubic}
\rho_{\mathrm{long}}^{\mathrm{cubic}} & = & A + \frac{1}{2}\left( B-C \right) +
C(\bm{j}\cdot \bm{m})^2
\nonumber \\
& + & \frac{1}{2} \left( C-B \right) (\bm{n}\cdot \bm{m})^2, \nonumber \\
\rho_{\mathrm{trans}}^{\mathrm{cubic}} & = & B(\bm{j}\cdot \bm{m})(\bm{t}\cdot \bm{m})
- D (\bm{n}\cdot \bm{m}).\end{aligned}$$ [*(113)A surface and $\bm{j}$ $\parallel$ \[33$\bar{\it 2}$\]*]{} $$\begin{aligned}
\label{rho113_33m2_cubic}
\rho_{\mathrm{long}}^{\mathrm{cubic}} & = & A + \frac{9}{22} \left( B-C \right)
\nonumber \\
& + & \frac{1}{121}\left( 126C - 5B \right) (\bm{j}\cdot \bm{m})^2 \nonumber \\
& + & \frac{45}{242} \left( C-B \right) (\bm{n}\cdot \bm{m})^2 \nonumber \\
& + & \frac{15\sqrt{2}}{121} \left( B-C \right) (\bm{j}\cdot \bm{m})(\bm{n}\cdot \bm{m}),
\nonumber \\
\rho_{\mathrm{trans}}^{\mathrm{cubic}} & = &
\frac{1}{11}\left( 9B + 2C \right) (\bm{j}\cdot \bm{m})(\bm{t}\cdot \bm{m}) \nonumber \\
& + & \frac{3\sqrt{2}}{11} \left( B-C \right) (\bm{t}\cdot \bm{m})(\bm{n}\cdot \bm{m})
\nonumber \\
& - & D (\bm{n}\cdot \bm{m}).\end{aligned}$$
The $(\bm{j}\cdot \bm{m})^2$ terms of $\rho_{\mathrm{long}}^{\mathrm{cubic}}$ give rise to a dependence of the sheet resistivity on the relative orientation between magnetization $\bm{M}$ and current density $\bm{J}$, commonly referred to as AMR. Microscopically, it is explained by a strong spin-orbit coupling in the semiconductor valence band. Experimentally observed differences in the in-plane and out-of-plane AMR, often defined as $$\begin{aligned}
\label{amr_ipop}
\mathrm{AMR}_{\mathrm{ip}} &=&
\frac{\rho_{\mathrm{long}}(\bm{m}\parallel\bm{j})-\rho_{\mathrm{long}}(\bm{m}\parallel\bm{t})}
{\rho_{\mathrm{long}}(\bm{m}\parallel\bm{t})}, \nonumber \\
\mathrm{AMR}_{\mathrm{op}} &=&
\frac{\rho_{\mathrm{long}}(\bm{m}\parallel\bm{j})-\rho_{\mathrm{long}}(\bm{m}\parallel\bm{n})}
{\rho_{\mathrm{long}}(\bm{m}\parallel\bm{n})},\end{aligned}$$ have been ascribed to biaxial strain in the layer.[@Jun02; @Bax02; @Mat04; @Wan05a] According to Eqs. (\[rho001\_110\_cubic\]) and (\[rho113\_33m2\_cubic\]), however, such differences may be expected even in the case of perfect cubic symmetry due to the $(\bm{n}\cdot \bm{m})^2$ terms of $\rho_{\mathrm{long}}^{\mathrm{cubic}}$. As will be shown below, a strain-induced tetragonal distortion leads to further $(\bm{n}\cdot \bm{m})^2$ terms, additionally affecting the difference between AMR$_{\mathrm{ip}}$ and AMR$_{\mathrm{op}}$. The PHE, represented by the $(\bm{j}\cdot \bm{m})(\bm{t}\cdot \bm{m})$ terms of $\rho_{\mathrm{trans}}$, is closely related to the AMR and describes the appearance of a transverse voltage in the presence of an in-plane magnetic field, or more precisely, of an in-plane magnetization. From the summation terms in Eqs. (\[rho\_general\_cubic\]) it becomes clear that $\rho_{\mathrm{long}}$ and $\rho_{\mathrm{trans}}$, and thus AMR and PHE, not only depend on the relative orientation between $\bm{m}$ and $\bm{j}$, but also on the orientations of $\bm{m}$, $\bm{j}$, and $\bm{t}$ with respect to the crystal axes.
So far, quantitative studies on the angular dependences of the AMR and the PHE in (Ga,Mn)As were restricted to in-plane configurations with $\bm{n}\cdot \bm{m}$ = 0 and the discussions were based on the well-known expressions[@McG75; @Jan57] $$\begin{aligned}
\label{amr_phe}
\rho_{\mathrm{long}} &=& \rho_{\perp} + (\rho_{\parallel}-\rho_{\perp}) \cos^2\phi_j\,,
\nonumber \\
\rho_{\mathrm{trans}} &=& (\rho_{\parallel}-\rho_{\perp}) \sin\phi_j \cos\phi_j,\end{aligned}$$ with $\phi_j$ denoting the angle between $\bm{j}$ and $\bm{M}$. These expressions, however, only hold for polycrystalline films, whereas (Ga,Mn)As layers are normally of high crystalline quality with a uniform crystallographic orientation of the layer. Accordingly, the expressions given in Eqs. (\[rho001\_100\_cubic\])–(\[rho113\_33m2\_cubic\]) are incompatible with Eqs. (\[amr\_phe\]) and cannot be brought into agreement by simply setting $\bm{n}\cdot \bm{m}$ = 0. In fact, Eqs. (\[amr\_phe\]) result from Eqs. (\[rho\_general\_cubic\]) by averaging the summation terms over all possible crystal orientations in space[@Bir60] with $\bm{M}$ lying in the plane spanned by $\bm{j}$ and $\bm{t}$ $$\begin{aligned}
\label{aver_sums}
\sum_{i}\overline{j_i^2 m_i^2} & = & \frac{1}{5}\left[2(\bm{j}\cdot \bm{m})^2 + 1\right],
\nonumber \\
\sum_{i}\overline{t_ij_i m_i^2} & = & \frac{2}{5}(\bm{j}\cdot \bm{m})(\bm{t}\cdot \bm{m}).\end{aligned}$$ Inserting the averaged terms into Eqs. (\[rho\_general\_cubic\]) and using the relations $\bm{j}\cdot \bm{m}$ = $\cos\phi_j$ and $\bm{t}\cdot \bm{m}$ = $\sin\phi_j$ (the latter only holds for $\bm{n}\cdot \bm{m}$ = 0), we obtain $$\begin{aligned}
\label{rho_poly}
\rho_{\mathrm{long}}^{\mathrm{poly}} & = & A + \frac{1}{5}\left( B-C \right) +
\frac{1}{5} \left( 2B+ 3C \right)\cos^2\phi_j, \nonumber \\
\rho_{\mathrm{trans}}^{\mathrm{poly}} & = & \frac{1}{5} \left( 2B+ 3C \right)
\sin\phi_j \cos\phi_j.\end{aligned}$$ These equations are formally identical to Eqs. (\[amr\_phe\]) and apply to polycrystalline materials. A comparison between Eqs. (\[amr\_phe\]) and Eqs. (\[rho\_poly\]) allows us to connect the quantities $\rho_{\parallel}$ and $\rho_{\perp}$ to the components of the galvanomagnetic tensors for cubic symmetry up to second order $$\begin{aligned}
\label{rhos_rhop}
\rho_{\parallel} & = & A + \frac{1}{5}\left( 3B + 2C \right) \nonumber \\
& = & a_{11} + \frac{1}{5}\left(3a_{1111} + 2a_{1122} + 2a_{2323}\right), \nonumber \\
\rho_{\perp} & = & A + \frac{1}{5}\left( B-C \right) \nonumber \\
& = & a_{11} + \frac{1}{5}\left(a_{1111} + 4a_{1122} - a_{2323}\right) .\end{aligned}$$ Thus, in general Eqs. (\[amr\_phe\]) are not appropriate to describe the in-plane AMR and the PHE in single-crystalline (Ga,Mn)As layers. Only in the limiting case where $B=C$, Eqs. (\[rho001\_100\_cubic\])–(\[rho113\_33m2\_cubic\]) simplify to Eqs. (\[amr\_phe\]).
Analyzing the angle-dependent magnetotransport data presented in Section IV, it turns out that additional terms proportional to $(\bm{n}\cdot \bm{m})^2$ have to be introduced in the expressions of $\rho_{\mathrm{long}}^{\mathrm{cubic}}$ to achieve a satisfactory description of the experimental results.[@Lim06] They are supposed to originate from a distortion of the crystal lattice due to compressive strain in the (Ga,Mn)As layer. To account for such strain-induced effects in a correct way, we extend our model to a tetragonal distortion of the cubic lattice along the \[001\] direction. As will be shown below, our FMR and magnetotransport data suggest that this applies not only for the (001)- but also for the (113)A-oriented sample.
In the case of a tetragonal lattice distortion along \[001\], the symmetry reduces to D$_{\mathrm{2d}}$ and the series expansion of $\bar{\rho}$ in Eq. (\[series\_expansion\]) yields further contributions which can be subsumed into an additional term $\Delta \bar{\rho}$. The resistivity tensor then reads as $$\label{rho_tetragonal}
\bar{\rho}^{\,\mathrm{tetra}} = \bar{\rho}^{\,\mathrm{cubic}} + \Delta\bar{\rho},$$ with $$\begin{aligned}
\label{delta_rho}
\Delta\bar{\rho} & = &
\left( \begin{array}{ccc}
0 & 0 & 0 \\
0 & 0 & 0 \\
0 & 0 & a \\
\end{array} \right)
+
\left( \begin{array}{ccc}
0 & d m_z & 0 \\
-d m_z & 0 & 0 \\
0 & 0 & 0 \\
\end{array} \right) \nonumber \\
& + &
\left( \begin{array}{ccc}
b_1 m_z^2 & c m_xm_y & 0 \\
c m_xm_y & b_1 m_z^2 & 0 \\
0 & 0 & b_2 m_z^2 \\
\end{array} \right).\end{aligned}$$ The additional resistivity parameters are given by $$\begin{aligned}
\label{resistivity_parameters_2}
a &=& a_{33}-a_{11}+a_{1133}-a_{1122} , \nonumber \\
b_1 &=& a_{3311}-a_{1122}, \nonumber \\
b_2 &=& a_{3333}-a_{1111}-a_{1133}+a_{1122} , \nonumber \\
c &=& a_{1212}-a_{2323}, \nonumber \\
d &=& a_{312} - a_{123}.\end{aligned}$$ Accordingly, Eqs. (\[rho001\_100\_cubic\])–(\[rho113\_33m2\_cubic\]) have to be rewritten as:\
[*(001) surface and $\bm{j}$ $\parallel$ \[100\]*]{} $$\begin{aligned}
\label{rho001_100_tetragonal}
\rho_{\mathrm{long}}^{\mathrm{tetra}} & = & A + B (\bm{j}\cdot \bm{m})^2 +
b_1 (\bm{n}\cdot \bm{m})^2 \nonumber \\
\rho_{\mathrm{trans}}^{\mathrm{tetra}} & = & (C+c) (\bm{j}\cdot \bm{m})(\bm{t}\cdot \bm{m})
\nonumber \\
& - & (D+d) (\bm{n}\cdot \bm{m}).\end{aligned}$$ [*(001) surface and $\bm{j}$ $\parallel$ \[110\]*]{} $$\begin{aligned}
\label{rho001_110_tetragonal}
\rho_{\mathrm{long}}^{\mathrm{tetra}} & = & A + \frac{1}{2}( B-C-c ) +
(C+c) (\bm{j}\cdot \bm{m})^2 \nonumber \\
&+& \left[ \frac{1}{2} ( C-B+c ) + b_1 \right] (\bm{n}\cdot \bm{m})^2 \nonumber \\
\rho_{\mathrm{trans}}^{\mathrm{tetra}} & = & B (\bm{j}\cdot \bm{m})(\bm{t}\cdot \bm{m}) -
(D+d) (\bm{n}\cdot \bm{m}),\end{aligned}$$ [*(113)A surface and $\bm{j}$ $\parallel$ \[33$\bar{\it 2}$\]*]{} $$\begin{aligned}
\label{rho113_33m2_tetragonal}
\rho_{\mathrm{long}}^{\mathrm{tetra}} & = & A + \frac{2}{11}a + \frac{9}{22} \left( B-C-c \right)
\nonumber \\
& + & \frac{1}{121}\left( 126C - 5B + b + 90c\right) (\bm{j}\cdot \bm{m})^2 \nonumber \\
& + & \frac{9}{242} \left( 5C - 5B + b + 13c \right) (\bm{n}\cdot \bm{m})^2 \nonumber \\
& + & \frac{3\sqrt{2}}{121} \left( 5B - 5C - b + 9c \right) (\bm{j}\cdot \bm{m})(\bm{n}\cdot \bm{m}),
\nonumber \\
\rho_{\mathrm{trans}}^{\mathrm{tetra}} & = &
\frac{1}{11}\left( 9B + 2C \right) (\bm{j}\cdot \bm{m})(\bm{t}\cdot \bm{m}) \nonumber \\
& + & \frac{3\sqrt{2}}{11} \left( B-C \right) (\bm{t}\cdot \bm{m})(\bm{n}\cdot \bm{m}) \nonumber \\
& - & \left( D + \frac{9}{11}d \right) (\bm{n}\cdot \bm{m}) \nonumber \\
& + & \frac{3\sqrt{2}}{11} d (\bm{j}\cdot \bm{m}),\end{aligned}$$ where $b = 18b_1 + 4b_2$. For perfect cubic symmetry the parameters $a$, $b_1$, $b_2$, $c$, and $d$ vanish and Eqs. (\[rho001\_100\_tetragonal\])–(\[rho113\_33m2\_tetragonal\]) reduce to Eqs. (\[rho001\_100\_cubic\])–(\[rho113\_33m2\_cubic\]). It should be emphasized that the expressions for $\rho_{\mathrm{long}}$ and $\rho_{\mathrm{trans}}$ derived above generally apply to ferromagnets of cubic or tetragonal symmetry, provided that the angular dependence of the resistivity tensor is exclusively determined by the direction cosines of the magnetization. Effects correlated with the magnitude $B$ of the magnetic induction $\bm{B}$, such as the negative magnetoresistance, can be easily taken into account by considering $B$-dependent resistivity parameters.
Magnetic anisotropy
-------------------
The pronounced MA in (Ga,Mn)As is associated with a density of the free enthalpy[@comment] $G$ being highly anisotropic with respect to the orientation of $\bm{M}$. The direction of $\bm{M}$, i.e., the vector $\bm{m}$ which enters the equations for $\rho_{\mathrm{long}}$ and $\rho_{\mathrm{trans}}$ given above, aligns in such a way that $G$ takes its minimum. In addition to the single-domain model, we assume that the magnitude $M$ of the magnetization is nearly constant under the given experimental conditions while its orientation $\bm{m}$ is strongly affected by the applied magnetic field $\bm{H}$. Instead of $G$ we therefore consider the normalized quantity $G_M$ = $G/M$, allowing for a more concise description of the MA. For a biaxially strained (Ga,Mn)As film grown on GaAs(001) substrate, it can be written as[@Liu05] $$\begin{aligned}
\label{FE_001}
G^{001}_{M} &=& -\mu_0\bm{H} \cdot \bm{m}+ B_{c\parallel}\left( m_x^4 + m_y^4 \right)
+ B_{c\perp}m_z^4
\nonumber \\
&+& B_{001}(\bm{n}\cdot\bm{m})^2 + B_{\bar{1}10}(\bm{t}\cdot\bm{m})^2,\end{aligned}$$ with $\bm{n}$ $\parallel$ \[001\] and $\bm{t}$ $\parallel$ \[$\bar{1}$10\]. The terms refer, respectively, to the Zeeman energy, to the cubic anisotropy under tetragonal distortion, to an effective uniaxial anisotropy perpendicular to the film including demagnetization and magnetoelastic effects, and to a uniaxial in-plane contribution whose origin is still under discussion.[@Saw05; @Wel04; @Ham06] The anisotropy parameters $B_i$ introduced in Eq. (\[FE\_001\]) are in SI units. Expressed by the anisotropy fields $H_i$ and $4\pi M_{eff}$ used in Refs. and , they read as $B_{c\parallel}$ = $-\mu_0H_{4\parallel}/4$, $B_{c\perp}$ = $-\mu_0H_{4\perp}/4$, $B_{\bar{1}10}$ = $-\mu_0H_{2\parallel}/2$, and $B_{001}$ = $\mu_{0}4\pi M_{eff}/2$. Note that by using the trivial identity $$\label{identity}
|\bm{m}|^2 = (\bm{j}\cdot\bm{m})^2 + (\bm{t}\cdot\bm{m})^2 + (\bm{n}\cdot\bm{m})^2 =1,$$ Eq. (\[FE\_001\]) can be easily converted to a completely equivalent expression where the in-plane contribution along $[\bar{1}10]$ is formally replaced by a contribution along $[110]$. The only consequence is a redefinition of the anisotropy parameters and the addition of a constant term which does not alter the physical information provided by $G_M$. In Fig. \[FE\_3D\_001\], the free enthalpy is visualized by a 3D plot, calculated for a weak magnetic field $\mu_0H$ = 0.15 T and a set of anisotropy parameters with arbitrarily chosen values $B_{c\parallel}$ = $B_{c\perp}$ = -0.1 T, $B_{001}$ = 0.15 T, and . The direction of $\bm{M}$ was calculated, as throughout the present work, by numerically minimizing $G_M$ with respect to $\theta$ and $\varphi$ (see Fig. \[coordinates\]). In doing so, we are able to trace the motion of $\bm{M}$, starting from a given orientation, while sweeping or rotating $\bm{H}$. Figure \[Sim\_mag\] shows as an example the simulated polar ($\theta_H$,$\theta$) and azimuth ($\varphi_H$,$\varphi$) angles of $\bm{H}$ (dashed line) and $\bm{M}$ (solid line), respectively, while $\bm{H}$ is rotated within the (111) plane. For the simulation the same field strength and the same anisotropy parameters have been chosen as for the 3D plot in Fig. \[FE\_3D\_001\]. While $\bm{H}$ smoothly rotates within the (111) plane, $\bm{M}$ remains very close to the (001) plane ($\theta \approx 90^{\circ}$) and undergoes sudden jumps in $\varphi$ whenever the minimum of $G_M$ discontiuously changes its position.
![\[FE\_3D\_001\] $G_M$ as a function of $\bm{m}$, calculated for a given magnetic field $\mu_0H$ = 0.15 T and a set of anisotropy parameters with arbitrarily chosen values $B_{c\parallel}$ = $B_{c\perp}$ = -0.1 T, $B_{001}$ = 0.15 T, and . The equilibrium position of $\bm{M}$ is determined by the minimum of $G_M$.](fig2.eps)
![\[Sim\_mag\] Simulated polar (left axis) and azimuth (right axis) angles of the magnetic field $\bm{H}$ (dashed lines) and the magnetization $\bm{M}$ (solid lines) for $\bm{H}$ rotated in the (111) plane. The same field strength and anisotropy parameters have been used as in Fig. \[FE\_3D\_001\].](fig3.eps)
In the case of the (Ga,Mn)As films grown on GaAs(113)A substrates, the best fits to the experimental data (see Section IV B) are achieved for a normalized free-enthalpy density of the form[@Lim06] $$\begin{aligned}
\label{FE_113}
G^{113}_{M} &=& -\mu_0\bm{H} \cdot \bm{m}+ B_{c\parallel}\left( m_x^4 + m_y^4 \right)
+ B_{c\perp}m_z^4 \nonumber \\
&+& B_{113}(\bm{n}\cdot\bm{m})^2 + B_{\bar{1}10}(\bm{t}\cdot\bm{m})^2 \nonumber \\
&+& B_{001} m_z^2.\end{aligned}$$ The first five terms correspond to those already presented in Eq. (\[FE\_001\]). The sixth term, which has to be additionally introduced to obtain an optimal agreement between experiment and theory, is an inclined uniaxial contribution along \[001\], i.e., neither parallel nor perpendicular to the film.[@Bih06] We attribute it to a lattice distortion along the \[001\] direction.
Ferromagnetic resonance
-----------------------
A highly efficient and widely used tool to study MA is ferromagentic resonance spectroscopy.[@Goe03] Most recently, a detailed review on FMR in (Ga,Mn)As has been given by Liu and Furdyna.[@Liu06] In the FMR experiments, the total magnetic moment of the Mn-ion/hole spin complex and thus the magnetization $\bm{M}$ precesses around its equilibrium position (which in general is not identical to the orientation of $\bm{H}$) at the Larmor frequency $\omega_L$. Sweeping the magnitude of the magnetic field $\bm{H}$ at a fixed microwave frequency $\omega_{HF}$, the resonance condition $\omega_L$ = $\omega_{HF}$ is fulfilled at the resonance field $H_{res}$ which strongly depends on the orientation of $\bm{H}$ due to MA. The resonance condition is given by[@Far98] $$\label{resonance_condition}
\left(\frac{\omega_{HF}}{\gamma}\right)^2 = \frac{1}{\sin^2\theta}
\left[\frac{\partial^2G_M}{\partial\theta^2}\frac{\partial^2G_M}{\partial\varphi^2}
-\left(\frac{\partial^2G_M}{\partial\theta\partial\varphi}\right)^2
\right],$$ where $\gamma = g\mu_B\hbar^{-1}$ denotes the gyromagnetic ratio, $g$ the $g$-factor, $\mu_B$ the Bohr magneton, and $\hbar$ the Planck constant. At any given direction of $\bm{H}$, the resonance field is obtained by solving Eq. (\[resonance\_condition\]) at the equilibrium position of $\bm{M}$, i.e., for $\partial G_M/\partial \varphi$ = 0 and $\partial G_M/\partial \theta$ = 0. The anisotropy parameters can then be derived from a fit to the measured $H_{res}$ recorded as a function of field orientation. In the present work, most of the calculations were carried out numerically.
Results and discussion
======================
The longitudinal and transverse resistivities of the (001) and (113)A (Ga,Mn)As layers were measured as a function of the magnetic field orientation at fixed field strengths $\mu_0H$ = 0.1, 0.25, and 0.7 T. In order to probe the anisotropy in all three directions in space, the applied magnetic field $\bm{H}$ was rotated within three different crystallographic planes perpendicular to $\bm{n}$, $\bm{j}$, and $\bm{t}$, respectively, as shown in Fig. \[planes\]. In the following, the measured angular dependences of $\rho_{\mathrm{long}}$ and $\rho_{\mathrm{trans}}$ are discussed and values for the resistivity and anisotropy parameters are derived by fits to the experimental data using the theoretical formalism presented in Section III. The results of FMR measurements, carried out on the same samples, are presented for reference. Note, however, that it is not the aim of the present study to yield a detailed or complete set of anisotropy parameters. In fact, the work is meant to provide a comprehensive theoretical tool for the description of the resistivies in arbitrarily oriented (Ga,Mn)As layers and to demonstrate the potential of angle-dependent magnetotransport studies for the investigation of magnetic anisotropy.
![\[planes\] The angle-dependent magnetotransport measurements were carried out for $\bm{H}$ rotated within (a) the layer plane, (b) a plane perpendicular to the current direction $\bm{j}$, and (c) a plane spanned by $\bm{j}$ and the normal vector $\bm{n}$.](fig4.eps)
![\[mag\_110\] Calculated polar (left axis) and azimuth (right axis) angles of the magnetic field $\bm{H}$ (dashed lines) and the magnetization $\bm{M}$ (solid lines) for $\bm{H}$ rotated in the (110) plane of the (001) (Ga,Mn)As sample at 0.1 T. The resistivity and anisotropy parameters used in the calculation were derived from a fit to the angle-dependent resistivities shown in Fig. \[mt\_001\].](fig6.eps)
(001) orientation
-----------------
The anisotropy of the (001)-oriented sample was probed by rotating $\bm{H}$ within the (001), (110), and ($\bar{1}$10) planes. The corresponding angular dependences of $\rho_{\mathrm{long}}$ and $\rho_{\mathrm{trans}}$, measured with the current direction along \[110\], are shown in Fig. \[mt\_001\]. At 0.7 T the Zeeman energy dominates the free enthalpy and MA only plays a minor role. As a consequence, $\bm{M}$ is expected to nearly align with the applied magnetic field and to continuously follow the motion of $\bm{H}$. In fact, the curves of $\rho_{\mathrm{long}}$ and $\rho_{\mathrm{trans}}$ at 0.7 T are smooth and largely reflect the anisotropy of the resistivity tensor. With decreasing magnetic field the influence of the MA increases and the orientation of $\bm{M}$ deviates more and more from the field direction. Accordingly, jumps and kinks occur in the curves at 0.25 and 0.1 T, arising from sudden movements of $\bm{M}$ caused by discontinuous displacements of the minimum of the free enthalpy. The observed angular dependences of $\rho_{\mathrm{long}}$ and $\rho_{\mathrm{trans}}$ can be understood in great detail by modeling the measured curves within the theoretical framework presented in Section III. For this purpose, the resistivity and anisotropy parameters from Eqs. (\[resistivity\_parameters\_1\]), (\[resistivity\_parameters\_2\]), and (\[FE\_001\]) were determined by an iterative fit procedure. Starting with an initial guess for the anisotropy parameters, the resistivity parameters were obtained by fitting Eqs. (\[rho001\_110\_tetragonal\]) to the experimental data recorded at 0.7 T. Then the anisotropy parameters were modified for an optimal agreement at 0.25 and 0.1 T, and the whole procedure was repeated until no further improvement of the fit could be achieved. The unit vector $\bm{m}$ at any given magnetic field $\bm{H}$ was calculated by numerically minimizing $G^{001}_{M}$ in Eq. (\[FE\_001\]) with respect to the polar and azimuth angles of $\bm{M}$. With the exception of $A$, the resistivity parameters turned out to be field independent within the accuracy of the fit and are given by $B$ = $-2.3\times 10^{-4}$ $\Omega$ cm, $C+c$ = $-1.7\times 10^{-4}$ $\Omega$ cm, $b_1$ = $0.9\times 10^{-4}$ $\Omega$ cm, and $D+d$ = $-4.4\times 10^{-4}$ $\Omega$ cm. The resistivity parameter $A$ was found to decrease from $7.21\times 10^{-3}$ $\Omega$ cm at 0.1 T to $7.08\times 10^{-3}$ $\Omega$ cm at 0.7 T, reflecting the negative-magnetoresistance behavior of $\rho_{\mathrm{long}}$.[@Ohn98; @Mat04] For the anisotropy parameters we obtained the values $B_{c\parallel}$ = $-0.015$ T, $B_{c\perp}$ = 0 T, $B_{001}$ = 0.17 T, and $B_{\bar{1}10}$ = 0.002 T. The theoretical curves calculated with these parameters are in excellent agreement with the experiment and are drawn as solid lines in Fig. \[mt\_001\].
Once the anisotropy parameters are known, the orientations of the easy axes can be determined by minimizing $G^{001}_{M}$ with respect to $\bm{m}$ at zero magnetic field. The easy axes are found to lie within the (001) layer plane ($\theta$ = 90$^{\circ}$) at the azimuth angles $\varphi_1$ = 1.9$^{\circ}$ and $\varphi_2$ = 88.1$^{\circ}$ (see Fig. \[coordinates\]). The slight deviation from the cubic \[100\] and \[010\] axes towards the \[110\] direction arises from the positive value of $B_{\bar{1}10}$.
Using the resistivity and anisotropy parameters given above, plots similar to Fig. \[Sim\_mag\] can be drawn for each configuration and field strength, revealing in great detail the motion of $\bm{M}$. Figure \[mag\_110\] shows as an example the polar ($\theta_H$,$\theta$) and azimuth ($\varphi_H$,$\varphi$) angles of $\bm{H}$ and $\bm{M}$, respectively, plotted as a function of the angle of rotation $\alpha$ for $\bm{H}$ rotated in the (110) plane at 0.1 T \[see Fig. \[mt\_001\](b)\]. While $\theta_H$ passes through all values between $0^{\circ}$ and $180^{\circ}$ (dashed line, left axis), $\bm{M}$ remains very close to the (001) plane with $75^{\circ} < \theta < 105^{\circ}$ (solid line, left axis). For $40^{\circ} < \alpha < 140^{\circ}$ and $220^{\circ} < \alpha < 320^{\circ}$, where $\bm{H}$ is closer to the (001) plane than to the \[001\] axis, the azimuth angles $\varphi_H$ and $\varphi$ of $\bm{H}$ and $\bm{M}$, respectively, almost perfectly coincide. When $\bm{H}$ approaches the hard \[001\] axis, however, $\bm{M}$ tends towards the easy \[100\] axis (azimuth angles $0^{\circ}$ and $180^{\circ}$) or to the easy \[010\] axis (azimuth angles $-90^{\circ}$ and $90^{\circ}$). At $\alpha$ = $180^{\circ}$ and $\alpha$ = $360^{\circ}$, where $\bm{H}$ exactly aligns with the \[00$\bar{1}$\] and \[001\] directions, respectively, the azimuth angle of $\bm{M}$ undergoes a sudden jump by $90^{\circ}$.
Considerable information can also be obtained by comparing the measured angular dependences with those expected for the limiting case where $\bm{M}$ perfectly aligns with $\bm{H}$. To this end, $\rho_{\mathrm{long}}$ and $\rho_{\mathrm{trans}}$ were calculated with $\bm{m}$ replaced by the vector $\bm{h}=\bm{H}/H$ in Eqs. (\[rho001\_110\_tetragonal\]). The resulting curves are depicted by the dashed lines in Fig. \[mt\_001\]. For $\bm{H}$ rotated within the (001) plane \[Fig. \[mt\_001\](a)\], the linear and quadratic $(\bm{n}\cdot \bm{m})$ terms in Eqs. (\[rho001\_110\_tetragonal\]) vanish and we obtain the well-known $\cos^2\phi_j$ and $\cos\phi_j\sin\phi_j$ dependences of $\rho_{\mathrm{long}}$ and $\rho_{\mathrm{trans}}$, respectively. At 0.7 T the dashed curves coincide with the solid curves, meaning that for $\bm{H}$ rotated within the layer plane the magnetization almost perfectly follows the motion of the magnetic field. For lower fields, $\bm{M}$ remains in the layer plane since \[001\] is a hard axis, but it increasingly deviates from $\bm{H}$ towards the easy \[100\] and \[010\] axes. At 0.1 T it abruptly switches whenever $\bm{H}$ approaches the somewhat harder \[110\] and \[$\bar{1}$10\] axes, leading to the kinks observed for $\rho_{\mathrm{long}}$.
The rotation of $\bm{H}$ within a plane perpendicular to the layer is accompanied by significant differences in the orientations of $\bm{H}$ and $\bm{M}$, even for 0.7 T. This is clearly demonstrated in Figs. \[mt\_001\](b) and (c), where the dashed curves represent the $(\bm{n}\cdot \bm{m})^2 = \cos^2\phi_n$ and $(\bm{n}\cdot \bm{m})=\cos\phi_n$ dependences of $\rho_{\mathrm{long}}$ and $\rho_{\mathrm{trans}}$ in Eqs. (\[rho001\_110\_tetragonal\]), respectively, with $\phi_n$ denoting the angle between $\bm{m}$ and $\bm{n}$. At 0.7 T, $\bm{H}$ and $\bm{M}$ coincide whenever $\bm{H}$ is orientated parallel or perpendicular to the layer plane. At lower fields this is no longer true and $\bm{M}$ avoids the perpendicular direction by tending towards the easy \[100\] and \[010\] axes (see Fig. \[mag\_110\]). Accordingly, the differences between the minimum and maximum values of $\rho_{\mathrm{long}}$ and $\rho_{\mathrm{trans}}$ are drastically reduced at 0.25 and 0.1 T.
It should be emphasized that the magnetotransport measurements give clear evidence for the tetragonal distortion of the (Ga,Mn)As layer: First, the parameters $B_{c\parallel}$ and $B_{c\perp}$, representing the in-plane and out-of-plane contributions to the cubic-anisotropy term in $G^{001}_{M}$, significantly differ. Second, Eqs. (\[rho001\_110\_cubic\]), which have been derived for the case of perfect cubic symmetry, correctly reproduce the measured amplitudes of $\rho_{\mathrm{long}}$ for the in-plane configuration in Fig. \[mt\_001\](a), but do not so for the two out-of-plane configurations in Figs. \[mt\_001\](b) and (c). Moreover, it is worth noting that the fits presented in Fig. \[mt\_001\] could be improved even further by taking into account higher-order terms in the series expansion of the resistivity tensor in Eq. (\[series\_expansion\]). However, since the agreement achieved in second order is more than satisfactory and since the mathematical expressions for the resistivities would become much more complicated, higher-order terms have not been considered in the present study.
The results of FMR measurements are presented in Fig. \[fmr\_001\]. It shows the measured and simulated angular dependences of the resonance field $H_{\mathrm{res}}$ for $\bm{H}$ rotated within the (001), ($\bar{1}$10), and (110) planes. The dashed lines, reproducing only roughly the experimental curves, were numerically calculated using Eq. (\[FE\_001\]), Eq. (\[resonance\_condition\]), $g$ = 2.0, and the anisotropy parameters derived from the magnetotransport measurements. The agreement between experiment and theory is significantly improved using $g$ = 1.9 and the slightly higher values $B_{c\parallel}$ = -0.02 T, $B_{c\perp}$ = 0 T, $B_{001}$ = 0.24 T, and $B_{\bar{1}10}$ = 0.002 T, which were obtained by a least squares fit based on Eqs. (\[FE\_001\]) and (\[resonance\_condition\]). The values agree within 30% with the anisotropy parameters determined from magnetotransport. The reason for the remaining difference between the two sets of parameters is not yet clear. Inevitable sample heating up to 150$^\circ$C for less than 30 min during the Hall-bar preparation as well as the different lateral sizes of our samples (shape anisotropy) are not expected to account for the observed variation.
![\[fmr\_001\] Angle-dependent FMR fields of the (001) (Ga,Mn)As sample at 5 K for $\bm{H}$ rotated in the (001), ($\bar{1}$10), and (110) planes. The solid lines represent the result of a least squares fit, the dashed lines were calculated using $g$ = 2.0 and the anisotropy parameters estimated from the magnetotransport data.](fig7.eps)
![\[fmr\_113\] Angle-dependent FMR fields of the (113)A (Ga,Mn)As sample at 5 K for $\bm{H}$ rotated in the (33$\bar{2}$), (1$\bar{1}$0), and (113)A planes. The solid and dashed lines represent the results of least squares fits with and without considering a uniaxial term along \[001\] in $G_{M}^{113}$, respectively.](fig8.eps)
(113)A orientation
------------------
As already mentioned in Section III B, our experimental data suggest the existence of a lattice distortion along \[001\] even in the (Ga,Mn)As film grown on GaAs(113)A substrate. This is demonstrated in Fig. \[fmr\_113\], which shows the measured and simulated angular dependences of the FMR field $H_{\mathrm{res}}$ for $\bm{H}$ rotated within the (33$\bar{2}$), (1$\bar{1}0)$, and (113) planes. The solid and dashed lines depict the results of least squares fits using Eqs. (\[FE\_113\]) and (\[resonance\_condition\]) with and without considering the uniaxial term $B_{001} m_z^2$, respectively. Even though no perfect simulation of the measured curves could be achieved, the solid curve is much closer to the experimental data than the dashed one. Optimal agreement is obtained for $g=2.0$ and the values $B_{c\parallel}$ = $B_{c\perp}$ = -0.046 T, $B_{113}$ = 0.032 T, and $B_{001}$ = 0.053 T of the anisotropy parameters (solid line in Fig. \[fmr\_113\]). Similar to the case of the (001) layer, the uniaxial in-plane contribution is almost negligible with $B_{\bar{1}10}$ = -0.005 T.
The magnetotransport data, measured with the current direction along \[33$\bar{2}$\], are depicted in Fig. \[mt\_113\]. The figure shows the experimental and simulated angular dependences of $\rho_{\mathrm{long}}$ and $\rho_{\mathrm{trans}}$ for $\bm{H}$ rotated within the (113), (33$\bar{2}$), and ($\bar{1}$10) planes. Using the fit procedure described in Section IV A, the values of the resistivity parameters in Eqs. (\[rho113\_33m2\_tetragonal\]) are obtained as $B$ = $-8.1\times 10^{-4}$ $\Omega$ cm, $C$ = $-8.7\times 10^{-4}$ $\Omega$ cm, $b$ = $7.2\times 10^{-3}$ $\Omega$ cm, $c$ = $5.1\times 10^{-4}$ $\Omega$ cm, $D$ = $-2.1\times 10^{-3}$ $\Omega$ cm, and $d$ = $2.1\times 10^{-4}$ $\Omega$ cm. The sum $A+2a/11$ decreases from $21.61\times 10^{-3}$ $\Omega$ cm at 0.1 T to $20.79\times 10^{-3}$ $\Omega$ cm at 0.7 T. For the anisotropy parameters we obtained the values $B_{c\parallel}$ = -0.046 T, $B_{c\perp}$ = -0.03 T, $B_{113}$ = 0.018 T, $B_{001}$ = 0.02 T, and $B_{\bar{1}10}$ = -0.008 T. Similar to the (001) sample, they are smaller than the values determined by the FMR study, however, the discrepancy is less than a factor of about two. The calculated curves, represented by the solid lines in Fig. \[mt\_113\], are in excellent agreement with the measured data. Again, the dashed curves simulate the case of a magnetization which perfectly aligns with $\bm{H}$. A comparison between the dashed and solid curves reveals that the magnetization at 0.7 T almost perfectly follows the motion of the magnetic field in the two out-of-plane configurations \[Figs. \[mt\_113\](b) and (c)\], whereas for $\bm{H}$ rotated within the layer plane \[Fig. \[mt\_113\](a)\] $\bm{M}$ significantly deviates from $\bm{H}$. According to the model calculations, the latter behavior arises from the cubic terms and the uniaxial \[001\] contribution in $G^{113}_{M}$, resulting in a deflection of $\bm{M}$ towards the (001) plane. The asymmetry of $\rho_{\mathrm{trans}}$ in Fig. \[mt\_113\](a) partly results from this deflection and partly from the last term in Eqs. (\[rho113\_33m2\_tetragonal\]) which originates from the tetragonal distortion. A similar asymmetry has been observed by Muduli et al.[@Mud05] in Fe$_3$Si films grown on GaAs(113)A substrates. There, the asymmetry has been explained by third-order terms in $\rho_{\mathrm{trans}}$. We cannot rule out that in the (113)A (Ga,Mn)As sample under study higher-order terms contribute to the asymmetry, too. However, since the present model, including terms up to second order, fully accounts for the observed angular dependence of $\rho_{\mathrm{long}}$ and $\rho_{\mathrm{trans}}$, it has not been considered as mandatory to include them.
Using the anisotropy parameters obtained from the curve fits, the orientations of the easy axes were determined by minimizing $G^{113}_{M}$ with respect to $\bm{m}$ at zero magnetic field. We find the easy axes in the (113)A layer at the angles (see Fig. \[coordinates\]) $\theta_1$ = 92$^{\circ}$, $\varphi_1$ = -3.4$^{\circ}$, and $\theta_2$ = 92$^{\circ}$, $\varphi_2$ = 93.4$^{\circ}$, i.e., very close to the \[100\] and \[010\] axes, in qualitative agreement with the results presented in Refs. and .
Summary
=======
A series expansion of the resistivity tensor with respect to the magnetization components yields general expressions for the longitudinal and transverse resistivites in single-crystalline ferromagnets with cubic and tetragonal symmetry. The expressions, applicable to (Ga,Mn)As layers with arbitrary surface index, were used to quantitatively model the angular dependences of the resistivities, measured in (001) and (113)A (Ga,Mn)As films as a function of magnetic field orientation. Whereas the curves at 0.7 T largely reflect the anisotropy of the resistivity tensor, the curves at 0.25 and 0.1 T are strongly affected by magnetic anisotropy, allowing access to anisotropy parameters. The magnetotransport data and comparative ferromagnetic resonance studies reveal an inclined uniaxial anisotropy along \[001\] in the (113)A-oriented (Ga,Mn)As layers in addition to the usual in- and out-of-plane contributions known from (001) layers.
This work was supported by the Deutsche Forschungsgemeinschaft (Li 988/4 and SFB 631).
[99]{} H. Ohno, Science [**281**]{}, 951 (1998). T. Jungwirth, K. Y. Wang, J. Mašek, K. W. Edmonds, J. K[ö]{}nig, J. Sinova, M. Polini, N. A. Goncharuk, A. H. MacDonald, M. Sawicki, A. W. Rushforth, R. P. Campion, L. X. Zhao, C. T. Foxon, and B. L. Gallagher, Phys. Rev. B [**72**]{}, 165204 (2005), and references therein. T. Dietl, H. Ohno, and F. Matsukura, Phys. Rev. B [**63**]{}, 195205 (2001). T. Jungwirth, J. Sinova, K. Y. Wang, K. W. Edmonds, R. P. Campion, B. L. Gallagher, C. T. Foxon, Q. Niu, and A. H. MacDonald, Appl. Phys. Lett. [**83**]{}, 320 (2003). T. Jungwirth, M. Abolfath, J. Sinova, J. Kučera, and A. H. MacDonald, Appl. Phys. Lett. [**81**]{}, 4029 (2002). D. V. Baxter, D. Ruzmetov, J. Scherschligt, Y. Sasaki, X. Liu, J. K. Furdyna, and C. H. Mielke, Phys. Rev. B [**65**]{}, 212407 (2001). F. Matsukura, M. Sawicki, T. Dietl, D. Chiba, and H. Ohno, Physica E [**21**]{}, 1032 (2004). K. Y. Wang, K. W. Edmonds, R. P. Campion, L. X. Zhao, C. T. Foxon, and B. L. Gallagher, Phys. Rev. B [**72**]{}, 085201 (2005). S. T. B. Goennenwein, S. Russo, A. F. Morpurgo, T. M. Klapwijk, W. Van Roy, and J. De Boeck, Phys. Rev. B [**71**]{}, 193306 (2005). H. X. Tang, R. K. Kawakami, D. D. Awschalom, and M. L. Roukes, Phys. Rev. Lett. [**90**]{}, 107201 (2003). U. Welp, V. K. Vlasko-Vlasov, X. Liu, J. K. Furdyna, and T. Wojtowicz, Phys. Rev. Lett. [**90**]{}, 167206 (2003). X. Liu, Y. Sasaki, and J. K. Furdyna, Phys. Rev. B [**67**]{}, 205204 (2005). X. Liu, W. L. Lim, L. V. Titova, M. Dobrowolska, J. K. Furdyna, M. Kutrowski, and T. Wojtowicz, J. Appl. Phys. [**98**]{}, 063904 (2005). M. Sawicki, K.-Y. Wang, K. W. Edmonds, R. P. Campion, C. R. Staddon, N. R. S. Farley, C. T. Foxon, E. Papis, E. Kamińska, A. Piotrowska, T. Dietl, and B. L. Gallagher, Phys. Rev. B [**71**]{}, 121302(R) (2005). K. Y. Wang, M. Sawicki, K. W. Edmonds, R. P. Campion, S. Maat, C. T. Foxon, B. L. Gallagher, and T. Dietl, Phys. Rev. Lett. [**95**]{}, 217204 (2005). C. Bihler, H. Huebl, M. S. Brandt, S. T. B. Goennenwein, M. Reinwald, U. Wurstbauer, M. D[ö]{}ppe, D. Weiss, and W. Wegscheider, Appl. Phys. Lett. [**89**]{}, 012507 (2006). W. Limmer, M. Glunk, J. Daeubler, T. Hummel, W. Schoch, R. Sauer, C. Bihler, H. Huebl, M. S. Brandt, and S. T. B. Goennenwein, accepted for publication in Microelectron. J. J. Daeubler, M. Glunk, W. Schoch, W. Limmer, and R. Sauer, Appl. Phys. Lett. [**88**]{}, 051904 (2006). T. Omiya, F. Matsukura, A. Shen, Y. Ohno, and H. Ohno, Physica E [**10**]{}, 206 (2001). M. Reinwald, U. Wurstbauer, M. D[ö]{}ppe, W. Kipferl, K. Wagenhuber, H.-P. Tranitz, D. Weiss, and W. Wegscheider, J. Cryst. Growth [**278**]{}, 690 (2005). K. Y. Wang, K. W. Edmonds, L. X. Zhao, M. Sawicki, R. P. Campion, B. L. Gallagher, and C. T. Foxon, Phys. Rev. B [**72**]{}, 115207 (2005). E. C. Stoner and E. P. Wohlfarth, Philos. Trans. R. Soc. London, Ser. A [**240**]{}, 74 (1948). R. R. Birss, [*Symmetry and Magnetism*]{}, North-Holland, Amsterdam, 1966. P. K. Muduli, K.-J. Friedland, J. Herfort, H.-P. Sch[ö]{}nherr, and K. H. Ploog, Phys. Rev. B [**72**]{}, 104430 (2005). R. R. Birss, Proc. Phys. Soc. [**75**]{}, 8 (1960). T. McGuire and R. Potter, IEEE Trans. Magn. [**11**]{}, 1018 (1975). J. P. Jan, in [*Solid State Physics*]{}, edited by F. Seitz and D. Turnbull (Academic Press, New York, 1957), Vol. 5, pp. 1-96. In most publications the energy of the magnetic system is referred to as free energy. Since the independent variable is represented by the magnetic field $\bm{H}$ and not by the magnetization $\bm{M}$, this energy should be correctly denoted as free enthalpy. U. Welp, V. K. Vlasko-Vlasov, A. Menzel, H. D. You, X. Liu, J. K. Furdyna, and T. Wojtowicz, Appl. Phys. Lett. [**85**]{}, 260 (2004). K. Hamaya, T. Watanabe, T. Taniyama, A. Oiwa, Y. Kitamoto, and Y. Yamazaki, Phys. Rev. B [**74**]{}, 045201 (2006). X. Liu and J. K. Furdyna, J. Phys.: Condens. Matter [**18**]{}, R245 (2006). S. T. B. Goennenwein, T. Graf, T. Wassner, M. S. Brandt, M. Stutzmann, J. B. Philipp, R. Gross, M. Krieger, K. Z[ü]{}rn, P. Ziemann, A. Koeder, S. Frank, W. Schoch, and A. Waag, Appl. Phys. Lett. [**82**]{}, 730 (2003). M. Farle, Rep. Prog. Phys. [**61**]{}, 755 (1998).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A brief review is given of all the Hamiltonians and effective potentials calculated hitherto covering the post-Newtonian (pN) dynamics of a two body system. A method is presented to compare (conservative) reduced Hamiltonians with nonreduced potentials directly at least up to the next-to-leading-pN order.'
author:
- Steven Hergt
- Jan Steinhoff
- Gerhard Schäfer
title: 'On the comparison of results regarding the post-Newtonian approximate treatment of the dynamics of extended spinning compact binaries'
---
Post-Newtonian modeling and results of the two body dynamics
============================================================
The post-Newtonian (pN) treatment of the dynamics of two bodies in general relativity has to incorporate both spin and tidal force induced mass multipoles of the constituents of the physical systems [@Binnington:Poisson:2009]. In our analysis we will focus on the spin multipole degrees of freedom and present some illuminating results, that will be useful for future research of gravitational wave date extraction from such a system. One easy way to start to model such a system is by choosing the representation by Tulczyjew’s singular stress-energy tensor[@Tulczyjew:1959] $T_{\mu\nu}$, with the greek coordinate indices running from 0 to 3, in the following way $$\begin{aligned}
\sqrt{-g} T^{\mu\nu}(x^{\sigma})& =& \int d \tau \bigg[
u^{(\mu} p^{\nu)} \delta_{(4)}+\left(u^{(\mu}S^{\nu)\alpha}\delta_{(4)}\right)_{||\alpha}+\frac{1}{3}R_{\alpha\beta\rho}^{(\mu}J^{\nu)\rho\beta\alpha}\delta_{(4)}\nonumber\\
&&\quad-\frac{2}{3}\left(J^{\mu\alpha\beta\nu}\delta_{(4)}\right)_{||(\alpha\beta)}+\dots\bigg]
\,,\quad u^{\mu} = \frac{d z^{\mu}}{d \tau}\,, \quad \delta_{(4)} = \delta(z^{\sigma} - x^{\sigma})\end{aligned}$$ with the body’s 4-velocity $u^{\mu}$, the 4-momentum $p_{\mu}$, the antisymmetric spin tensor $S^{\mu\nu}$ modeling the pole-dipole structure and Dixon’s reduced quadrupole moment tensor $J^{\mu\alpha\beta\nu}$ modeling first order finite size effects while possessing the same symmetries as the Riemann tensor $R^{\mu\alpha\beta\nu}$. It follows a decomposition of $J^{\mu\alpha\beta\nu}$ into stress, flow and the symmetric trace-free mass quadrupole $Q^{}_{\mu\nu}$. The latter is given by the ansatz with a vector $f_{\mu}$ to which the spin is orthogonal, $S^{\mu\nu}f_{\nu}=0$ $$Q_{\mu\nu}=\frac{C_{Q}}{m}\left(S_{\mu\rho}S_{\nu}^{\;\;\rho}-\frac{1}{3}P_{\mu\nu}S^{\rho\sigma}S_{\rho\sigma}\right)\,,\quad P^{\mu\nu}=g^{\mu\nu}-\frac{1}{f_{\rho}f^{\rho}}f^{\mu}f^{\nu}\,.$$ and is parametrized only by $C_{Q}$ in the Newtonian limit and quadratic level in spin fully encoding the rotational deformation. For black holes one has $C_{Q}=1$ [@Thorne:1980] while for neutron star models $C_{Q}$ depends on the equations of state [@Laarakkers:Poisson:1999] and varies between $4.3\dots7.4$. The next step to perform explicit pN calculation is complex in various ways. We compare two prominent methods. One method aims at calculating a Hamiltonian. This is achieved by a 3+1 decomposition of Einstein’s field equations and the energy-momentum tensor from Eq. (1) leading to constraints which have to be fulfilled at all times on the 3-dimensional hypersurfaces orthogonal to the time direction. We then use the ADM formalism as outlined in [@Steinhoff:2011] to find the canonical set of variables ($\hat{\vct{z}}_{I}, \hat{\vct{p}}_{I},\hat{\vct{S}}_{I}$) with the body label $I=1,2$ fulfilling their standard canonical Poisson bracket relations $\{\hat{z}^i_{I},\hat{p}_{Jj}\}=\delta_{ij}\delta_{IJ}$ and $\{\hat{S}_{I(i)},\hat{S}_{I(j)}\}=\epsilon_{ijk}\hat{S}_{I(k)}$ with $i,j,k$ running from 1 to 3 and the round brackets around them indicate the components of local Lorentz indices $a,b,\dots$ from the beginning of the alphabet, so $a\in\{(0),(i)\}$. The spin tensor $S_{ab}$ defined in a local Lorentz frame is therefore connected to the coordinate frame by a vierbein transformation $S_{ab}=e_{a\mu}e_{b\nu}S^{\mu\nu}$. The ADM formalism also leads to a formula for calculating the Hamiltonian in full reduced phase space by imposing the ADM$TT$ or transverse traceless gauge to the 3-metric on the 3-hypersurface and by choosing the correct (canonical) spin supplementary condition (SSC) which fixes the center of the object. By expansion of the constraints in pN powers of $\frac{v^2}{c^2}\sim \frac{Gm}{rc^2}$ one ends up with a perturbative scheme to calculate Hamiltonians to formally arbitrary pN orders. The general Hamiltonian being the generator for the equations of motion of the binary therefore intrinsically adopts the post-Newtonians expansion of the field equations. As the spin is of pN order $1/c$ or $1/c^2$ depending on its strenght the formal labeling is such that we call the first post-Newtonian spin Hamiltonians not 1.5pN or 2.5pN according to formal counting rules but just the leading order (LO) ones and the higher corrections we call next-to-leading (NLO) and next-to-next-to-leading order (N$^2$LO). In table \[tab1\] we give a list of all known pN Hamiltonians for the case of maximally rotating objects where $|\vct{S}|=\frac{Gm^2\chi}{c}$ with the dimensionless spin $\chi=1$.
[r@r\*[15]{}[@[$\;$]{}c]{}]{} order && 1.0 && 1.5 && 2.0 && 2.5 && 3.0 && 3.5 && 4.0 && 4.5 & $H^{\text{N}}$ PM & $+$ & $H^{\text{1PN}}$ &&& $+$ & $H^{\text{2PN}}$ & $+$ & $H^{\text{2.5PN}}$ & $+$ & $H^{\text{3PN}}$ & $+$ & $H^{\text{3.5PN}}$ & $+$ & $(H^{\text{4PN}})$ & $+$ & $\{H^{\text{4.5PN}}\}$ SO &&& $+$ & $H^{\text{LO}}_{\text{SO}}$ && & $+$ & $H^{\text{NLO}}_{\text{SO}}$ && & $+$ & $H^{\text{N$^2$LO}}_{\text{SO}}$ & $+$ & $H^{\text{LO,R}}_{\text{SO}}$ & $+$ & $(H^{\text{N$^3$LO}}_{\text{SO}})$ S$_1^2$ &&&&& $+$ & $H^{\text{LO}}_{\text{S$_1^2$}}$ && & $+$ & $H^{\text{NLO}}_{\text{S$_1^2$}}$ && & $+$ & $(H^{\text{N$^2$LO}}_{\text{S$_1^2$}})$ & $+$ & $\{H^{\text{LO,R}}_{\text{S$_1^2$}}\}$ S$_1$S$_2$ &&&&& $+$ & $H^{\text{LO}}_{\text{S$_1$S$_2$}}$ && & $+$ & $H^{\text{NLO}}_{\text{S$_1$S$_2$}}$ && & $+$ & $H^{\text{N$^2$LO}}_{\text{S$_1$S$_2$}}$ & $+$ & $H^{\text{LO,R}}_{\text{S$_1$S$_2$}}$ spin$^3$ &&&&&&&&&&& $+$ & $[H_{S^3}^{\text{LO}}]$ & && $+$ & $(H_{S^3}^{\text{NLO}})$ spin$^4$ &&&&&&&&&&&&& $+$ & $[H_{S^4}^{\text{LO}}]$ &&\
$\vdots$ &&&&&&&&&&&&&&&& $\ddots$\
$\{.\}$ EOM known$[.]$ for Black Holes only $(.)$ not known (yet)
$H^{N}$ is the Newtonian Hamiltonian, PM means point mass, $H^{nPN}$ with $n\in\{1,2,\dots\}$ are the conservative pure point mass Hamiltonians, $H^{\frac{n}{2}PN}$ are the dissipative (radiative) pure point mass Hamiltonians, SO refers to spin-orbit coupling, $S_{1}S_{2}$ to spin(1)-spin(2) coupling and $S_{1}^2$ to spin quadrupole coupling involving the constant $C_{Q}$. LO,R in the index indicates the dissipative counterpart to leading order, so $H_{S_{1}^2}^{LO,R}$ is the radiative counterpart to the conservative part $H_{S_{1}^2}^{LO}$. Obviously, the radiative part is much higher in pN order then the conservative part, but nevertheless they are important to cover the dynamics to 4.5pN order consistently; up until now the radiation field is known to 2.5pN order only. One other method to arrive at pN equations of motion is the derivation of effective potentials which are subtly related to Hamiltonians by a Legendre transformation. This derivation is most effectivly achieved by sophisticated methods from Effective Field Theory (EFT) that uses full knowledge from quantum field theoretical calculations. Up until now, pN potentials have been calculated to 3pN order [@Foffa:Sturani:2011] for point masses and to NNLO for spin(1)-spin(2) coupling [@Levi:2011].
Comparison between Effective Field Theory potentials and ADM Hamiltonians
=========================================================================
Effective potentials are part of a Lagrangian with the Newtonian kinetic energy $T_{N}$ $$L_{eff}=T_{N}-V_{eff}=\frac{m_{1}}{2}v_{1}^2+\frac{m_{2}}{2}v_{2}^2-V_{eff}.$$ The conservative effective potential $V_{eff}$ for two interacting bodies is pN expanded up to next-to-leading order (NLO) spin effects in the following way $$V_{eff} = V_{PM}
+ V^{LO}_{SO} + V^{LO}_{S_1^2}
+ V^{LO}_{S_2^2} + V^{LO}_{S_1S_2}
+ V^{NLO}_{SO} + V^{NLO}_{S_1^2}
+ V^{NLO}_{S_2^2}
+ V^{NLO}_{S_1S_2}\,.$$ One key difference between EFT potentials and ADM Hamiltonians is that in most cases the potentials still depend on the $S^{(0)(i)}$-components of the spin tensor, which have to be fixed by choosing an appropriate SSC. For a direct comparison a formal Legendre transformation of the nonreduced potentials is conducted yielding the effective Hamiltonian $H_{eff}$, which is been followed by a reduction process in phase space in order to arrive at a canonical set of variables, see [@Hergt:Steinhoff:Schafer:2011] for details. This ‘canonicalization’ is most transparently accomplished by reducing the following effective action $$S_{eff}=\int\dd t\, L_{eff}=\int\dd t\left(p_{1i}\dot z^{i}_{1}+p_{2i}\dot z^{i}_{2}-\frac{1}{2}S_{1ab}\Omega_{1}^{ab}-\frac{1}{2}S_{2ab}\Omega_{2}^{ab}
-H_{eff}\left(\vct{z}_{I},\vct{p}_{I},S_{Iab}\right)\right)\,.$$ Here we have defined the angular velocity tensor $\Omega^{ab}\equiv\Lambda_{A}^{\;\;a}\dot{\Lambda}^{Ab}$ rendering $\Omega^{ab}$ antisymmetric and $\Lambda_{A\mu}\Lambda^{A}_{\;\;\nu}=g_{\mu\nu}$, $\Lambda_{Aa}\Lambda^{A}_{\;\;b}=\eta_{ab}$ with $(A,B,\dots)\in\{[0],[i]\}$ being the body-fixed frame labels. The reduced action has to read $$\hat{S}_{eff}=\int\dd t\left(\hat{p}_{1i}\dot{\hat{z}}^{i}_{1}+\hat{p}_{2i}\dot {\hat{z}}^{i}_{2}-\frac{1}{2}\hat{S}_{1(i)(j)}
\hat{\Omega}_{1}^{(i)(j)}-\frac{1}{2}\hat{S}_{2(i)(j)}\hat{\Omega}_{2}^{(i)(j)}-H_{can}\left(\hat{\vct{z}}_{I},\hat{\vct{p}}_{I},\hat{\vct{S}}_{I}\right)\right)$$ with $\hat{\Omega}^{(i)(j)}=\hat{\Lambda}_{[k]}^{\;\;\;(i)}\dot{\hat{\Lambda}}^{[k](j)}$ given by a nonlinear shift of $\Lambda^{[k](i)}$ to $\hat{\Lambda}^{[k](i)}$ so that $\hat{\Lambda}^{[k](i)}\hat{\Lambda}^{[k](j)}=\delta_{ij}$. This reduction is achieved by inserting the covariant SSC $S_{ab}u^b=0$ as well as its conjugate condition $\Lambda^{[i]a}u_{a}=0$ into $\frac{1}{2}S_{ab}\Omega^{ab}$ and performing a pN approximate variable transformation of spin and position reading $$\begin{aligned}
z_{1}^{i}&=&~\hat{z}_{1}^{i}-\bigg[\frac{1}{2m_{1}^2}p_{1k}\hat{S}_{1(i)(k)}\left(1-\frac{\vct{p}_{1}^2}{4m_{1}^2}\right)
-G\frac{m_{2}}{m_{1}^2}\frac{p_{1k}\hat{S}_{1(i)(k)}}{\hat{r}_{12}}+\frac{3}{2}G\frac{p_{2k}\hat{S}_{1(i)(k)}}{m_{1}\hat{r}_{12}}\nonumber\\
&&+\frac{G}{2}\frac{\hat{n}_{12}^{k}(\vct{\hat{n}}_{12}\cdot\vct{p}_{2})\hat{S}_{1(i)(k)}}{m_{1}\hat{r}_{12}}+G\frac{m_{2}}{m_{1}^2}\frac{\hat{S}_{1(k)(l)}\hat{S}_{1(i)(l)}\hat{n}_{12}^{k}}{\hat{r}_{12}^2}+G\frac{\hat{n}_{12}^{k}
\hat{S}_{1(i)(l)}\hat{S}_{2(k)(l)}}{m_{1}\hat{r}_{12}^2}\bigg]\,,\end{aligned}$$ $$\begin{aligned}
S_{1(i)(j)}&=&~\hat{S}_{1(i)(j)}-\bigg[\frac{p_{1[i}\hat{S}_{1(j)](k)}p_{1k}}{m_{1}^2}\left(1-\frac{\vct{p}_{1}^2}{4m_{1}^2}\right)-\frac{2Gm_{2}}{m_{1}^2\hat{r}_{12}}p_{1[i}\hat{S}_{1(j)](k)}p_{1k}\nonumber\\
&&+\frac{3G}{m_{1}\hat{r}_{12}}p_{1[i}\hat{S}_{1(j)](k)}p_{2k}+\frac{G}{m_{1}\hat{r}_{12}}p_{1[i}\hat{S}_{1(j)](k)}\hat{n}_{12}^{k}(\vct{\hat{n}}_{12}\cdot\vct{p}_{2})\\
&&+\frac{2Gm_{2}}{m_{1}^2\hat{r}_{12}^2}p_{1[i}\hat{S}_{1(j)](l)}\hat{S}_{1(k)(l)}\hat{n}_{12}^{k}+\frac{2G}{m_{1}
\hat{r}_{12}^2}p_{1[i}\hat{S}_{1(j)](l)}\hat{S}_{2(k)(l)}\hat{n}_{12}^{k}\bigg]\,.\nonumber\end{aligned}$$ Those formulas are valid to transform the potentials at least to NLO to their canonical Hamiltonian counterpart, which enabled us to obtain an overall agreement of all EFT NLO potentials with their corresponding ADM Hamiltonian as displayed in table \[tab2\] up to canonical transformations indicated by $\approx$, see [@Hergt:Steinhoff:Schafer:2011] for a thorough investigation.
------------------------- ----------------------------------------------------------------------- ----------- ----------------------------- ---------------------------------------------------------------------------------------------
$ V_{NLO}^{SO}$ Levi[@Levi:2010]
$H_{NLO ADM}^{SO}$ Damour/Jaranowski/Schäfer[@Damour:Jaranowski:Schafer:2008:1; @Steinhoff:Schafer:Hergt:2008]
$ V_{NLO}^{SO}$ Porto[@Porto:2010]
$\approx$
$ V_{NLO}^{S_{1}S_{2}}$ Porto/Rothstein[@Porto:Rothstein:2008:1; @Porto:Rothstein:2008:1:err] $ H_{NLO ADM}^{S_{1}S_{2}}$ Steinhoff/Hergt/Schäfer[@Steinhoff:Hergt:Schafer:2008:2; @Steinhoff:Schafer:Hergt:2008]
$ V_{NLO}^{S_{1}^2}$ Porto/Rothstein[@Porto:Rothstein:2008:2; @Porto:Rothstein:2008:2:err] $ H_{NLO ADM}^{S_{1}^2}$ Hergt/Steinhoff/Schäfer [@Hergt:Steinhoff:Schafer:2010:1]
------------------------- ----------------------------------------------------------------------- ----------- ----------------------------- ---------------------------------------------------------------------------------------------
: \[tab2\] Agreement between EFT potentials and ADM Hamiltonians
Acknowledgments {#acknowledgments .unnumbered}
===============
This work is supported by the Deutsche Forschungsgemeinschaft (DFG) through SFB/TR7 “Gravitational Wave Astronomy,” project STE 2017/1-1, and GRK 1523, and by the FCT (Portugal) through PTDC project CTEAST/098034/2008.
References {#references .unnumbered}
==========
[9]{}
Binnington T and Poisson E 2009 [[*Phys. Rev. D*]{} [ **80**]{} (2009) 084018]{}
Tulczyjew W M 1959 [*Acta Phys. Pol.*]{} [**18**]{} 393–409.
Thorne K S 1980 [[*Rev. Mod. Phys.*]{} [ **52**]{} 299–339]{} Laarakkers W G and Poisson E 1999 [[*Astrophys. J.*]{} [ **512**]{} 282–7]{}
Steinhoff J Canonical formulation of spin in general relativity 2011 [[*Ann. Phys. (Berlin)*]{} [**523**]{} 296–353]{}
Hartung J and Steinhoff J 2011 [[*Ann. Phys. (Berlin)*]{} [**523**]{} 783–90]{} Hartung J and Steinhoff J 2011 [[*Ann. Phys. (Berlin)*]{} [**523**]{} 919–24]{}
Foffa S and Sturani R 2011 [[*Phys. Rev. D*]{} [ **84**]{} 044031]{}
Levi M 2011 Binary dynamics from spin1-spin2 coupling at fourth post-Newtonian order [*Preprint*]{} gr-qc/1107.4322
Hergt S, Steinhoff J and Sch[ä]{}fer G 2012 [[*Ann. Phys. (N.Y.)*]{} ]{} doi:10.1016/j.aop.2012.02.006
Levi M 2010 [[*Phys. Rev. D*]{} [ **82**]{} (2010) 104004]{} Damour T, Jaranowski P and Sch[ä]{}fer G 2008 [[*Phys. Rev. D*]{} [**77**]{} 064032]{} Steinhoff J, Sch[ä]{}fer G and Hergt S 2008 [[*Phys. Rev. D*]{} [ **77**]{} 104018]{} Porto R A 2010 [[*Class. Quant. Grav.*]{} [**27**]{} 205001]{} Porto R A and Rothstein I Z 2008 [[*Phys. Rev. D*]{} [**78**]{} 044012]{} Porto R A and Rothstein I Z 2010 [[*Phys. Rev. D*]{} [ **81**]{} 029904(E)]{}
Steinhoff J, Hergt S and Sch[ä]{}fer G 2008 [[*Phys. Rev. D*]{} [ **77**]{} 081501(R)]{} Porto R A and Rothstein I Z 2008 [[*Phys. Rev. D*]{} [ **78**]{} 044013]{} Porto R A and Rothstein I Z 2010 [[*Phys. Rev. D*]{} [ **81**]{} 029905(E)]{}
Hergt S, Steinhoff J and Sch[ä]{}fer G 2010 [[*Class. Quant. Grav.*]{} [**27**]{} 135007]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper extends the matrix based approach to the setting of multiple subdivision schemes studied in [@Sauer2012]. Multiple subdivision schemes, in contrast to stationary and non-stationary schemes, allow for level dependent subdivision weights and for level dependent choice of the dilation matrices. The latter property of multiple subdivision makes the standard definition of the transition matrices, crucial ingredient of the matrix approach in the stationary and non-stationary settings, inapplicable. We show how to avoid this obstacle and characterize the convergence of multiple subdivision schemes in terms of the joint spectral radius of certain square matrices derived from subdivision weights. We illustrate our results with several examples.'
address: 'Fakultät für Mathematik, Universität Wien, Oskar-Morgenstern-Platz 1, 1090 Wien (Austria)'
author:
- Maria Charina
- Thomas Mejstrik
title: 'Multiple multivariate subdivision schemes: matrix and operator approaches'
---
multiple subdivision schemes ,convergence ,joint spectral radius ,restricted spectral radius\
[*Classification (MSC):* 65D17, 15A60]{}
Introduction
============
The main contribution of this paper is the adaptation of the well known joint spectral radius approach (matrix based approach) to the setting of multiple subdivision. The connection between stationary (level independent) subdivision and joint spectral radius techniques was established in [@DL]. In [@CCGP16], the authors introduced the matrix approach into the setting of non-stationary subdivision (with level dependent weights). In both cases, the essential ingredient of the spectral radius technique are the so-called transition matrices whose entries depend on the subdivision weights and whose structure is inherited from the dilation matrix. The main challenge of adapting the matrix approach to the case of level dependent dilation matrices is in combining their properties into an appropriate structure of the corresponding transition matrices.
Subdivision schemes are recursive algorithms for generating meshes in ${{\mathbb R}}^d$, usually $d=2,3$. If the scheme is convergent, then the sequence of meshes converges to a smooth limit curve or surface. The vertices of these meshes are computed by means of weighted local averages of the vertices of the mesh from the previous level of the subdivision recursion. The topology of the regular mesh is characterized by the integer dilation matrix $M \in {{\mathbb Z}}^{s \times
s}$ all of whose eigenvalues are larger than $1$ in the absolute value. In the case of a *multiple subdivision scheme*, the subdivision weights of local averages and the dilation matrices may vary depending on the level of the subdivision recursion. Multiple subdivision schemes were introduced and studied in [@Sauer2012].
The theory of subdivision schemes has had an impact on several applied areas of mathematics and engineering and, in return, has been greatly influenced by applications. First subdivision schemes with level independent subdivision weights and dilation matrices appeared in the ‘60s and are related to the wavelet and frame theory whose applications are e.g. in signal and image processing and in progressive geometry processing targeting faster data transfer via internet. Recently, isogeometric analysis and biological imaging promoted subdivision schemes with level-dependent weights by exploiting their ability to generate and reproduce exponential polynomials. Multiple subdivision schemes are building blocks for processing of images with anisotropic directional features [@Cotronei2015; @KuSa] and for multigrid methods for solving anisotropic PDEs [@CharinaDRT2].
The most important properties of curves or surfaces generated by subdivision are their shape and smoothness. In the case of level independent or dependent subdivision weights, these properties are well understood. The subdivision shapes are characterized in terms of algebraic properties of subdivision symbols [@CDM; @CCRomani; @JePlo; @Jia; @JJ; @MoeSau]. The smoothness of subdivision is characterized either using the joint spectral radius [@CHM; @CCGP16; @CP2017; @CJR; @DL; @H1] or restricted spectral radius techniques [@CDM; @CCS04; @DynLevin02]. Recent advances [@GP1; @GP2] in the exact computation of the joint spectral radius of compact sets of square matrices provide efficient methods for checking both Hölder and Sobolev regularity of subdivision. The study of the properties of multiple subdivision is at its very beginning. The convergence analysis of multiple subdivision in terms of the restricted spectral radius is given in [@Sauer2012]. Our main result, Theorem \[th:main\], relates the convergence analysis of multiple subdivision and the joint spectral radius techniques and allows us to use the efficient methods from [@GP1].
The paper is organized as follows. In section \[sec:background\] we recall the basic facts about subdivision and multiple subdivision in particular. Section \[sec\_transition\_matrices\_supports\] is devoted to the construction of transition matrices with certain important invariance properties. These properties are crucial for our comparison, see section \[sec:JSRvRSR\], of the restricted and joint spectral radius techniques in the context of multiple subdivision. The applications of our theoretical results are given in section \[sec:examples\].
Background and notation {#sec:background}
=======================
The so-called matrix (or, the joint spectral radius) approach studies the spectral properties of finite or compact sets of square matrices derived from the subdivision masks, see e.g. [@CHM; @CCGP16; @CP2017; @CJR; @CH; @DL; @H1; @HJ].
\[def\_JSR\] The *joint spectral radius* of a finite set $\mathcal{A}$ of square matrices $A_r \in \mathcal{A}$ is defined by $$\label{equ_jsr}
\rho(\mathcal{A})=\lim_{n\rightarrow\infty}\max_{A_r \in\mathcal{A}} \big\| \prod_{r=1}^n A_r \;\big\|^{1/n}.$$
The limit in the Definition \[def\_JSR\] exists and is independent of the matrix norm [@Rota60 Proposition 1]. The joint spectral radius quantifies the joint expanding properties of the matrices in $\mathcal{A}$.
Properties of the dilation matrices {#subsec:background_dilation}
-----------------------------------
In the context of multiple subdivision the concept of the joint spectral radius is also used to describe the joint expanding properties of several dilation matrices. This is a generalization of the standard requirement on the single dilation matrix to be expanding.
\[def\_jointly\_expanding\] A finite set of invertible matrices $\{M_j\in{{\mathbb Z}}^{s\times s}\;:\; j=1,\ldots,J\}$ is *jointly expanding* if $$\rho(\{M_j^{-1}\;:\; j=1,\ldots,J\})<1.$$
Every dilation matrix $M_j$ has a possibly different digit set, which we define next.
\[def\_digit\_set\] Let $j \in \{1, \ldots, J\}$. A *digit set $D_j\subset{{\mathbb Z}}^s$* corresponding to a dilation matrix $M_j$ is a complete set of representatives of the quotient group ${{\mathbb Z}}^s/M_j{{\mathbb Z}}^s=\{\alpha+M_j{{\mathbb Z}}^s:\alpha\in{{\mathbb Z}}^s\}$, i.e. $D_j\simeq {{\mathbb Z}}^s/M_j{{\mathbb Z}}^s$. The elements of a digit set are called *digits*.
We settle for the standard choice $D_j= {{\mathbb Z}}^s \cap M_j [0, 1)^s$, $j=1,\ldots,J$, implying that $0 \in D_j$. This choice of the digit sets does not necessarily lead to a tiling, but rather to a covering of ${{\mathbb R}}^s$, see e.g. [@CHM section 2.2.2]. We would like to emphasize that our results in section \[sec:JSRvRSR\] do not depend on the tiling property of the attractors.
\[def\_attractor\] Let $\left\{ M_j\in{{\mathbb Z}}^{s\times s} \;:\; j=1,\ldots,J\right\}$ be jointly expanding with corresponding digit sets $D=\{D_j\subset{{\mathbb Z}}^s\;:\; j=1,\ldots,J\}$. We define the *attractors* (subsets of ${{\mathbb R}}^s$) associated to $\boldsymbol{j}=(j_\ell)_{\ell \in {{\mathbb N}}}$, $j_{\ell} \in \{1,\ldots,J\}$ $$\label{equ_attractor}
K_{D\!,\,\boldsymbol{j}}=\operatorname{clos} \left(
M_{j_1}^{-1}D_{\!j_1}+M_{j_1}^{-1}M_{j_{2}}^{-1}D_{\!j_2}+\cdots \right)
= \operatorname{clos} \left(\sum_{r=1}^\infty \Big(\prod_{\ell=1}^{r} M_{j_{\ell}}^{-1}\Big)\;D_{\!j_{r}} \right).$$
Note that the structure of the attractor $K_{D\!,\,\boldsymbol{j}}$ depends on the order of the indices in $\boldsymbol{j}$.
The following properties of the attractors are reminiscent of the stationary and non-stationary settings.
\[thm\_attractor\] Let $\left\{ M_j\in{{\mathbb Z}}^{s\times s} \;:\; j=1,\ldots,J\,\right\}$ be jointly expanding with corresponding digit sets $D=\{D_j\subset{{\mathbb Z}}^s\;:\; j=1,\ldots,J\,\}$ and $\boldsymbol{j}=(j_\ell)_{\ell\in{{\mathbb N}}}$, $j_\ell\in\{1,\ldots,J\}$. Then $K_{D\!,\,\boldsymbol{j}}$ is compact.
The boundedness of $K_{D\!,\,\boldsymbol{j}}$ follows, by [@Rota60 Proposition 1], due to the existence of a matrix norm $\|\cdot\|$ such that $\displaystyle C_1=\max_{j=1,\ldots,J}\|M_j^{-1}\|<1$, and the fact that the sets $D_j$, $j=1,\ldots,J$, are finite, i.e. bounded by $0<C_2< \infty$. Indeed, for every $x \in K_{D\!,\,\boldsymbol{j}}$, we have $$\|x\|= \| M_{j_1}^{-1} d_{j_1} + M_{j_1}^{-1} M_{j_2}^{-1} d_{j_2}+\cdots\|
\leq
C_2 \sum_{\ell=1}^\infty C_1^{\ell} = \frac{C_2 C_1}{1-C_1},
\quad
d_{j_\ell} \in D_{\!j_\ell}.
\qedhere$$
Multiple subdivision and its properties {#subsec:background_subdivision}
---------------------------------------
The definition of subdivision operators associated to finite sets of finitely supported masks and jointly expanding dilation matrices is done analogously to the stationary or the non-stationary case.
\[def\_subdivision\_operator\] Let $j \in \{1, \ldots,J\}$, $J \in {{\mathbb N}}$. For a *mask* $a_j\in\ell_0({{\mathbb Z}}^s)$ and a dilation matrix $M_j \in{{\mathbb Z}}^{s\times s}$, the *subdivision operator* $S_j:\ell({{\mathbb Z}}^s)\rightarrow \ell({{\mathbb Z}}^s)$ defined by the pair $(a_j,M_j)$ is given by $$\label{equ_subdivision_opertor}
S_j c(\alpha)=\sum_{\beta\in{{\mathbb Z}}^s} a_j(\alpha - M_j\beta)c(\beta),\quad\alpha\in{{\mathbb Z}}^s.$$
Without loss of generality, we assume that $0\in\operatorname{supp}(a_j)$, $j=1,\ldots,J$.
The concept of multiple subdivision schemes was introduced in [@Sauer2012].
\[def\_subdivision\_scheme\] Given $\{a_j \in \ell_0({{\mathbb Z}}^s)\;:\; j=1,\ldots,J\}$ and jointly expanding $\{M_j\in{{\mathbb Z}}^{s\times s}\;:\; j=1,\ldots,J\}$.
$(i)$
: We define the *finite set $\mathcal{S}$ of subdivision operators* $S_j$ by $$\label{def:setS}
\mathcal{S}=\left\{S_j=(a_j,M_j)\;:\; a_j\in \ell_0({{\mathbb Z}}^s),\; M_j\in {{\mathbb Z}}^{s\times s},\;j=1,\ldots,J\right\}.$$
$(ii)$
: A sequence $(S_{j_\ell})_{\ell \in {{\mathbb N}}}\in\mathcal{S}^{{\mathbb N}}$, $j_\ell\in \{1,\ldots,J\}$, is called a *(multiple) subdivision scheme*.
The concept of multiple subdivision generalizes stationary and non-stationary settings. Indeed, the set $\mathcal{S}^{{\mathbb N}}$ of all possible (multiple) subdivision schemes contains stationary subdivision schemes – the sequences $(S)_{\ell \in {{\mathbb N}}} \in \mathcal{S}^{{\mathbb N}}$ with $S \in \mathcal{S}$ defined by the pair $(a,M)$. The set $\mathcal{S}^{{\mathbb N}}$ also includes certain non-stationary subdivision schemes – the sequences $(S_{j_\ell})_{\ell \in {{\mathbb N}}} \in \mathcal{S}^{{\mathbb N}}$ with the subdivision operators $S_{j_\ell} \in \mathcal{S}$ defined by the pairs $(a_{j_\ell},M)$.
\[def\_ss\_conv\_to\_Wkp\_functions\] Let $\mathcal{S}$ be a finite set of subdivision operators.
$(i)$
: We say that a (multiple) subdivision scheme $(S_{j_\ell})_{\ell \in {{\mathbb N}}}\in\mathcal{S}^{{\mathbb N}}$ is *convergent* if for every sequence $c\in\ell_\infty({{\mathbb Z}}^s)$ there exists a *function* $g_{c,\boldsymbol{j}}\in C({{\mathbb R}}^s)$ (which is non-zero for at least one sequence $c$) such that $$\label{equ_ss_conv_to_C0_functions}
\lim_{n\rightarrow\infty} \Big\| g_{c,\boldsymbol{j}}(M_{j_1}^{-1}\cdots M_{j_n}^{-1}\cdot) - S_{j_n}\cdots S_{j_1}c\Big\| _{\ell_\infty} = 0, \quad \quad
\boldsymbol{j}=(j_\ell)_{\ell \in {{\mathbb N}}}.$$
$(ii)$
: We say that $\mathcal{S}^{{\mathbb N}}$ is convergent, if every subdivision scheme in $\mathcal{S}^{{\mathbb N}}$ is convergent.
For the limit function $g_{c,\boldsymbol{j}}$ in Definition \[def\_ss\_conv\_to\_Wkp\_functions\] we write $$\label{equ_function_as_limit_of_subdiv_op}
g_{c,\boldsymbol{j}} = \lim_{n\rightarrow \infty} S_{j_n}\cdots S_{j_2} S_{j_1} c, \quad
\boldsymbol{j}=(j_\ell)_{\ell \in {{\mathbb N}}}.$$
The necessary conditions, the sum rules of order one, for convergence of stationary subdivision schemes in $\mathcal{S}^{{\mathbb N}}$ are well known, see e.g. [@CDM; @DynLevin02; @JePlo; @JJ].
\[thm\_sum\_rule\_zero\] Let $\mathcal{S}^{{\mathbb N}}$ be convergent. Then every stationary subdivision scheme defined by the pair $(a_j,M_j)$, $j \in \{1, \ldots,J\}$ is convergent and its mask $a_j$ satisfies the sum rules of order one, $$\label{equ_sum_rule_zero}
\sum_{\beta\in{{\mathbb Z}}^s} a_j(M_j\beta+\alpha)=1, \quad\alpha\in{{\mathbb Z}}^s.$$
The result of Lemma \[thm\_sum\_rule\_zero\] gives rise to the following assumption.
**Assumption S:** [*We assume that the masks $a_j$, $j=1,\ldots,J$, in $\mathcal{S}$ satisfy sum rules of order one.*]{}
Furthermore, if $\mathcal{S}^{{\mathbb N}}$ is convergent, then every (multiple) subdivision scheme in $\mathcal{S}^{{\mathbb N}}$ possesses a sequence of basic limit functions. Similarly to the non-stationary setting, the concept of refinability is defined for the basic limit functions generated by the certain (multiple) subdivision schemes $(S_{j_\ell})_{\ell \ge r}\in\mathcal{S}^{{\mathbb N}}$, $r \in {{\mathbb N}}$, related by the shift in the ordering of the corresponding subdivision operators. To indicate this shift we introduce the following sequence $\boldsymbol{j}^{[r]}$.
Let $\boldsymbol{j}=(j_\ell)_{\ell\in{{\mathbb N}}}$, $j_\ell\in\{1,\ldots,J\}$. For $r\in{{\mathbb N}}$ we define a shifted sequence $$\boldsymbol{j}^{[r]}=(j_r,j_{r+1},j_{r+2},\cdots)=(j_{\ell+r-1})_{\ell\in{{\mathbb N}}}.$$
\[def\_basic\_limit\_function\] For a (multiple) convergent subdivision scheme $(S_{j_\ell})_{\ell \in {{\mathbb N}}}\in\mathcal{S}^{{\mathbb N}}$, we define the sequence of *basic limit functions* $$\label{equ_basic_limit_function}
\phi_{\boldsymbol{j}^{[r]}} =
\lim_{n \rightarrow \infty} S_{j_{r+n}}\cdots S_{j_{r+1}} S_{j_r} \delta
,\quad
\delta(\alpha)=\left\{\begin{array}{rr} 1, & \alpha=0 \\ 0, & \text{otherwise} \end{array}\right.
,\quad
r\in{{\mathbb N}}.$$ If the scheme $(S_{j_\ell})_{\ell \in {{\mathbb N}}}$ is stationary, i.e. $S_{j_\ell}=S$ for all $\ell \in {{\mathbb N}}$, then $\phi_{\boldsymbol{j}^{[r]}}=\phi$ for all $r \in {{\mathbb N}}$.
Note that $\phi_{\boldsymbol{j}^{[r]}}$, $r=2,3, \ldots$, by themselves are limits of certain subdivision schemes in $\mathcal{S}^{{\mathbb N}}$.
The proof of the mutual refinability of the functions $\phi_{\boldsymbol{j}^{[r]}}$, $r \in {{\mathbb N}}$, is analogous to stationary or non-stationary settings [@CDM Theorem 2.1].
\[thm\_blf\_properties\] Let $(S_{j_\ell})_{\ell \in {{\mathbb N}}}\in\mathcal{S}^{{\mathbb N}}$ be a convergent subdivision scheme. Then its basic limit functions $\phi_{\boldsymbol{j}^{[r]}}$, $r \in {{\mathbb N}}$, are mutually refinable, i.e. they satisfy the system of refinement equations$$\label{equ_refinement_equation_1}
\phi_{\boldsymbol{j}^{[r]}}(x)=
\sum_{\alpha\in{{\mathbb Z}}^s} a_{j_r}(\alpha)
\; \phi_{\boldsymbol{j}^{[r+1]}} (M_{j_r} x -\alpha)
, \quad x\in{{\mathbb R}}^s
,\quad r\in{{\mathbb N}}.$$
For a given $c\in\ell_\infty({{\mathbb Z}}^s)$, the limit function $g_{c,\boldsymbol{j}}$ in of the subdivision scheme $(S_{j_\ell})_{\ell \in {{\mathbb N}}}\in\mathcal{S}^{{\mathbb N}}$ can be written as a linear combination of the integer shifts of the corresponding basic limit function $\phi_{\boldsymbol{j}^{[1]}}$. Thus, the convergence analysis of $\mathcal{S}^{{\mathbb N}}$ is equivalent to the analysis of uniform continuity of the corresponding basic limit functions. In section \[sec\_transition\_matrices\_supports\], we show how to rewrite in an equivalent vector-valued form, where the summation in is replaced by a matrix vector multiplication. To do that we need to gain more insight about the structure of the supports of the basic limit functions. See e.g. [@CohenDyn] for details in the stationary and non-stationary settings.
The straightforward observation that the compact sets $$\label{def:KA}
K_{A,\,{\boldsymbol{j}^{[r]}}}=
\operatorname{clos} \left(\sum_{\ell=r}^\infty \Big(\prod_{i=1}^{\ell} M_{j_{i}}^{-1}\Big)\;\operatorname{supp}(a_{j_{\ell}}) \right), \quad A={\{\operatorname{supp}(a_j) \;:\; j=1, \ldots,J\}}, \quad r \in {{\mathbb N}},$$ determine the re-parametrization (see e.g. [@CCRomani (1.2)] in the non-stationary case) for the subdivision sequences that approximate the values of $(\phi_{\boldsymbol{j}^{[r]}})_{{r \in {{\mathbb N}}}}$ implies the following result.
\[thm\_supp\_refinement\_equation\] Let $(S_{j_\ell})_{\ell \in {{\mathbb N}}}\in\mathcal{S}^{{\mathbb N}}$ be a convergent subdivision scheme and $A={\{\operatorname{supp}(a_j) \;:\; j=1, \ldots,J\}}$. Then the supports of the corresponding basic limit functions $\phi_{\boldsymbol{j}^{[r]}}$, $r \in {{\mathbb N}}$, satisfy $$\label{equ_support_1}
\operatorname{supp}(\phi_{\boldsymbol{j}^{[r]}}) \subseteq K_{A,\,{\boldsymbol{j}^{[r]}}}, \quad r \in {{\mathbb N}}.$$ Moreover, if the mask entries $a_j(\alpha)>0$, $\alpha \in \operatorname{supp}(a_j)$, $j=1, \ldots,J$, then $\operatorname{supp}(\phi_{\boldsymbol{j}^{[r]}}) = K_{A,\,{\boldsymbol{j}^{[r]}}}$, $r \in {{\mathbb N}}$.
Example \[ex\_supp\_of\_blf\] shows that, for different subdivision schemes in $\mathcal{S}^{{\mathbb N}}$, the supports of the corresponding basic limit functions may have a completely different structure. Similar observation has been already made in the context of non-stationary schemes in e.g. [@CCGP16a].
\[ex\_supp\_of\_blf\] We consider the set $\mathcal{S}=\{S_j=(a_j,M_j)\;:\; j=1,2\}$ from [@Cotronei2015 section 4] with the dilation matrices $$M_1=\left(\begin{array}{rr}1 &1\\ 1& -2\end{array}\right), \quad
M_2=\left(\begin{array}{rr}2 &-1\\ 1& -2\end{array}\right),$$ and the masks $$a_j(0,-2)=a_j(0,2)=\frac{1}{3}, \quad a_j(0,-1)=a_j(0,1)=\frac{2}{3}, \quad a_j(0,0)=1, \quad j=1,2.$$ The matrices $M_1$ and $M_2$ are jointly expanding, due to $\|M_{j_1}^{-1}M_{j_2}^{-1}\|_2<1$ for all $j_1, j_2 \in \{1,2\}$. The supports of the basic limit functions $$\phi_1=\lim_{n \rightarrow \infty} \prod_{\ell=1}^n (S_2^{2\ell} S_1) \delta, \quad
\phi_2=\lim_{n \rightarrow \infty} \prod_{\ell=2}^n (S_2^{2\ell} S_1) S_2 S_2 \delta \quad
\text{and} \quad
\phi_3=\lim_{n \rightarrow \infty} \prod_{\ell=2}^n (S_2^{2\ell} S_1) S_2 \delta$$ are given in Figure \[fig\_supp\_of\_blf\]. The Matlab code to produce the figures is
S=getS('2_ex_CGRS');
blf({[1 2 2 1 2 2 2 2 1 2 2 2 2 2 2 1],[2]},S,'iterations',9)
blf({[ 2 2 1 2 2 2 2 1 2 2 2 2 2 2 1],[2]},S,'iterations',9)
blf({[ 2 1 2 2 2 2 1 2 2 2 2 2 2 1],[2]},S,'iterations',9)
axis equal; axis([-2.8 2.9 -3.2 3.2]);
![Supports of the basic limit functions from Example \[ex\_supp\_of\_blf\].[]{data-label="fig_supp_of_blf"}](suppblf "fig:"){width=".7\textwidth" height="5cm"}\
Transition matrices and matrix refinement {#sec_transition_matrices_supports}
=========================================
In section \[sec\_transition\_matrices\_supports\], we construct the transition matrices in the setting of multiple multivariate subdivision. In particular, in Lemma \[alg\_invariantomega\], we provide an algorithm for the construction of the minimal, invariant subspace of the transition operators from Definition \[def\_transition\_operator\]. The corresponding finite set $\Omega_C$ leads to transition matrices of minimal size and, thus, is more suitable for computations in section \[sec:examples\]. Furthermore, Lemma \[thm\_invariantomega\] together with Lemma \[lem:Support\_OmegaZ\_Attractor\] guarantees the existence of a matrix vector form of the refinement equation . This explains the special role of the finite set $\Omega_{{\mathbb Z}}$ constructed in Lemma \[thm\_invariantomega\]. The set $\Omega_{{\mathbb Z}}$ would be also suitable for defining the transition matrices, but it cannot always be computed and would lead to transition matrices of a larger size.
Similarly to the stationary and non-stationary settings, there are two important ingredients of our construction: the transition operators and their common finite dimensional invariant subspaces.
\[def\_transition\_operator\] Let $j \in \{1, \ldots, J\}$. For the subdivision mask $a_j \in \ell_0({{\mathbb Z}}^s)$ and the dilation matrix $M_j \in {{\mathbb Z}}^{s \times s}$ with the digit set $D_j\simeq {{\mathbb Z}}^s/M_j{{\mathbb Z}}^s$, we define the *transition operator* $\mathcal{T}_{d,j}: \ell({{\mathbb Z}}^s)\rightarrow\ell({{\mathbb Z}}^s)$ by $$\label{equ_transition_operator}
(\mathcal{T}_{d,j} c)(\alpha)=\sum_{\beta\in{{\mathbb Z}}^s} a_j(M_j\alpha-\beta+d) \;c(\beta), \quad d\in D_j, \quad c \in\ell({{\mathbb Z}}^s),\quad \alpha\in{{\mathbb Z}}^s.$$ The set of all transition operators is denoted by $$\mathcal{T}=\{\mathcal{T}_{d,j} \;:\; d \in D_j, \; j=1,\ldots,J\}.$$
The result of Lemma \[thm\_invariantomega\] ensures the existence of a common finite dimensional invariant subspace of the transition operators.
Let $\Omega\subset{{\mathbb Z}}^s$. The set $\ell(\Omega)=\{c\in\ell({{\mathbb Z}}^s) : \operatorname{supp} c \subseteq\Omega\}$ is the set of all sequences $c\in\ell({{\mathbb Z}}^s)$ supported on $\Omega$.
\[thm\_invariantomega\] There exists a finite set $\Omega_{{\mathbb Z}}\subset{{\mathbb Z}}^s$ such that $\ell(\Omega_{{\mathbb Z}})$ is invariant under all operators in $\mathcal{T}$.
By [@Rota60 Proposition 1] and due to the assumption that the dilation matrices are jointly expanding, there exists a matrix norm $\|\cdot\|$ such that the inverses of $M_j$ are contractive on ${{\mathbb R}}^s$ w.r.t. this norm. Let $\mathcal{X}$ be the set of all non-empty, compact subsets of ${{\mathbb R}}^s$. By [@Federer 2.10.21], the space $(\mathcal{X},d_H)$, where $d_H$ is the Hausdorff metric w.r.t $\|\cdot\|$, is a complete metric space. The mappings $$\label{eq:def_contractiveMj}
\mathcal{M}_j: \mathcal{X} \rightarrow \mathcal{X}, \quad
\mathcal{M}_j(X)=M_j^{-1}(\operatorname{supp}a_j+X-D_j),\quad X\in \mathcal{X}, \quad
j=1,\ldots,J,$$ are contractive. By the results in [@Hut81 section 3.1], there exists a unique $\Omega_{{\mathbb R}}\in \mathcal{X}$ such that $$\label{equ_one_invariant_vector_3}
\Omega_{{\mathbb R}}=\bigcup_{j=1,\ldots,J} \mathcal{M}_{j}(\Omega_{{\mathbb R}}).$$ Define $$\Omega_{{\mathbb Z}}=\Omega_{{\mathbb R}}\cap{{\mathbb Z}}^s.$$ Let $d \in D_j$ for $j \in\{1,\ldots,J\}$. We show that $\mathcal{T}_{d,j}:\ell(\Omega_{{\mathbb Z}})\rightarrow\ell(\Omega_{{\mathbb Z}})$. Indeed, if $v \in \ell(\Omega_{{\mathbb Z}})$, then by $ \displaystyle \operatorname{supp}(\mathcal{T}_{d,j} v)\subseteq
\bigcup_{j} M^{-1}_j(\operatorname{supp}a_j+\Omega_{{\mathbb Z}}-D_j)\subseteq
\bigg(\bigcup_{j} M^{-1}_j(\operatorname{supp}a_j+\Omega_{{\mathbb R}}-D_j)\bigg) \cap {{\mathbb Z}}^s= \Omega_{{\mathbb Z}}$.
The result of Lemma \[thm\_invariantomega\] allows us to associate each transition operator $\mathcal{T}_{d,j}$ with a certain square matrix.
\[def:transition\_matrices\] Let $\Omega \subset {{\mathbb Z}}^s$ be finite and such that $\ell(\Omega)$ is $\mathcal{T}$ invariant. For the operators in $\mathcal{T}$ we define the *transition matrices* $$T_{d,j,\Omega}=\big(a_j(M_j\alpha-\beta+d)\big)_{\alpha,\beta\in\Omega}, \quad d \in D_j, \quad j=1, \ldots,J.$$ We note that $\alpha$ and $\beta$ are the respective row and column indices.
In the rest of the paper we use two other sets $\Omega \subset {{\mathbb Z}}^s$ such that $\ell(\Omega)$ is invariant under all operators in $\mathcal{T}$: the set $\Omega_C$ from Lemma \[alg\_invariantomega\] for numerical computation in section \[sec:examples\]; and the larger set $\Omega_V$ from Proposition \[thm\_VkequVbark\_existence\] for the theoretical analysis in section \[sec:JSRvRSR\].
The sum rules of order one for the masks $a_j$, Assumption **S**, become conditions on the spectral properties of the transition matrices.
\[thm\_invariantone\] Let $\mathcal{S}$ be a finite set of subdivision operators whose masks satisfy Assumption $\operatorname{\hbox{\bf S}}$ and $\mathcal{T}_{d,j} \in \mathcal{T}$.
$(i)$
: If $\mathcal{T}_{d,j}:\ell(\Omega)
\rightarrow\ell(\Omega)$, $\Omega \subset {{\mathbb Z}}^s$, then the transition matrix $T_{d,j,\Omega}$ satisfies $(1,\ldots,1)\, T_{d,j,\Omega}=(1,\ldots,1)$.
$(ii)$
: If all entries of $T_{d,j,\Omega}$ are non-negative, then $(1,\ldots,1)\, T_{d,j,\Omega}=(1,\ldots,1)$ implies that $\mathcal{T}_{d,j}:\ell(\Omega)\rightarrow\ell(\Omega)$.
$(i)$ Invariance of $\ell(\Omega)$ under $\mathcal{T}_{d,j}$, $d \in D_j$, $j \in \{1, \ldots,J\}$, implies, by Definition \[def\_transition\_operator\], that $a_j(M_j\alpha-\beta+d)=0$, whenever $\alpha \notin \Omega$ and $\beta\in \Omega$. By Assumption **S**, we conclude that the entries in each column of the corresponding transition matrix $\mathcal{T}_{d,j}$ sum up to one, since $$1=\sum_{\alpha\in{{\mathbb Z}}^s} a_j(M_j\alpha-\beta+d)=\sum_{\alpha\in\Omega} a_j(M_j\alpha-\beta+d), \quad \beta \in \Omega.$$ $(ii)$ Assume that $a_j(\alpha)\geq 0$ for all $\alpha\in{{\mathbb Z}}^s$, $j \in \{1,\ldots,J\}$. Due to $(1,\ldots,1)\, T_{d,j,\Omega}=(1,\ldots,1)$, $d \in D_j$, we get $$1=\sum_{\alpha\in\Omega}a_j(M_j\alpha-\beta+d),\quad \beta\in\Omega.$$ Assumption **S**, i.e. the sum rules, implies that $a_j(M_j\alpha-\beta+d)=0$ for all $\alpha\not\in\Omega$, $\beta\in \Omega$. Thus, $\mathcal{T}_{d,j}:\ell(\Omega)\rightarrow\ell(\Omega)$.
\[rem\_invariantomega\] For the computation of the joint spectral radius in section \[sec:examples\], the approximations (via the fixed point iteration [@Hut81 section 3.1 $(viii)$]) of $\Omega_{{\mathbb Z}}$ defined in Lemma \[thm\_invariantomega\] are of no practical use. The following straightforward observation leads to an algorithm (see Lemma \[alg\_invariantomega\]) for explicit computation of $\Omega_C \subseteq \Omega_{{\mathbb Z}}$ with desired invariance properties as in Lemma \[thm\_invariantone\]. Since $0\in\operatorname{supp}(a_j)$ and $0\in D_j$ for all $j=1,\ldots,J$, we conclude from [@Hut81 section 3.1 $(iii)$] that $0\in\Omega_{{\mathbb Z}}$. Note that this set $\Omega_C$ is, by construction, the smallest $\mathcal{T}$ invariant set which contains $0$.
\[alg\_invariantomega\] The following algorithm constructs a finite set $\Omega_C \subset {{\mathbb Z}}^s$ such that $\ell(\Omega_C)$ is $\mathcal{T}$ invariant.
&\_0={0}, i=0 &\
&\
&i=i+1\
&\_i=\_[i-1]{}\
& j=1,…,J \
&\_[i,j]{}=(M\_j\^[-1]{}(a\_j + \_i - D))[[Z]{}]{}\^s\
&\_i=\_i \_[i,j]{}\
&\
& \_[i]{}=\_[i-1]{}\
&\_C=\_i
We first prove that the algorithm terminates after finitely many steps. More precisely, we show, by induction on $i$, that the sets $(\Omega_i)_{i=0}^n$, $n \in {{\mathbb N}}$, are increasing, nested subsets of the finite set $\Omega_{{\mathbb Z}}$ determined in Lemma \[thm\_invariantomega\]. Thus, $n$ is finite. Indeed, by Remark \[rem\_invariantomega\], $\Omega_0=\{0\}\subset\Omega_{{\mathbb Z}}$. Assume that $\Omega_i\subseteq\Omega_{{\mathbb Z}}$, $i\le n$. By Lemma \[thm\_invariantomega\], $\Omega_{{\mathbb Z}}$ is invariant under all operators $M_j^{-1}(\operatorname{supp}(a_j)+\cdot-D)$, thus, we get $\Omega_{i,j}\subseteq \Omega_{{\mathbb Z}}$ for all $j=1,\ldots,J$. Therefore, $\Omega_{i}=\Omega_{i-1} \cup \bigcup_{j} \Omega_{i,j} \subseteq\Omega_{{\mathbb Z}}$. Due to the stopping criterion, we get increasing, nested sets $\Omega_0\subset \Omega_1\subset \Omega_2 \subset \cdots \Omega_n \subseteq\Omega_{{\mathbb Z}}$. Moreover, for $\Omega_C=\Omega_n=\Omega_{n-1}$, due to , we get $\mathcal{T}_d:\ell(\Omega_{n-1}) \rightarrow \ell(\Omega_{n})$, $d\in D_j$, $j\in\{1,\ldots,J\}$. Thus, the claim follows.
The choice of $\Omega_0=\{0\}$ in Lemma \[alg\_invariantomega\] is not crucial. Given any $\Omega_0\subset{{\mathbb Z}}^s$, finite, the algorithm in Lemma \[alg\_invariantomega\] constructs a set $\Omega_C$ such that $\ell(\Omega_C)$ is $\mathcal{T}$ invariant and $\Omega_0\subseteq\Omega_C$. This follows directly by artificially enlarging the sets $\operatorname{supp}a_j$ such that $\Omega_0\subseteq\Omega_{{\mathbb R}}$, with $\Omega_{{{\mathbb R}}}$ from Lemma \[thm\_invariantomega\].
In some cases, the sets $\Omega_{{\mathbb Z}}$ defined in Lemma \[thm\_invariantomega\] and $\Omega_C$ constructed in Lemma \[alg\_invariantomega\] coincide.
$(i)$ Let $M=2$, $D=\{0,1\}$ and $\operatorname{supp}a=\{0,1,2\}$. Then $$\Omega_{{\mathbb R}}= M^{-1}(\operatorname{supp}a-D)+M^{-2}(\operatorname{supp}a-D)+M^{-3}(\operatorname{supp}a-D)+\cdots=[-1,2]$$ Thus, $\Omega_{{\mathbb Z}}=\{-1,0,1,2\}$. The algorithm in Lemma \[alg\_invariantomega\] generates $\Omega_C=\{0,1\}$.
$(ii)$ Let $M=-2$, $D=\{-1,~0\}$ and $\operatorname{supp}a=\{0,1,2,3\}$. Then, by [@CHM Proposition 2.7], $\Omega_{{\mathbb R}}=K_{A,\,\boldsymbol{j}}-K_{D\!,\,\boldsymbol{j}}=
[-\frac{5}{3},\frac{1}{3}]-[-\frac{1}{3},\frac{2}{3}]=
[-\frac{7}{3},\frac{2}{3}]$. Thus, $\Omega_{{\mathbb Z}}=\{-2,-1,0\}$. The algorithm in Lemma \[alg\_invariantomega\] produces the same set.
Similarly to the stationary and non-stationary settings, the supports of the basic limit functions $(\phi_{\boldsymbol{j}^{[r]}})_{r \in {{\mathbb N}}}$ can be covered by the integer shifts of the corresponding attractors $K_{D\!,\,\boldsymbol{j}^{[r]}}$ in Definition \[def\_attractor\]. This leads to a standard matrix form of the refinement equations used for analysing the existence and regularity of refinable functions in the stationary and non-stationary settings. The results of section \[sec:JSRvRSR\] however do not rely on such representations and the remaining part of this section is merely for a curious reader.
\[lem:Support\_OmegaZ\_Attractor\] Let $\Omega_{{\mathbb Z}}\subset {{\mathbb Z}}^s$ be as in Lemma \[thm\_invariantomega\]. Assume that the subdivision scheme $(S_{j_\ell})_{\ell \in {{\mathbb N}}}\in\mathcal{S}^{{\mathbb N}}$ is convergent. Then $\operatorname{supp}(\phi_{\boldsymbol{j}^{[r]}}) \subseteq \Omega_{{\mathbb Z}}+ K_{D\!,\,\boldsymbol{j}^{[r]}}$ for all $r\in{{\mathbb N}}$.
Without loss of generality we assume $r=1$, i.e. $\boldsymbol{j}=\boldsymbol{j}^{[1]}$. Recall, from Lemma \[thm\_invariantomega\], that $\Omega_{{\mathbb R}}\subset {{\mathbb R}}^s$ is compact and is the unique solution of the fixed point equation in . We show first that $K_{A,\,\boldsymbol{j}}-K_{D\!,\,\boldsymbol{j}} \subseteq \Omega_{{\mathbb R}}$ with $K_{A,\,\boldsymbol{j}}$ in , $K_{D\!,\,\boldsymbol{j}}$ in and $\boldsymbol{j}=(j_\ell)_{\ell \in {{\mathbb N}}}$. By the results in [@Hut81 section 3.1], the set $\Omega_{{\mathbb R}}$ is also the closure (in the Hausdorff metric) of the fixed points of the compositions of the contractive mappings ${\cal M}_{r_i}$ defined in . More precisely, $$\Omega_{{\mathbb R}}=
\operatorname{clos} \left\{
\Omega_{\ell_1,\ldots, \ell_k} \in\mathcal{X} \;:\;
\Omega_{\ell_1,\ldots, \ell_k} =
{\cal M}_{\ell_1} \circ \ldots \circ {\cal M}_{\ell_k}(\Omega_{\ell_1,\ldots, \ell_k}),
\ \ell_i \in \{1, \ldots,J\}, \
i=1,\ldots,k, \ k \in {{\mathbb N}}\, \right\}$$ and $\displaystyle \lim_{k \rightarrow \infty} \Omega_{\ell_1,\ldots, \ell_k}$ exist and belong to $\Omega_{{\mathbb R}}$. Thus, for the specific ordering in $\boldsymbol{j}$, by and due to $$\begin{aligned}
\Omega_{{\mathbb R}}\supseteq {\cal M}_{j_1} \circ \ldots \circ {\cal M}_{j_k}(\Omega_{j_1,\ldots, j_k})
= M_{j_1}^{-1} \operatorname{supp}a_{j_1} + M_{j_1}^{-1} M_{j_2}^{-1} \operatorname{supp}a_{j_2}+ \ldots + M_{j_1}^{-1} \ldots M_{j_k}^{-1}
\operatorname{supp}a_{j_k} +&\\
+ M_{j_1}^{-1} \ldots M_{j_k}^{-1}\Omega_{j_1,\ldots, j_k}
- M_{j_1}^{-1} D_{\!j_1} - M_{j_1}^{-1} M_{j_2}^{-1} D_{\!j_2}+ \ldots - M_{j_1}^{-1} \ldots M_{j_k}^{-1} D_{\!j_k},&\end{aligned}$$ we get $K_{A,\,\boldsymbol{j}}-K_{D\!,\,\boldsymbol{j}} \subseteq \Omega_{{\mathbb R}}$. Now we are ready to prove the claim. By [@Groe2006 Lemma 1], ${{\mathbb R}}^s=K_{D\!,\,\boldsymbol{j}}+{{\mathbb Z}}^s$. Thus, for $x\in\operatorname{supp}(\phi_{\boldsymbol{j}})=K_{A,\,\boldsymbol{j}}$, there exists $\alpha\in{{\mathbb Z}}^s$ such that $x\in K_{D\!,\,\boldsymbol{j}}+\alpha$. Therefore, $\alpha\in x-K_{D\!,\,\boldsymbol{j}}\subseteq K_{A,\,\boldsymbol{j}}-K_{D\!,\,\boldsymbol{j}}\subseteq\Omega_{{\mathbb R}}$, i.e. $\alpha \in \Omega_{{\mathbb Z}}=\Omega_{{\mathbb R}}\cap {{\mathbb Z}}^s$. This implies that $x \in K_{D\!,\,\boldsymbol{j}}+\Omega_{{\mathbb Z}}$.
Lemma \[lem:Support\_OmegaZ\_Attractor\] generalizes the result [@CHM Proposition 2.7]. We conjecture that the result of Lemma \[lem:Support\_OmegaZ\_Attractor\] is true for an arbitrary finite $\Omega \subset {{\mathbb Z}}^s$, such that $\ell(\Omega)$ is $\mathcal{T}$ invariant, e.g. the set $\Omega_C$ from Lemma \[alg\_invariantomega\].
\[lemma:supp\_phi\] Let $\mathcal{S}^{{{\mathbb N}}}$ be convergent and $\Omega\subset{{\mathbb Z}}^s$ be finite and such that $\ell(\Omega)$ is $\mathcal{T}$ invariant. Then $\operatorname{supp}(\phi_{\boldsymbol{j}^{[r]}}) \subseteq \Omega+K_{D\!,\,\boldsymbol{j}^{[r]}}$ for all $r\in{{\mathbb N}}$.
Comparison of matrix and operator approaches: convergence of multiple subdivision schemes {#sec:JSRvRSR}
=========================================================================================
The goal of this section is to unify the matrix (joint spectral radius) and operator (restricted spectral radius) approach in the setting of multiple subdivision schemes, see Theorem \[th:main\]. It generalizes similar results in [@Charina; @CCS05] that were proven in the stationary setting for the case of the dilation matrix $M=2I$.
One of the standard tools for checking the regularity of subdivision schemes is the so-called restricted spectral radius (see e.g. [@CDM; @Charina; @CCS04; @Sauer2012]) that measures the spectral properties of the difference subdivision operators restricted to a certain subspace of $\ell({{\mathbb Z}}^s)$.
The concept of the restricted spectral radius relies on the difference operators and difference subdivision schemes operating on the sequences in $\ell({{\mathbb Z}}^s)$. By $\ell({{\mathbb Z}}^s,{{\mathbb R}}^s)$ we denote the space of vector-valued (with values in ${{\mathbb R}}^s$) sequences indexed by ${{\mathbb Z}}^s$.
Let $e_\ell$, $1\leq \ell \leq s$, be the standard unit vectors of ${{\mathbb R}}^s$. We define
$(i)$
: *the $\ell$-th backward difference operator* $\nabla_{\!\ell}:\ell({{\mathbb Z}}^s)\rightarrow\ell({{\mathbb Z}}^s)$ by $\nabla_{\!\ell}\, c=c-c(\cdot-e_\ell)$, $c \in \ell({{\mathbb Z}}^s)$.
$(ii)$
: *the backward difference operator* $\nabla:\ell({{\mathbb Z}}^s)\rightarrow\ell({{\mathbb Z}}^s,{{\mathbb R}}^{s})$ by $ \displaystyle \nabla= \left(\begin{array}{cccc}
\nabla_{\!1}&
\nabla_{\!2}&
\dots&
\nabla_{\!s} \end{array} \right)^{\!T}$.
The existence of difference subdivision operators (see e.g. for details [@CDM; @MoeSau]) is ensured by Assumption **S**.
Let $\mathcal{S}$ be a finite set of subdivision operators whose masks satisfy Assumption $\operatorname{\hbox{\bf S}}$. For $S \in \mathcal{S}$, *a difference subdivision operator* $S':\ell({{\mathbb Z}}^s) \rightarrow \ell({{\mathbb Z}}^s, {{\mathbb R}}^s)$ is defined by $$\label{def:Sprime}
\nabla S=S' \nabla \, .$$ By $\mathcal{S}'$ we denote a set of the difference operators $S'$ associated to the set $\mathcal{S}$ of subdivision operators.
In the setting of the multiple subdivision, we use the following definition of the restricted spectral radius given in [@Sauer2012 section 3, “*normalized joint spectral radius*”].
\[def:RSR\] Let $\mathcal{S}$ be a finite set of subdivision operators whose masks satisfy Assumption **S**. The *restricted norm* of $S' \in \mathcal{S}'$ is defined by $$\label{equ_def_RSR_norm}
\|S'|_{\nabla}\|_\infty= \max_{\substack{ \|\nabla c \|_\infty=1}} \| S' \nabla c\|_{\infty}.$$ The *restricted spectral radius* of $\mathcal{S}'$ is defined by $$\label{equ_RSR}
\begin{aligned}
\rho(\mathcal{S}'|_\nabla)=& \limsup_{n\rightarrow\infty} \sup_{ S'_{j_\ell} \in \mathcal{S}'}
\left\|S'_{j_n}\cdots S'_{j_1}|_{\nabla}\right\|_\infty^{1/n}.
\end{aligned}$$
The main result of this section, Theorem \[th:main\], leads to a characterization of convergence of $\mathcal{S}^{{\mathbb N}}$ in terms of the joint spectral radius of the transition matrices (Definition \[def:transition\_matrices\]) restricted to a common invariant subspace. This characterization follows from Theorem \[th:main\] and the following result.
[[@Sauer2012 Theorem 2]]{}\[thm\_RSR\_convergence\] $\mathcal{S}^{{\mathbb N}}$ is convergent if and only if there exists $\mathcal{S}'$ such that $\rho(\mathcal{S}'|_\nabla)<1$.
Theorem \[th:main\] allows us to use the invariant polytope algorithm from [@GP1] for the computation of the joint spectral radius, when checking the convergence of multiple subdivision schemes, see section \[sec:examples\]. The proof of Theorem \[th:main\] is similar to the one of [@Charina Proposition 4.6], see also [@CCS05]. The crucial differences between stationary and multiple cases are pointed out in Propositions \[thm\_different\_RN\] and \[thm\_VkequVbark\_existence\]. Proposition \[thm\_different\_RN\] is a generalization of [@Charina Proposition 4.1].
\[thm\_different\_RN\] Let $(S'_{j_\ell})_{\ell \in {{\mathbb N}}}\in\mathcal{S'}^{{\mathbb N}}$ be a difference subdivision scheme. Then $$\label{equ_rth_restricted_normalized_norm}
\|S'_{j_n}\cdots S'_{j_1}|_{\nabla}\|_\infty=\max_{\substack{\nabla c\in\ell_\infty(([-1,1]^s-K)\cap {{\mathbb Z}}^s)\\
\|\nabla c\|_\infty = 1}}\, \max_{\alpha\in M_{j_n}\cdots M_{j_1}[0,1)^s\cap{{\mathbb Z}}^s}\,
\left\|S'_{j_n}\cdots S'_{j_1} \nabla c(\alpha)\right\|_{\infty}, \quad n \in {{\mathbb N}},$$ where $$\label{def:K}
K=\bigcup_{j=1,\ldots,J} M_j^{-1}(\operatorname{supp}a_j + K).$$
By definition of $S'_{j_\ell}$ we get $$\|S'_{j_n}\cdots S'_{j_1}|_{\nabla}\|_\infty
=\max_{\substack{ \|\nabla c\|_\infty = 1}}\,
\sup_{\alpha\in{{\mathbb Z}}^s}\, \left\| \sum_{\beta\in{{\mathbb Z}}^s} (S'_{j_n}\cdots S'_{j_1}\delta I)(\alpha - M_{j_n}\cdots M_{j_1} \beta)\nabla c(\beta)
\right\|_\infty.$$ Due to the periodicity of the subdivision (i.e. $|M_{j_n}\cdots M_{j_1}[0,1)^s\cap{{\mathbb Z}}^s|$ different subdivision rules at the $n$-th level of subdivision recursion), it suffices to take $\alpha\in M_{j_n}\cdots M_{j_1}[0,1)^s\cap{{\mathbb Z}}^s$. Finally, by [@Charina Remark 3.7] $$\operatorname{supp}(S'_{j_n}\cdots S'_{j_1}\delta I) \subseteq
M_{j_n}\cdots M_{j_2} \operatorname{supp}a_{j_1}+ \cdots +
M_{j_n} \operatorname{supp}a_{j_{n-1}}+ \operatorname{supp}a_{j_n}.$$ For $\alpha-M_{j_n}\cdots M_{j_1}\beta \in \operatorname{supp}(S'_{j_n}\cdots S'_{j_1}\delta I)$, by a similar argument as in the proof of Lemma \[lem:Support\_OmegaZ\_Attractor\] and due to $0\in\operatorname{supp}a_j$, we obtain $$\begin{aligned}
\beta &\in [0,1)^s\cap M_{j_1}^{-1}\cdots M_{j_n}^{-1}{{\mathbb Z}}^s - M_{j_1}^{-1} \operatorname{supp}a_{j_1} - \cdots -M_{j_1}^{-1}\cdots
M_{j_n}^{-1} \operatorname{supp}a_{j_n}\\
& \subseteq ([-1,1]^s-K) \cap {{\mathbb Z}}^s,\end{aligned}$$ where $K \subset {{\mathbb R}}^s$ is the unique compact set satisfying the fixed point equation .
Sufficient conditions for continuity of refinable functions or characterizations of continuity of basic limit functions of subdivision schemes are usually formulated in terms of the spectral properties of restrictions of transition matrices to $V_\Omega$ in or restrictions of difference subdivision operators to $\tilde{V}_\Omega$ in , respectively.
\[def\_V0\] Let $\Omega\subset{{\mathbb Z}}^s$ be finite with $n=|\Omega|$. We define the linear spaces $$\label{equ_def_Vk}
V_\Omega=\big\{ v\in {{\mathbb R}}^n\;:\; \sum_{\beta\in\Omega} v(\beta)=0 \big\},$$ $$\label{equ_def_Vkbar}
\tilde{V}_\Omega= \operatorname{span}\left\{ v \in \ell_0({{\mathbb Z}}^s) \;:\; v= \nabla \delta (\cdot-\beta), \ \beta\in {{\mathbb Z}}^s, \ \operatorname{supp}v \subseteq \Omega
\right\}.$$
In the rest of the paper, we view $\tilde{V}_\Omega$ as a subspace of ${{\mathbb R}}^n$ (or, equivalently, $V_\Omega$ as a subspace of $\ell_0({{\mathbb Z}}^s)$) and make use of the following properties of $V_\Omega$ and $\tilde{V}_\Omega$.
\[thm\_VkequVbark\_dim\] Let $\Omega\subset{{\mathbb Z}}^s$ be finite. Then $\tilde{V}_\Omega\subseteq V_\Omega$ and, if $\operatorname{dim}V_\Omega=
\operatorname{dim}\tilde{V}_\Omega$, then $V_\Omega = \tilde{V}_\Omega$.
We are now ready to formulate the main result of this section, Theorem \[th:main\]. The proof of Theorem \[th:main\] is given in subsection \[subsec:proof\_main\].
\[th:main\] Let $\mathcal{S}$ be a finite set of subdivision operators whose masks satisfy Assumption $\operatorname{\hbox{\bf S}}$. Assume that there exists a finite set $\Omega \subset {{\mathbb Z}}^s$ such that
- $\ell(\Omega)$ is invariant under the transition operators in $\mathcal{T}$ and
- $V_\Omega = \tilde{V}_\Omega$.
Then $\rho(\mathcal{S}'|_\nabla)=\rho(\{T_{d,j,\Omega}|_{V_\Omega} \;:\; d \in D_j, \ j=1,\ldots,J\})$.
Example \[ex\_mult1\] shows that assumption $(ii)$ of Theorem \[th:main\] is indeed crucial. The natural candidate for such a set $\Omega$ would be the set $\Omega_C$ from Lemma \[alg\_invariantomega\]. The set $\Omega_C$, by Lemma \[alg\_invariantomega\], satisfies assumption $(i)$ of Theorem \[th:main\] and our numerical experiments show that in most cases $\Omega_C$ also satisfies the assumption $(ii)$. However, Example \[ex\_V0\_neq\_V0bar\] illustrates that the case $\tilde{V}_{\Omega_C} \subset V_{\Omega_C}$ occurs sometimes even in the stationary setting. In such cases, we choose $\Omega=\Omega_V$ from Proposition \[thm\_VkequVbark\_existence\].
\[ex\_V0\_neq\_V0bar\] Consider the dilation matrix $M=\left(\begin{array}{rr}-3& -4\\ 4& 4\end{array}\right)$ with the digit set $D=\{ (-k,k) : k=0,1,2,3 \}$ and choose any mask $a$ with $$\operatorname{supp}(a)=\left\{ \left(\begin{array}{c}1\\0\end{array}\right),
\left(\begin{array}{c}2\\1\end{array}\right),
\left(\begin{array}{c}3\\1\end{array}\right),
\left(\begin{array}{c}1\\2\end{array}\right),
\left(\begin{array}{c}0\\3\end{array}\right),
\left(\begin{array}{c}1\\4\end{array}\right),
\left(\begin{array}{c}3\\4\end{array}\right)
\right\}.$$ The set $\Omega_C$ constructed by the algorithm in Lemma \[alg\_invariantomega\] is drawn in Figure \[fig\_V0\_neq\_V0bar\].
Straightforward computation shows that $\operatorname{dim} V_{\Omega_C}=33 > \operatorname{dim} \tilde{V}_{\Omega_C}=32$. Thus, $\Omega_C$ will be inappropriate for further theoretical analysis. The problematic point is $(-2,1)$ which has no direct neighbour. See Remark \[rem:badOmegaC\] for more details. The Matlab code to produce Figure \[fig\_V0\_neq\_V0bar\] is
S=getS('2_ex_V0neqV0bar_1');
Om=constructOmega(S);
plotm(Om,'k.','MarkerSize',10)
axis equal; axis([-3 9 -10 2]);
![The set $\Omega_C$ from Example \[ex\_V0\_neq\_V0bar\].[]{data-label="fig_V0_neq_V0bar"}](V0_neq_V0bar){width=".99\textwidth" height="4cm"}
In Proposition \[thm\_VkequVbark\_existence\], we determine a finite set $\Omega_V \subset {{\mathbb Z}}^s$ such that $V_{\Omega_V} = \tilde{V}_{\Omega_V}$. The structure of $\Omega_V$ is adapted to the definition of the restricted spectral radius and makes the link between the two spectral radii more evident. The definition of the set $\Omega_V$ is straightforward in comparison to the set $\Omega_C$ from Lemma \[alg\_invariantomega\], but the latter is by far more efficient for computations in section \[sec:examples\].
For simplicity of presentation and without loss of generality we make the following assumption.
**Assumption N:** [ *We assume that $$\label{equ_assumption_N}
\|M_j^{-1}\|_2<1,\ j=1, \ldots,J.$$*]{}
The above assumption is true for a variety of dilation matrices considered in the literature, but is not true e.g. for the dilation matrix $M=\left(\begin{array}{rr} 1&-2\\2&-1 \end{array} \right)$ of the $\sqrt{3}$-subdivision. Nevertheless, Definition \[def\_jointly\_expanding\], norm equivalences and [@Rota60] guarantee the existence of $n \in {{\mathbb N}}$ such that any product of $n$ matrices from the set $\{M_j^{-1} \;:\; j=1,\ldots,J\}$ satisfies Assumption **N**. Indeed, for this matrix $\|M^{-2}\|_2<1$. If $n>1$, we then study the convergence of multiple subdivision defined by $\mathcal{S}^n$ instead of $\mathcal{S}$ in .
\[thm\_VkequVbark\_existence\] For the set $\mathcal{T}$ of transition operators, there exists a finite set $\Omega_V \subset{{\mathbb Z}}^s$ such that
1. $\ell(\Omega_V)$ is invariant under operators in $\mathcal{T}$ and
2. $V_{\Omega_V} = \tilde{V}_{\Omega_V}$.
By Assumption **N** in , $$C_M=\max_{j=1,\ldots,J} \|{M_j^{-1}}\|_2 <1.$$ Due to the finite support of the masks and finiteness of digit sets, we get finite constants $$C_a=\max \{\|\alpha\|_2\;:\; \alpha\in\operatorname{supp}(a_j),\ j=1,\ldots,J\}
\quad\text{and}\quad
C_D=\max \{ \|d\|_2\;:\; d\in D_j, \ j=1,\ldots,J\}.$$ We define $$\Omega_V =\big\{x \in {{\mathbb R}}^s\;:\; \|x\|_2 \le \frac{C_a+C_D}{1-C_M} \big\}\cap{{\mathbb Z}}^s.$$
$(i)$ Let $d\in D_j$, $j=1,\ldots,J$. By , $\mathcal{T}_{d,j} v(\alpha) \not =0$, if $\alpha\in{{\mathbb Z}}^s$ is such that $M_j\alpha-\beta+d \in\operatorname{supp}a_j$ for some $\beta\in\operatorname{supp}v\subseteq \Omega_V$, or, equivalently, $\alpha \in M_j^{-1}(\operatorname{supp}a_j-d+\Omega_V)$. Thus, we obtain $$\label{eq:aux}
\|{\alpha\|}_2\leq C_M \left(C_a+C_d+\frac{C_a+C_D}{1-C_M}\right)=\frac{C_a+C_D}{1-C_M} (C_M(1-C_M) + C_M)\le
\frac{C_a+C_D}{1-C_M},$$ since $-C_M^2+2C_M \le 1$, implying $\mathcal{T}_{d,j} v \in \ell(\Omega_V)$.
$(ii)$ The dimension of $V_{\Omega_V}$ is $|\Omega_V|-1$, due to $V_{\Omega_V}$ being orthogonal to the vector of all ones. To determine the dimension of $\tilde{V}_{\Omega_V}$, we consider the graph $G=(\Omega_V,W)$ with the set of edges $$W=\{ (w_1, w_2) \in \Omega_V^2 \;:\; \|w_1-w_2\|_1=1\}.$$ Using this point of view, every sequence of the form $\nabla_{\!l} \delta(\cdot-\beta)\in\tilde{V}_{\Omega_V}$, $\beta\in{{\mathbb Z}}^s$, $l\in\{1,\ldots,s\}$, is associated uniquely to an edge in $W$. The graph $G$ is connected, thus, there exists a corresponding spanning tree consisting of $|\Omega_V|-1$ edges from $W$ [@Diestel2005 Theorem 1.5.1]. Since any spanning tree does not contain cycles, the set of edges of the spanning tree corresponds to a set of linearly independent sequences in $\tilde{V}_{\Omega_V}$. Thus, $\operatorname{dim} \tilde{V}_{\Omega_V}=|\Omega_V|-1$.
\[rem:badOmegaC\] The proof of Proposition \[thm\_VkequVbark\_existence\] explains the phenomenon occurring in Example \[ex\_V0\_neq\_V0bar\]. The graph corresponding to the set $\Omega_C$ from this example consists of two connected components. This fact forces $\operatorname{dim} V_{\Omega_C}> \operatorname{dim} \tilde{V}_{\Omega_C}$.
Proof of Theorem \[th:main\] {#subsec:proof_main}
----------------------------
The proof of Theorem \[th:main\] generalizes the proofs of [@Charina Proposition 4.6] from the stationary setting, thus, we only sketch the steps of the proof.
Assumption **S**, i.e. the sum rules, for $\mathcal{S}=\{S_j\ : \ j=1, \ldots,J\}$ guarantees the existence of the difference subdivision operators $S_j'$ in $\mathcal{S}'=\{S'_j\ : \ j=1, \ldots,J\}$. Moreover, Assumption **S** and $(ii)$, by Lemma \[thm\_invariantone\] part $(i)$ and [@Jia Theorem 5.2], ensure that $V_{\Omega}$ is a common invariant subspace of the transition matrices in $\{ T_{d,j,\Omega} \ :\ d \in D_j, \ j=1,\ldots,J\}$. Thus, the restrictions $T_{d,j,\Omega}|_{V_\Omega}$ of the matrices in $\{T_{d,j,\Omega} \ :\ d \in D_j, \ j=1,\ldots,J\}$ to $V_{\Omega}$ are well defined.
Note that, by the definitions of the joint and restricted spectral radii, the claim follows from (with $T_\ell=T_{d_\ell,j_\ell,\Omega}$ for $d_\ell \in D_{\!j_\ell}$, $j_\ell \in \{1, \ldots, J\}$ and $\ell=1, \ldots, n$, $n \in {{\mathbb N}}$) $$C_1 \ \max_{T_\ell \in \mathcal{T}} \
\| T_n \cdots T_1|_{V_\Omega}\|_\infty
\leq
\|S'_{j_n} \cdots S'_{j_1} |_\nabla\|_\infty
\leq
C_2 \ \max_{T_\ell \in \mathcal{T}} \
\|T_n \cdots T_1|_{V_\Omega}\|_\infty$$ with some constants $C_1, C_2 >0$. To determine $C_1$ and $C_2$, we first use the assumption $(ii)$ and an argument similar to the one of [@Charina Lemma 4.5] which implies that $$C_3 \ \max_{T_\ell \in \mathcal{T}} \
\| T_n \cdots T_1|_{V_\Omega}\|_\infty
\leq
\| S'_{j_n} \cdots S'_{j_1} \nabla \delta \|_\infty
\leq
C_4 \ \max_{T_\ell \in \mathcal{T}} \
\| T_n \cdots T_1|_{V_\Omega}\|_\infty$$ with $C_3=|\Omega|^{-2}$ and $C_4=1$. Then, for $K$ in , define $\Omega_K=([-1,1]^s-K) \cap {{\mathbb Z}}^s$. Due to $\delta \in \ell_\infty(\Omega_K)$ and $\|\nabla \delta\|_\infty=1$, by Proposition \[thm\_different\_RN\], we get $$\|S'_{j_n} \cdots S'_{j_1} |_\nabla\|_\infty \ge \| S'_{j_n} \cdots S'_{j_1} \nabla \delta \|_\infty.$$ Thus, $C_1=C_3$. Moreover, using
$${\nabla} c=
{\nabla}\left(\sum_{\alpha\in{{\mathbb Z}}^s} c(\alpha)\delta(\cdot-\alpha) \right)=
\sum_{\alpha\in{{\mathbb Z}}^s} c(\alpha) {\nabla}\delta(\cdot-\alpha),$$
for the maximizing sequence $c \in \ell_\infty(\Omega_K)$ from Proposition \[thm\_different\_RN\], we obtain $$\|S'_{j_n} \cdots S'_{j_1} |_\nabla\|_\infty \le C_5 \| S'_{j_n} \cdots S'_{j_1} \nabla \delta \|_\infty$$ with $C_5=|\Omega_K| \cdot \|c\|_\infty$. Thus, $C_2= C_5$.
Examples {#sec:examples}
========
Example \[ex\_mult1\] shows that already in the univariate, stationary case the assumption $V_{\Omega} =\tilde{V}_{\Omega}$ in Theorem \[th:main\] is crucial.
\[ex\_mult1\] We consider the stationary subdivision scheme with dilation factor $M=2$ and mask $a \in \ell_0({{\mathbb Z}})$ whose non-zero elements are given by $$a(0)=\frac{1}{2}, \quad a(3)=1 \quad \text{and}\quad a(6)=\frac{1}{2}.$$ It is well known that this subdivision scheme does not converge, although there is a continuous, piecewise linear, compactly supported on $[0,6]$, solution $\phi$ of the corresponding refinement equation. For illustration purposes, we choose the digit set $D=\{0,3\}$. The algorithm in Lemma \[alg\_invariantomega\] generates the set $\Omega_C=\{0,3\}$ and, by Definition \[def\_V0\], we have $\operatorname{dim}V_{\Omega_C} =2 > \operatorname{dim}\tilde{V}_{\Omega_C}=0$. The set $\Omega_V$, with the property $\operatorname{dim}V_{\Omega_V} = \operatorname{dim}\tilde{V}_{\Omega_V}$, can be chosen, in this case, to be $\Omega_V=\{-2,\ldots,5\}$. We make this choice for simplicity reasons, the set $\Omega_V$ from Proposition \[thm\_VkequVbark\_existence\] would be of size $61$. The corresponding transition matrices $T_{d,\Omega_V}$, $d \in D$, have the following block form $$T_{0,\Omega_V}=\frac{1}{2}
\left(\! \begin{array}{cccc} T_{0,\Omega_C}&0 \\0& T_{0,\Omega'}\end{array}\!\right)
\quad\text{and}\quad
T_{3,\Omega_V}=\frac{1}{2}
\left(\! \begin{array}{cccc} T_{3,\Omega_C}&0 \\0& T_{3,\Omega'}\end{array}\!\right),
\quad
\Omega'=\Omega_V \setminus \Omega_C =
\{-2,-1,1,2,4,5\},$$ with $$T_{0,\Omega_C}=\frac{1}{2}\left( \begin{array}{cc}1&0\\1&2\end{array}\right), \quad
T_{3,\Omega_C}=\frac{1}{2}\left( \begin{array}{cc}2&1\\0&1\end{array}\right),$$ and $$T_{0,\Omega'}=\frac{1}{2}\left( \begin{array}{cccccc}0&0&0&0&0&0\\1&0&0&0&0&0\\0&2&0&1&0&0
\\1&0&2&0&1&0\\0&0&0&1&0&2\\0&0&0&0&1&0\end{array}\right), \quad
T_{3,\Omega'}=\frac{1}{2}\left( \begin{array}{cccccc}0&1&0&0&0&0\\2&0&1&0&0&0\\1&0&2&0&1&0
\\0&0&1&0&2&0\\0&0&0&0&0&1\\0&0&0&0&0&0\end{array}\right).$$ By Lemma \[thm\_invariantone\] part $(ii)$, the space $\ell(\Omega_V)$ is $\mathcal{T}$ invariant. Thus, by Theorems \[thm\_RSR\_convergence\] and \[th:main\], due to $\operatorname{dim}V_{\Omega_V} = \operatorname{dim}\tilde{V}_{\Omega_V}$ and $\rho(\{T_{d,\Omega_V}|_{V_{\Omega_V}} : d \in D\})=1$, we get the correct answer that the scheme is not convergent. On the contrary, $\rho(\{T_{d,\Omega_C}|_{V_{\Omega_C}} : d \in D\})=1/2$ is misleading. Here, we used the invariant polytope algorithm from [@GP1] for our computations.
Multivariate example \[ex\_mult2\] illustrates the properties of multiple subdivision $\mathcal{S}^{{\mathbb N}}$ in the case $V_{\Omega_C}=\tilde{V}_{\Omega_C}$.
\[ex\_mult2\] We consider the set $\mathcal{S}=\{(a_j,M_j)\ : \ j=1,2\}$ of subdivision operators from Example \[ex\_supp\_of\_blf\] with the corresponding digit sets $$D_1=\left\{
\left(\begin{array}{r}0\\0\end{array}\right),
\left(\begin{array}{r}1\\\!-1\end{array}\right),
\left(\begin{array}{r}1\\0\end{array}\right)
\right\}
\quad\text{and}\quad
D_2=\left\{
\left(\begin{array}{r}0\\0\end{array}\right),
\left(\begin{array}{r}0\\\!-1\end{array}\right),
\left(\begin{array}{r}1\\0\end{array}\right)
\right\}.$$
\[fig:Omega\_eample\_section\] ![Set $\Omega_C$ from Example \[ex\_mult2\].[]{data-label="fig_mult"}](Om_mult "fig:"){width=".99\textwidth" height="4cm"}
Note that the masks $a_1=a_2$ satisfy Assumption **S**, i.e. the sum rules. The set $\Omega_C$ computed by the algorithm in Lemma \[alg\_invariantomega\] is given on Figure \[fig:Omega\_eample\_section\]. Note that $\operatorname{dim}V_{\Omega_C}=\operatorname{dim}\tilde{V}_{\Omega_C}$. By the invariant polytope algorithm from [@GP1] we obtain $$\rho(\{T_{d,\Omega_C,j}|_{V_{\Omega_C}},\ d\in D_{j},\ j=1,2\})=
\rho\left(
T_{{\tiny\left(\!\!\!\!\begin{array}{c}0\\0\end{array}\!\!\!\!\right)}, \Omega_C,1}
T_{{\tiny\left(\!\!\!\!\begin{array}{c}1\\0\end{array}\!\!\!\!\right)}, \Omega_C,2}
\right)^{1/2} =
0.8971\!\ldots
$$ Therefore, by Theorems \[thm\_RSR\_convergence\] and \[th:main\], $\mathcal{S}^{{\mathbb N}}$ is convergent. By [@CP2017 Theorem 1 and Remark 3] and [@CHM Proposition 3.27], the critical Hölder exponent $\alpha$ of the stationary subdivision scheme $(S_1)_{\ell \in {{\mathbb N}}}$ with the anisotropic dilation matrix $M_1$ satisfies $\alpha \in [0.3446\!\ldots,1]$. For the stationary subdivision scheme $(S_2)_{\ell \in {{\mathbb N}}}$ with the isotropic dilation matrix $M_2$, we obtain $\alpha=1$. For the stationary subdivision scheme $(S_1 S_2)_{\ell \in {{\mathbb N}}}$ with the isotropic dilation matrix $M_1M_2$, we get $\alpha=0.1977\!\ldots\ $.
Acknowledgement
===============
Both authors are sponsored by the Austrian Science Foundation (FWF) grant P28287-N35.
Appendix
========
For completeness, we provide the MATLAB code of the algorithm described in Lemma \[alg\_invariantomega\] in copy-paste-able format. The code, together with everything needed to execute the code snippets in this paper, is available for download from <http://tommsch.com> .
function [ Om ] = Omega(a, M, D, Om)
% a, M, D: cell vector of masks, dilation matrices and digit sets (as column vectors)
% Om: (Optional) the starting set
% ex: Omega({[1:3 2 1]/3,[1:3 2 1]/3},{[2 -1;1 -2],[1 1;1 -2]},{[0:2;0 0 0],[0:2;0 0 0]})
% Out: -[4 4 4 3 3 3 3 3 2 2 2 2 2 1 1 1 1 0 0;4 3 2 4 3 2 1 0 3 2 1 0 -1 2 1 0 -1 0 -1];
dim=size(M{1},1); %the dimension
if(nargin==3); Om=zeros(dim,1); end %if Omega is not given, set it to zero
while(true)
sizebefore=size(Om,2); %used to check if elements where added to Omega
for j=1:size(a,1) %iterate through all subdivision operators
OmN=M{j}\setplus(supp(a{j},dim),Om,-D{j}); %compute new possible entries
OmN=round(OmN(:,sum(abs(OmN-round(OmN)),1)<.5/abs(det(M{j})))); %round to integers
Om=unique([Om OmN]','rows')'; %remove duplicates
end
if(size(Om,2)==sizebefore); break; end %if no elements were added, terminate
end
function [ X ] = setplus( varargin )
% setplus(A,B) = { x=a+b : a in A, b in B}, operates column wise
% ex: setplus([1 2; 1 0],[0 -1; -1 -1]); %Output: [0 1 1 2;0 -1 0 -1]
sze=size(varargin,2); %number of sets
X=varargin{sze}; %the output set
for i=sze-1:-1:1 %iterate through all sets
A=varargin{i}; %the set to be added
X=repmat(A,1,size(X,2))+reshape(repmat(X,size(A,2),1),size(A,1),[]); %add the set
X=unique(X','rows')'; %remove duplicates
end
function [ L ] = supp(a, dim)
% returns the support of an array. First entry is supposed to have index (0,0,...,0)
% ex: supp([1 1; 0 1],2) %Output: [0 0 1; 0 1 1];
L=zeros(dim,nnz(a)); %output variable
CO=cell(1,dim); %dummy-variable to do calculation with indices
j=1; %index-variable for the columns of D
for i=1:numel(a) %iterate through all elements of the masks
if(a(i)~=0) %if the element is nonzero, save the indices
[CO{:}]=ind2sub(size(a),i); %get the indices
L(:,j)=[CO{:}]'-1; %change to zero-based indexing, add converted cell to vector
j=j+1; %increase counter
end
end
References
==========
C. A. Cabrelli, C. Heil, U. M. Molter, , Memoirs Amer. Math. Soc., 170 (2004), no. 807.
A. S. Cavaretta, W. Dahmen, C.A. Micchelli, , Memoirs Amer. Math. Soc., 93 (1991), no. 453.
M. Charina, , Appl. Comp. Harm. Anal., 32 (2012), 86–108.
M. Charina, C. Conti, N. Guglielmi, V. Yu. Protasov, , Numer. Math., published online (2016), 1–40.
M. Charina, C. Conti, N. Guglielmi, V. Yu. Protasov, , Appl. Math. Comput., 271 (2016), 20–27.
M. Charina, C. Conti, L. Romani, , Numer. Math., 127 (2014), 223–254.
M. Charina, C. Conti, T. Sauer, , Numer. Algorithms, 39 (2005), 97–113.
M. Charina, C. Conti, T. Sauer, , in Approximation Theory XI: Gatlinburg 2004, C. K. Chui, M. Neamtu and L. Schumaker (eds.), Nashboro Press, Brentwood, TN., 101–122.
M. Charina, M. Donatelli, L. Romani, V. Turati, , arXiv:1708.03469
M. Charina, V. Yu. Protasov, , Appl. Comp. Harm. Anal., (2017), $https://doi.org/10.1016/j.acha.2017.12.003$.
D. R. Chen, R. Q. Jia, S. D. Riemenschneider, , Appl. Comp. Harm. Anal., 12 (2002), 128–149.
A. Cohen, N. Dyn, , SIAM J. Math. Anal., 27 (1996), 1745–1769.
D. Collela, C. Heil, , SIAM J. Matrix Anal. Appl., 15 (1994), 496–518.
M.Cotronei, D. Ghisi, M. Rossini, T. Sauer, , Adv. Comput. Math., 41 (2015), 709–726.
I. Daubechies, J. Lagarias, , SIAM J. Math. Anal., 23 (1992), 1031–1079.
R. Diestel , Springer, Berlin Heidelberg, 2005.
N. Dyn, D. Levin, , Acta Numer., 11 (2002), 73–144.
H. Federer, , Springer-Verlag, New York, 1969.
K. Gröchenig, W. R. Madych, , IEEE Trans. Inf. Theor., 38 (2006), 556–568.
N. Guglielmi, V.Yu. Protasov, , Found. Comput. Math., 13 (2013), 37–97.
N. Guglielmi, V.Yu. Protasov, , SIAM J. Matrix Anal. Appl., 37 (2016), 18–52.
B. Han, , SIAM J. Matr. Anal. Appl., 24 (2003), 693–714.
B. Han, R-Q. Jia, , SIAM J. Math. Anal., 29 (1998), 1177–1199.
J. Hutchinson, , Indiana Univ. Math. J, 30 (1981), 713–747.
K. Jetter, G. Plonka, , N. Dyn, D. Leviatan, D. Levin, and A. Pinkus (eds.), [*Multivariate Approximation and Applications*]{}, Cambridge University Press, Cambridge, 2001, 73–111.
R.-Q. Jia, , Math. Comput., 67 (1998), 647–665.
R.-Q. Jia, Q. Jiang, , in: Wavelet Analysis and Applications, Studies in Advanced Mathematics 25, American Mathematical Society, Providence, RI, 2002, 155–178.
G. Kutyniok, T. Sauer, , SIAM J. Math. Anal., 41 (2009), 1436–1471.
H. M. Möller, T. Sauer, , Adv. Comput. Math., 20 (2004), 205–228.
G.C. Rota, G. Strang, , Kon. Nederl. Acad. Wet. Proc., 63 (1960), 379–381.
T. Sauer, , in Curves and Surfaces: 7th International Conference, Avignon, France, June 24 - 30, 2010, Revised Selected Papers, 2012, Springer, Berlin, Heidelberg, 612–628.
| {
"pile_set_name": "ArXiv"
} |
---
bibliography:
- 'refs.bib'
title: Automatically Generating Features for Learning Program Analysis Heuristics
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The DFT/vdW-WF method, recently developed to include the Van der Waals interactions in Density Functional Theory (DFT) using the Maximally Localized Wannier functions, is applied to the study of the adsorption of rare-gas atoms (Ne, Ar, Kr, and Xe) on the Cu(111) and Pb(111) surfaces, at three high-symmetry sites. We evaluate the equilibrium binding energies and distances, and the induced work-function changes and dipole moments. We find that, for Ne, Ar, and Kr on the Cu(111) surface the different adsorption configurations are characterized by very similar binding energies, while the favored adsorption site for Xe on Cu(111) is on top of a Cu atom, in agreement with previous theoretical calculations and experimental findings, and in common with other close-packed metal surfaces. Instead, the favored site is always the hollow one on the Pb(111) surface, which therefore represents an interesting system where the investigation of high-coordination sites is possible. Moreover, the Pb(111) substrate is subject, upon rare-gas adsorption, to a significantly smaller change in the work function (and to a correspondingly smaller induced dipole moment) than Cu(111). The role of the chosen reference DFT functional and of different Van der Waals corrections, and their dependence on different rare-gas adatoms, are also discussed.'
author:
- 'Pier Luigi Silvestrelli, Alberto Ambrosetti, Sonja Grubisiĉ,[@present] and Francesco Ancilotto'
title: 'Adsorption of rare-gas atoms on Cu(111) and Pb(111) surfaces by van der Waals-corrected Density Functional Theory'
---
Introduction
============
Understanding adsorption processes on solid surfaces is essential to design and optimize countless material applications, and to interpret, for instance, scattering experiments and atomic-force microscopy. In particular, the adsorption of rare-gas (RG) atoms on metal surfaces is prototypical[@Bruch] for physisorption processes. Basically, the weak binding of physisorbed closed electron-shell atoms, such as RG atoms, is due to an equilibrium between attractive, long-range van der Waals (vdW) interactions and short-range Pauli repulsion acting between the electronic charge densities of the substrate and the adatoms.[@Vidali]
Up to now RG adsorption on many close-packed metal surfaces, such as Ag(111), Al(111), Cu(111), Pd(111), Pt(111),.. have been extensively studied both experimentally[@Gottlieb; @Seyller; @Narloch; @Diehl] and theoretically,[@Diehl; @Silva; @DaSilva05; @DaSilva; @Betancourt; @Lazic; @Righi; @Sun] while, to our knowledge, Pb has not, but for the experimental measurements of Ferralis [*et al.*]{}[@Ferralis] and, very recently, the theoretical investigation of Zhang [*et al.*]{},[@Zhang] who studied the tribological properties of Ne and Kr on the Pb(111) surface. The Pb surface is important for practical applications: for instance, there is considerable interest in the frictional (tribological) properties of gases on Pb at low temperatures; in particular, Pb is used[@Ferralis; @Zhang; @Bruschi] as a material for the electrodes and as adsorption surfaces in nanofriction experiments because it is easy to grow a very uniform film already at room temperature and to remove the surface contaminants deposited over time on the electrodes, thanks to its large diffusion coefficient. The Pb(111) surface also exhibits interesting and unusual properties: for instance, one striking finding is the drastic difference between the sliding friction of Ne and Kr mono- or multilayers.[@Bruschi; @Zhang]
In principle, due to the non-directional character of the vdW interactions that should be the dominant one in physisorption processes, surface sites that maximize the coordination of the RG adsorbate atom were expected to be the preferred ones, so that it was usually assumed that the adsorbate occupies the maximally coordinated [*hollow*]{} site. This assumption was also based on the expectation that the atom in the [*hollow*]{} site would be closer to the surface, thus experiencing a more attractive potential; behind this is the notion that the repulsive potential at the surface is proportional to the atomic charge density and the natural assumption is that the charge density is highest at the locations of the atoms, thus making the [*top*]{} site energetically unfavored. Calculations where the total adatom-substrate interaction is described by the sum of empirical binary potentials, which are widely used and often give reasonable results for adsorption energies, seem to confirm this expectation since the highly coordinated [*hollow*]{} sites naturally emerge as the preferred adsorption sites for the adatoms. However this picture has been questioned by many experimental[@Gottlieb; @Seyller; @Narloch] and theoretical[@DaSilva05; @DaSilva; @Betancourt; @Lazic] recent studies which indicate that the actual scenario is more complex: in particular, for Xe and Kr a general tendency is found[@Diehl; @DaSilva05; @DaSilva; @Betancourt; @Lazic] for adsorption on metallic surfaces in the low-coordination [*top*]{} sites (this behavior was attributed[@Diehl; @Bagus] to the delocalization of charge density that increases the Pauli repulsion effect at the [*hollow*]{} sites relative to the [*top*]{} site and lifts the potential well upwards both in energy and height); for Ar the situation seems to be less clear:[@DaSilva] for instance, comparison of theoretical and experimental results[@Diehl] would suggest that the [*hollow*]{} sites is still favored for Ar on Ag(111).
The importance of polarization effects to determine the favored adsorption sites was pointed out by Da Silva [*et al.*]{},[@DaSilva] who studied the interaction of RG adatoms with the Pd(111) surface: in fact, for instance, for Xe the polarization is larger in the on-top site, i.e. the larger induced dipole moment increases the attractive interaction between Xe and the metal surface. Therefore, the dominant mechanisms appear to be polarization-induced attraction and site-dependent Pauli repulsion. The latter, being weaker for the on-top site, stabilizes on-top adsorption.[@Silva]
In spite of this recent substantial progress, the understanding of the interaction of RGs with metal surfaces is not complete yet.[@Diehl] It is not clear, for instance, whether a system exists where high-coordinated site are always preferred. Moreover, there have been relatively few studies of adsorption geometries for the smaller RGs, although these are probably better candidates for the observation of high-coordination sites, due to their reduced polarizability with respect to that of Xe or Kr: in fact, the considerable mismatch between the lattice constants of the smaller RGs and that of most metal surfaces cause most commensurate structures to have multiple atoms per unit cell, so that the characterization and interpretation of such systems is quite complex.
Density Functional Theory (DFT) is a well-established computational approach to study the structural and electronic properties of condensed matter systems from first principles, and, in particular, to elucidate complex surface processes such as adsorptions, catalytic reactions, and diffusive motions. Although current density functionals are able to describe quantitatively condensed matter systems at much lower computational cost than other first principles methods, they fail[@Kohn] to properly describe dispersion interactions. Dispersion forces originate from correlated charge oscillations in separate fragments of matter and the most important component is represented by the $R^{-6}$ vdW interaction,[@london] originating from correlated instantaneous dipole fluctuations, which plays a fundamental role in adsorption processes of fragments weakly interacting with a substrate (“physisorbed”).
This is clearly the case for the present systems which can be divided into well separated fragments (RG atoms and the metal substrate) with negligible electron-density overlap. The local or semilocal character of the most commonly employed exchange-correlation functionals makes DFT methods unable to correctly predict binding energies and equilibrium distances within both the local density (LDA) and the generalized gradient (GGA) approximations.[@Riley] As a consequence, the basic results often depend, even at a qualitative level, on the adopted DFT functional: for instance, in their ab initio study of the interaction of RG adatoms with the Pd(111) surface, Da Silva [*et al.*]{}[@DaSilva] found that the on-top site preference is obtained by the LDA for all RG adatoms, while the GGA functionals (in the PBE and PW91 schemes) yield the on-top site preference for Xe, Kr, and He adatoms, but the [*hollow*]{} site for Ne and Ar. Typically, in many physisorbed systems GGAs give only a shallow and flat adsorption well at large atom-substrate separations, while the LDA binding energy turns out to be not far from the experimental adsorption energy; however, since it is well known that LDA tends to overestimate the binding in systems with inhomogeneous electron density (and to underestimate the equilibrium distances), the reasonable performances of LDA must be considered as accidental. Therefore, a theoretical approach beyond the DFT-LDA/GGA framework, that is able to properly describe vdW effects is required to provide more quantitative results.[@DaSilva]
In the last few years a variety of practical methods have been proposed to make DFT calculations able to accurately describe vdW effects (for a recent review, see, for instance, ref. ). We have investigated by such a method the adsorption of RG atoms on the Cu(111) and Pb(111) surfaces. Cu(111) has been chosen because of the many experimental and theoretical data available (especially for Xe-Cu(111)), which can be compared with ours in such a way to validate the present approach; as mentioned above, the less studied Pb(111) surface could be interesting because, given the relatively large Pb lattice constant (and hence nearest-neighbor surface Pb-Pb distance) it represents a good candidate for a system where RG atoms are preferably adsorbed on [*hollow*]{} sites (the lattice constant of Pb is 4.95 Å, compared to 4.09 Å for Ag, 4.05 Å for Al, 3.92 Å for Pt, 3.89 Å for Pd, and 3.61 Å for Cu).
Method
======
In this study we include vdW effects within a standard DFT approach by using the method proposed in refs. (where further details can be found), hereafter referred to as DFT/vdW-WF, by introducing an additional term in the exchange-correlation functional as originally proposed by Andersson [*et al.*]{}[@andersson] to describe the interactions between separate fragments. This contribution, which effectively accounts for the dispersion forces both in the uniform electron gas and separate atom limits, has the form : $$E_{vdW}=-\sum_{n,l}f_{nl}(r_{nl})\frac{C_{6nl}}{r_{nl}^6}
\label{Evdw}$$ with (in a.u.) $$C_{6nl}=\frac{3}{16\pi^{3/2}}\int_{|\mathbf{r'}|<r'_c}d\mathbf{r'}
\int_{|\mathbf{r}|<r_c}d\mathbf{r}
\frac{\sqrt{\rho_n(r)\rho_l(r')}}{\sqrt{\rho_n(r)}+\sqrt{\rho_l(r')}}\,.
\label{vdw}$$ In the above formulas $r_{nl}$ is the distance between the two separate fragments $n$ and $l$, and $\rho_n(r)$ is the $n$-th fragment electronic density. The cutoff $r_c$ is introduced to remove the divergence of the integral, taking into account that, at small momentum values, the interaction is highly damped.[@andersson]
In our approach all the fragment densities are conveniently rewritten in terms of the Maximally Localized Wannier Functions (MLWFs), {$w_n(r)$}, i.e. $\rho_n(r)=w^2_n(r)$. The MLWFs can be obtained from the occupied Kohn-Sham orbitals, generated by a standard DFT calculation, by means of a unitary transformation which minimizes the functional[@wannier] $$\Omega=\sum_n S^2_n=
\sum_n \left( <w_n|r^2|w_n>-<w_n|\mathbf{r}|w_n>^2\right).$$ The unitary transformation conserves the total density, which is however partitioned into single localized fragments, each of them being characterized by its spread $S_n$ and center of mass position $r_n$. It is therefore possible to express the vdW correction (see eqs. and ) as a sum of single contributions coming from each pair of Wannier functions belonging to different fragments, by approximating the shape of the $n$-th Wannier function[@silvmetodo] with an H-like exponential.
The DFT/vdW-WF method has been already successfully applied to several systems, including small molecules, bulk, and surfaces;[@silvprl; @silvsurf; @silvmetodo; @silvhb; @silvinter; @ambrosetti] in particular it allowed us to study the interaction of Ar with graphite and of Ar, He, and H$_2$ with Al surfaces,[@silvsurf; @silvmetodo] of water with the Cl- and H-terminated Si(111) surfaces,[@silvinter] and of RG atoms and water with graphite and graphene.[@ambrosetti]
We here apply the DFT/vdW-WF method to the case of adsorption of Ne, Ar, Kr, and Xe atoms on the Cu(111) and Pb(111) surfaces. All calculations have been performed with the Quantum-ESPRESSO[@ESPRESSO] ab initio package (MLWFs have been generated as a post-processing calculation using the WanT package[@WanT]). Similarly to DaSilva [*et al.*]{},[@DaSilva] we modeled the clean and RG-covered metal surfaces using a periodically-repeated hexagonal supercell, with a $(\sqrt{3}\times \sqrt{3})R30^{\circ}$ structure and a surface slab made of 15 metal (Cu or Pb) atoms distributed over 5 layers (repeated slabs were separated along the direction orthogonal to the surface by a vacuum region of about 24 Å). The Brillouin Zone has been sampled using a $6\times6\times1$ $k$-point mesh. In this model system the RG coverage is 1/3, i.e. one RG adatom for each 3 metal atoms in the topmost surface layer. The $(\sqrt{3}\times \sqrt{3})R30^{\circ}$ structure has been indeed observed at low temperature by LEED for the case of Xe adsorption on Cu(111) and Pd(111)[@Seyller] (actually, this is the simplest commensurate structure for RG monolayers on close-packed metal surfaces and the only one for which good experimental data exist), and it was adopted in most of the previous ab initio studies[@Silva; @DaSilva05; @DaSilva; @Righi; @Lazic; @Zhang]. Since the lateral interactions between RG adatoms do not play a critical role in the RG adsorption site preference,[@DaSilva05] for the sake of simplicity, we have used the same structure also for the other RGs (Ne, Ar, and Kr) and in the case of adsorption on Pb(111) as well.
The Pb or Cu surface atoms were kept frozen (of course after a preliminary relaxation of the outermost layers of the clean metal surfaces) and only the vertical coordinates of the RG atoms, perpendicular to the surface, were optimized, this procedure being justified by the fact that only minor surface atom displacements are observed upon physisorption.[@DaSilva05; @Zhang; @Abad] Moreover, the RG atoms were adsorbed on both sides of the slab: in this way the surface dipole generated by adsorption on the upper surface of the slab is cancelled by the dipole appearing on the lower surface, thus greatly reducing the spurious dipole-dipole interactions between the periodically repeated images (previous DFT-based calculations have shown that these choices are appropriate[@DaSilva; @Sun]). Note that, apparently, in their recent study of Ne and Kr on Pb(111), Zhang [*et al.*]{}[@Zhang] have instead considered adsorption on a single side of the metal slab; the effect of such a choice is non-negligible: in fact, for instance, in the case of Xe on Pb(111), we find that the (absolute value of the) binding energy is reduced by 7 meV (about 4 %) with respect to that obtained when Xe is adsorbed on both sides of the slab. The results of ref. are thus likely affected by the artificial dipole-dipole interactions discussed above.
We have carried out calculations for various separations of the RG atoms adsorbed on high-symmetry sites, namely [*hollow*]{} (on the center of the triangle formed by the 3 surface metal atoms contained in the supercell), [*top*]{} (on the top of a metal atom), and [*bridge*]{} (intermediate between two nearest-neighbor metal atoms). Actually, two kinds of [*hollow*]{} sites are present: HCP [*hollows*]{}, characterized by having atoms directly beneath them in the next layer of atoms, and FCC [*hollows*]{} where this condition does not apply; however the HCP-[*hollow*]{} and the FCC-[*hollow*]{} sites can be considered equivalent for adsorption because of the small differences in the adsorption properties (for instance, Righi and Ferrario,[@Righi] using LDA, found a difference of less than 1 meV in the adsorption energy and of 0.01 Å in the equilibrium distance for RGs adsorbed on Cu(111)). For a better accuracy, as done in previous applications on adsorption processes,[@silvsurf; @silvmetodo; @silvinter; @ambrosetti] we have also included the interactions of the MLWFs of the physisorbed fragments not only with the MLWFs of the underlying surface, within the reference supercell, but also with a sufficient number of periodically-repeated surface MLWFs (in any case, given the $R^{-6}$ decay of the vdW interactions, the convergence with the number of repeated images is rapidly achieved). Electron-ion interactions were described using norm-conserving pseudopotentials: in the case of Pb and Cu we have explicitly included 14 and 11 valence electrons per atom, respectively (those coming from the $5d^{10}$, $6s^2$, $6p^2$ atomic orbitals for Pb, and $3d^{10}$, $4s^1$ for Cu). As a reference DFT functional we chose PW91[@PW91] because it is widely used in ab initio DFT calculations of solids and surfaces and, in particular, was adopted in some previous simulations[@Lazic] of Xe interacting with the Cu(111) surface, which facilitates comparison with the results of the present calculations (note that typically PW91 gives similar results to that obtained by PBE,[@pbe] which represents another popular GGA functional). Using the PW91 functional in test calculations with bulk Pb and Cu, for the equilibrium properties the agreement with experimental estimates is comparable to that found in previous DFT calculations.[@DaSilva05; @Lazic; @Zhang]
By generating the MLWFs for the Cu(111) and Pb(111) substrates, we observe a clear quantitative separation between the spreads of the MLWFs describing $d$-like orbitals and those of the (much more delocalized) MLWFs describing the $s$- and $p$-like orbitals; moreover, given the high valence-electron density, screening effects are certainly relevant in these metal surfaces. Therefore, at variance with previous calculations,[@silvsurf; @silvmetodo; @silvinter; @ambrosetti] we have applied the DFT/vdW-WF correction by explicitly considering only the more localized MLWFs corresponding to the $d$-like orbitals, while the $s$- and $p$-like electrons are supposed to give a screening-effect[@screening] contribution, which is taken into account by multiplying the vdW correction (the C$_6$ coefficients) by a simple Thomas-Fermi factor: $ f_{TF} = e^{-2(z-z_s)/r_{_{TF}}}$ where $r_{_{TF}}$ is the Thomas-Fermi screening length relative to the electronic density of a uniform electron gas (”jellium model”) equal to the average density of the $s$- and $p$-like electrons of the present systems, $z_s$ is the average vertical position of the topmost Cu or Pb atoms, and $z$ is the vertical position, measured with respect to $z_s$, of the adatom. In practice it turns out that only the topmost metal layer gives a relevant contribution, while the effects of the other ones is dramatically reduced by the exponential factor, in line with the common expectation about screening effects in metal surfaces.[@screening] This observation can be exploited to considerably reduce the computational cost of the vdW correction since only the topmost MLWFs must be really taken into account.
Results and Discussion
======================
In Tables I-VI results are reported for all the systems under consideration, for adsorption on [*hollow*]{}, [*top*]{}, and [*bridge*]{} sites. The [*binding energy*]{}, $E_b$, is defined as $$E_b=1/2(E_{tot}-(E_s+2E_{RG}))$$ where $E_{s,RG}$ represent the energies of the isolated fragments (the substrate and the RG atoms) and $E_{tot}$ is the energy of the interacting system, including the vdW-correction term (the factors 2 and 1/2 are due to the adsorption of RG atoms on both sides of the slab).
One should point out that the experimentally measured [*adsorption energy*]{}, $E_a$, includes not only the interaction of RG atoms with the substrate but also lateral, vdW, RG-RG interactions;[@Sun] however in most of previous calculations the mostly attractive lateral interaction contribution was not considered. As pointed out, for instance, by Lee [*et al.*]{},[@Lee] who studied $n$-butane on transition-metal surfaces (another typical weak physisorption system where vdW interaction is the only attractive force between the nonpolar molecule and the substrate) lateral adatom-adatom interaction energies can be as large as 25% of the total adsorption energy at full coverage. $E_a$ is here defined as:
$E_a = E_b + (E_l - E_f) $,
where $E_l$ is the total energy (per atom) of the 2D RG lattice (that is as in the adsorption configurations but without the substrate) and $E_f$ is the energy of an isolated RG atom. Clearly the quantity in parenthesis in the above formula represents the lateral adatom-adatom interaction energy (per atom). Note that, in their DFT study of Ne and Kr on Pb(111), Zhang [*et al.*]{}[@Zhang] seem instead to identify $E_a$ with $E_b$.
$E_b$ has been evaluated for several adsorbate-substrate distances; then the equilibrium distances and the corresponding binding energies have been obtained by fitting the calculated points with the function: $A\,e^{-Bz}-C_3/(z-z_0)^3$, $A$, $B$, $C_3$, and $z_0$ being adjustable parameters (as illustrated for the Xe-Cu(111) and Xe-Pb(111) cases in Figs. 1 and 2). Typical uncertainties in the fit are of the order of $0.05$ Å$ $ for the distances and a few meV for the minimum binding energies. Our results are compared to available theoretical and experimental estimates and to corresponding data obtained using a pure PW91 functional, the simple LDA functional, and the ”seamless” vdW-DF method of Langreth [*et al.*]{}[@Dion] (note that the vdW-DF method has been also used in the recent DFT study of Zhang [*et al.*]{}[@Zhang]). As can be seen in Figs. 1 and 2, and in Tables I and II, the effect of the vdW correction computed by DFT/vdW-WF is a much stronger bonding than with a pure PW91 scheme, with the formation of a clear minimum in the binding energy curve at a shorter equilibrium distance. In spite of the clear shortcomings of the pure PW91 scheme, in general the preferred adsorption site seems to be correctly determined by the latter, although the differences between the binding energies of the different adsorption sites are very small.
We have also computed $E_a$ (assuming a full monolayer coverage of RGs) in the case of Xe on Cu(111), where a RG overlayer in the $(\sqrt{3}\times \sqrt{3})R30^{\circ}$ structure is experimentally found[@Seyller] and in the case of Xe on Pb(111), where the formation of a commensurate Xe monolayer was also observed.[@Ferralis] As can be seen in Table III, all the methods, but pure PW91, correctly predict a smaller $E_a$ on Pb(111) than on Cu(111), although the quantitative results considerably depend on the adopted scheme: in fact, pure PW91 clearly underestimates $E_a$, DFT/vdW-WF and vdW-DF give comparable results, while LDA is close to DFT/vdW-WF and vdW-DF for Xe on Cu(111) but underestimates for Xe on Pb(111): this can be explained by the fact that LDA is not able to describe properly the lateral interactions of Xe adatoms which are further from each other on Pb(111) than on Cu(111).
Concerning the adsorption on the Cu(111) surface (see Table I), all the methods used predict that the [*top*]{} configuration is energetically favored in the case of Xe, while for Ne, Ar, and Kr the differences among the binding energies of the different adsorption configurations are quite small (using vdW-DF the same is true also for Xe); since these differences are probably comparable to the expected accuracy of the calculations, a precise assignment of the favored adsorption site is not possible. In contrast, the [*hollow*]{} configuration is instead clearly favored by all the methods (see Table II) in the case of the adsorption on Pb(111) of all the considered RG atoms (actually, with DFT/vdW-WF, for Ar on Pb(111) the [*bridge*]{} site appears to be lower in energy: however, given the small difference with the energy of the [*hollow*]{} site, this result should not be overemphasized). Our results for Pb(111) are in qualitative agreement with those of Zhang [*et al.*]{}[@Zhang] who predict that Ne and Kr indeed prefer high-coordination [*hollow*]{} sites. Note that the energy difference between the [*hollow*]{} and [*top*]{} sites increases by subsequently considering the PW91, vdW-DF, DFT/vdW-WF, and LDA methods (see also Da Silva [*et al.*]{}[@DaSilva05]).
Interestingly, in the case where several experimental reference values are available, namely Xe on Cu(111), our DFT/vdW-WF method performs better (considering both the binding and adsorption energy, and the equilibrium distance, see Tables I, III, and IV) than all the other schemes: in fact LDA gives reasonable binding energies but underestimates the equilibrium distances, while vdW-DF underestimates the binding energies and overestimates the equilibrium distances, in line with the behavior reported for systems including a metallic surface.[@Vanin] Also note that, at a variance with the experimental findings, the vdW-DF method predicts that the [*top*]{} site (see Table I) is only marginally favored (and the distance only marginally different) than the [*hollow*]{} ones; in general, for all the RG atoms on Cu(111) vdW-DF gives almost identical binding energies for the [*top*]{} and [*hollow*]{} adsorption sites. In the case of RGs on Pb(111) the [*hollow*]{} structure is favored also by vdW-DF, although the difference in the binding energy with respect to the [*top*]{} site is smaller than with the present DFT/vdW-WF scheme (the difference was instead larger, see the last column of Table II, in the study of Zhang [*et al.*]{}[@Zhang], who used vdW-DF but with a reference DFT functional differing from ours by the exchange term). In the case of Ar on Cu(111) and on Pb(111), we observe that our computed binding energies compare favorably with the estimates obtained, using a simple Lennard-Jones potential, by Cheng [*et al.*]{},[@Cheng] who predicted a binding energy between 70 and 85 meV for Ar on noble metals.
As expected, we find that, both for adsorption on Cu(111) and Pb(111), the binding energy increases by going from Ne to Xe, in line with the increasing polarizability of this atom sequence. In particular, for several close-packed transition-metal surfaces the binding energy of Xe is found[@DaSilva] to be about 2 to 3 times larger than that of Kr, and Ar, respectively, a behavior which is well reproduced by our DF/vdW-WF method (the factors are 1.5 and 3, and 1.6 and 2.5, for adsorption on Cu(111) and on Pb(111), respectively). This general behavior is also in line with the results of Zhang [*et al.*]{}[@Zhang]
Our energetic results are not far from the “best estimate” reported by Vidali [*et al.*]{} [@Vidali] for Xe on Cu(111), i.e. a binding energy of -183 $\pm 10$ meV at a distance of 3.60 $\pm 0.08$Å (these values represent averages over different theoretical/experimental estimates). In their tables Vidali [*et al.*]{}[@Vidali] also report for Ar on Cu(111) a binding energy of -85 meV at a distance of 3.53 Å and for Kr on Cu(111) a binding energy of -119 meV, in fair agreement with our results. Lazic [*et al.*]{}[@Lazic] studied the adsorption of Xe on Cu(111) by a DFT approach where vdW corrections were included using the method of Andersson [*et al.*]{},[@andersson] using PW91 and PBE as reference DFT functionals (see the last column in Tables I and IV). As can be seen, our results are much closer to the experimental estimate than those of Lazic [*et al.*]{},[@Lazic] which tend to overestimate the binding energy and underestimate the equilibrium distance. The Xe-adsorbed Cu(111) surface has been also recently investigated by Sun and Yamauchi[@Sun] using DFT with semiempirical vdW corrections: they found reasonable equilibrium distances, however the computed binding energy was very overestimated (it was even larger than that obtained by LDA) and the favored adsorption site was incorrectly predicted to be the [*hollow*]{} site, probably due to the use of semiempirical pair potentials which favor close-packed structures and high coordinated sites (see discussion above).
From Tables I, III, and IV, on can also see that the binding energies are reasonably reproduced by the LDA scheme for RGs on Cu(111), a behavior common to several physisorption systems. However, as already outlined above, this agreement should be considered accidental: the well-known LDA overbinding, due to the overestimate of the long-range part of the exchange contribution, somehow mimics the missing vdW interactions; the equilibrium distances predicted by LDA are clearly underestimated since LDA cannot reproduce the $R^{-6}$ behavior in the interaction potential. For RGs on Pb(111), the LDA binding energies are instead underestimated as a consequence (as discussed above) of the larger equilibrium distances than for RGs on Cu(111).
As already found elsewhere,[@DaSilva05; @DaSilva] for all the used schemes, the binding energies correlate with the RG-metal distance: in fact, for a given RG, the configurations having the strongest binding are characterized by the shortest RG-substrate distance. Moreover, all the methods predict that Ar and Xe adatoms get closer to the Cu(111) surface when adsorbed on [*top*]{} site, as found in some previous studies.[@DaSilva05; @DaSilva; @Righi] Remarkably, this behavior cannot be reproduced[@Diehl; @DaSilva] using a hard-sphere model, indicating that there is a significant interaction between the Ar and Xe atoms and the Cu(111) surface so that a simple stacking (hard-sphere) model of weakly or noninteracting spheres is not valid (for comparison, in Tables IV and V we also list the sums of the RG atoms and metal atom vdW literature radii). Instead, for adsorption on Pb(111), the adatoms in the [*hollow*]{} site are closer to the surface than in the [*top*]{} one, in line with the usual behavior. These results can be easily elucidated by analyzing the parameters of the adopted fitting function (see above), $A\,e^{-Bz}-C_3/(z-z_0)^3$ : we find that, as a general rule, at the equilibrium distance, the repulsive potential term is weaker on the favored adsorption site (for instance the [*top*]{} site for Xe on Cu(111) and the [*hollow*]{} one for Xe on Pb(111)), in agreement with the results of Da Silva [*et al.*]{}[@DaSilva05]
Ferralis [*et al.*]{}[@Ferralis] studied the structural and thermal properties of Xe on the Pb(111) surface by LEED. They observed the formation of a Xe monolayer with an incommensurate hexagonal structure with a lattice parameter similar to that found in bulk Xe (4.33 Å); this structure is aligned with the substrate lattice but has a larger unit cell, similarly to the case of Xe on Ag(111), which is also an aligned incommensurate monolayer. They also found that the heat of adsorption for the first Xe layer is -191 $\pm 10$ meV with an, overlayer-substrate spacing of 3.95 $\pm 0.10$ Å. Looking at Table III we found that our computed $E_a$ (-205.5 meV for the [*hollow*]{} adsorption site) is close to the experimental value and in better agreement than with the other methods, although our model structure is not exactly the same observed experimentally; moreover, also the Xe-Pb(111) distance (3.93 Å) is in excellent agreement (see Table V) with that estimated by Ferralis [*et al.*]{},[@Ferralis] which gives further support to the reliability of our DF/vdW-WF method. As expected, it has been found[@Ferralis] that a hard-sphere model is unable to give a good description of adsorption of Xe on Pb(111). For Xe-Pb(111) the heat of adsorption is lower than for Xe on any surface measured so far,[@Ferralis] with the possible exception of Al(110) and for alkali metals; a low heat of adsorption is not particularly surprising since the Pb atoms are much larger than most other metals (the vdW radius of Pb is 2.02 Å, compared to 1.72 Å for Ag, 1.72 Å for Pt, 1.63 Å for Pd, and 1.40 Å for Cu), implying that the repulsive Xe-Pb interaction prevents the Xe from approaching the deeper part of the attractive holding potential. It must be noted that Ferralis [*et al.*]{}[@Ferralis] were unable to determine the preferred adsorption site, the lack of satellite intensities in the LEED patterns indicating that the overlayer is quite uniform and the corrugation is small.
An important quantity which often provides revealing details of the bonding mechanism in adsorption processes is represented by the electron density difference, $\Delta n({\bf r})= n_{RG/s}({\bf r})-n_s({\bf r})-n_{RG}({\bf r})$, obtained from the electron density (at the equilibrium geometry) of the RG on the substrate, of the clean substrate, and the isolated RG monolayer, respectively. Our approach in this respect is not fully self-consistent because we use the electron density obtained at a pure PW91 level, that is without vdW corrections, however, the effects due to the lack of self-consistency are expected to be negligible because the rather weak and diffuse vdW interactions should not substantially change the electronic charge distribution.[@Langreth07] Plots of $\Delta n({\bf r})$ for Xe on Cu(111) and Xe on Pb(111), both in the [*hollow*]{} and [*top*]{} site (see Figs. 3 and 4), show that, in agreement with what found previously[@DaSilva] for RGs on Pd(111), the electron density redistribution is stronger on the Cu atoms for the Xe on the [*top*]{} site than for the [*hollow*]{}; both sites exhibit a significant depletion of electron density centered about the Xe atom together with a slight density accumulation close to the center of the Xe atom, this effect being attributed[@DaSilva] to orthogonalization of Xe states to the states of the substrate atoms. Moreover, for Xe in the on-top site, there is a significant electron density accumulation between the Xe atom and the topmost surface layer. Interestingly, there is a clear tendency of Xe to induce a much larger charge delocalization on the Cu(111) surface than on Pb(111), in line with the delocalization mechanism invoked[@Diehl; @Bagus] to explain the preference for the [*top*]{} adsorption site on Cu(111).
Since polarization effects are assumed to play a key role in determining the favored adsorption sites,[@DaSilva; @Righi] we have also computed the change of the work function, $\Delta W$, of the Cu(111) and Pb(111) substrate upon adsorption of RG atoms. The work functions have been calculated as the difference between the averaged electrostatic Coulomb potential at the midpoint of the vacuum region of the slab and the Fermi energy: [@Binggeli] for the clean Cu(111) and Pb(111) surfaces we estimate a work function of 4.85 and 3.86 eV, respectively, in excellent agreement with the reference values, that are in the range from 4.90 to 5.01 eV[@Fall] for Cu(111), and 3.83 eV for Pb(111).[@Sun08] $\Delta W$ can be related to the dipole moment induced in the substrate by the presence of the RG adatom, $\Delta \mu$, using the Helmholtz equation:[@Schmidt]
$$\Delta \mu = {1 \over {12\pi}} {{A_{(1\times 1)}}\over {\Theta}}
\Delta W\;,
\label{deltamu}$$
where $ A_{(1\times 1)}$ is the area of the $(1\times 1)$ surface unti cell (in Å$^2$) and $\Theta$ is the RG coverage; if $\Delta W$ is given in eV, then $\Delta \mu$ is in debyes. In our case $\Theta = 1/3$, so that $\Delta \mu = \sqrt{3} a_0^2/16\pi \Delta W$, where $a_0$ is the Cu or Pb lattice constant. Our computed $\Delta W$ and $\Delta \mu$ values are listed in Table VI. In agreement with previous ab initio calculations,[@DaSilva; @DaSilva05] we find that the RG adsorption induces a decrease in the work function, thus indicating that the RG atoms behave as adsorbates with an effective positive charge; note that this is consistent with the depletion of the electron density about the Xe atom discussed above, which corresponds to an induced surface dipole moment that points out of the surface. For Xe on Cu(111) our estimated $\Delta W$ and $\Delta \mu$ values (see Table VI) agree well with the experimental estimates[@Zeppenfeld] of -0.60 eV and -0.24 D, respectively. As can be seen in Table VI, the absolute value of $\Delta \mu$ increases from Ne to Xe, because the corresponding electronic polarizabilities increase, and is larger for the optimal adsorption site, for instance the [*top*]{} for Xe on Cu(111) and the [*hollow*]{} for Xe on Pb(111). Moreover, it is considerably larger on Cu(111) than on Pb(111) in line with the energetic analysis reported above, that indicated a stronger interaction of RGs with the Cu(111) surface than with Pb(111).
Zhang [*et al.*]{}[@Zhang] explain the much larger mobility of Ne overlayers on Pb(111), as observed in friction experiments, than of Kr overlayers on the basis of the different activation energies which characterize the lateral motion of Ne and Kr atoms on the Pb(111) surface. The activation energies for a monolayer can be directly calculated from the difference in the binding energy of the adatom between the favored ([*hollow*]{}) site and the transition state, which is expected to correspond to the [*bridge*]{} site.[@Zhang] Considering the differences between the binding energy of the [*hollow*]{} and [*bridge*]{} configurations for Ne and Kr on Pb(111), we qualitatively confirm the trend observed by Zhang [*et al.*]{},[@Zhang] being our estimated activation energies (1.3 meV for Ne and 6.0 meV for Kr) of the same order of magnitude as those reported in ref. (0.7 meV for Ne and 2.5 meV for Kr). However, such small energy values are comparable to (or even smaller than) the expected accuracy of the computed binding energies, thus making quantitative estimates of the hopping probabilities[@Zhang] (which depend exponentially on the aforementioned activation energies) rather questionable.
![Electron density difference of Xe on Cu(111) in (a) [*hollow*]{} and (b) [*top*]{} site shown in a plane perpendicular to the surface, within the range of $\pm 1 \times 10^{-4} e/{\rm bohr}^3$. Red (light grey) and blue (dark grey) represent electron accumulation and depletion, respectively. The green and orange spheres indicate the Xe and Cu atoms, respectively.[]{data-label="fig3"}](Fig3)
[-0.8cm]{}
![Electron density difference of Xe on Pb(111) in (a) [*hollow*]{} and (b) [*top*]{} site shown in a plane perpendicular to the surface, within the range of $\pm 1 \times 10^{-4} e/{\rm bohr}^3$. Red (light grey) and blue (dark grey) represent electron accumulation and depletion, respectively. The green and grey spheres indicate the Xe and Pb atoms, respectively.[]{data-label="fig4"}](Fig4)
[-0.8cm]{}
Conclusions
===========
In summary, by analyzing the results of our study of the adsorption of RG atoms on the Cu(111) and Pb(111) surfaces, one can conclude that the inclusion of the vdW corrections by the DFT/vdW-WF method systematically improves upon the estimates for the binding energies as obtained by a standard GGA approach. In particular, using a pure PW91 functional the binding is underestimated in all cases, while equilibrium distances are overestimated. For all the system considered the vdW correction term represents the dominant part of the binding energy, although, particularly for RG adsorption on Pb(111), the pure PW91 approach gives a substantial contribution. However, vdW interactions appear not to play a critical role in the adsorption site preference (the same result has been obtained by Zhang [*et al.*]{}[@Zhang] studying the interaction of Ne and Kr on Pb(111)): Xe on Cu(111) clearly prefers the [*top*]{} site, while for Ne, Ar, an Kr on Cu(111) the differences in binding energies relative to different adsorption sites are so small that is not easy to attribute a definitive preference; instead, the [*hollow*]{} configuration tends to be preferred for adsorption of all the considered RGs on Pb(111), in agreement with previous calculations and experimental observations.[@Ferralis; @Zhang] Moreover, the Pb(111) substrate is subject, upon rare-gas adsorption, to a significantly smaller change in the work function, and to a correspondingly smaller (in absolute value) induced dipole moment, than Cu(111). Given these relevant peculiarities of the Pb(111) surface, where the [*hollow*]{} site is undoubtedly favored for adsorption of RG atoms, this surface would represent an ideal substrate to study, both theoretically and experimentally, high-coordination adsorption sites.
Acknowledgements
================
We thank very much F. Costanzo and F. Toigo for useful discussions.
[9]{} Present Address: Center for Chemistry, IHTM, University of Belgrade, Njegoševa 12, 11001, Belgrade, Serbia, and Scuola Normale Superiore, piazza dei Cavalieri 7, I-56126 Pisa, Italy. L. W. Bruch, M. W. Cole, and E. Zaremba, [*Physical Adsorption: Forces and Phenomena*]{} (Clarendon Press, Oxford, 1997). G. Vidali, G. Ihm, H. Y. Kim, M. W. Cole, Surf. Sci. Rep. [**12**]{}, 133 (1991). J. M. Gottlieb, Phys. Rev. B [**42**]{}, 5377 (1990). Th. Seyller, M. Caragiu, R. D. Diehl, P. Kaukasoina, M. Lindroos, Chem. Phys. Lett. [**291**]{}, 567 (1998); M. Caragiu, Th. Seyller, R. D. Diehl, Phys. Rev. B [**66**]{}, 195411 (2002). B. Narloch, D. Menzel, Chem. Phys. Lett. [**290**]{}, 163 (1997). R. D. Diehl, Th. Seyller, M. Caragiu, G. S. Leatherman, N. Ferralis, K. Pussi, P. Kaukasoina, M. Lindroos, J. Phys.: Condens. Matter [**16**]{}, S2839 (2004). J. L. F. Da Silva, C. Stampfl, M. Scheffler, Phys. Rev. Lett. [**90**]{}, 066104 (2003). J. L. F. Da Silva, C. Stampfl, M. Scheffler, Phys. Rev. B [**72**]{}, 075424 (2005). J. L. F. Da Silva, C. Stampfl, Phys. Rev. B [**77**]{}, 045401 (2008). A. E. Betancourt, D. M. Bird, J. Phys.: Condens. Matter [**12**]{}, 7077 (2000). P. Lazic, Z. Crljen, R. Brako, B. Gumhalter, Phys. Rev. B [**72**]{}, 245407 (2005). M. C. Righi, M. Ferrario, J. Phys.: Condens. Matter [**19**]{}, 305008 (2007). X. Sun, Y. Yamauchi, J. Appl. Phys. [**110**]{}, 103701 (2011). N. Ferralis, H. I. Li, K. J. Hanna, J. Stevens, H. Shin, F. M. Pan, R. D. Diehl, J. Phys.: Condens. Matter [**19**]{}, 056011 (2007). Y. N. Zhang, F. Hanke, V. Bortolani, M. Persson, R. Q. Wu, Phys. Rev. Lett. [**106**]{}, 236103 (2011). L. Bruschi, M Pierno, G. Fois, F. Ancilotto, G. Mistura, C. Boragno, F. Buatier de Mongeot, U. Valbusa, Phys. Rev. B [**81**]{}, 115419 (2010). P. S. Bagus, V. Staemmler, C. Wöll, Phys. Rev. Lett. [**89**]{}, 096104 (2002). See, for instance, W. Kohn, Y. Meir, D. E. Makarov, Phys. Rev. Lett. [**80**]{}, 4153 (1998). R. Eisenhitz, F. London, Z. Phys. [**60**]{}, 491 (1930). K. E. Riley, M. Pitoňák, P. Jurečka, P. Hobza, Chem. Rev. [**110**]{}, 5023 (2010). P. L. Silvestrelli, Phys. Rev. Lett [**100**]{}, 053002 (2008). P. L. Silvestrelli, K. Benyahia, S. Grubisiĉ, F. Ancilotto, F. Toigo, J. Chem. Phys. [**130**]{}, 074702 (2009). P. L. Silvestrelli, J. Phys. Chem. A [**113**]{}, 5224 (2009). Y. Andersson, D. C. Langreth, B. I. Lundqvist, Phys. Rev. Lett. [**76**]{}, 102 (1996). N. Marzari, D. Vanderbilt, Phys. Rev. B [**56**]{}, 12847 (1997). P. L. Silvestrelli, Chem. Phys. Lett. [**475**]{}, 285 (2009). P. L. Silvestrelli, F. Toigo, F. Ancilotto, J. Phys. Chem. C [**113**]{}, 17124 (2009). A. Ambrosetti, P.L. Silvestrelli, J. Phys. Chem. C [**115**]{}, 3695 (2011). S. Baroni [*et al.*]{}, www.quantum-espresso.org . WanT code by A. Ferretti [*et al.*]{}, www.wannier-transport.org . E. Abad, Y. J. Dappe, J. I. Martínez, F. Flores, J. Ortega, J. Chem. Phys. [**134**]{}, 044701 (2011). J. P Perdew, Y. Wang, Phys. Rev. B [**45**]{}, 13244 (1992). J. P. Perdew, K. Burke, M. Ernzerhof, Phys. Rev. Lett. [**77**]{}, 3865 (1996). R. R. Rehr, E. Zaremba, W. Kohn, Phys. Rev. B [**12**]{}, 2062 (1975); J. Tao, J. P. Perdew, A. Ruzsinszky, Phys. Rev. B [**81**]{}, 233102 (2010). K. Lee, Y. Morikawa, D. C. Langreth, Phys. Rev. B [**82**]{}, 155461 (2010). M. Dion, H. Rydberg, E. Schröder, D. C. Langreth, B. I. Lundqvist, Phys. Rev. Lett. [**92**]{}, 246401 (2004); G. Roman-Perez, J. M. Soler, Phys. Rev. Lett. [**103**]{}, 096102 (2009). M. Vanin, J. J. Mortensen, A. K. Kelkkanen, J. M. Garcia-Lastra, K. S. Thygesen, K. W. Jacobsen, Phys. Rev. B [**81**]{}, 081408 (2010). E. Cheng, M. W. Cole, W. F. Saam, J. Treiner, Phys. Rev. B [**48**]{}, 18214 (1993). T. Thonhauser, V. R. Cooper, S. Li, A. Puzder, P. Hyldgaard, D. C. Langreth, Phys. Rev. B [**76**]{}, 125112 (2007). C. J. Fall, N. Binggeli, A. Baldereschi, J. Phys.: Condens. Matter [**11**]{}, 2689 (1999). C. J. Fall, N. Binggeli, A. Baldereschi, Phys. Rev. B [**61**]{}, 8489 (2000). B. Sun, P. Zhang, Z. Wang, S. Duan, X.-G. Zhao, X. Ma, Q.-K. Xue, Phys. Rev. B [**78**]{}, 035421 (2008). L. D. Schmidt, R. Gomer, J. Chem. Phys. [**45**]{}, 1605 (1966). P. Zeppenfeld, [*Physics of Covered Solid Surfaces*]{}, Börnstein, New Series, Group III, Vol. 42: Numerical data and functional relationships in science and technology, Subvol. A: Adsorbed layers on surfaces (Springer-Verlag, Berlin, 2001), p. 67.
------------------------- ------- ------------ -------- -------- -----------------------------------------
system PW91 DFT/vdW-WF LDA vdW-DF ref.
Ne-Cu(111) [*hollow*]{} -17.6 -31.6 -55.7 -56.1 —
Ne-Cu(111) [*top*]{} -17.5 -31.1 -55.4 -55.9 —
Ne-Cu(111) [*bridge*]{} -17.6 -31.0 -55.3 -56.1 —
Ar-Cu(111) [*hollow*]{} -13.0 -67.8 -88.9 -106.6 —
Ar-Cu(111) [*top*]{} -13.0 -71.9 -94.5 -106.3 -85$^a$
Ar-Cu(111) [*bridge*]{} -13.0 -70.6 -89.4 -106.4 —
Kr-Cu(111) [*hollow*]{} -20.3 -134.2 -117.6 -135.7 —
Kr-Cu(111) [*top*]{} -20.3 -131.1 -126.0 -135.8 -119$^a$
Kr-Cu(111) [*bridge*]{} -20.3 -130.0 -118.4 -135.7 —
Xe-Cu(111) [*hollow*]{} -22.9 -194.5 -199.3 -167.4 -276$^b$, -268$^c$
Xe-Cu(111) [*top*]{} -23.1 -208.1 -221.9 -167.7 -280$^b$, -183$^a$, -277$^c$ (-190$^c$)
Xe-Cu(111) [*bridge*]{} -17.1 -191.2 -201.0 -167.4 -278$^b$
------------------------- ------- ------------ -------- -------- -----------------------------------------
: Binding energy, $E_b$ in meV, of RG atoms on the Cu(111) surface computed using the standard DFT-PW91 calculation, and including the vdW corrections using our DFT/vdW-WF method, compared to the LDA result, the vdW-DF method by Langreth [*et al.*]{}[@Dion] and available theoretical and experimental (in parenthesis) reference data.
\[table1\]
------------------------- ------- ------------ -------- -------- ------------
system PW91 DFT/vdW-WF LDA vdW-DF ref.
Ne-Pb(111) [*hollow*]{} -31.2 -59.8 -49.4 -71.4 -51.6$^a$
Ne-Pb(111) [*top*]{} -27.8 -49.1 -42.9 -63.3 -46.8$^a$
Ne-Pb(111) [*bridge*]{} -19.8 -58.5 -49.1 -64.6 —
Ar-Pb(111) [*hollow*]{} -23.5 -82.4 -78.3 -100.8 —
Ar-Pb(111) [*top*]{} -22.1 -75.0 -64.2 -95.3 —
Ar-Pb(111) [*bridge*]{} -22.7 -84.5 -76.6 -100.1 —
Kr-Pb(111) [*hollow*]{} -30.8 -132.8 -98.8 -136.9 -134.9$^a$
Kr-Pb(111) [*top*]{} -29.1 -109.8 -81.6 -130.9 -125.1$^a$
Kr-Pb(111) [*bridge*]{} -24.0 -126.8 -96.7 -136.1 —
Xe-Pb(111) [*hollow*]{} -59.6 -193.5 -142.0 -192.2 -172.6$^a$
Xe-Pb(111) [*top*]{} -56.3 -186.4 -116.1 -186.4 —
Xe-Pb(111) [*bridge*]{} -52.7 -188.9 -138.6 -191.2 —
------------------------- ------- ------------ -------- -------- ------------
: Binding energy, $E_b$ in meV, of RG atoms on the Pb(111) surface computed using the standard DFT-PW91 calculation, and including the vdW corrections using our DFT/vdW-WF method, compared to the LDA result, the vdW-DF method by Langreth [*et al.*]{}[@Dion] and available theoretical and experimental (in parenthesis) reference data.
\[table2\]
------------------------- ------- ------------ -------- -------- ------------
system PW91 DFT/vdW-WF LDA vdW-DF ref.
Xe-Cu(111) [*hollow*]{} -51.4 -289.3 -297.3 -268.9 —
Xe-Cu(111) [*top*]{} -51.6 -302.9 -319.9 -269.2 (-227$^a$)
Xe-Cu(111) [*bridge*]{} -45.6 -286.0 -299.0 -268.9 —
Xe-Pb(111) [*hollow*]{} -62.5 -205.5 -147.9 -252.2 (-191$^a$)
Xe-Pb(111) [*top*]{} -59.2 -198.4 -122.0 -246.4 —
Xe-Pb(111) [*bridge*]{} -55.6 -200.9 -146.9 -251.2 —
------------------------- ------- ------------ -------- -------- ------------
: Adsorption energy ($E_a$, see text for the definition), in meV, of Xe atoms on the Cu(111) and Pb(111) surfaces computed using the standard DFT-PW91 calculation, and including the vdW corrections using our DFT/vdW-WF method, compared to the LDA result, the vdW-DF method by Langreth [*et al.*]{}[@Dion] and available experimental (in parenthesis) reference data.
\[table3\]
------------------------- ------ ------------ ------ -------- ---------------------------------------- ------
system PW91 DFT/vdW-WF LDA vdW-DF ref. $s$
Ne-Cu(111) [*hollow*]{} 3.90 3.59 3.10 3.70 — 2.94
Ne-Cu(111) [*top*]{} 3.90 3.57 3.09 3.68 — 2.94
Ne-Cu(111) [*bridge*]{} 3.90 3.60 3.10 3.68 — 2.94
Ar-Cu(111) [*hollow*]{} 4.50 3.48 3.19 3.90 — 3.28
Ar-Cu(111) [*top*]{} 4.50 3.45 3.15 3.86 3.53$^a$ 3.28
Ar-Cu(111) [*bridge*]{} 4.50 3.43 3.19 3.86 — 3.28
Kr-Cu(111) [*hollow*]{} 4.50 3.32 3.21 3.99 — 3.42
Kr-Cu(111) [*top*]{} 4.50 3.36 3.17 3.99 — 3.42
Kr-Cu(111) [*bridge*]{} 4.50 3.35 3.20 3.99 — 3.42
Xe-Cu(111) [*hollow*]{} 4.70 3.42 3.00 4.10 3.40$^b$, 3.31$^c$ 3.56
Xe-Cu(111) [*top*]{} 4.40 3.36 2.90 4.09 3.45$^b$, 3.2$^d$, 3.25$^c$ (3.60$^e$) 3.56
Xe-Cu(111) [*bridge*]{} 4.70 3.41 3.00 4.10 — 3.56
------------------------- ------ ------------ ------ -------- ---------------------------------------- ------
: Equilibrium RG adatom-surface distance, in Å, on the Cu(111) surface computed using the standard DFT-PW91 calculation, and including the vdW corrections using our DFT/vdW-WF method, compared to the LDA result, the vdW-DF method by Langreth [*et al.*]{}[@Dion] and available theoretical and experimental (in parenthesis) reference data; the sum, $s$, of the vdW radii of the RG atom and the Cu atom is also reported.
\[table4\]
------------------------- ------ ------------ ------ -------- ------------ ------
system PW91 DFT/vdW-WF LDA vdW-DF ref. $s$
Ne-Pb(111) [*hollow*]{} 3.80 3.41 3.10 3.70 3.5$^a$ 3.56
Ne-Pb(111) [*top*]{} 4.00 3.68 3.40 3.90 3.8$^a$ 3.56
Ne-Pb(111) [*bridge*]{} 3.80 3.36 3.27 3.50 — 3.56
Ar-Pb(111) [*hollow*]{} 4.40 3.68 3.40 4.00 — 3.90
Ar-Pb(111) [*top*]{} 4.40 4.04 3.60 4.22 — 3.90
Ar-Pb(111) [*bridge*]{} 4.50 3.77 3.43 4.10 — 3.90
Kr-Pb(111) [*hollow*]{} 4.40 3.69 3.40 4.14 3.8$^a$ 4.04
Kr-Pb(111) [*top*]{} 4.40 3.98 3.70 4.24 3.9$^a$ 4.04
Kr-Pb(111) [*bridge*]{} 4.30 3.79 3.51 4.13 — 4.04
Xe-Pb(111) [*hollow*]{} 4.30 3.93 3.50 4.30 (3.95$^b$) 4.18
Xe-Pb(111) [*top*]{} 4.50 4.02 3.70 4.30 — 4.18
Xe-Pb(111) [*bridge*]{} 4.70 3.93 3.55 4.31 — 4.18
------------------------- ------ ------------ ------ -------- ------------ ------
: Equilibrium RG adatom-surface distance, in Å, on the Pb(111) surface computed using the standard DFT-PW91 calculation, and including the vdW corrections using our DFT/vdW-WF method, compared to the LDA result, the vdW-DF method by Langreth [*et al.*]{}[@Dion] compared to the LDA result, and available theoretical and and available theoretical and experimental (in parenthesis) reference data; the sum, $s$, of the vdW radii of the RG atom and the Pb atom is also reported.
\[table5\]
------------ --------------- ---------------
system [*hollow*]{} [*top*]{}
Ne-Cu(111) -0.04 (-0.02) -0.03 (-0.01)
Ar-Cu(111) -0.28 (-0.13) -0.37 (-0.17)
Kr-Cu(111) -0.54 (-0.24) -0.37 (-0.17)
Xe-Cu(111) -0.53 (-0.24) -0.57 (-0.26)
Ne-Pb(111) -0.03 (-0.03) -0.03 (-0.03)
Ar-Pb(111) -0.10 (-0.08) -0.03 (-0.03)
Kr-Pb(111) -0.11 (-0.09) -0.05 (-0.04)
Xe-Pb(111) -0.13 (-0.11) -0.04 (-0.03)
------------ --------------- ---------------
: Work-function change, in eV, and induced dipole moment (in parenthesis), in debyes, for RGs adatoms on the Cu(111) and Pb(111) surfaces, at equilibrium geometries.
\[table6\]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'One of the most unusual military aircraft programs V / STOL was the Avro VZ-9 “Avrocar”. Designed to be a real flying saucer, the Avrocar was one of the few V / STOL to be developed in complete secrecy. Despite significant changes in the design, during flight tests, the Avrocar was unable to achieve its objectives, and the program was eventually canceled after an expenditure of 10 million US dollars between 1954 and 1961. But the concept of a lift fan, driven by a turbojet engine is not dead, and lives today as a key component of Lockheed X-35 Joint Strike Fighter contender. Was held in a data research and information related to Avrocar project carried out during the Second World War, which was directly linked to advances in aircraft that were built after it, and correlate them with the turbo fan engines used today.'
author:
- |
**Desire Francine G. Fedrigo\***\
Panoramic Residence, Rua Luísa, 388s, ap. 05,\
Vila Portuguesa, Tangará da Serra/MT, 78300-000, Brasil\
**Ricardo Gobato**\
Secretaria de Estado da Educação do Paraná (SEED/PR),\
Av. Maringá, 290, Jardim Dom Bosco,\
Londrina/PR, 86060-000, Brasil\
**Alekssander Gobato**\
Faculdade Pitágoras Londrina,\
Rua Edwy Taques de Araújo, 1100,\
Gleba Palhano, Londrina/PR, 86047-500, Brasil\
\***\*Corresponding author**: [email protected]
bibliography:
- 'journals.bib'
title: '[Avrocar: a real flying saucer]{}'
---
**Keywords**: [AVRO Project, A. V. Roe Canada, Flying Saucer, Lockheed X-35, Military aircraft, Turbo Fan Engines, US Air Force, US Army.]{}
Introduction
============
One of the most unusual military aircraft programs V / STOL was the Avro VZ-9 “Avrocar" (Figure 1). Designed to be a real flying saucer, the Avrocar was one of the few V / STOL to be developed in complete secrecy. Despite significant changes in the design, during flight tests, the Avrocar was unable to achieve its objectives, and the program was eventually canceled after an expenditure of 10 million US dollars between 1954 and 1961. [@2015]
![[]{data-label="fig:sweeps"}](Figure1)
Material and methods
--------------------
Based on books, documents available in digital media and in public libraries and using conventional qualitative research methods descriptive nature the work is presented. Noting also of the 40s books, related to the project Avrocar and their respective manufacturing, development of aircraft and engines that followed the same turbo fan concept prototype after World War II.
Project development Avrocar
===========================
In 1952, a project team led by Jack Frost, the Avro Aircraft, Canada, began work on the design of an aircraft V / TOL supersonic with a circular wing. The Canadian Defense Research Board funded the project, with a contract of \$ 400,000. The Capacity V / TOL was to be done by the air ducts of the fan and the engine exhaust to the periphery of the platform, diverting the flow of air down. Near the ground, this provides an effect of “air mattress", where the weight exceeds the pressure due to increased pressure on the underside of the aircraft. This phenomenon was confirmed in a wind tunnel test. In transit to forward flight, the airflow was gradually redistributed back. Frost was convinced that a circular wing thin disk, was the ideal format to take advantage of both the effect of “air cushion" close to the ground (for STOL) and to achieve supersonic speeds.
![. [@Gobato2011][]{data-label="fig:sweeps"}](Figure2)
In 1954, the Canadian government abandoned the project to be very expensive, but the progress that had been made was enough to interest the United States Air Force (USAF). Concern about the vulnerability of air bases on the front line of Europe, the Cold War, has increased the interest of the Air Force V / TOL aircraft. For three quarters of a million dollars, the contract was signed by the Air Force in 1955 for further study. In 1956, the Avro was sufficiently satisfied with the results to release 2.5 million dollars to build a prototype of the search plane. In March 1957, the Air Force approved additional funding, and the aircraft became, officially, the “System 606A Arms". These efforts remained highly classified as top secret until July 1960. One of the most promising proposals was the 606A (Figure 2) that would have a circular thin wing 35 meters in diameter, with a maximum weight of 27,000 pounds and a speed of more than Mach 1.4. A large turbofan engine was driven by the exhaust gas flow turbojets Armstrong Siddeley six Viper. Numerous wind tunnel testing in both the Avro as the USAF test base at Wright Field, Ohio, were conducted and a test platform, full-scale, the propulsion system was built. The Avrocar 59-4975 after modifications, was tested without the canopies and incorporating the perimeter “focusing" ring c. 1961 (Figure 3). Tests showed that the heat was so oppressive that all instruments were baked brown after only a few flights (Figure 3).
![[]{data-label="fig:sweeps"}](Figure3)
In 1958, the Avro made a series of presentations to the Army and United States Air Force, after which the Avro began a project of an aircraft for the US military, which has been given the official designation VZ-9. It was baptized with the name of Avrocar (Figure 4 and Figure 5).
![[]{data-label="fig:sweeps"}](Figure4)
![[]{data-label="fig:sweeps"}](Figure4a)
The Avrocar (Figures 7-9) should be a circular platform aircraft, which to present both hovered flight and horizontal flight at high speed, with capacity V / TOL. The Army was interested in the survival of troops on the battlefield and in order to improve its air capacity, they were studying alternatives to their aircraft and helicopters then in service. The Air Force supported the Avrocar program because it would demonstrate many of the features of the 606 project, in less time and at a much lower cost. A two million dollar contract, to be managed by the Air Force was made in the Avro to build and test an Avrocar. The initial performance requirements for Avrocar were: be able to sustain ten minutes flight in ground effect and carry a load of 1000 pounds, 25 miles away. The work began in earnest, and a contract 1.77 million dollars was granted to build a second copy of Avrocar in March 1959. The first Avrocar left the factory in May 1959. At launch, the projected performance was very beyond the initial requirement, and reach a top speed of 225 kt, ceiling 10,000 feet, 130 miles of range with 1,000 kg payload, and hover out of ground effect, with 2428 kg of payload. The maximum take off weight, with transition to forward flight, out of ground effect, it was estimated at 5,650 pounds and the maximum weight, with a transition in ground effect, it would be 6,970 pounds. The Avrocar was about 18 meters in diameter, three meters of disc thickness, and two separate cockpits. The cockpit was located on the left front side of the aircraft, with another crew member on the right. A third compartment on the back was due to charge storage. The Avrocar would be raised by the flow of a turbo-fan engine five meters core diameter, called exhaust turbo-rotor, composed of three Continental J-69 turbojet with 920 pounds of thrust each, whose air flow was channeled to the outer edge of the turbo rotor. It had 124 small straws. Each engine was connected to its own fuel tanks and oil. The fuel tanks were not connected, although it had planned for a later version.
The main pilot control consisted of a side-stick, which provided control pitching and side when moved forwards and backwards or sideways. The yaw control was obtained twisting the side-stick. This action controlled low flows and high pressure around the aircraft, causing it to rotate around the vertical axis. In forward flight, the Avrocar was statically unstable, with a pressure center far ahead of the center of gravity. An automatic stabilization system was employed, using the gyroscopic action of the turbo-rotor. The turbo-rotor, in turn, was not rigidly attached to the vehicle, but mounted on a kind of hinge that allowed him freedom of movement. Control cables are set on the basis of the turbo rotor, to enable its control. The interest of the Army in Avrocar program was great. One of the authors surveyed (Lindenbaum) recalls a trip that made Washington in the late 1950s, to request additional funding for a study on reducing the aerodynamic drag of helicopter Bell UH-1. Although funding had been approved, he heard a note of an army general, that the Bell UH-1 Huey helicopter would be the last that the army would buy, since the helicopter would be replaced by Avrocar. From June to October 1959, the first Avrocar was tested on a static platform, hovering. The hot gases re-circulated the turbo-rotor reduced the thrust. Excessive losses in the duct system also became apparent, and these defects have never been solved, despite major changes in the design. The maximum load reached, out of ground effect, was 3,150 pounds. With a zero fuel weight (ZFW) of 4285 pounds, the Avrocar was therefore unable to hover out of ground effect. Following these tests, the vehicle was sent to NASA for assessing wind tunnel Ames Research Center. The second Avrocar left the factory in August 1959. On September 29, the first attempt of sustained flight was made with the Avrocar pinned to the ground by cables. After that the vehicle took off, an uncontrollable oscillation occurred with each wheel alternately bouncing on the ground. The pilot immediately shut down all engines. Subsequently, several alternative schemes, these “captive" flights have been tested and numerous changes were made to the springs of spoilers and control of rotor shaft base. These first captive flight revealed then a new problem, called “hubcapping" which has never been fully resolved. The hubcapping was rapid and unpredictable swings in pitch and roll axes. It resulted in an unstable air mattress if the vehicle exceeds a critical height (Figure 6). [@Force2009; @Campagn1998; @DesireFrancineGobato2011; @Stevens2003]
![[]{data-label="fig:sweeps"}](Figure5)
The critical point was found about two meters from the ground. Inputs the controls were ineffective in dampening the oscillation. Fifty-two holes were drilled at the vehicle bottom, located radially and three meters from the center. These were used to provide a flow of air attempting to center and stabilize the air mattress. Such a device never reached the expected success.
![[]{data-label="fig:sweeps"}](Figure6)
![[]{data-label="fig:sweeps"}](Figure6a)
![[]{data-label="fig:sweeps"}](Figure6b)
The first completely free flight occurred on November 12, 1959, and the control system based on adjustable nozzle air outlet, proved unacceptable. After five flights, the test was temporarily interrupted in December 5, 1959, when the Avrocar had recorded 18.5 hours of testing in captive and free flights. A new control system, centered on a set in a ring, was installed later in December. The tip opening to the upper surface was covered, and spoilers were replaced by a flat ring the underside of the vehicle. Lateral changes in position of the ring increased the weight on one side of the vehicle while reducing weight of the opposite side. Flight tests resumed in January 1960, with this system. The trial flight of the Air Force was held on April 4, 1960, with Major Walter Hodgson controls. The maximum speed reached was 30 Kts, and above this speed, an uncontrollable oscillation manifested. The cockpit was cramped, noisy and became unbearably hot during a flight of 15 minutes. Later that month, a trial was conducted in a wind tunnel at NASA’s Ames Research Center in Moffet Field. This test discovered that the control ring centered system provided sufficient buoyancy to allow a flight out of the ground effect, but large angles of attack were needed to generate aerodynamic lift. In late April, however, the Avrocar the initial program ended. Shortly after the program was declassified by the USAF HQ. The Avro was convinced that the concept was still viable, and proposed a new program to rework the main failures of propulsion and control system (Figure 10).
![[]{data-label="fig:sweeps"}](Figure7)
The USAF made a new contract for the period July 1960 to July 1961 for the modification and testing of both vehicles. A new control nozzle is installed at the rear of the vehicle.
A second test in the wind tunnel with the new configuration was done at NASA Ames RC in April 1961. It has been found that an adequate control was available for the transition to a speed of about 100 Kts, and the flight was possible at this speed. However, the vehicle was still unstable. It was hoped that the change of flow over the rear of the vehicle lift the nose up, reducing the instability. Unfortunately, this was not the case. A tail “T" has been added (Figure 11), but it has proved totally ineffective. NASA believes that this failure resulted from the fact that the tail was in a region of “downwash" too high caused by the propulsion system. In any case, it became clear that the Avrocar as configured, could not sustain flight high speed.
![[]{data-label="fig:sweeps"}](Figure8)
On June 9, 1961, the final evaluation and second flight of the Air Force Avrocar was conducted in Avro facilities. During these tests, the vehicle reached a top speed of 20 Kts and showed the ability to cross a ditch six meters wide and 18 centimeters deep. Fly above the critical time was impossible. The flight test report summarized a number of control problems. For example, a large asymmetry in directional control was present. Five seconds were required to turn the aircraft 90 degrees to the left, while eleven seconds were required for a right turn of 90 degrees.
![[]{data-label="fig:sweeps"}](Figure9)
The Avro submitted proposals for radical change in vehicle to address the key problems. Frost (team member) has developed two new models, one with a large vertical tail and another with winglets vertical (Figure 12). Both models would use two turbojets GE-85 J 2700 pounds of thrust, rather than the three original J-69 turbines and the turbo rotor diameter would be increased from five to six feet. The proposals were rejected, and the program was officially closed in December 1961. The second Avrocar had recorded about 75 hours of flight. The Figure 13, shows the Avrocar at the National Museum of the United States Air Force in Dayton, Ohio. [@Gobato2011; @Force2009; @Campagn2003; @Stevens2003; @Gobato2011a; @Zuk2002; @Zuk2006]
![[]{data-label="fig:sweeps"}](Figure12)
Discussions
===========
The concept of ground effect produced by a fan at takeoff and landing did not die with the Avrocar. In 1963, Bell Aerospace initiated studies of a landing system for air mattress (ACLS), which was later patented. These studies were directed by Desmond T. Conde, former chief of aerodynamics for Avrocar. A ACLS replace conventional landing gear with a large internal rubber tube-like structure which surrounds a region of higher air pressure. In August 1967, the concept was proved by Bell, with successful tests on a LA-4 (amphibian). The development was funded by the Air Force Flight Dynamics Laboratory, and a much larger system is designed to test on a Fairchild C-119 (weight of 64,000 lb).
The Brazilian government joined the program and a De Havilland C-115 Buffalo aircraft weighing 41,000 pounds, was selected for further testing. With the XC-8A designation, this aircraft flew with the ACLS in March 1975. The ACLS was considered but rejected as an option for the program from a STOL aircraft / Midfielder Forward, a program that ended up producing the Boeing YC-prototypes 14 and McDonnell Douglas YC-15. The latter was adopted and evolved into the Boeing C-17 transport, which went into series production. The concept of a lift fan, driven by a turbojet engine is not dead, and lives today as a key component of Lockheed X-35 Joint Strike Fighter contender. While Avrocar was in development, Peter Kappus, General Electric has developed independently by a booster fan propulsion system, which has evolved to Ryan VZ GE-11 (later XV-5) “Vertifan". This vehicle, discussed in the previous two editions of Vertiflite magazine (March / April 1990 March / April 1996), paved the way for further study of the “fans elevators", or supportive fans, as the study of supersonic fighters sponsored by DARPA which included both fans driven gas (McDonnell Douglas) and mechanically driven shaft (Lockheed).
Conclusions
===========
Through research carried out during the work it was found that despite the possible failure Avrocar the project was responsible for the breakthrough in the aeronautical environment and their van turbo engines, in which the project was considered essential to the development and improvement of existing fan motors. It is hoped that with this work can be broken the view that many drivers have about the project, which is considered as a bad investment of time since it is not succeeded in the course of its development but it sure your project was in history and served as a study base for future projects. The concept of ground effect produced by a fan at takeoff and landing did not die with the Avrocar. In 1963, Bell Aerospace initiated studies of a landing system for air mattress (ACLS), which was later patented. These studies were directed by the former head of Avrocar project, Desmond T. Conde. The Avrocar also ended up producing the prototype Boeing YC-14 and McDonnell Douglas YC-15. The latter was adopted and evolved into the Boeing C-17 transport, which went into series production. This concept lives today as a key component of Lockheed X-35 Joint Strike Fighter contender. While Avrocar was in development, Peter Kappus, General Electric has developed independently by a booster fan propulsion system, which has evolved to Ryan VZ GE-11 (later XV-5) “Vertifan".
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We report on spectral and intensity variability analysis from a [[*Chandra*]{}]{}/ACIS-S observation of the transient, type-I X-ray bursting low-mass X-ray binary Cen X-4. The quiescent X-ray spectrum during this observation is statistically identical to one observed previously with [[[*BeppoSAX*]{}]{}]{}, and close, but not identical, to one observed previously with [[*ASCA*]{}]{}. The X-ray spectrum is best described as a pure Hydrogen atmosphere thermal spectrum plus a power-law component that dominates the spectrum above 2 keV. The best-fit radius of the neutron star is $r=$12.92.6 $(d/1.2 \; {\rm kpc})$ km if the interstellar absorption is fixed at the value implied by the optical reddening. Allowing the interstellar absorption to be a free parameter yields $r=$19 $(d/1.2 {\rm kpc})$ km (90% confidence). The thermal spectrum from the neutron star surface is inconsistent with a solar metallicity. We find a 3$\sigma$ upper-limit of root-mean-square variability $\leq 18\%$ (0.2-2.0 keV; 0.0001-1 Hz) during the observation. On the other hand, the 0.5-10.0 keV luminosity decreased by 408% in the 4.9 years between the [[*ASCA*]{}]{} and [[*Chandra*]{}]{}observations. This variability can be attributed to the power-law component. Moreover, we limit the variation in thermal temperature to 10% over these 4.9 years. The stability of the thermal temperature and emission area radius supports the interpretation that the quiescent thermal emission is due to the hot neutron star core.'
author:
- |
Robert E. Rutledge, Lars Bildsten, Edward F. Brown, George G. Pavlov,\
and Vyacheslav E. Zavlin
title: 'The Quiescent X-Ray Spectrum of the Neutron Star in 4 Observed with [[*Chandra*]{}]{}/ACIS-S'
---
internalcite
citex\[\#1\]\#2[@fileswauxout citeacite[forciteb:=\#2]{}[\#1]{}]{}
cite\#1\#2[citeleft\#1@tempswa , \#2citeright]{} biblabel\#1
1608[[4U 1608$-$522]{}]{} 4[[Cen X$-$4]{}]{} 1808[[SAX J1808.4$-$3658]{}]{}
Introduction
============
Measuring both the mass ($M$) and the radius ($R$) of a neutron star (NS) would strongly constrain the nuclear equation of state. To distinguish between the competing models, several measurements of a few percent accuracy in $M$ and $R$ are required. While the masses of several NSs have been constrained to 10% or better (see for a recent review) from pulsar timing of Doppler shifts, measuring the NS radii has proven to be difficult. There are at least five ways to measure $R$ using X-ray emission from accreting NSs: (1) the spectral evolution of radius-expansion type-I X-ray bursts; (2) the measurements of $M/R$ from the gravitational red-shift of metallic spectral lines during type I X-ray bursts or $\gamma$-ray lines during accretion; (3) inferring constraints from NS kHz quasi-periodic oscillations; (4) pressure broadening and red-shift of photospheric metal lines; and (5) spectral analysis of transiently accreting NSs in quiescence.
Efforts to measure the NS radius through the spectral evolution of radius-expansion type-I X-ray bursts (for a review, see ) based on theoretical non-Planckian spectra [@jvp82; @london86; @pavlov91; @madej91; @tit94] have been somewhat successful. Performing this measurement with the emission from a type-I X-ray burst has the advantage that the observed luminosity originates from the NS photosphere and not the surrounding accretion disk. The measured NS radii range from 6 to 15 km. The reliability of these measurements is limited by systematic uncertainties in the emergent spectrum, the fraction of the NS surface area involved in the burst and the elemental composition of the photosphere, as well as the distance to the NS [@jvp87b; @damen89; @damen90].
Attempts to measure NS photospheric metal lines as expected from model atmospheres [@foster87] have had mixed success. Absorption lines observed in the tails of type-I X-ray bursts [@waki84; @nakamura88; @magnier89] have not been confirmed through repeated observation with more sensitive instrumentation, resulting in metallicity limits of $Z<0.4$[$Z_\odot$]{} [@day92]. Moreover, the interpretation of these lines as due to absorption in the photosphere is in doubt [@madej90]. Repeated observations of such lines, and their identification with a known transition would measure $M/R$. Gamma-ray lines from the accretion process are also a possibility [@bildsten92; @bildsten93], though the levels of emission are still below the ability of current instruments.
Observations of kilo-Hz quasi-periodic oscillations from accreting neutron stars, interpreted as the orbital period at or above the marginally stable orbit [@kluzniak90; @kaaret97; @kluzniak98] have been used to constrain both $M$ and $R$ using observations from eight accreting neutron stars, finding $M\approx$2[$M_\odot$]{} [@zhang97b]. The NS radius is then less than the marginally stable orbit for a mass of this size ($<$18 km for a non-rotating star).
Another method of measuring $R$ for a slowly rotating NS is the approach described by Paerels [-@paerels97], with metal lines from an emergent NS X-ray spectrum. Measuring the pressure broadening ($\propto M/R^2$) and photospheric red-shift ($\propto M/R$) yields both $M$ and $R$. This method is intriguing because it is [*independent of distance*]{}, which can be uncertain by up to a factor of 2 in the isolated transient NSs, although it can be determined to $\approx$5% for the low-luminosity X-ray sources in globular clusters [@hertz83; @verbunt84]. Once several metal lines are identified in the photospheric spectrum, this method is very likely to provide a number of $M$ and $R$ measurements.
To these approaches has recently been added a new means of measuring $R$ with X-ray spectroscopic observations of transiently accreting, low magnetic-field ($<$ G) NSs in quiescence [@brown98; @rutledge99; @rutledge00]. Until this work, the spectral observations had been interpreted as black-body radiation [@jvp87; @garcia94; @verbunt95; @asai96b; @asai98; @campana98a; @garcia99; @campana00], yielding effective temperatures of $kT_{\infty}$=0.2-0.3 keV. For most sources, the thermal component has a black-body emission area radius much smaller than a NS (${\mbox{$\lesssim$}}1$ km).
As described by Brown, Bildsten & Rutledge [-@brown98 BBR98 hereafter], a black-body spectrum is not appropriate for weakly magnetic[^1] transiently accreting NSs with $kT_{\rm
eff}<5{\mbox{$\times 10^{6}$}} K$. At accretion rates $\ll 10^{12} {\rm g \ s^{-1}}$ the accreting metals gravitationally settle faster than they are supplied, and the atmosphere is nearly pure hydrogen [@bildsten92]. The dominant opacity is free-free absorption ($\propto \nu^{-3}$) which results in a spectrum in which higher energy photons escape from deeper in the NS atmosphere, where the temperature is higher [@pavlov78; @rajagopal96; @zavlin96]. This hardens the spectrum which, (mis-)interpreted as a black-body, results in systematically higher temperatures and [*lower*]{} emission area radii. When the [[*ASCA*]{}]{} and [[*ROSAT*]{}]{} observations are fit with the more appropriate thermal H atmosphere spectrum, the emission area radii are $\approx$10 km [@rutledge99; @rutledge00], confirming that this emission is mostly thermal. These objects have thus become the focus of our efforts to measure the NS surface area.
If the emission area radius of the thermal part of the quiescent X-ray spectrum is the NS radius, it should be constant from observation to observation. In addition, if accretion onto the NS makes no contribution to the quiescent luminosity, the temperature should also be constant on timescales $\ll$ yrs, the core cooling timescale (see for exceptions). In the most sensitive observations, an additional power-law component is observed above a few keV [@asai96b; @asai98; @campana98a; @campana00]. If produced through interaction between the accretion and magnetosphere [@campana98a], it might vary on short ($\sim$msec) timescales, and perhaps show a pulsation at the NS spin frequency.
In this paper, we present X-ray spectroscopic analysis of 4 in quiescence. 4 is a transient, type-I X-ray bursting low-mass X-ray binary (for reviews on transients, see ). Its distance is estimated to be 1.20.3 kpc on the basis of two observed radius-expansion bursts [@chev89]; formally, this is a distance upper-limit. It is in a $P_{\rm orb}= 0.629$ day binary with a K dwarf [@chev89; @mcclintock90], and has been observed in outburst twice (1969 and 1979; see discussion in § \[sec:recur\]). We compare the observed X-ray spectrum with a number of models, in particular with the H atmosphere model. We also compare the spectrum with spectra of the same source in archived observations. We place limits on changes in the thermal spectrum of the NS on year long timescales. In §\[sec:anal\], we present the analysis of the [[*Chandra*]{}]{} data. In §\[sec:multi\], we compare the [[*Chandra*]{}]{} data with past observations, and characterize the spectral differences between this and past X-ray observations. We discuss these results and their implications for this and other sources and conclude in §\[sec:discuss\].
Observation and Analysis {#sec:anal}
========================
The observation occurred 23 Jun 2000 01:23:01-04:32:54 TT for a total exposure time of 9561.4 seconds with [[*Chandra*]{}]{}/ACIS-S3 (backside illuminated). The source position was aimed 4 off-axis, with a 1/8 subarray used with 0.4 sec exposures. The time between successive frames was 0.44104 sec, which gives $\approx$ 10% dead-time. We analyzed data from the pre-processed Level 2 FITS data provided with the standard data products. The X-ray source appears at the known optical position of Cen X-4 (1 ; ). Only one other point source (previously unknown) was found in the field with the [[*Chandra*]{}]{} Interactive Analysis of Observations Software (CIAO) point-source detection tool [*celldetect*]{} [^2], with a S/N of 3.5, roughly 1 away.
We extracted data from a circle 10 pixels in radius about the 4X-ray source position, and background from an annulus with inner and outer radius of 13 and 50 pixels, respectively. There were a total of 2714 good counts in the source region, and 575 in the background region; we expect 25 background counts in the source region ($\approx$1% of the total counts in the source region).
[[*Chandra*]{}]{} Spectral Analysis
-----------------------------------
The temperature of the focal-plane instruments has been decreased during [[*Chandra*]{}]{}’s lifetime; this alters the energy response of the ACIS-S chips. The data were taken with a focal-plane chip temperature of $-120{\rm C}$, and we used the corresponding response files for this temperature, according to the standard ACIS-S analysis.
We binned the photons between 0.5 and 1.5 keV into bins of width $\approx$130 eV wide, comparable to the energy resolution (FWHM) in the BI-S3 detector. Above 1.5 keV, we binned the data so that there were 40 counts per bin (wider than the spectral resolution). We used spectral data in the energy range 0.5-10 keV. While there are significant counts below 0.5 keV, the ACIS-S energy calibration is presently not reliable below this energy; we will re-examine this analysis when the energy calibration is refined below 0.5 keV. In addition, the current [[*Chandra*]{}]{} response files for the ACIS-S-BI underestimate the detector area by up to 20% near 0.5 keV (N. Schulz, priv. comm.). To account for the calibration uncertainties, we included a 25% systematic uncertainty in the 0.5-0.6 keV energy range, and a 5% systematic uncertainty in 0.6-0.7 keV energy range. This is in addition to a 4% systematic uncertainty across all energies that accounts for other calibration uncertainties.
We fit the data in XSPEC v11 [@xspec], with several spectral models (powerlaw, H atmosphere, blackbody, Raymond-Smith, or thermal bremsstrahlung), all with galactic absorption () as a free parameter, except where noted. The H atmosphere spectrum is not a standard XSPEC model, but has been described and calculated elsewhere [@zavlin96]. The hydrogen column density for 4 was estimated during quiescence to be $<$0.20 (; = [[cm]{}]{}); its optical reddening is $E(B-V)=0.1$ [@blair84]. These values are consistent with an approximate optical reddening/hydrogen column density ratio found from observing the halos of X-ray sources [@gorenstein75; @predehl95], which imply an equivalent hydrogen column density of =0.055.
No single component model fit the data acceptably (prob $\ll
10^{-6}$). In the black-body and H atmosphere fits, the high-energy ($>$2 keV) powerlaw spectral component reported previously [@asai96b; @campana00] is apparent. In general, the single component models failed due to the presence of this high-energy component. We then fit the data with two-component models, where the second component is a power-law that accounts for the $>$2 keV emission. The best-fit absorbed H atmosphere spectrum has parameters listed in Table \[tab:chandra\]. We also provide the best fit parameters with held fixed at the optically implied value (=0.055), and the resulting intrinsic (that is, unabsorbed) spectrum in Figure \[fig:chandraspec\].
While a double power-law model was statistically acceptable, the low energy component has a steep power-law slope (photon index $\alpha$=6.2, 90% confidence; all uncertainties and upper-limits are 90% confidence unless otherwise stated) and considerably higher absorption than observed previously from this source or implied by optical observations (=0.50). We reject the model on this basis.
A Raymond-Smith plasma model [@raymondsmith] with an underlying power law is also statistically acceptable, however, with an abundance limited to $Z<$1.4$Z_{\rm sol}$. [^3] In addition, it has been shown previously that the X-ray-to-optical flux ratio is substantially greater than is typical for stellar coronal sources [@bildsten00]. We therefore reject a model which ascribes this emission to the corona of the stellar companion. Though not physically motivated, a thermal bremsstrahlung spectrum acceptably fits the data, with a volume emission measure of $\int n_e\; n_I dV = 2.3{\mbox{$\times 10^{56}$}} (d/1.2 {\rm kpc})$ [[cm]{}]{} and a temperature of 0.34 keV.
To enable comparison with previous work, we also fit the data with an absorbed blackbody and powerlaw. The best fit with a fixed =0.26 [@campana00] is rejected (=3.1/10 dof; prob=5). With all parameters free, we find $<$0.071 (90%); [$kT_{\rm eff}$]{}=0.175; r=1.1 $(d/1.2 \; {\rm
kpc})$ km; $\alpha=$1.2 (=0.75/9 dof; prob=0.67).
Solar Metallicity Atmosphere
----------------------------
We fit the [[*Chandra*]{}]{} data with a solar metallicity atmospheric model for a 10 km, 1.4[$M_\odot$]{} NS, with relative abundances given by Grevesse & Noels [-@grevesse93]. A galactic absorption and solar metallicity atmosphere model failed to fit the spectrum (=18/11 dof), due to the high energy ($>$2 keV) excess attributed above to the power-law component, and to spectral curvature below 1 keV. With an additional power-law component, the best-fit spectrum is still unacceptable (prob = 3), largely due to the Fe L edge near 0.534 keV (for a 1.4[$M_\odot$]{}, 10 km NS). While the model atmospheres assume no rotation and are static, the major discrepancy between the model and data is this absorption edge, which will not be significantly altered by these effects.
A solar metallicity NS atmosphere spectrum is therefore strongly rejected. More detailed investigation regarding sub-solar metallicity spectra is in progress, and will be presented in forthcoming work. It would be highly useful to have response matrices refined down to 0.3 keV for this work. These will be valuable limits, as accretion at a rate high enough to explain the quiescent thermal emission would enrich the atmospheric metal content to a detectable level (BBR98).
Comparison between the [[*Chandra*]{}]{}, [[[*BeppoSAX*]{}]{}]{} and [[*ASCA*]{}]{} X-ray spectra {#sec:multi}
-------------------------------------------------------------------------------------------------
4 has been detected on 6 occasions in quiescence, with [*Einstein*]{} and [*EXOSAT*]{} [@jvp87], twice with [[*ASCA*]{}]{}[@asai96b; @asai98], once with [[*ROSAT*]{}]{}/HRI [@campana97] and once with BeppoSAX [@campana00]. The measured luminosities and blackbody spectral parameters from the published observations are listed in Table \[tab:prevwork\].
In this section, we compare the [[*Chandra*]{}]{} X-ray spectrum with that previously obtained with [[[*BeppoSAX*]{}]{}]{} [@campana00] and with the [[*ASCA*]{}]{}/SIS+GIS spectrum [@asai96a; @asai98; @rutledge99]. We do not compare with one other existing [[*ASCA*]{}]{}/GIS observation as it was performed largely with the GIS, with lower S/N than the first [[*ASCA*]{}]{} observation, and is comparable to the first [[*ASCA*]{}]{} observation [@asai98]. We have previously re-analyzed the [[*ASCA*]{}]{} observation [@rutledge99] and use the same resulting spectrum here.
### Re-Analysis of [[[*BeppoSAX*]{}]{}]{} observation {#sec:sax}
Campana (2000; C00 hereafter) interpreted the quiescent X-ray spectrum (measured with [[*ASCA*]{}]{} and [[[*BeppoSAX*]{}]{}]{}) of 4 as a black-body, and found a mean blackbody radius 3.1 km for the thermal component (the best-fit blackbody radius with [[[*BeppoSAX*]{}]{}]{} data alone was r=10 km). However, was held fixed at the best-fit value of =0.26, which is higher than the optically implied value. We undertook a reanalysis of the [[[*BeppoSAX*]{}]{}]{} observation of 4, previously analyzed by C00. There are two major differences between our spectral analysis of the LECS data (0.3-2 keV) and theirs:
1. For the LECS background data, C00 used the standard “blank fields” background which are taken from several “blank” areas of the sky [@parmar99]. However, 4 is located in a region of the sky with higher than average low-energy background ($l=332.24\deg$, $b=+23.88\deg$), according to the ROSAT/All-Sky-Survey Soft X-ray Background maps ($\approx$0.75 keV; , Fig 6b). Thus, the “blank fields” background underestimates soft X-ray background in the 4 observation. We used an alternative method of background estimation described by Parmar ( the “annulus” method; ), in which the background intensity is assumed to be proportional to the countrate in an annulus in the LECS detector, with a spectrum identical to that found with the “blank fields” method. By this method, the 0.3-1.0 keV background counts account for 60% of the total counts in the 4 source region, whereas they only account for 34% in the “blank fields” method; the counting uncertainty in the source region is 10%, and therefore this difference in background countrate is significant. As the “annulus” method scales with the local background countrate, it is a more realistic estimation of the local low-energy ($<$2 keV) background, and so we adopt it.
2. For the LECS data, we used a larger extraction radius than C00 (8 vs. 4); the LECS point-spread-function becomes broad toward the lower energies (80% encircled energy at 1.5 keV is 3.0, and at 0.28 keV is 6.1; the 95% radius is 5.5 and 8.5 at 1.5 and 0.28 keV, respectively). The smaller extraction radius used by C00 may have biased the resulting upwards. This collected 647 counts from the LECS in 21524 sec (much greater than the 233 counts found by C00). In the 0.3-1.0 keV, there are 84 counts in the source region. We estimate that using an 8 radius instead of a 4 radius increases the observed source countrate by 30%, which is larger than the Poisson uncertainty of 10%.
For the MECS data, we used a 4 extraction radius, which extracted 651 counts (comparable to the 632 found by C00). We used the standard MECS background file (MECS23\_bkg.evt) extracting background counts using the same region as for the source fields. The source background subtracted countrates are (5.31.6) c/s (0.1-3.1 keV) and (3.30.4) c/s (1.7-9.0 keV) in the LECS and MECS instruments, respectively. The LECS data has lower S/N than reported by C00 due to the higher background; we find a slightly higher MECS countrate than found previously.
For comparison with previous work, we fit the [[[*BeppoSAX*]{}]{}]{} LECS+MECS data with an absorbed black-body plus power-law spectrum. We obtained an acceptable fit (=0.65/1 dof; prob=0.42), with the following best-fit parameters: $<$0.95, $\alpha$=1.6, [$kT_{\rm eff}$]{}=0.10 keV. We were unable to put any reasonable limits on the BB radius ($>$0.4 km, unbounded from above), as the best-fit spectrum merely would increase the column density to compensate for larger and larger areas. When we hold the column density fixed at the best-fit value, we find $r_{BB}$=7 km; when we hold the column density fixed at the optically implied value (=0.055), we find $r_{BB}$=2 km. These values are comparable to those found with [[*ASCA*]{}]{} [@asai96b; @rutledge99], and below those found by C00. We attribute this difference to our more accurate background subtraction and larger extraction radius for the LECS data. The unabsorbed total flux is 2.7 (4.7 , 0.5-10.0 keV); the unabsorbed flux of the BB component is 2.2 (0.5-10.0 keV), and of the power-law component it is 0.44 (0.5-10.0 keV). As we find in the next section, the [[[*BeppoSAX*]{}]{}]{} and [[*Chandra*]{}]{} spectra (and their corresponding fluxes) are, within statistics, the same.
### Joint Spectral Fitting of [[*Chandra*]{}]{}, [[[*BeppoSAX*]{}]{}]{} and [[*ASCA*]{}]{} X-ray Spectra
We performed a joint spectral fit of the [[*Chandra*]{}]{}, [[*ASCA*]{}]{}, and [[[*BeppoSAX*]{}]{}]{}data. We assumed a 4% systematic uncertainty in the detector responses for all instruments, and held fixed at its optically implied value. We used an absorbed H atmosphere+powerlaw spectrum. The best spectral fit rejects a single model to account for all three observed spectra (=3.3/53 dof; prob=5). A joint [[*Chandra*]{}]{}/[[[*BeppoSAX*]{}]{}]{} fit provided a statistically acceptable fit, indicating that the [[*Chandra*]{}]{} and [[[*BeppoSAX*]{}]{}]{} spectra are statistically identical. A joint [[*Chandra*]{}]{}/[[*ASCA*]{}]{} fit was not acceptable. The [[*ASCA*]{}]{} data are, by themselves, acceptably fit by this same assumed spectrum [@rutledge99].
We jointly fit the [[*Chandra*]{}]{} and [[*ASCA*]{}]{} spectra, permitting one of the five parameters (, $r$, $\alpha$, and power-law normalization ) to be different between the two spectra, each in turn. None of these fits are formally statistically acceptable. However, while a change in $\alpha$ is soundly rejected (prob$\leq$4), and a change in $r$ is of low probability (prob =0.001), changes in the or are found to be marginal (both prob = 0.01). This is in the range where systematic uncertainty in the detector responses (in particular, the area as a function of energy) become important. For example, if we increase the systematic uncertainty from 4% to 8% (which, for example, could be due to a systematic offset in the absolute flux calibrations between the two detectors), the best fits for a changing or become statistically acceptable.
Thus, while the [[*Chandra*]{}]{} and [[*ASCA*]{}]{} spectra are significantly different, we cannot unequivocally state whether this difference is due to a change in the or the , a combination of these with other parameters, or a systematic difference between the absolute calibrations of the two instruments. However, we provide the best-fit spectral parameters in Table \[tab:jointfit\], for a changing power-law normalization and thermal temperature. These then serve as an upper-limit for variability in these parameters between the two observations. Between the [[*Chandra*]{}]{} and [[*ASCA*]{}]{} epochs, then, we find that the thermal temperature decreased (at most) from =0.085 to 0.077 keV; or the power-law normalization decreased from 15.2 to 5.9 phot [[cm]{}]{} [[s]{}]{} at 1 keV. Moreover, the interpretation of these spectra impacts the best-fit power-law slope, which is $\alpha$=1.0 if it is the temperature which changes, or $\alpha$=1.7 if it is the normalization which changes.
Finally, we fit a spectral model to the [[*Chandra*]{}]{}, [[[*BeppoSAX*]{}]{}]{}, and [[*ASCA*]{}]{}data assuming that the power-law slope and flux changes, but that the thermal component does not. The resulting best-fit parameters and uncertainties are shown in Table \[tab:3fit\]. The best-fit is statistically acceptable. The power-law component (taken to the be the same for the [[*Chandra*]{}]{} and [[[*BeppoSAX*]{}]{}]{} observations) is flatter in the [[*Chandra*]{}]{}+[[[*BeppoSAX*]{}]{}]{}observations than during the [[*ASCA*]{}]{} observation (1.00.4 vs. 1.90.3), and has a lower flux, by about a factor of two.
We also fit these spectra assuming that the thermal $r$ and vary while the power-law component does not (although this is not theoretically motivated). An acceptable fit is found (=1.33/52 dof; prob=0.06), with [[*Chandra*]{}]{}+[[[*BeppoSAX*]{}]{}]{}values of ($r$=13.4 km , =0.074 keV) and [[*ASCA*]{}]{} values of ($r$=7.1 km, =0.10keV).
Intensity Variability
---------------------
Variability in the luminosity of 4 has been observed over timescales from days to years. Van Paradijs [-@jvp87 JVP87 hereafter] detected 4 using [*EXOSAT*]{}/LE1 with the CMA instrument and LEXAN 3000 Å filter [@taylor81], with a countrate of (5.41.2) c/s. JVP87 assumed a thermal bremsstrahlung spectrum with temperature between 1 and 5 keV and a column density =0.066, which, for this countrate, corresponded to a flux at earth of (2.2-5.6) ${\rm erg \ s^{-1} \ cm^{-2}}$ . Using the same assumptions with an earlier [*Einstein*]{}/IPC observation, they found a flux at earth between (1.1-1.5) ${\rm erg \ s^{-1} \
cm^{-2}}$, concluding that the X-ray luminosity must have increased between the Einstein and [*EXOSAT* ]{} observations by 50-500%. Campana [-@campana97] measured a flux variation of a factor of 3 in $<$4 days using [[*ROSAT*]{}]{}/HRI.
We started by looking for variability during our observation. We first separated the data into two energy bins: pulse-invariant (PI) bins 2-136 (0.2-2 keV) and PI bins $>$136 ($>$2 keV). We included PI bins in which the area and energy calibration is presently not reliable; while these are not useful (presently) for flux and spectral calculations, they can be used for investigations in variability.
There were a total of $N_{\rm phot}=2553$ counts in PI bins 2-136 ($\approx$0.2-2 keV). We used 23903 time bins which were 0.44104 sec in length which each had exposure of only 0.40 sec, for a total observation time of $T_{\rm obs}=10542$ sec. We performed a Fourier transform, producing a power density spectrum (PDS) of 11951 frequency bins across the 0.0001-1.13 Hz frequency range, normalized according to Leahy [-@leahy83]. We logarithmically rebinned this data, and fit the PDS with functions $P(\nu)$, to extract the integrated r.m.s power. The PDS is acceptably described as counting noise (constant power of 2.0; =1.41/16 dof; prob=0.13); we therefore observe no variability in this data. Using a power-law distribution with the slope of $\alpha=1$ held fixed, with an underlying Poisson level ($P=2$) also held fixed, we find a 3$\sigma$ upper limit on the root-mean-square variability of $<$18% (0.0001-1 Hz). For a flat power-law ($\alpha=0$), the 3$\sigma$ upper limit is $<$10%.
There are a total of 151 counts in PI bins $>$136 (2.0 keV) in this observation. We performed an identical PDS analysis as for the low energy counts, although in the fit to the PDS we held the exponent fixed at $\alpha$=1.0 (producing an acceptable fit; =1.46/15 dof). The 3$\sigma$ upper-limit to the RMS variability $>$2.0 keV is $<$50% (0.0001-1 Hz).
We also looked for longer term variability by investigating the hypothesis that the [[*Chandra*]{}]{} and [[*ASCA*]{}]{} spectra were identical in all parameters, but different in absolute normalization. We found that the spectra were describable in this way (=1.05/47 dof; prob=0.37), with the [[*ASCA*]{}]{} spectrum a factor 1.650.12 (90% uncertainty) more luminous than the [[*Chandra*]{}]{}spectrum. We interpret this as fitting a well-constrained spectrum from [[*Chandra*]{}]{} (signal-to-noise ratio S/N$\approx$100) to the low S/N data of [[*ASCA*]{}]{} (S/N$\approx$4). This provides a rough measure of the luminosity difference between the [[*ASCA*]{}]{} and [[*Chandra*]{}]{} observations – a decrease of 408% over 4.9 yr.
Summary, Conclusions and Implications {#sec:discuss}
=====================================
\[sec:con\]
We have analyzed the X-ray spectrum and intensity variability of 4 observed with [[*Chandra*]{}]{}, and compared it with earlier observations by [[*ASCA*]{}]{} and [[[*BeppoSAX*]{}]{}]{}. The [[*Chandra*]{}]{} X-ray spectrum is inconsistent with all single component models we applied, requiring at least a two component model, with a power-law component that dominates the spectrum above 2 keV and a softer (most likely thermal) component which dominates the spectrum below this energy. The intensity variability during the [[*Chandra*]{}]{} observation is $<$18% (3$\sigma$) on timescales between 1-10,000 sec in the 0.2-2.0 keV energy band dominated by the thermal component. The limit on variability is weaker above 2 keV ($<$50% r.m.s., 3$\sigma$).
Cen X-4’s luminosity decreased between the [[*ASCA*]{}]{} and [[*Chandra*]{}]{}observations (4.9 yrs) by 408% (0.5-10.0 keV). We attribute this variation to changes in the power-law component. This is comparable to the factor of $\sim$3 decrease found over a few days by Campana [-@campana97]. We cannot statistically exclude that the power-law component remains constant, and the thermal component varies. Thus, 4 is variable in quiescence on timescales longer than hours.
The Thermal Component
---------------------
Rutledge [-@rutledge99; -@rutledge00] have shown that the soft emission seen in all NS transients in quiescence is best explained as thermal emission from a pure hydrogen NS atmosphere. We have found the same in this [[*Chandra*]{}]{} observation. Indeed, comparing X-ray spectra taken with [[*Chandra*]{}]{}, [[*ASCA*]{}]{} and [[[*BeppoSAX*]{}]{}]{} at different times over 4.9 years, we find that this thermal component is consistent with being constant, with a best-fit radius of $r$=12.92.6 $(d/1.2 {\rm
kpc})$ km, and a temperature =0.0760.007 keV. These results improve the precision of the quiescent spectrum by a factor of $\sim$3 over previous work [@rutledge99]. We place an upper-limit on the amount of temperature variability across 4.9 years of ${\mbox{$\lesssim$}}$10% (Table \[tab:jointfit\]). The agreement between this radius and that measured during Type I X-ray bursts from other systems is strong confirmation that the emitting area is the NS surface.
Two energy sources for the thermal emission have been discussed: accretion in quiescence at a low-rate [@jvp87; @menou99] and re-emission of heat deposited into the crust during the large accretion events (BBR98). The required accretion rate for the bolometric luminosity of $7\times 10^{32} \ {\rm erg \ s^{-1}}$ is $\dot M_q\approx 6\times 10^{-14} M_\odot \; {\rm yr^{-1}}$, adequate to keep the metal content in the atmosphere comparable or larger than that in the accreting material [@bildsten92]. If the metal content in the accreting material is very sub-solar (that is, effectively pure hydrogen), accretion at these rates would give thermal emission much like we observe [@zampieri95]. However, solar metallicity accretion is likely ruled out, largely due to the lack of absorption from the Fe L edge near 0.5 keV. Tighter constraints on the photospheric metallicity will be the subject of future work with these data and will allow us to more thoroughly constrain the active accretion hypothesis, which does not have a specific way of predicting $\dot M_q$.
The other possible mechanism is re-emission of heat deposited in the inner crust due to pycnonuclear reactions, electron captures and neutron emissions [@haensel90] during accretion events (BBR98). The layers where this heat is deposited are in close thermal contact with the NS core. The reactions then heat the core to a temperature of $\sim$ K over $\sim$ yr (BBR98; ). The NS reaches an equilibrium where the time averaged nuclear heating equals the quiescent thermal emission, $L_q\approx (1 \ {\rm MeV}/{\rm nucleon}) \langle \dot M \rangle$, implying a time averaged accretion rate for this NS of $\langle \dot
M\rangle\approx 10^{-11}M_\odot \; { \rm yr^{-1}}$.
\[sec:recur\] We can compare this $\langle \dot M\rangle$ to that implied by the outbursts. The outburst in 1969 had a total fluence of $\approx 3 \;
{\rm ergs \ cm^{-2}}$ [@chen97]. There has not been another large outburst of this magnitude recorded since; the small outburst in 1979 had a fluence two orders of magnitude smaller. It is uncertain to measure the time-averaged accretion rate on the basis of a single outburst. If a large outburst like that in 1969 occurs every 100 years, the time-averaged accretion flux would be $\approx 10^{-9} \;
{\rm ergs \ cm^{-2} \ s^{-1}}$, which at the 1.2 kpc distance gives $\langle \dot M \rangle \approx 10^{-11} M_\odot {\rm yr^{-1}}$, comparable to that implied by the level of the quiescent thermal emission. However, this is a low accretion rate for a binary at this long orbital period, where the likely driver of mass transfer is nuclear evolution of the binary (see ). So, this is an outstanding puzzle for this source. If outbursts like that seen in 1969 occur more frequently than 100 years, the time averaged accretion rate would be higher and in conflict with that implied by the quiescent thermal emission through the mechanism of BBR98, unless the fraction of deposited energy re-emitted as photons is much less than unity. Possible solutions to this, involving enhanced neutrino emission from the core, have been proposed by Colpi [-@colpi00b] and Ushomirsky & Rutledge [-@rutledge01].
If the thermal emission is due to a hot neutron star core, then the measured effective temperature of $0.076\pm 0.007{\mbox{$\rm\,keV$}}$ tells us the internal NS temperature. For example, if the outer layer consists of light elements to a density $6\times10^{5}{\mbox{$\rm\,g\,cm^{-3}$}}$, then using the fit of Potekhin [-@potekhin97], we find [$T_{\rm core}$]{}$ =
(5.1{\mbox{$^{+ 1.2}_{- 1.1}$}}){\mbox{$\times 10^{7}$}}{\mbox{$\rm\,K$}}$. For these core temperatures, the modified URCA neutrino luminosity is orders of magnitude less than the photon luminosity. The NS core would not be in a thermal steady-state if pion condensation occurs, as the resultant neutrino luminosity [@umeda94] would be much larger than the heating supplied by reactions in the deep crust. As a result, if some enhanced cooling mechanism were operating in the core, then we would expect [$T_{\rm core}$]{}$\lesssim 10^7{\mbox{$\rm\,K$}}$ and the crust and surface temperature should be decreasing on a timescale of $\sim 1{\mbox{$\rm\,yr$}}$ (BBR98; ).
The lack of luminosity variability in general, and the stability of the thermal component over a $\sim$5 year timescale supports the model proposed by BBR98 for the thermal component. It remains to be unequivocally demonstrated, however, that the variability observed is restricted to the power-law component, and that the variability of the temperature of the thermal component is limited to 1%, as expected from variations in the [$T_{\rm core}$]{} from outburst to outburst [@colpi00b].
The Power-Law Spectral Component
--------------------------------
The origin of the hard power-law component remains unclear. Many ideas have been put forward, ranging from emission due to an active pulsar wind colliding with the accretion disk, x-ray emission from a turned-on radio pulsar [@stella94] and accretion onto the magnetosphere (for a review, see ).
There is no direct knowledge of either the NS spin or magnetic field in Cen X-4, though the presence of Type I bursts points to $B<10^{10}
G$. None of the above models make specific predictions about either the level of emission or its spectral shape which explain the present observations.
Moreover, the limitation to $<$1% rms of coherent intensity pulsations in the transiently accreting NS Aql X-1 (another transient, type-I X-ray bursting source) following a rapid decrease in flux [@chandler00] which was interpreted as due to magnetic inhibition of the accretion flow [@campana98a; @zhang98a] does not support the idea that accretion onto the compact object is ultimately restricted by a “propeller effect”, favoring instead an interpretation of the end of transient outbursts as due to an end of the disk instability, such as occurs in dwarf novae [@king98]. In fact, abrupt declines in dwarf novae outburst fluxes are observed, and are explained within the disk instability model [@cannizzo94], without invoking magnetic inhibition. The absence, then, of a magnetosphere in transiently accreting NSs would preclude attributing the power-law to magnetospheric accretion, and energy production would have to be attributed to another site within the NS binary system.
Spectral comparison between the [[*ASCA*]{}]{} and [[*Chandra*]{}]{} observation indicate that a change in (only) the power-law normalization is marginally acceptable, and therefore that a change in the spectral slope (from $\alpha=1.9{\mbox{$\pm$}}0.3$ to $\alpha=1.0{\mbox{$\pm$}}0.4$) is also required. It is unclear how a change in the spectral slope could be produced while limiting the luminosity decrease to less than a factor of three in any of the proposed emission mechanisms [@campana98b].
There might be some help in understanding the power-law component from other wavelength bands, where it was found recently that, in quiescence, the energy density ($\nu F_\nu $) is nearly flat from UV through the X-ray energy range, in marked contrast to , in which the energy density falls in the UV band [@mcclintock00]. This flat-spectrum has been interpreted as due to a shock at the splash-point of accretion from the companion [@menou01]. However, while we also find a flat spectrum for the [[*ASCA*]{}]{} observation (photon power-law slope of $\alpha=1.9{\mbox{$\pm$}}0.4$) we find an [*increasing*]{} energy density during the [[*Chandra*]{}]{} observation ($\alpha=1.0{\mbox{$\pm$}}0.4$). Clearly, further observational study of the power-law component is required to understand even its most basic aspects.
Future Work on Neutron Star Radii from Thermal Emission
-------------------------------------------------------
The dominant systematic uncertainty in the measurement of the radius of the neutron star in 4 is the distance to the object. This will be dramatically improved by the Space Interferometric Mission, which can measure the parallax of the 18$.^m$5 counterpart to $\sim10\; \mu$asec, or $<$1% distance uncertainty at 1.2 kpc. This would make the dominant uncertainty in the measured thermal emission radius, followed by any systematic uncertainty in the modelled spectrum. For example, a planned 50 ksec XMM observation will permit constraint of $r/D$ to $\sim$2% while simultaneously measuring .
It has been suggested that some fraction of the low-luminosity X-ray sources in globular clusters are transiently accreting NSs in quiescence [@verbunt84; @rutledge00]. These make excellent targets for NS radius measurements, as there are multiple objects per observing field, all at the same distance and interstellar column density (BBR98). [[*Chandra*]{}]{} imaging observations of globular clusters [@grindlay00; @pooley00] indicate source densities in excess of several per square arcmin (down to luminosities of $L_X\sim$ ) with over 100 X-ray sources in the core of 47 Tuc alone. It is still unanswered as to what fraction of these are quiescent NS’s. The high source density will complicate X-ray spectroscopy with XMM or [[*Con-X*]{}]{} (with angular resolution $\sim$15), making [[*Chandra*]{}]{}imaging spectroscopy the best way to pursue this science.
The authors are grateful to the [[*Chandra*]{}]{} Observatory team for producing this exquisite observatory. R.R. thanks Andy Fabian for a useful conversation regarding historical measurements of photospheric absorption lines. The authors thank Dany Page and Andrew Cumming for comments on the text prior to submission. This research was partially supported by the National Science Foundation under Grant No. PHY99-07949 and by NASA through grant NAG 5-8658, NAG 5-7017 and the [[*Chandra*]{}]{} Guest Observer program through grant NAS 8-39073. L. B. is a Cottrell Scholar of the Research Corporation. E. F. B. acknowledges support from an Enrico Fermi Fellowship.
Arnaud, K. A., 1996, in G. Jacoby & J. Barnes (eds.), [*Astronomical Data Analysis Software and Systems V.*]{}, Vol. 101, p. 17, ASP Conf. Series
, K., [Dotani]{}, T., [Hoshi]{}, R., [Tanaka]{}, Y., [Robinson]{}, C. R., & [Terada]{}, K., 1998, , 611
, K., [Dotani]{}, T., [Kunieda]{}, H., & [Kawai]{}, N., 1996a, , L27
, K., [Dotani]{}, T., [Mitsuda]{}, K., [Hoshi]{}, R., [Vaughan]{}, B., [Tanaka]{}, Y., & [Inoue]{}, H., 1996b, , 257
, L. & [Rutledge]{}, R. E., 2000, , 908
, L., [Salpeter]{}, E. E., & [Wasserman]{}, I., 1992, , 143
, L., [Salpeter]{}, E. E., & [Wasserman]{}, I., 1993, , 615
, W. P., [Raymond]{}, J. C., [Dupree]{}, A. K., [Wu]{}, C. C., [Holm]{}, A. V., & [Swank]{}, J. H., 1984, , 270
, E. F., Bildsten, L., & Rutledge, R. E., 1998, , L95,
, S., [Colpi]{}, M., [Mereghetti]{}, S., [Stella]{}, L., & [Tavani]{}, M., 1998a, , 279
, S., [Mereghetti]{}, S., [Stella]{}, L., & [Colpi]{}, M., 1997, , 941
, S., [Stella]{}, L., [Mereghetti]{}, S., [Colpi]{}, M., [Tavani]{}, M., [Ricci]{}, D., [Fiume]{}, D. D., & [Belloni]{}, T., 1998b, , L65
, S., [Stella]{}, L., [Mereghetti]{}, S., & [Cremonesi]{}, D., 2000, , 583
, C. R., [McClintock]{}, J. E., & [Grindlay]{}, J. E., 1980, , L55
, J. K., 1994, , 389
Chandler, A. & Rutledge, R. E., 2000, , in press
, W., [Shrader]{}, C. R., & [Livio]{}, M., 1997, , 312
, C., [Ilovaisky]{}, S. A., [Van Paradijs]{}, J., [Pedersen]{}, H., & [Van der Klis]{}, M., 1989, , 114
, M., [Geppert]{}, U., [Page]{}, D., & [Possenti]{}, A., 2001, , submitted, astro-ph/001572
, A. P., [Hutchings]{}, J. B., [Schmidtke]{}, P. C., [Hartwick]{}, F. D. A., [Crampton]{}, D., & [Thompson]{}, I. B., 1988, , 1231
, E., [Jansen]{}, F., [Penninx]{}, W., [Oosterbroek]{}, T., [van Paradijs]{}, J., & [Lewin]{}, H. G., 1989, , 523
, E., [Magnier]{}, E., [Lewin]{}, W. H. G., [Tan]{}, J., [Penninx]{}, W., & [van Paradijs]{}, J., 1990, , 103
, C. S. R., [Fabian]{}, A. C., & [Ross]{}, R. R., 1992, , 471
, A. J., [Fabian]{}, A. C., & [Ross]{}, R. R., 1987, , 259
, M. R., 1994, , 407
, M. R. & [Callanan]{}, P. J., 1999, , 1390
, P., 1975, , 95
, N. & [Noels]{}, A., 1993, in E. Prantzos, M. [Vaugioni-Flam]{}, & M. Gasse (eds.), [*Origin and Evolution of the Elements*]{}, p. 14, Cambridge University Press
, J., [Edmonds]{}, P., [Heinke]{}, C., & [Murray]{}, S., 2000, , 4505
, P. & [Zdunik]{}, J. L., 1990, , 431
, P. & [Grindlay]{}, J. E., 1983, , L83
, P., [Ford]{}, E. C., & [Chen]{}, K., 1997, , L27
, A. R. & [Ritter]{}, H., 1998, , L42
, W., 1998, , L37
, W., [Michelson]{}, P., & [Wagoner]{}, R. V., 1990, , 538
, D. A., [Darbro]{}, W., [Elsner]{}, R. F., [Weisskopf]{}, M. C., [Kahn]{}, S., [Sutherland]{}, P. G., & [Grindlay]{}, J. E., 1983, , 160
Lewin, W., [Van Paradijs]{}, J., & Taam, R., 1993, , 223
, R. A., [Howard]{}, W. M., & [Taam]{}, R. E., 1986, , 170
, J., 1990, , 223
, J., 1991, , 161
, E., [Lewin]{}, W. H. G., [van Paradijs]{}, J., [Tan]{}, J., [Penninx]{}, W., & [Damen]{}, E., 1989, , 729
, E. L., [Rebolo]{}, R., [Casares]{}, J., & [Charles]{}, P. A., 1994, , 791
, J. E. & [Remillard]{}, R. A., 1990, , 386
, J. E. & [Remillard]{}, R. A., 2000, , 956
, K., [Esin]{}, A. A., [Narayan]{}, R., [Garcia]{}, M. R., [Lasota]{}, J. P., & [McClintock]{}, J. E., 1999, , 276
, K. & McClintock, J. E., 2001, , submitted, astro-ph/0010430
, N., [Inoue]{}, H., & [Tanaka]{}, Y., 1988, , 209
, F., 1997, , L47
, A. N., [Oosterbroek]{}, T., [Orr]{}, A., [Guainazzi]{}, M., [Shane]{}, N., [Freyberg]{}, M. J., [Ricci]{}, D., & [Malizia]{}, A., 1999, , 407
, G. G. & [Shibanov]{}, I. A., 1978, , 214
, G. G., [Shibanov]{}, I. A., & [Zavlin]{}, V. E., 1991, , 193
, D., [Lewin]{}, W. H. G., [Verbunt]{}, F., [Fox]{}, D. W., [Margon]{}, B., [Kaspi]{}, V. M., [van der Klis]{}, M., & [Miller]{}, J., 2000, , 2901
, A. Y., [Chabrier]{}, G., & [Yakovlev]{}, D. G., 1997, , 415
, P. & [Schmitt]{}, J. H. M. M., 1995, , 889
, M. & [Romani]{}, R. W., 1996, , 327
, J. C. & [Smith]{}, B. W., 1977, , 419
, R. E., [Bildsten]{}, L., [Brown]{}, E. F., [Pavlov]{}, G. G., & [Zavlin]{}, V. E., 1999, , 945
, R. E., [Bildsten]{}, L., [Brown]{}, E. F., [Pavlov]{}, G. G., & [Zavlin]{}, V. E., 2000, , 985
, S. L., [Egger]{}, R., [Freyberg]{}, M. J., [McCammon]{}, D., [Plucinsky]{}, P. P., [Sanders]{}, W. T., [Schmitt]{}, J. H. M. M., [Truemper]{}, J., & [Voges]{}, W., 1997, , 125
, L., [Campana]{}, S., [Colpi]{}, M., & [Mereghetti]{}, S. andx [Tavani]{}, M., 1994, , L47
Tanaka, Y. & Lewin, W., 1995, in W. Lewin, J. [Van Paradijs]{}, & E. [Van Den Heuvel]{} (eds.), [ *X-Ray Binaries*]{}, Vol. 1, p. 126, Cambridge University Press
, Y. & [Shibazaki]{}, N., 1996, , 607
, B. G., [Andresen]{}, R. D., [Peacock]{}, A., & [Zobl]{}, R., 1981, , 479
, S. E. & [Chakrabarty]{}, D., 1999, , 288
, L., 1994, , 340
, H., [Nomoto]{}, K., [Tsuruta]{}, S., [Muto]{}, T., & [Tatsumi]{}, T., 1994, , 309
Ushomirsky, G. & [Rutledge]{}, R. E., 2001, in preparation
, J., 1982, , 51
, J. & [Lewin]{}, W. H. G., 1987, , L20
, J., [Verbunt]{}, F., [Shafer]{}, R. A., & [Arnaud]{}, K. A., 1987, , 47
, F., [Bunk]{}, W., [Hasinger]{}, G., & [Johnston]{}, H. M., 1995, , 732
, F., [Elson]{}, R., & [Van Paradijs]{}, J., 1984, , 899
, I., [Inoue]{}, H., [Koyama]{}, K., [Matsuoka]{}, M., [Murakami]{}, T., [Ogawara]{}, Y., [Ohashi]{}, T., [Tanaka]{}, Y., [Hayakawa]{}, S., [Tawara]{}, Y., [Miyamoto]{}, S., [Tsunemi]{}, H., & [Kondo]{}, I., 1984, , 819
, R. F., [Rappaport]{}, S., & [Savonije]{}, G. J., 1983, , 678
, L., [Turolla]{}, R., [Zane]{}, S., & [Treves]{}, A., 1995, , 849
, V. E., [Pavlov]{}, G. G., & [Shibanov]{}, Y. A., 1996, , 141
, S. N., [Yu]{}, W., & [Zhang]{}, W., 1998, , L71
, W., [Strohmayer]{}, T. E., & [Swank]{}, J. H., 1997, , L167
(55,10)(-400,35) (0,0)[Figure \[fig:chandraspec\]]{}
[lc]{}
\
& 0.10\
$\alpha$ & 1.00.4\
& 5\
$kT$ (keV) & 0.0670.019\
$r$ & 19\
Total Model Flux & 13\
/dof (prob) & 0.94/9 (0.49)\
\
& (0.055)\
$\alpha$ & 1.0\
& 4.6\
$kT$ (keV) & 0.074\
$r$ & 13.3 km\
Total Model Flux & 11.3\
/dof (prob) & 0.83/10 (0.50)\
[llcccc]{}
1 & 1980 Jul 28& 0.32& 2-3 & & 0.066\
& & & (0.5-4.5 keV) & &\
1 & 1986 Feb 21 & & 4-11 & & 0.066\
& & & (0.5-4.5 keV) & &\
2 & 1994 Feb 27-28 & 0.16 & 4 & 1.90.3 & $<$0.2\
& & & (0.5-10 keV) & &\
3 & 1995 Aug 16-26 & & 7 & &\
& & & (0.1-2.4 keV)& &\
4 & 1997 Feb 4-5 & 0.130.02 & 3 & 2.50.5 & 0.30.1\
& & & (0.5-10 keV) & &\
5 & 1999 Feb 9 &\
& & & (0.5-10 keV) & &\
6 & 2000 June 23 & 0.176 & 1.7 & 1.2& $<$0.06\
& & & (0.5-10 keV) & &\
[ccccc|ccccc]{}
(0.055) & 1.00.4 & 5.7 & & 12 & & 0.0770.007 & 0.085 & 1.5/47 & 0.01\
(0.055) & 1.7 & & 0.076 & 12 & & 3.5 & 9.1 & 1.5/47 & 0.01\
[ccccc|ccccc]{}
(0.055) & 0.0760.007 & 12.92.6 & 1.00.4 & 4.4& & 1.90.3 & 8.61.6 & 1.0/52 & 0.43\
[^1]: In low magnetic field NSs ($\le$ G), the magnetic field plays no role in determining the opacity of the atmosphere at energies above 0.1 keV, and thus can be neglected [@zavlin96].
[^2]: available at http://asc.harvard.edu
[^3]: While the metallicity of the companion has not been observationally constrained, Martin [-@martin94a], while investigating Li abundance, found that the Ca I $\lambda6717$ line is consistent with solar metallicity, and that an assumed solar Fe abundance in their spectral fits to the Li I doublet is consistent with the data. In optical spectral studies of the companion [@cowley88; @chev89; @mcclintock90; @mcclintock00], while metal absorption features are observed, the metallicity of the companion is not estimated; however, it does not appear to be so substantially sub-solar.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We propose and study a number of layer methods for stochastic Navier-Stokes equations (SNSE) with spatial periodic boundary conditions and additive noise. The methods are constructed using conditional probabilistic representations of solutions to SNSE and exploiting ideas of the weak sense numerical integration of stochastic differential equations. We prove some convergence results for the proposed methods. Results of numerical experiments on two model problems are presented.
**Keywords** Navier-Stokes equations, Oseen-Stokes equations, Helmholtz-Hodge-Leray decomposition, conditional Feynman-Kac formula, weak approximation of stochastic differential equations layer methods.
**AMS 2000 subject classification.** 65C30, 60H15, 60H35
author:
- 'G.N. Milstein[^1]'
- 'M.V. Tretyakov[^2]'
title: 'Layer methods for Navier-Stokes equations with additive noise'
---
Introduction
============
Let $(\Omega,\mathcal{F},P)$ be a probability space and $(w(t),\mathcal{F}%
_{t}^{w})=((w_{1}(t),\ldots,w_{q}(t))^{\top},\mathcal{F}_{t}^{w})$ be a $q$-dimensional standard Wiener process, where $\mathcal{F}_{t}^{w},\ 0\leq
t\leq T,$ is an increasing family of $\sigma$-subalgebras of $\mathcal{F}$ induced by $w(t).$ We consider the system of stochastic Navier-Stokes equations (SNSE) with additive noise for velocity $v$ and pressure $p$ in a viscous incompressible flow:$$\begin{gathered}
dv(t)=\left[ \frac{\sigma^{2}}{2}\Delta v-(v,\nabla)v-\nabla p+f(t,x)\right]
dt+\sum_{r=1}^{q}\gamma_{r}(t,x)dw_{r}(t),\label{NS1}\\
\ \ 0\leq t\leq T,\ x\in \mathbf{R}^{n}, \nonumber \\
\operatorname{div}\ v=0, \label{NS2}%\end{gathered}$$ with spatial periodic conditions$$\begin{aligned}
v(t,x+Le_{i}) & =v(t,x),\ p(t,x+Le_{i})=p(t,x),\label{NS4}\\
0 & \leq t\leq T,\ \ i=1,\ldots,n,\nonumber\end{aligned}$$ and the initial condition$$v(0,x)=\varphi(x). \label{NS3}%$$ In (\[NS1\])-(\[NS2\]) we have $v\in\mathbf{R}^{n}$,$\ p$ is a scalar, $f$ $\in\mathbf{R}^{n},$ $\gamma_{r}\in\mathbf{R}^{n}$; $\{e_{i}\}$ is the canonical basis in $\mathbf{R}^{n}$ and $L>0$ is the period (for simplicity in writing, the periods in all the directions are taken the same). The functions $f=f(t,x)$ and $\gamma_{r}(t,x)$ are supposed to be spatial periodic as well. Further, we require that $\gamma_{r}(t,x)$ are divergence free: $$\operatorname{div}\gamma_{r}(t,x)=0,\ r=1,\ldots,q. \label{NS03}%$$
SNSE can be useful for explaining the turbulence phenomenon (see [@turb1; @Flandoli; @RozNS04] and references therein). They have complicated dynamics and some interesting properties (e.g., ergodicity of solutions [@HaMa06; @HasNS; @DaPr; @MaPa06]). At the same time, rather little has been done in numerics for SNSE. Let us cite [@HRoz07], where algorithms based on Wiener Chaos expansion are considered, and quite recent works [@BCP10; @CP; @D], where splitting schemes with finite element or Galerkin approximations are applied. Here we suggest to exploit some probabilistic representations of solutions to SNSE for constructing numerical methods of the layer type. The proposed methods are promised to be effective, reliable numerical methods for studying SNSE. Layer methods for deterministic semilinear and quasilinear partial differential equations of parabolic type were proposed in [@M1; @quasic] (see also [@MT1; @french]), and for deterministic NSEs they were first considered in [@BM] and further developed in [@NS5; @NSB]. Layer methods for linear and semilinear stochastic partial differential equations (SPDE) were constructed and analyzed in [@spde].
The rest of the paper is organized as follows. In Section \[Secpre\] we introduce additional notation and write down probabilistic representations for linearized SNSE (i.e., stochastic Oseen-Stokes equations) and for the SNSE (\[NS1\])-(\[NS3\]) which we use in Section \[secLayer\] for constructing layer methods for the SNSE. Three layer methods are given in Section \[secLayer\] together with discussion of their implementation. Numerical error analysis is done in Section \[secER\]. Results of numerical experiments on two test models are presented in Section \[secnum\].
Preliminaries\[Secpre\]
=======================
In this section we recall the required function spaces [@CM; @RT; @T; @RozNS04; @RozNS05] and write probabilistic representations of solutions to linearized SNSE and to SNSE resting on results from [@KryR; @Kun; @Pard; @R].
Function spaces, the Helmholtz-Hodge-Leray decomposition, and notation
----------------------------------------------------------------------
Let $\{e_{i}\}$ be the canonical basis in $\mathbf{R}^{n}.$ We shall consider spatial periodic $n$-vector functions $u(x)=(u^{1}(x),\ldots,u^{n}(x))^{\top}$ in $\mathbf{R}^{n}:$ $u(x+Le_{i})=u(x),\ i=1,\ldots,n,$ where $L>0$ is the period in $i$th direction. Denote by $Q=(0,L)^{n}$ the cube of the period (of course, one may consider different periods $L_{1},\ldots,L_{n}$ in the different directions). We denote by $\mathbf{L}^{2}(Q)$ the Hilbert space of functions on $Q$ with the scalar product and the norm$$(u,v)=\int_{Q}\sum_{i=1}^{n}u^{i}(x)v^{i}(x)dx,\ \Vert u\Vert=(u,u)^{1/2}.$$ We keep the notation $|\cdot|$ for the absolute value of numbers and for the length of $n$-dimensional vectors, for example,$$|u(x)|=[(u^{1}(x))^{2}+\cdots+(u^{n}(x))^{2}]^{1/2}.$$
We denote by $\mathbf{H}_{p}^{m}(Q),\ m=0,1,\ldots,$ the Sobolev space of functions which are in $\mathbf{L}^{2}(Q),$ together with all their derivatives of order less than or equal to $m,$ and which are periodic functions with the period $Q.$ The space $\mathbf{H}_{p}^{m}(Q)$ is a Hilbert space with the scalar product and the norm$$(u,v)_{m}=\int_{Q}\sum_{i=1}^{n}\sum_{[\alpha^{i}]\leq m}D^{\alpha^{i}}%
u^{i}(x)D^{\alpha^{i}}v^{i}(x)dx,\ \Vert u\Vert_{m}=[(u,u)_{m}]^{1/2},$$ where $\alpha^{i}=(\alpha_{1}^{i},\ldots,\alpha_{n}^{i}),\ \alpha_{j}^{i}%
\in\{0,\ldots,m\},\ [\alpha^{i}]=\alpha_{1}^{i}+\cdots+\alpha_{n}^{i},$ and$$D^{\alpha^{i}}=D_{1}^{\alpha_{1}^{i}}\cdots D_{n}^{\alpha_{n}^{i}}%
=\frac{\partial^{\lbrack\alpha^{i}]}}{(\partial x^{1})^{\alpha_{1}^{i}}%
\cdots(\partial x^{n})^{\alpha_{n}^{i}}}\ ,\ i=1,\ldots,n.$$ Note that $\mathbf{H}_{p}^{0}(Q)=\mathbf{L}^{2}(Q).$
Introduce the Hilbert subspaces of $\mathbf{H}_{p}^{m}(Q):$$$\begin{aligned}
\mathbf{V}_{p}^{m} & =\{v:\ v\in\mathbf{H}_{p}^{m}(Q),\ \operatorname{div}%
v=0\},\ m>0,\\
\mathbf{V}_{p}^{0} & =\text{the closure of }\mathbf{V}_{p}^{m},\ m>0\text{
in }\mathbf{L}^{2}(Q).\end{aligned}$$ Clearly,$$\mathbf{V}_{p}^{m_{1}}=\text{the closure of }\mathbf{V}_{p}^{m_{2}}\text{ in
}\mathbf{H}_{p}^{m_{1}}(Q)\text{ for any}\ m_{2}\geq m_{1}.$$
Denote by $P$ the orthogonal projection in $\mathbf{H}_{p}^{m}(Q)$ onto $\mathbf{V}_{p}^{m}$ (we omit $m$ in the notation $P$ here). The operator $P$ is often called the Leray projection. Due to the Helmholtz-Hodge-Leray decomposition, any function $u\in\mathbf{H}_{p}^{m}(Q)$ can be represented as $$u=Pu+\nabla g,\ \operatorname{div}Pu=0,$$ where $g=g(x)$ is a scalar $Q$-periodic function such that $\nabla
g\in\mathbf{H}_{p}^{m}(Q).$ It is natural to introduce the notation $P^{\bot
}u:=\nabla g$ and hence write $$u=Pu+P^{\bot}u$$ with $$P^{\bot}u\in(\mathbf{V}_{p}^{m})^{\bot}=\{v:\ v\in\mathbf{H}_{p}%
^{m}(Q),\ v=\nabla g\}.$$
Let$$\begin{gathered}
u(x)=\sum_{\mathbf{n}\in\mathbf{Z}^{n}}u_{\mathbf{n}}e^{i(2\pi/L)(\mathbf{n}%
,x)},\ g(x)=\sum_{\mathbf{n}\in\mathbf{Z}^{n}}g_{\mathbf{n}}e^{i(2\pi
/L)(\mathbf{n},x)},\ g_{\mathbf{0}}=0,\label{N00}\\
Pu(x)=\sum_{\mathbf{n}\in\mathbf{Z}^{n}}(Pu)_{\mathbf{n}}e^{i(2\pi
/L)(\mathbf{n},x)},\ P^{\bot}u(x)=\nabla g(x)=\sum_{\mathbf{n}\in
\mathbf{Z}^{n}}(P^{\bot}u)_{\mathbf{n}}e^{i(2\pi/L)(\mathbf{n},x)}\nonumber\end{gathered}$$ be the Fourier expansions of $u,$ $g,$ $Pu,$ and $P^{\bot}u=\nabla g.$ Here $u_{\mathbf{n}},$ $(Pu)_{\mathbf{n}},\ $and $(P^{\bot}u)_{\mathbf{n}}=(\nabla
g)_{\mathbf{n}}$ are $n$-dimensional vectors and $g_{\mathbf{n}}$ are scalars. We note that $g_{\mathbf{0}}$ can be any real number but for definiteness we set $g_{\mathbf{0}}=0.$ The coefficients $(Pu)_{\mathbf{n}},\ (P^{\bot
}u)_{\mathbf{n}}$, and $g_{\mathbf{n}}$ can be easily expressed in terms of $u_{\mathbf{n}}:$ $$\begin{aligned}
(Pu)_{\mathbf{n}} & =u_{\mathbf{n}}-\frac{u_{\mathbf{n}}^{\top}\mathbf{n}%
}{|\mathbf{n}|^{2}}\mathbf{n,\ }(P^{\bot}u)_{\mathbf{n}}=i\frac{2\pi}%
{L}g_{\mathbf{n}}\mathbf{n=}\frac{u_{\mathbf{n}}^{\top}\mathbf{n}}%
{|\mathbf{n}|^{2}}\mathbf{n,\ }\label{N01}\\
g_{\mathbf{n}} & =-i\frac{L}{2\pi}\frac{u_{\mathbf{n}}^{\top}\mathbf{n}%
}{|\mathbf{n}|^{2}},\ \mathbf{n\neq0,\ }g_{\mathbf{0}}=0.\nonumber\end{aligned}$$
We have$$\nabla e^{i(2\pi/L)(\mathbf{n},x)}=\mathbf{n}e^{i(2\pi/L)(\mathbf{n},x)}\cdot
i\frac{2\pi}{L},$$ hence $u_{\mathbf{n}}e^{i(2\pi/L)(\mathbf{n},x)}\in\mathbf{V}_{p}^{m}$ if and only if $(u_{\mathbf{n}},\mathbf{n)}=0.$ We obtain from here that the orthogonal basis of the subspace $(\mathbf{V}_{p}^{m})^{\bot}$ consists of $\mathbf{n}e^{i(2\pi/L)(\mathbf{n},x)},\ \mathbf{n}\in\mathbf{Z}%
^{n},\ \mathbf{n\neq0}$; and an orthogonal basis of $\mathbf{V}_{p}^{m}$ consists of $_{k}u_{\mathbf{n}}e^{i(2\pi/L)(\mathbf{n},x)},$ $k=1,\ldots
,n-1,\ \mathbf{n}\in\mathbf{Z}^{n},\ $where under $\mathbf{n\neq0}$ the vectors $_{k}u_{\mathbf{n}}$ are orthogonal to $\mathbf{n:}$** **$\mathbf{(}_{k}u_{\mathbf{n}},\mathbf{n)}=0,\ k=1,\ldots,n-1,$ and they are orthogonal among themselves: $\mathbf{(}_{k}u_{\mathbf{n}},\ _{m}%
u_{\mathbf{n}}\mathbf{)}=0,$ $k,m=1,\ldots,n-1,$ $m\neq k,$ and finally, for $\mathbf{n=0,}$ the vectors $_{k}u_{\mathbf{0}},\ k=1,\ldots,n,$ are orthogonal.
In what follows we suppose that the below assumptions hold.$\medskip$
**Assumptions 2.1.** *We assume that the coefficients* $f(t,x)$ *and* $\gamma_{r}(s,x),$ $r=1,\ldots,q,$ *are sufficiently smooth* *and* *the problem* (\[NS1\])-(\[NS3\]) *has a unique classical solution* $v(t,x),\ p(t,x),$ $(t,x)\in\lbrack0,T]\times\mathbf{R}^{n},$ *which* *has continuous derivatives in the space variable* $x$ *up to some order, and the solution and the derivatives have uniformly in* $(t,x)$ *bounded moments of a sufficiently high order* $m,$ $2\leq m<m_{0},$ *where* $m_{0}>2$ *is a positive number or* $m_{0}=\infty
$*.* $\medskip$
The solution $v(t,x),\ p(t,x),$ $(t,x)\in\lbrack0,T]\times\mathbf{R}^{n},$ to (\[NS1\])-(\[NS3\]) is $\mathcal{F}_{t}^{w}$-adaptive, $v(t,\cdot
)\in\mathbf{V}_{p}^{m}$ and $\nabla p(t,\cdot)\in(\mathbf{V}_{p}^{m})^{\bot}$ for every $t\in\lbrack0,T]$ and $\omega\in\Omega.$
Assumptions of this kind are rather usual for works dedicated to numerics. They are rested on results concerning regularity of solutions (see, e.g., the corresponding theory for deterministic NSE in [@RT; @T]). Unfortunately, we could not find explicit results on the classical solution for SNSE in literature. At the same time, the question about existence of the unique sufficiently regular (with respect to $x)$ solution of the SNSE (\[NS1\])-(\[NS3\]) on a time interval $[0,T]$ is analogous to the one in the deterministic case. Indeed, the following remark reduces this problem of regularity for the SNSE to regularity of solutions to NSE with random coefficients which is close to the theory of deterministic NSE treated in [@RT; @T].
Let $\Gamma(t,x)=\sum_{r=1}^{q}\int_{0}^{t}\gamma_{r}(s,x)dw_{r}(s).$ Then $V(t,x)=v(t,x)+\Gamma(t,x)$ together with $p(t,x)$ solves the following ‘usual’ NSE with random coefficients: $$\begin{gathered}
\frac{\partial}{\partial t}V=\frac{\sigma^{2}}{2}\Delta V-(V-\Gamma
(t,x),\nabla)(V-\Gamma(t,x))-\nabla p+f(t,x)-\frac{\sigma^{2}}{2}\Delta
\Gamma(t,x),\\
0\leq t\leq T,\ x\in\mathbf{R}^{n},\\
\operatorname{div}\ V=0,\end{gathered}$$ with spatial periodic conditions$$\begin{aligned}
V(t,x+Le_{i}) & =V(t,x),\ p(t,x+Le_{i})=p(t,x),\ \\
0 & \leq t\leq T,\ \ i=1,\ldots,n,\end{aligned}$$ and the initial condition$$V(0,x)=\varphi(x).$$
Probabilistic representations of solutions to linearized SNSE\[prepOseen\]
--------------------------------------------------------------------------
We start with considering a linearized version of the SNSE (\[NS1\])-(\[NS3\]), i.e., the stochastic Oseen-Stokes equations (see [@MikStokes]): $$\begin{gathered}
dv_{a}(t)=\left[ \frac{\sigma^{2}}{2}\Delta v_{a}-(a,\nabla)v_{a}-\nabla
p_{a}+f(t,x)\right] dt+\sum_{r=1}^{q}\gamma_{r}(t,x)dw_{r}(t), \label{os1}\\
\ \ 0\leq t\leq T,\ x\in\mathbf{R}^{n}, \nonumber \\
\operatorname{div}\ v_{a}=0, \label{os2}%\end{gathered}$$ with spatial periodic conditions$$\begin{aligned}
v_{a}(t,x+Le_{i}) & =v_{a}(t,x),\ p_{a}(t,x+Le_{i})=p_{a}(t,x),\ \label{0s3}%
\\
0 & \leq t\leq T,\ i=1,\ldots,n,\nonumber\end{aligned}$$ and the initial condition$$v_{a}(0,x)=\varphi(x), \label{os4}%$$ where $a=a(t,x)$ is an $n$-dimensional vector $a=(a^{1},\ldots,a^{n}%
)^{\intercal}$ with $a^{i}$ being $Q$-periodic deterministic functions which have continuous derivatives with respect to $x$ up to some order; and the rest of the notation is the same as in (\[NS1\])-(\[NS3\]).
We re-write the problem (\[os1\])-(\[os4\]) with positive direction of time into the problem with negative direction of time which is more convenient for making use of probabilistic representations. To this end, introduce the new time variable $s=T-t$ and the functions $u_{a}(s,x):=v_{a}(T-s,x),$ $\tilde{a}(s,x):=a(T-s,x),$ $\tilde{f}(s,x):=f(T-s,x),$ $\tilde{\gamma}%
_{r}(s,x):=\gamma_{r}(T-s,x),$ and $\tilde{p}_{a}(s,x):=p_{a}(T-s,x).$
Further, we recall the definition of a backward Ito integral [@R]. Introduce the backward Wiener processes $$\tilde{w}_{r}(t):=w_{r}(T)-w_{r}(T-t),\ \ r=1,\ldots,q,\ \ 0\leq t\leq T,
\label{bs4}%$$ and a decreasing family of $\sigma$-subalgebras $\mathcal{F}_{t,T}^{w},$ $0\leq t\leq T,$ induced by the increments $w_{r}(T)-w_{r}(t^{\prime}),$ $r=1,\ldots,q,$ $t^{\prime}\geq t$. The increasing family of $\sigma$-subalgebras $\mathcal{F}_{t}^{\tilde{w}}$ induced by $\tilde{w}_{r}(s^{\prime}),$ $s^{\prime}\leq t,$ coincides with $\mathcal{F}_{T-t,T}^{w},$ while $\mathcal{F}_{t,T}^{\tilde{w}}$ is induced by the increments $\tilde{w}%
_{r}(T)-\tilde{w}_{r}(t^{\prime}),$ $r=1,\ldots,q,$ $t^{\prime}\geq t$, and coincides with $\mathcal{F}_{T-t}^{w}.$ The backward Ito integral with respect to $\tilde{w}_{r}(s)$ is defined as the Ito integral with respect to $w_{r}(s)$: $$\int_{t}^{t^{\prime}}\psi(t^{\prime\prime})\ast d\tilde{w}_{r}(t^{\prime
\prime}):=\int_{T-t^{\prime}}^{T-t}\psi(T-t^{\prime\prime})dw_{r}%
(t^{\prime\prime}),\ \ 0\leq t\leq t^{\prime}\leq T, \label{bs5}%$$ where $\psi(T-t),$ $t\leq T,$ is an $\mathcal{F}_{t}^{w}$-adapted square-integrable function and $\psi(t)$ is $\mathcal{F}_{t}^{\tilde{w}}%
$-adapted. Note that $w_{r}(t)=\tilde{w}_{r}(T)-\tilde{w}_{r}(T-t),$ $r=1,\ldots,q,$ $0\leq t\leq T.$
The backward stochastic Oseen-Stokes equations can be written as $$\begin{gathered}
-du_{a}(s)=\left[ \frac{\sigma^{2}}{2}\Delta u_{a}-(\tilde{a},\nabla
)u_{a}-\nabla\tilde{p}_{a}+\tilde{f}(s,x)\right] ds+\sum_{r=1}^{q}%
\tilde{\gamma}_{r}(s,x)\ast d\tilde{w}_{r}(s),\ \ \label{os11}\\
0\leq s\leq T,\ x\in\mathbf{R}^{n},\nonumber\\
\operatorname{div}\ u_{a}=0, \label{os12}%\end{gathered}$$ with spatial periodic conditions$$\begin{aligned}
u_{a}(s,x+Le_{i}) & =u_{a}(s,x),\ \tilde{p}_{a}(s,x+Le_{i})=\tilde{p}%
_{a}(s,x),\ \label{os13}\\
0 & \leq s\leq T,\ i=1,\ldots,n,\nonumber\end{aligned}$$ and the terminal condition$$u_{a}(T,x)=\varphi(x). \label{os14}%$$ We note that (\[bs5\]) implies $$\int_{s}^{T}\tilde{\gamma}_{r}(s^{\prime},x)\ast d\tilde{w}_{r}(s^{\prime
})=\int_{0}^{T-s}\gamma_{r}(s^{\prime},x)dw_{r}(s^{\prime}).$$ The processes $u_{a}(s,x),$ $\tilde{p}_{a}(s,x)$ are $\mathcal{F}%
_{s,T}^{\tilde{w}}$-adapted (and $\mathcal{F}_{T-s}^{w}$-adapted), they depend on $\tilde{w}_{r}(T)-\tilde{w}_{r}(s^{\prime})=w_{r}(T-s^{\prime}),$ $s\leq
s^{\prime}\leq T.$
Let $u_{a}(s,x),$ $\tilde{p}_{a}(s,x)$ be a solution of the problem (\[os11\])-(\[os14\]). For the function $u_{a}(s,x)$, one can use the following probabilistic representation of solutions to the Cauchy problem for linear SPDE of parabolic type (the conditional Feynman-Kac formula or the averaging over characteristics formula, see, e.g., [@R] and [@spde]):$$u_{a}(s,x)=E^{\tilde{w}}\left[ \varphi(X_{s,x}(T))Y_{s,x,1}(T)+Z_{s,x,1,0}%
(T)\right] ,\ 0\leq s\leq T, \label{FBD5}%$$ where $X_{s,x}(s^{\prime}),\ Y_{s,x,y}(s^{\prime}),\ Z_{s,x,y,z}(s^{\prime
}),\ s^{\prime}\geq s,$ solves the system of Ito stochastic differential equations:$$\begin{gathered}
dX=(-\tilde{a}(s^{\prime},X)-\sigma\mu(s^{\prime},X))ds^{\prime}+\sigma
dW(s^{\prime}),\ X(s)=x,\label{BDF0}\\
dY=\mu^{\intercal}(s^{\prime},X)YdW(s^{\prime}),\ Y(s)=y,\label{BDF1}\\
dZ=(-\nabla\tilde{p}_{a}(s^{\prime},X)+\tilde{f}(s^{\prime},X))Yds^{\prime
}+F(s^{\prime},X)YdW(s^{\prime})\label{BDF2}\\
+\sum_{r=1}^{q}\tilde{\gamma}_{r}(s^{\prime},X)Yd\tilde{w}_{r}(s^{\prime
}),\ Z(s)=z.\nonumber\end{gathered}$$ In (\[FBD5\])-(\[BDF2\]), $W(s)$ is a standard $n$-dimensional Wiener process independent of $\tilde{w}_{r}(s)$ on the probability space $(\Omega,\mathcal{F},P)$; $Y$ is a scalar, and $Z$ is an $n$-dimensional column-vector;$\ \mu(s,x)$ is an arbitrary $n$-dimensional spatial periodic vector function and $F(s,x)$ is an arbitrary $n\times n$-dimensional spatial periodic matrix function, which are sufficiently smooth in $s,x$; the expectation $E^{\tilde{w}}$ in (\[FBD5\]) is taken over the realizations of $W(s),$ $t\leq s\leq T,$ for a fixed $\tilde{w}_{r}(s^{\prime}),$ $r=1,\ldots,q,$ $s\leq s^{\prime}\leq T,$ in other words, $E^{\tilde{w}%
}\left( \cdot\right) $ means the conditional expectation: $$E\left( \cdot|\tilde{w}_{r}(s^{\prime})-\tilde{w}_{r}(s),\text{ }%
r=1,\ldots,q,\text{ }s\leq s^{\prime}\leq T\right) .$$
The probabilistic representation like (\[FBD5\])-(\[BDF2\]) for the Cauchy problem (\[os11\]), (\[os14\]) is obtained (see, e.g., [@R]) for linear SPDEs with deterministic coefficients. However here $\tilde{p}%
_{a}(s,x)$ is a part of solution of problem (\[os11\])-(\[os14\]) and it is random (more precisely it is $\mathcal{F}_{s,T}^{\tilde{w}}$-adapted). In this case the representation (\[FBD5\])-(\[BDF2\]) can be rigorously justified in the following way. The solution $u_{a}$ of (\[os11\]), (\[os14\]) can be represented in the form of the sum$$u_{a}=u_{a}^{(0)}+u_{a}^{(1)},$$ where $u_{a}^{(0)}$ satisfies the Cauchy problem for the backward deterministic linear parabolic PDE with random parameters:$$\begin{aligned}
-\frac{\partial u_{a}^{(0)}}{\partial s} & =\frac{\sigma^{2}}{2}\Delta
u_{a}^{(0)}-(\tilde{a},\nabla)u_{a}^{(0)}-\nabla\tilde{p}_{a},\label{Fr1}\\
u_{a}^{(0)}(T,x) & =0,\nonumber\end{aligned}$$ and $u_{a}^{(1)}$ satisfies the Cauchy problem for the backward stochastic linear parabolic PDE with deterministic parameters:$$\begin{aligned}
-du_{a}^{(1)}(s) & =\left[ \frac{\sigma^{2}}{2}\Delta u_{a}^{(1)}%
-(\tilde{a},\nabla)u_{a}^{(1)}+\tilde{f}(s,x)\right] ds+\sum_{r=1}^{q}%
\tilde{\gamma}_{r}(s,x)\ast d\tilde{w}_{r}(s),\ \label{Fr2}\\
u_{a}^{(1)}(T,x) & =\varphi(x).\ \nonumber\end{aligned}$$
Clearly,$$u_{a}^{(0)}(s,x)=E^{\tilde{w}}\left[ Z_{s,x,1,0}^{(0)}(T)\right]
=-E^{\tilde{w}}\int_{s}^{T}\nabla\tilde{p}_{a}(s^{\prime},X_{s,x}(s^{\prime
}))\ Y_{s,x,1}(s^{\prime})ds^{\prime}.$$ The Feynman-Kac formula for $u_{a}^{(1)}$ coincides with (\[FBD5\])-(\[BDF2\]) under $\nabla\tilde{p}_{a}(s,x)$ $=$ $0$.
Let $\mathcal{F}_{s,t}^{W}$ be a $\sigma$-algebra induced by $W_{r}%
(s^{\prime})-W_{r}(s),$ $r=1,\ldots,n,\ s\leq s^{\prime}\leq t.$ We note that $\nabla\tilde{p}_{a}(s^{\prime},X_{s,x}(s^{\prime}))$ in (\[BDF2\]) is $\mathcal{F}_{s,s^{\prime}}^{W}\vee\mathcal{F}_{s^{\prime},T}^{\tilde{w}}%
$-adapted, where the family of $\sigma$-algebras $\mathcal{F}_{s,s^{\prime}%
}^{W}\vee\mathcal{F}_{s^{\prime},T}^{\tilde{w}}$ is neither increasing nor decreasing in $s^{\prime}$. Consequently, $Z_{s,x,y,z}(s^{\prime})$ is measurable with respect to $\mathcal{F}_{s,s^{\prime}}^{W}\vee\mathcal{F}%
_{s^{\prime},T}^{\tilde{w}}$ for every $s^{\prime}\in\lbrack s,T].$ Since $\tilde{\gamma}_{r}(s^{\prime},X_{s,x}(s^{\prime}))Y(s^{\prime})$ are independent of $\tilde{w}_{r},$ the Ito integral in (\[BDF2\]) is well defined.
\[Rem\_ant\]We remark that within the non-anticipating stochastic calculus the probabilistic representation $(\ref{FBD5})$-$(\ref{BDF2})$ for the linear problem $(\ref{os11})$-$(\ref{os14})$ cannot be carried over to the backward SNSE problem by changing the coefficient $\tilde{a}(s,x)$ to $u(s,x)$ since then the integrand $\tilde{\gamma}_{r}(s^{\prime},X_{s,x}(s^{\prime
}))Y(s^{\prime})$ would be $\mathcal{F}_{s,s^{\prime}}^{W}\vee\mathcal{F}%
_{s^{\prime},T}^{\tilde{w}}$-measurable. Nevertheless, the representation $(\ref{FBD5})$-$(\ref{BDF2})$ allows us to derive layer methods for the stochastic Oseen-Stokes equations $(\ref{os11})$-$(\ref{os14})$, and then, using them as a guidance, one can obtain layer methods for the SNSE $(\ref{NS1})$-$(\ref{NS3})$ as well $($see Sections 3.1 and 3.2$)$.
For deriving layer methods, we also use some direct probabilistic representations for solutions of the SNSE. In Sections \[prepDirect\] and \[prepDoub\] we give two such representations. The first one follows from a specific probabilistic representation for a linear SPDE which differs from $(\ref{FBD5})$-$(\ref{BDF2})$ and the second one uses backward doubly stochastic differential equations [@PP].
A direct probabilistic representation for solutions of SNSE\[prepDirect\]
-------------------------------------------------------------------------
As in the case of the stochastic Oseen-Stokes equations, we re-write the SNSE problem (\[NS1\])-(\[NS3\]) with positive direction of time into the problem with negative direction of time. Again introduce the new time variable $s=T-t$ and the functions $u(s,x):=v(T-s,x)$, $\tilde{f}(s,x):=f(T-s,x),$ $\tilde{\gamma}_{r}(s,x):=\gamma_{r}(T-s,x),$ and $\tilde{p}(s,x):=p(T-s,x).$ The corresponding backward SNSE take the form: $$\begin{gathered}
-du=(\frac{\sigma^{2}}{2}\Delta u-(u,\nabla)u-\nabla\tilde{p}+\tilde
{f})ds+\sum_{r=1}^{q}\tilde{\gamma}_{r}(s,x)\ast d\tilde{w}_{r}%
(s),\ u(T,x)=\varphi(x),\label{D1}\\
\operatorname{div}u=0\ , \label{D01}%\end{gathered}$$ with spatial periodic conditions for $u$ and $\tilde{p}$.
Introduce $F(s,x,u,\nabla u):=-(u,\nabla)u-\nabla\tilde{p}+\tilde{f}$ and write (\[D1\]) as $$-du=\left( \frac{\sigma^{2}}{2}\Delta u+F(s,x,u,\nabla u)\right)
ds+\sum_{r=1}^{q}\tilde{\gamma}_{r}(s,x)\ast d\tilde{w}_{r}%
(s),\ u(T,x)=\varphi(x). \label{D03}%$$ Let us assume that the solution $u(s,x)=u(s,x,\omega)$ to (\[D1\])-(\[D01\]) is known. We substitute it in the $F(s,x,u,\nabla u)$ which becomes a function $\tilde{F}(s,x,\omega)$ depending on $\omega$ as a parameter. Hence (\[D03\]) can be considered as a linear parabolic SPDE. For solutions of this linear SPDE, we can write the following probabilistic representation analogously to (\[FBD5\])-(\[BDF2\]) (we take $Y\equiv1$): $$\begin{gathered}
u(s,x)=E^{\tilde{w}}\varphi(X_{s,x}(T))\label{NS11n}\\
-E^{\tilde{w}}\left[ \int_{s}^{T}\{\nabla\tilde{p}(s^{\prime},X_{s,x}%
(s^{\prime}))-\tilde{f}(s^{\prime},X_{s,x}(s^{\prime})) \right. \nonumber \\
\Bigg. +(u(s^{\prime}%
,X_{s,x}(s^{\prime})),\nabla)u(s^{\prime},X_{s,x}(s^{\prime}))\}ds^{\prime
}\Bigg] \nonumber\\
+\sum_{r=1}^{q}E^{\tilde{w}}\left[ \int_{s}^{T}\tilde{\gamma}_{r}(s^{\prime
},X_{s,x}(s^{\prime}))d\tilde{w}_{r}(s^{\prime})\right] ,\nonumber\end{gathered}$$ where $X_{s,x}(s^{\prime}),$ $s^{\prime}\geq s,$ solves the system of stochastic differential equations $$dX=\sigma dW(s^{\prime}),\ X(s)=x, \label{NS12n}%$$ $W$ is a standard $n$-dimensional Wiener process independent of $\tilde{w}%
_{r}$ on the probability space $(\Omega,\mathcal{F},P).$
A probabilistic representation for solution of SNSE using backward doubly stochastic differential equations\[prepDoub\]
-----------------------------------------------------------------------------------------------------------------------
In connection with the backward SNSE (\[D1\])-(\[D01\]), we introduce the system of backward doubly stochastic differential equations [@PP]: $$\begin{aligned}
dX & =\sigma dW(s^{\prime}),\ \ X(s)=x,\label{D2} \displaybreak[0]\\
dU & =(\nabla\tilde{p}(s^{\prime},X)-\tilde{f}(s^{\prime},X)+\frac{1}%
{\sigma}\mathbb{Z}U)ds^{\prime}+\mathbb{Z}dW(s^{\prime})-\sum_{r=1}^{q}%
\tilde{\gamma}_{r}(s^{\prime},X)\ast d\tilde{w}_{r}(s^{\prime}),\label{D3} \displaybreak[0]\\
&U(T) =\varphi(X_{s,x}(T)). \label{D4}%\end{aligned}$$ In (\[D2\])-(\[D4\]) $X,$ $U,$ $W$ are column vectors of dimension $n$ and $\mathbb{Z}$ is a matrix of dimension $n\times n,\ W(s)$ and $\tilde
{w}(s),\ 0\leq s\leq T,$ are mutually independent standard Wiener processes on the probability space $(\Omega,\mathcal{F},P)$. We recall that the triple $\{X_{s,x}(s^{\prime}),U_{s,x}(s^{\prime}),\mathbb{Z}_{s,x}(s^{\prime}),s\leq
s^{\prime}\leq T\}$ is a solution of (\[D2\])-(\[D4\]) if $X_{s,x}%
(s^{\prime})$ satisfies (\[D2\]), $(U_{s,x}(s^{\prime}),\mathbb{Z}%
_{s,x}(s^{\prime}))$ for each $s^{\prime}$ is $\mathcal{F}_{s,s^{\prime}}%
^{W}\vee\mathcal{F}_{s^{\prime},T}^{\tilde{w}}$-measurable, and$$\begin{gathered}
U_{s,x}(s^{\prime})=\varphi(X_{s,x}(T))-\int_{s^{\prime}}^{T}(\nabla\tilde
{p}(s^{\prime\prime},X_{s,x}(s^{\prime\prime}))-\tilde{f}(s^{\prime\prime
},X_{s,x}(s^{\prime\prime})) \nonumber \\
+\frac{1}{\sigma}\mathbb{Z}_{s,x}(s^{\prime\prime
})U_{s,x}(s^{\prime\prime}))ds^{\prime\prime}\label{D45}\\
-\int_{s^{\prime}}^{T}\mathbb{Z}_{s,x}(s^{\prime\prime})dW(s^{\prime\prime
})+\int_{s^{\prime}}^{T}\sum_{r=1}^{q}\tilde{\gamma}_{r}(s^{\prime\prime
},X_{s,x}(s^{\prime\prime}))\ast d\tilde{w}_{r}(s^{\prime\prime}),\ s\leq
s^{\prime}\leq T.\nonumber\end{gathered}$$
Let $u(s,x)$ be a solution of the problem (\[D1\]), i.e., $$\begin{aligned}
u(s,x) & =\varphi(x)+\int_{s}^{T}(\frac{\sigma^{2}}{2}\Delta u(s^{\prime
},x)-(u,\nabla)u(s^{\prime},x)-\nabla\tilde{p}(s^{\prime},x)+\tilde
{f}(s^{\prime},x))ds^{\prime}\label{D5}\\
& +\sum_{r=1}^{q}\int_{s}^{T}\tilde{\gamma}_{r}(s^{\prime},x)\ast d\tilde
{w}_{r}(s^{\prime}).\nonumber\end{aligned}$$ It is known (see [@PP]) that then$$\begin{aligned}
X(s^{\prime}) & =X_{s,x}(s^{\prime}),\ U(s^{\prime})=U_{s,x}(s^{\prime
})=u(s^{\prime},X_{s,x}(s^{\prime})),\label{D55}\\
\mathbb{Z}(s^{\prime}) & =\mathbb{Z}_{s,x}(s^{\prime})=\{\mathbb{Z}%
^{k,j}(s^{\prime})\}=\sigma\cdot\left\{ \frac{\partial u^{k}}{\partial x^{j}%
}(s^{\prime},X_{s,x}(s^{\prime}))\right\} ,\ k,j=1,\ldots,n,\nonumber\end{aligned}$$ is a solution of (\[D2\])-(\[D4\]).
Conversely, if $X_{s,x}(s^{\prime}),$ $U_{s,x}(s^{\prime}),$ $\mathbb{Z}%
_{s,x}(s^{\prime})$ is a solution of the system of backward doubly stochastic differential equations (\[D2\])-(\[D4\]) then it can be verified that $$u(s,x)=U_{s,x}(s) \label{D9}%$$ is the solution of (\[D1\]) (see [@PP]). The condition (\[D01\]) is satisfied by choosing an appropriate pressure $\tilde{p}$.
We note that $u(s,x)$ is $\mathcal{F}_{s,T}^{\tilde{w}}$-measurable and then using (\[D45\]) we get $$\begin{aligned}
u(s,x) & =U_{s,x}(s)=E[U_{s,x}(s)|\mathcal{F}_{s,T}^{\tilde{w}}%
]=E^{\tilde{w}}U_{s,x}(s)\label{D10} \displaybreak[0]\\
& =E^{\tilde{w}}\varphi(X_{s,x}(T))\nonumber \displaybreak[0]\\
& -E^{\tilde{w}}\int_{s}^{T}(\nabla\tilde{p}(s^{\prime},X_{s,x}(s^{\prime
}))-\tilde{f}(s^{\prime},X_{s,x}(s^{\prime}))+\frac{1}{\sigma}\mathbb{Z}%
_{s,x}(s^{\prime})U_{s,x}(s^{\prime}))ds^{\prime}\nonumber \displaybreak[0]\\
& +\sum_{r=1}^{q}E^{\tilde{w}}\int_{s}^{T}\tilde{\gamma}_{r}(s^{\prime
},X_{s,x}(s^{\prime}))\ast d\tilde{w}_{r}(s^{\prime}).\nonumber\end{aligned}$$ Due to smoothness of $\tilde{\gamma}_{r}(s,x)$ in $s$ and independence of $X$ and $\tilde{w},$ the equality$$\int_{s}^{T}\tilde{\gamma}_{r}(s^{\prime},X_{s,x}(s^{\prime}))\ast d\tilde
{w}_{r}(s^{\prime})=\int_{s}^{T}\tilde{\gamma}_{r}(s^{\prime},X_{s,x}%
(s^{\prime}))d\tilde{w}_{r}(s^{\prime})$$ holds. Hence the right-hand side of (\[D10\]) coincides with the right-hand side of the probabilistic representation (\[NS11n\]).
Layer methods\[secLayer\]
=========================
In this section we construct three layer methods based on the probabilistic representations from Sections \[prepOseen\] and \[prepDirect\]. In the case of deterministic NSE (i.e., when $\gamma_{r}=0$ in the SNSE (\[NS1\])-(\[NS3\])) these methods coincide with the ones presented in [@NS5].
On the basis of the probabilistic representation (\[FBD5\])-(\[BDF2\]) we, first, construct layer methods for the stochastic Oseen-Stokes equations and, second, using the obtained methods as a guidance, we construct the corresponding methods for the SNSE (this way of deriving numerical methods for nonlinear SPDEs was proposed in [@spde]). This is done in Sections \[secSL\] and \[secSimL\]. We underline that derivation of these methods does not rely on direct probabilistic representations for the SNSE themselves that would require the anticipating stochastic calculus (see Remark \[Rem\_ant\]) which is not developed satisfactorily from the numerical point of view. That is why we prefer to use the mimicry approach here.
In Section \[secDiL\] we derive a layer method based on the direct probabilistic representation for the SNSE from Section \[prepDirect\].
In Sections \[secSL\], \[secSimL\] and \[secDiL\] we deal with approximation of velocity $v(t,x)$ (i.e., a part of the solution $v(t,x),$ $p(t,x)$ to the SNSE) only. Since we consider here the spatial-periodic problem (\[NS1\])-(\[NS3\]), we can separate approximation of velocity $v(t,x)$ and pressure $p(t,x)$ in a constructive way. Approximation of pressure is considered in Section \[secPres\].
Let us introduce a uniform partition of the time interval $[0,T]:$ $0=t_{0}<t_{1}<\cdots<t_{N}=T$ and the time step $h=T/N$ (we restrict ourselves to the uniform partition for simplicity only).
A layer method based on the standard probabilistic representation\[secSL\]
--------------------------------------------------------------------------
Each choice of $\mu(s,x)$ and $F(s,x)$ in (\[FBD5\])-(\[BDF2\]) gives us a particular probabilistic representation for the solution of the stochastic Oseen-Stokes equations (\[os11\])-(\[os14\]) which can be used for deriving the corresponding layer method. In this and the next section we derive layer methods based on two of such probabilistic representations which can be, in a sense, viewed as limiting cases of (\[FBD5\])-(\[BDF2\]). If we put $\mu(s,x)=0$ and $F(s,x)=0$ in (\[FBD5\])-(\[BDF2\]), we obtain the standard probabilistic representation for the solution to the backward linear SPDE (\[os11\])-(\[os14\]) [@R]. This case is considered in this section. The case of $F(s,x)=0$ and $\mu(s,x)$ turning the equation (\[BDF0\]) for $X(s)$ into pure diffusion is treated in the next section.
Analogously to (\[FBD5\])-(\[BDF2\]) with $\mu(s,x)=0$ and $F(s,x)=0,$ we get the following local probabilistic representation of the solution to (\[os11\])-(\[os14\]): $$\begin{aligned}
u_{a}(t_{k},x) & =E^{\tilde{w}}\left[ u_{a}(t_{k+1},X_{t_{k},x}%
(t_{k+1}))-\int_{t_{k}}^{t_{k+1}}\nabla\tilde{p}_{a}(s,X_{t_{k},x}%
(s))ds\right. \label{spr1}\\
& \left. +\int_{t_{k}}^{t_{k+1}}\tilde{f}(s,X_{t_{k},x}(s))ds+\sum_{r=1}%
^{q}\int_{t_{k}}^{t_{k+1}}\tilde{\gamma}_{r}(s,X_{t_{k},x}(s))d\tilde{w}%
_{r}(s)\right] ,\nonumber\end{aligned}$$ where $$dX=-\tilde{a}(s,X)ds+\sigma dW(s),\ X(t_{k})=x. \label{spr2}%$$
A slightly modified explicit Euler scheme with the simplest noise simulation applied to (\[spr2\]) gives$$X_{t_{k},x}(t_{k+1})\simeq\bar{X}_{t_{k},x}(t_{k+1})=x-\tilde{a}%
(t_{k+1},x)h+\sigma\sqrt{h}\xi, \label{NS15}%$$ where $\xi=(\xi^{1},\ldots,\xi^{n})^{\top}$ and$\ \xi^{1},\ldots,\xi^{n}$ are i.i.d. random variables with the law $P(\xi^{i}=\pm1)=1/2.$ We substitute $\bar{X}_{t_{k},x}(t_{k+1})$ from (\[NS15\]) in (\[spr1\]) instead of $X_{t_{k},x}(t_{k+1})$, evaluate the expectation exactly, and thus obtain (recall that $\operatorname{div}\tilde{\gamma}_{r}=0$ and $\nabla\tilde{p}%
_{a}(s,x)\in(\mathbf{V}_{p}^{m})^{\bot}):$ $$\begin{gathered}
u_{a}(t_{k},x)=\breve{u}_{a}(t_{k+1},x)-\nabla\tilde{p}_{a}(t_{k+1}%
,x)h+\tilde{f}(t_{k+1},x)h\label{NS16}\\
+\sum_{r=1}^{q}\tilde{\gamma}_{r}(t_{k+1},x)\left( \tilde{w}_{r}%
(t_{k+1})-\tilde{w}_{r}(t_{k})\right) +\rho\nonumber\\
=P\breve{u}_{a}(t_{k+1},x)+P\tilde{f}(t_{k+1},x)h+P^{\bot}\breve{u}%
_{a}(t_{k+1},x)+P^{\bot}\tilde{f}(t_{k+1},x)h\nonumber\\
-\nabla\tilde{p}_{a}%
(t_{k+1},x)h
+\sum_{r=1}^{q}\tilde{\gamma}_{r}(t_{k+1},x)\Delta_{k}\tilde{w}_{r}%
+\rho,\nonumber\end{gathered}$$ where $\Delta_{k}\tilde{w}_{r}=\tilde{w}_{r}(t_{k+1})-\tilde{w}_{r}(t_{k}),$ $r=1,\ldots,q;$ $\rho=\rho(t_{k},x)$ is a remainder, and $$\breve{u}_{a}(t_{k+1},x)=E^{\tilde{w}}u_{a}(t_{k+1},\bar{X}_{k+1})=2^{-n}%
\sum_{j=1}^{2^{n}}u_{a}(t_{k+1},x-\tilde{a}(t_{k+1},x)h+\sigma\sqrt{h}\xi_{j})
\label{NS17}%$$ with $\xi_{1}=(1,1,\ldots,1)^{\top},\ \ldots,\ \xi_{2^{n}}=(-1,-1,\ldots
,-1)^{\top}.$ Taking into account that $u_{a}(t_{k},x)$ in (\[NS16\]) is divergence free, we get$$u_{a}(t_{k},x)=P\breve{u}_{a}(t_{k+1},x)+P\tilde{f}(t_{k+1},x)h+\sum_{r=1}%
^{q}\tilde{\gamma}_{r}(t_{k+1},x)\Delta_{k}\tilde{w}_{r}+P\rho. \label{NS165}%$$ Neglecting the remainder, we get the one-step approximation for $u_{a}%
(t_{k},x)$: $$\hat{u}_{a}(t_{k},x)=P\breve{u}_{a}(t_{k+1},x)+P\tilde{f}(t_{k+1}%
,x)h+\sum_{r=1}^{q}\tilde{\gamma}_{r}(t_{k+1},x)\Delta_{k}\tilde{w}_{r}.
\label{os15n}%$$
Re-writing $\hat{u}_{a}(t_{k},x)$ of (\[os15n\]) in the positive direction of time, we obtain the one-step approximation for the velocity $v_{a}%
(t_{k},x)$ of the forward-time stochastic Oseen-Stokes equations (\[os1\])-(\[os4\]): $$\hat{v}_{a}(t_{k+1},x)=P\breve{v}_{a}(t_{k},x)+Pf(t_{k},x)h+\sum_{r=1}%
^{q}\gamma_{r}(t_{k},x)\Delta_{k}w_{r}, \label{os15}%$$ where $\Delta_{k}w_{r}=w_{r}(t_{k+1})-w_{r}(t_{k}),$ $r=1,\ldots,q,$ and $$\breve{v}_{a}(t_{k},x)=2^{-n}\sum_{j=1}^{2^{n}}v_{a}(t_{k},x-a(t_{k}%
,x)h+\sigma\sqrt{h}\xi_{j}). \label{os17}%$$
Now let us turn our attention from the stochastic Oseen-Stokes equation to the *stochastic NSE* (\[NS1\])-(\[NS3\]).
Using the one-step approximation (\[os15\])-(\[os17\]) for the stochastic Oseen-Stokes equations (\[os1\])-(\[os4\]) as a guidance, we construct the one-step approximation for the SNSE (\[NS1\])-(\[NS3\]) by substituting $a(t_{k},x)$ with $v(t_{k},x):$ $$\hat{v}(t_{k+1},x)=P\breve{v}(t_{k},x)+Pf(t_{k},x)h+\sum_{r=1}^{q}\gamma
_{r}(t_{k},x)\Delta_{k}w_{r}, \label{NSA1}%$$ where $$\breve{v}(t_{k},x)=2^{-n}\sum_{j=1}^{2^{n}}v(t_{k},x-v(t_{k},x)h+\sigma
\sqrt{h}\xi_{j}). \label{NSA3}%$$ It is easy to see that under Assumptions 2.1 $\operatorname{div}\hat
{v}(t_{k+1},x)=0.$
The corresponding layer method for the SNSE (\[NS1\])-(\[NS3\]) has the form$$\begin{gathered}
\bar{v}(0,x)=\varphi(x),\ \bar{v}(t_{k+1},x)=P\breve{v}(t_{k},x)+Pf(t_{k}%
,x)h+\sum_{r=1}^{q}\gamma_{r}(t_{k},x)\Delta_{k}w_{r},\label{NS18}\\
k=0,\ldots,N-1,\nonumber\end{gathered}$$ where$$\breve{v}(t_{k},x)=2^{-n}\sum_{j=1}^{2^{n}}\bar{v}(t_{k},x-\bar{v}%
(t_{k},x)h+\sigma\sqrt{h}\xi_{j}). \label{NS19}%$$ We note that we use the same notation $\breve{v}(t_{k},x)$ for the functions appearing in the one-step approximation (\[NSA3\]) and in the layer method (\[NS19\]) but this does not cause any confusion.
Knowing the expansions $$\begin{aligned}
\breve{v}(t_{k},x) & =\sum_{\mathbf{n}\in\mathbf{Z}^{n}}\breve
{v}_{\mathbf{n}}(t_{k})e^{i(2\pi/L)(\mathbf{n},x)},\ \ \ f(t_{k}%
,x)=\sum_{\mathbf{n}\in\mathbf{Z}^{n}}f_{\mathbf{n}}(t_{k})e^{i(2\pi
/L)(\mathbf{n},x)},\label{NS20}\\
\gamma_{r}(t_{k},x) & =\sum_{\mathbf{n}\in\mathbf{Z}^{n}}\gamma
_{r,\mathbf{n}}(t_{k})e^{i(2\pi/L)(\mathbf{n},x)},\nonumber\end{aligned}$$ it is not difficult to find $\bar{v}(t_{k+1},x)$. Indeed, using (\[N00\]) and (\[N01\]), we obtain from (\[NS18\])-(\[NS19\]): $$\begin{gathered}
\bar{v}(t_{k+1},x)=\sum_{\mathbf{n}\in\mathbf{Z}^{n}}\bar{v}_{\mathbf{n}%
}(t_{k+1})e^{i(2\pi/L)(\mathbf{n},x)},\ \ \label{NS21}\\
\bar{v}_{\mathbf{n}}(t_{k+1})=\breve{v}_{\mathbf{n}}(t_{k})+f_{\mathbf{n}%
}(t_{k})h-\frac{\breve{v}_{\mathbf{n}}^{\top}(t_{k})\mathbf{n}}{|\mathbf{n}%
|^{2}}\mathbf{n}-h\frac{f_{\mathbf{n}}^{\top}(t_{k})\mathbf{n}}{|\mathbf{n}%
|^{2}}\mathbf{n}+\sum_{r=1}^{q}\gamma_{r,\mathbf{n}}(t_{k})\ \Delta_{k}%
w_{r}\mathbf{.}\nonumber\end{gathered}$$ We note that turning the layer method (\[NS18\])-(\[NS19\]) into a numerical algorithm requires to complement it with an interpolation in order to compute the terms $\bar{v}(t_{k},x-\bar{v}(t_{k},x)h+\sigma\sqrt{h}\xi
_{j})$ in (\[NS19\]) used for finding $\breve{v}_{\mathbf{n}}(t_{k})$ from (\[NS20\]), see the corresponding discussion in the case of deterministic NSE in [@NS5].
Layer methods based on the probabilistic representation with simplest characteristics\[secSimL\]
------------------------------------------------------------------------------------------------
If we put $\mu(s,x)=-\tilde{a}(s,x)/\sigma$ and $F(s,x)=0$ in (\[FBD5\])-(\[BDF2\]), we can obtain the following local probabilistic representation for the solution to the backward stochastic Oseen-Stokes equation (\[os11\])-(\[os14\]):$$\begin{gathered}
u_{a}(t_{k},x)=E^{\tilde{w}}[u_{a}(t_{k+1},X_{t_{k},x}(t_{k+1}))Y_{t_{k}%
,x,1}(t_{k+1})]\label{NS14}\\
+E^{\tilde{w}}\left[ -\int_{t_{k}}^{t_{k+1}}\nabla\tilde{p}_{a}(s,X_{t_{k}%
,x}(s))Y_{t_{k},x,1}(s)ds+\int_{t_{k}}^{t_{k+1}}\tilde{f}(s,X_{t_{k}%
,x}(s))Y_{t_{k},x,1}(s)ds\right. \nonumber\\
\left. +\sum_{r=1}^{q}\int_{t_{k}}^{t_{k+1}}\tilde{\gamma}_{r}(s,X_{t_{k}%
,x}(s))Y_{t_{k},x,1}(s)d\tilde{w}_{r}(s)\right] \ ,\nonumber\end{gathered}$$ where $X_{t,x}(s),$ $Y_{t,x,1}(s),$ $s\geq t,$ solve the system of stochastic differential equations $$\begin{aligned}
dX & =\sigma dW(s),\ X(t)=x,\label{NS12}\\
dY & =-\frac{1}{\sigma}Y\tilde{a}^{\top}(s,X)dW(s),\ Y(t)=1. \label{NS13}%\end{aligned}$$ We apply a slightly modified explicit Euler scheme with the simplest noise simulation to (\[NS12\])-(\[NS13\]): $$\bar{X}_{t_{k},x}(t_{k+1})=x+\sigma\sqrt{h}\xi,\ \bar{Y}_{t_{k},x,1}%
(t_{k+1})=1-\frac{1}{\sigma}\tilde{a}^{\top}(t_{k+1},x)\sqrt{h}\xi,
\label{NS30}%$$ where $\xi$ is the same as in (\[NS15\]). Approximating $X_{t_{k},x}%
(t_{k+1})$ and $Y_{t_{k},x,1}(t_{k+1})$ in (\[NS14\]) by $\bar{X}_{t_{k}%
,x}(t_{k+1})$ and $\bar{Y}_{t_{k},x,1}(t_{k+1})$ from (\[NS30\]), we obtain$$\begin{gathered}
u_{a}(t_{k},x)=E^{\tilde{w}}[u_{a}(t_{k+1},x+\sigma\sqrt{h}\xi)(1-\frac
{1}{\sigma}\tilde{a}^{\top}(t_{k+1},x)\sqrt{h}\xi)]-\nabla\tilde{p}%
_{a}(t_{k+1},x)h\label{NS31}\\
+\tilde{f}(t_{k+1},x)h+\sum_{r=1}^{q}\tilde{\gamma}_{r}(t_{k+1},x)\Delta
_{k}\tilde{w}_{r}+\rho\nonumber\\
=2^{-n}\sum_{q=1}^{2^{n}}u_{a}(t_{k+1},x+\sigma\sqrt{h}\xi_{q})-\frac{\sqrt
{h}}{\sigma}\breve{u}_{a}(t_{k+1},x)-\nabla\tilde{p}_{a}(t_{k+1},x)h \nonumber\\
+\tilde{f}(t_{k+1},x)h
+\sum_{r=1}^{q}\tilde{\gamma}_{r}(t_{k+1},x)\Delta_{k}\tilde{w}_{r}%
+\rho,\nonumber\end{gathered}$$ where $$\begin{aligned}
\breve{u}_{a}(t_{k+1},x) & =E^{\tilde{w}}[u_{a}(t_{k+1},x+\sigma\sqrt{h}%
\xi)\xi^{\top}]\tilde{a}(t_{k+1},x)\label{NS32}\\
& =2^{-n}\sum_{j=1}^{2^{n}}u_{a}(t_{k+1},x+\sigma\sqrt{h}\xi_{j})\xi
_{j}^{\top}\tilde{a}(t_{k+1},x)\nonumber\end{aligned}$$ and $\rho=\rho(t_{k},x)$ is a remainder.
Using the Helmholtz-Hodge-Leray decomposition and taking into account that $$\operatorname{div}u_{a}(t_{k+1},x+\sigma\sqrt{h}\xi_{q})=0,\text{\ \ }%
\operatorname{div}\gamma_{r}=0,$$ we get from (\[NS31\])-(\[NS32\]): $$\begin{gathered}
u_{a}(t_{k},x)=2^{-n}\sum_{j=1}^{2^{n}}u_{a}(t_{k+1},x+\sigma\sqrt{h}\xi
_{j})-\frac{\sqrt{h}}{\sigma}P\breve{u}_{a}(t_{k+1},x)+P\tilde{f}%
(t_{k+1},x)h\\
-\frac{\sqrt{h}}{\sigma}P^{\bot}\breve{u}_{a}(t_{k+1},x)+P^{\bot}\tilde
{f}(t_{k+1},x)h-\nabla\tilde{p}_{a}(t_{k+1},x)h\\
+\sum_{r=1}^{q}\tilde{\gamma}_{r}(t_{k+1},x)\Delta_{k}\tilde{w}_{r}+\rho,\end{gathered}$$ whence we obtain after applying the operator $P:$ $$\begin{gathered}
u_{a}(t_{k},x)=2^{-n}\sum_{j=1}^{2^{n}}u_{a}(t_{k+1},x+\sigma\sqrt{h}\xi
_{j})-\frac{\sqrt{h}}{\sigma}P\breve{u}_{a}(t_{k+1},x)+P\tilde{f}%
(t_{k+1},x)h\label{NS33} \displaybreak[0]\\
+\sum_{r=1}^{q}\tilde{\gamma}_{r}(t_{k+1},x)\Delta_{k}\tilde{w}_{r}%
+P\rho.\nonumber\end{gathered}$$ Dropping the remainder in (\[NS33\]) and re-writing the obtained approximation in the one with positive direction of time, we obtain the one-step approximation for the forward-time stochastic Oseen-Stokes equation (\[os1\])-(\[os4\]): $$\begin{gathered}
\hat{v}_{a}(t_{k+1},x)=2^{-n}\sum_{j=1}^{2^{n}}v_{a}(t_{k},x+\sigma\sqrt{h}%
\xi_{j})-\frac{\sqrt{h}}{\sigma}P\breve{v}_{a}(t_{k},x)+Pf(t_{k}%
,x)h\label{os18}\\
+\sum_{r=1}^{q}\gamma_{r}(t_{k},x)\Delta_{k}w_{r},\nonumber\end{gathered}$$ where$$\breve{v}_{a}(t_{k},x)=2^{-n}\sum_{j=1}^{2^{n}}v_{a}(t_{k},x+\sigma\sqrt{h}%
\xi_{j})\xi_{j}^{\top}a(t_{k},x). \label{os20}%$$ Using (\[os18\])-(\[os20\]) as a guidance, we arrive at the one-step approximation for the SNSE (\[NS1\])-(\[NS3\]): $$\begin{gathered}
\hat{v}(t_{k+1},x)=2^{-n}\sum_{q=1}^{2^{n}}v(t_{k},x+\sigma\sqrt{h}\xi
_{q})-\frac{\sqrt{h}}{\sigma}P\breve{v}(t_{k},x) \label{NSA4} \\
+Pf(t_{k},x)h+\sum_{r=1}%
^{q}\gamma_{r}(t_{k},x)\Delta_{k}w_{r}, \nonumber\end{gathered}$$ where$$\breve{v}(t_{k},x)=2^{-n}\sum_{j=1}^{2^{n}}v(t_{k},x+\sigma\sqrt{h}\xi_{j}%
)\xi_{j}^{\top}v(t_{k},x). \label{NSA6}%$$ It is easy to see that under Assumptions 2.1 $\operatorname{div}\hat
{v}(t_{k+1},x)=0.$ The corresponding layer method for the SNSE (\[NS1\])-(\[NS3\]) has the form$$\begin{gathered}
\bar{v}(0,x)=\varphi(x),\ \bar{v}(t_{k+1},x)=2^{-n}\sum_{j=1}^{2^{n}}\bar
{v}(t_{k},x+\sigma\sqrt{h}\xi_{j})-\frac{\sqrt{h}}{\sigma}P\breve{v}%
(t_{k},x)\label{NSM21} \displaybreak[0]\\
+Pf(t_{k},x)h+\sum_{r=1}^{q}\gamma_{r}(t_{k},x)\Delta_{k}w_{r},\ \ k=0,\ldots
,N-1,\nonumber\end{gathered}$$ where $$\breve{v}(t_{k},x)=2^{-n}\sum_{j=1}^{2^{n}}\bar{v}(t_{k},x+\sigma\sqrt{h}%
\xi_{j})\xi_{j}^{\top}\bar{v}(t_{k},x). \label{NSM23}%$$
Practical implementation of the layer method (\[NSM21\])-(\[NSM23\]) is straightforward and efficient. Let us write the corresponding numerical algorithm for simplicity in the two-dimensional ($n=2)$ case. We choose a positive integer $M$ as a cut-off frequency and write the approximate velocity at the time $t_{k+1}$ as the partial sum: $$\bar{v}(t_{k+1},x)=\sum_{n_{1}=-M}^{M-1}\sum_{n_{2}=-M}^{M-1}\bar
{v}_{\mathbf{n}}(t_{k+1})e^{i(2\pi/L)(\mathbf{n},x)}, \label{al1}%$$ where $\mathbf{n}=(n_{1},n_{2})^{\top}.$
We note that we use the same notation $\bar{v}(t_{k+1},x)$ for the partial sum in (\[al1\]) instead of writing $\bar{v}_{M}(t_{k+1},x)$ while in (\[NSM21\]) $\bar{v}(t_{k+1},x)$ denotes the approximate velocity containing all frequencies but this should not lead to any confusion.
Further, we have $$\frac{1}{4}\sum_{j=1}^{4}\bar{v}(t_{k},x+\sigma\sqrt{h}\xi_{j})=\sum
_{n_{1}=-M}^{M-1}\sum_{n_{2}=-M}^{M-1}\bar{v}_{\mathbf{n}}(t_{k}%
)e^{i(2\pi/L)(\mathbf{n},x)}\frac{1}{4}\sum_{j=1}^{4}e^{i(2\pi\sigma\sqrt
{h}/L)(\mathbf{n},\xi_{j})}. \label{al11}%$$ Then $$\begin{aligned}
\breve{v}(t_{k},x) & =\frac{1}{4}\sum_{j=1}^{4}\bar{v}(t_{k},x+\sigma
\sqrt{h}\xi_{j})\xi_{j}^{\top}\bar{v}(t_{k},x)\\
& =\sum_{n_{1}=-N}^{M-1}\sum_{n_{2}=-N}^{M-1}\bar{v}_{\mathbf{n}}%
(t_{k})e^{i(2\pi/L)(\mathbf{n},x)}\frac{1}{4}\sum_{j=1}^{4}e^{i(2\pi
\sigma\sqrt{h}/L)(\mathbf{n},\xi_{j})}\xi_{j}^{\top}\bar{v}(t_{k},x)\\
& =\sum_{n_{1}=-M}^{M-1}\sum_{n_{2}=-M}^{M-1}V_{\mathbf{n}}(t_{k}%
)e^{i(2\pi/L)(\mathbf{n},x)}\bar{v}(t_{k},x),\end{aligned}$$ where $$V_{\mathbf{n}}(t_{k})=\bar{v}_{\mathbf{n}}(t_{k})\cdot\frac{1}{4}\sum
_{j=1}^{4}e^{i(2\pi\sigma\sqrt{h}/L)(\mathbf{n},\xi_{j})}\xi_{j}^{\top}.$$ Note that $V_{\mathbf{n}}(t_{k})$ is a $2\times2$-matrix. Let $$V(t_{k},x):=\sum_{n_{1}=-M}^{M-1}\sum_{n_{2}=-M}^{M-1}V_{\mathbf{n}}%
(t_{k})e^{i(2\pi/L)(\mathbf{n},x)} \label{ext}%$$ then $$\breve{v}(t_{k},x)=V(t_{k},x)\bar{v}(t_{k},x).$$
We obtain the algorithm: $$\begin{aligned}
\bar{v}_{\mathbf{n}}(0) & =\varphi_{\mathbf{n}},\ \label{alg2} \displaybreak[0]\\
\bar{v}_{\mathbf{n}}(t_{k+1}) & =\bar{v}_{\mathbf{n}}(t_{k})-\frac{\sqrt{h}%
}{\sigma}\left( \breve{v}_{\mathbf{n}}(t_{k})-\frac{\breve{v}_{\mathbf{n}%
}^{\top}(t_{k})\mathbf{n}}{|\mathbf{n}|^{2}}\mathbf{n}\right) +f_{\mathbf{n}%
}(t_{k})h-h\frac{f_{\mathbf{n}}^{\top}(t_{k})\mathbf{n}}{|\mathbf{n}|^{2}%
}\mathbf{n} \nonumber \displaybreak[0]\\
&+\sum_{r=1}^{q}\gamma_{r,\mathbf{n}}(t_{k})\ \Delta_{k}%
w_{r},\nonumber\end{aligned}$$ where $$\breve{v}_{\mathbf{n}}(t_{k})=(\breve{v}(t_{k},x))_{\mathbf{n}}=\left(
V(t_{k},x)\bar{v}(t_{k},x)\right) _{\mathbf{n}}. \label{alg3}%$$ To find $\breve{v}_{\mathbf{n}}(t_{k})$ one can either multiply two partial sums of the form (\[al1\]) and (\[ext\]) or exploit fast Fourier transform in the usual fashion (see, e.g. [@Can98]) to speed up the algorithm. The algorithm (\[alg2\]) can be viewed as analogous to spectral methods. It is interesting that the layer method (\[NSM21\])-(\[NSM23\]) is, on the one hand, related to a finite difference scheme (see below) and on the other hand, to spectral methods.
\[remFD\]Let us discuss a relationship between the layer method $(\ref{NSM21})$-$(\ref{NSM23})$ and finite difference methods. For simplicity in writing, we give this illustration in the two-dimensional case. It is not difficult to notice that the two-dimensional analog of the layer approximation $(\ref{NSM21})$ can be re-written as the following finite difference scheme for the SNSE $(\ref{NS1})$-$(\ref{NS3})$: $$\begin{aligned}
& \frac{\bar{v}(t_{k+1},x)-\bar{v}(t_{k},x)}{h}\label{fd1}\\
& =\frac{\bar{v}(t_{k},x^{1}+\sigma\sqrt{h},x^{2}+\sigma\sqrt{h})+\bar
{v}(t_{k},x^{1}-\sigma\sqrt{h},x^{2}+\sigma\sqrt{h})-4\bar{v}(t_{k}%
,x^{1},x^{2})}{4h}\nonumber\\
& +\frac{\bar{v}(t_{k},x^{1}+\sigma\sqrt{h},x^{2}-\sigma\sqrt{h})+\bar
{v}(t_{k},x^{1}-\sigma\sqrt{h},x^{2}-\sigma\sqrt{h})}{4h}\nonumber\\
& -\frac{1}{\sigma\sqrt{h}}P\breve{v}(t_{k},x)+Pf(t_{k},x)+\sum_{r=1}%
^{q}\gamma_{r}(t_{k},x)\frac{\Delta w_{r}(t_{k+1})}{h}\nonumber\end{aligned}$$ with $$\begin{aligned}
\frac{\breve{v}(t_{k},x)}{\sigma\sqrt{h}} & =\bar{v}^{1}(t_{k},x)\frac
{\bar{v}(t_{k},x^{1}+\sigma\sqrt{h},x^{2}+\sigma\sqrt{h})-\bar{v}(t_{k}%
,x^{1}-\sigma\sqrt{h},x^{2}+\sigma\sqrt{h})}{4\sigma\sqrt{h}}\label{fd2}\\
& +\bar{v}^{1}(t_{k},x)\frac{\bar{v}(t_{k},x^{1}+\sigma\sqrt{h},x^{2}%
-\sigma\sqrt{h})-\bar{v}(t_{k},x^{1}-\sigma\sqrt{h},x^{2}-\sigma\sqrt{h}%
)}{4\sigma\sqrt{h}}\nonumber\\
& +\bar{v}^{2}(t_{k},x)\frac{\bar{v}(t_{k},x^{1}+\sigma\sqrt{h},x^{2}%
+\sigma\sqrt{h})-\bar{v}(t_{k},x^{1}+\sigma\sqrt{h},x^{2}-\sigma\sqrt{h}%
)}{4\sigma\sqrt{h}}\nonumber\\
& +\bar{v}^{2}(t_{k},x)\frac{\bar{v}(t_{k},x^{1}-\sigma\sqrt{h},x^{2}%
+\sigma\sqrt{h})-\bar{v}(t_{k},x^{1}-\sigma\sqrt{h},x^{2}-\sigma\sqrt{h}%
)}{4\sigma\sqrt{h}}\ \ .\nonumber\end{aligned}$$ As one can see, $\bar{v}(t_{k},\cdot)$ in the right-hand side of $(\ref{fd1})$ is evaluated at the nodes $(x^{1},x^{2}),$ $(x^{1}\pm\sigma\sqrt{h},x^{2}%
\pm\sigma\sqrt{h})$, which is typical for a standard explicit finite difference scheme with the space discretization step $h_{x}$ taken equal to $\sigma\sqrt{h}$ and $h$ being the time-discretization step. We also note that if in the approximation $(\ref{NS30})$ we choose a different random vector $\xi$ than in $(\ref{NS15})$ then we can obtain another layer method for the SNSE which can be again re-written as a finite difference scheme (see such a discussion in the case of the deterministic NSE in [@NS5]).
We recall [@M1; @MT1; @spde] that convergence theorems for layer methods (in comparison with the theory of finite difference methods) do not contain any conditions on stability of their approximations. In layer methods we do not need to a priori prescribe space nodes: they are obtained automatically depending on choice of a probabilistic representation and a numerical scheme. We note that our error analysis for the layer methods (see Section \[secER\]) immediately implies the same error estimates for the corresponding finite difference scheme $(\ref{fd1})$.
It is not difficult to see from $(\ref{fd2})$ that $$(\bar{v}(t_{k},x),\nabla)\bar{v}(t_{k},x)\approx\frac{\breve{v}(t_{k}%
,x)}{\sigma\sqrt{h}}\ . \label{fd3}%$$ If we put the exact $v(t_{k},x)$ in $(\ref{fd3})$ (both in its left and right-hand sides) instead of the approximate $\bar{v}(t_{k},x)$ then the accuracy of the approximation in $(\ref{fd3})$ is of order $O(h).$ This observation is helpful for understanding a relationship between the layer methods from this and the next section (see Remark \[Remfd2\] at the end of the next section).
A layer method based on the direct probabilistic representation\[secDiL\]
-------------------------------------------------------------------------
The local version of probabilistic representation (\[NS11n\])-(\[NS12n\]) for the solution to the backward SNSE (\[D1\])-(\[D01\]) has the form: $$\begin{gathered}
u(t_{k},x)=E^{\tilde{w}}u(t_{k+1},X_{t_{k},x}(t_{k+1}))\label{NS13n}\\
-E^{\tilde{w}}\left[ \int_{t_{k}}^{t_{k+1}}\{\nabla\tilde{p}(s^{\prime
},X_{t_{k},x}(s^{\prime}))-\tilde{f}(s^{\prime},X_{t_{k},x}(s^{\prime
})) \right. \nonumber \\
\Bigg.+(u(s^{\prime},X_{t_{k},x}(s^{\prime})),\nabla)u(s^{\prime},X_{t_{k}%
,x}(s^{\prime}))\}ds^{\prime}\Bigg] \nonumber\\
+\sum_{r=1}^{q}E^{\tilde{w}}\left[ \int_{t_{k}}^{t_{k+1}}\tilde{\gamma}%
_{r}(s^{\prime},X_{t_{k},x}(s^{\prime}))d\tilde{w}_{r}(s^{\prime})\right]
.\nonumber\end{gathered}$$
Using (\[NS13n\]), we construct the one-step approximation of the solution to the backward SNSE (\[D1\])-(\[D01\]):$$\begin{aligned}
u(t_{k},x) & =E^{\tilde{w}}u(t_{k+1},X_{t_{k},x}(t_{k+1}))-h\{\nabla
\tilde{p}(t_{k+1},x)-\tilde{f}(t_{k+1},x)\label{DL2} \displaybreak[0] \\
& +(u(t_{k+1},x),\nabla)u(t_{k+1},x)\}+\sum_{r=1}^{q}\tilde{\gamma}%
_{r}(t_{k+1},x)\Delta_{k}\tilde{w}_{r}+\rho\nonumber \displaybreak[0]\\
& =2^{-n}\sum_{j=1}^{2^{n}}u(t_{k+1},x+\sigma\sqrt{h}\xi_{j})\nonumber \displaybreak[0]\\
& -h\{\nabla\tilde{p}(t_{k+1},x)-\tilde{f}(t_{k+1},x)+(u(t_{k+1}%
,x),\nabla)u(t_{k+1},x)\}\nonumber \displaybreak[0]\\
& +\sum_{r=1}^{q}\tilde{\gamma}_{r}(t_{k+1},x)\Delta_{k}\tilde{w}_{r}%
+\rho,\nonumber\end{aligned}$$ where $\rho=\rho(t_{k},x)$ is a remainder.
Using the Helmholtz-Hodge-Leray decomposition and taking into account that $\operatorname{div}u(t_{k+1},x+\sigma\sqrt{h}\xi_{q})=0$ and $\operatorname{div}\gamma_{r}=0,$ we get from (\[DL2\]): $$\begin{gathered}
u(t_{k},x)=2^{-n}\sum_{j=1}^{2^{n}}u(t_{k+1},x+\sigma\sqrt{h}\xi
_{j})-P[(u(t_{k+1},x),\nabla)u(t_{k+1},x)]h
\label{DL30}\\
+P\tilde{f}(t_{k+1},x)h-P^{\bot}[(u(t_{k+1},x),\nabla)u(t_{k+1},x)]h+P^{\bot}\tilde{f}(t_{k+1}%
,x)h\nonumber\\
-\nabla\tilde{p}(t_{k+1},x)h+\sum_{r=1}^{q}\tilde{\gamma}_{r}(t_{k+1},x)\Delta_{k}\tilde{w}_{r}%
+\rho,\nonumber\end{gathered}$$ whence we obtain after applying the operator $P:$$$\begin{aligned}
u(t_{k},x) & =2^{-n}\sum_{j=1}^{2^{n}}u(t_{k+1},x+\sigma\sqrt{h}\xi
_{j})-P[(u(t_{k+1},x),\nabla)u(t_{k+1},x)]h
\label{DL3}\\
& +P\tilde{f}(t_{k+1},x)h+\sum_{r=1}^{q}\tilde{\gamma}_{r}(t_{k+1},x)\Delta_{k}\tilde{w}_{r}%
+P\rho.\nonumber\end{aligned}$$ We re-write (\[DL30\])-(\[DL3\]) for the forward-time SNSE (\[NS1\])-(\[NS3\]): $$\begin{gathered}
v(t_{k+1},x)=2^{-n}\sum_{j=1}^{2^{n}}v(t_{k},x+\sigma\sqrt{h}\xi
_{j})-P[(v(t_{k},x),\nabla)v(t_{k},x)]h\label{DL30n}\\
+Pf(t_{k},x)h-P^{\bot}[(v(t_{k},x),\nabla)v(t_{k},x)]h+P^{\bot}f(t_{k},x)h\nonumber\\
-\nabla
p(t_{k},x)h+\sum_{r=1}^{q}\gamma_{r}(t_{k},x)\Delta_{k}w_{r}+\rho\nonumber\end{gathered}$$ and $$\begin{aligned}
v(t_{k+1},x) & =2^{-n}\sum_{j=1}^{2^{n}}v(t_{k},x+\sigma\sqrt{h}\xi
_{j})-P\left[ (v(t_{k},x),\nabla)v(t_{k},x)\right] h\label{DL3n}\\
& +Pf(t_{k},x)h+\sum_{r=1}^{q}\gamma_{r}(t_{k},x)\Delta_{k}w_{r}%
+P\rho.\nonumber\end{aligned}$$ Dropping the remainder in (\[DL3n\]), we obtain the one-step approximation for the velocity $v(t_{k+1},x)$ in (\[NS1\])-(\[NS3\]): $$\begin{aligned}
\hat{v}(t_{k+1},x) & =2^{-n}\sum_{j=1}^{2^{n}}v(t_{k},x+\sigma\sqrt{h}%
\xi_{j})-P\left[ (v(t_{k},x),\nabla)v(t_{k},x)\right] h\label{DL5}\\
& +Pf(t_{k},x)h+\sum_{r=1}^{q}\gamma_{r}(t_{k},x)\Delta_{k}w_{r}.\nonumber\end{aligned}$$ It is easy to see that under Assumptions 2.1 $\operatorname{div}\hat
{v}(t_{k+1},x)=0.$ The corresponding layer method for the velocity of the SNSE (\[NS1\])-(\[NS3\]) has the form$$\begin{gathered}
\bar{v}(0,x)=\varphi(x), \label{DL7} \displaybreak[0]\\
\bar{v}(t_{k+1},x)=2^{-n}\sum_{j=1}^{2^{n}}\bar
{v}(t_{k},x+\sigma\sqrt{h}\xi_{j})-P\left[ (\bar{v}(t_{k},x),\nabla)\bar
{v}(t_{k},x)\right] h \nonumber \displaybreak[0]\\
+Pf(t_{k},x)h+\sum_{r=1}^{q}\gamma_{r}(t_{k},x)\Delta_{k}w_{r},\ \ k=0,\ldots
,N-1.\nonumber\end{gathered}$$ This method can be turned into a numerical algorithm analogously to how we constructed the numerical algorithm (\[alg2\]) based on the layer method (\[NSM21\]) in Section \[secSimL\].
\[Remfd2\]It is interesting to note (see also $(\ref{fd2})$ and $(\ref{fd3})$) the relationship between the methods $(\ref{NSM21})$ and $(\ref{DL7})$: $\sqrt{h}\breve{v}(t_{k},x)/\sigma$ from $(\ref{NSA4}%
)$-$(\ref{NSA6})$ is a finite-difference approximation of the term $(\bar
{v}(t_{k},x),\nabla)\bar{v}(t_{k},x)h$ in $(\ref{DL7})$. We remark that this finite difference naturally arises via the probabilistic approach. It is useful to have both methods in the arsenal of layer methods for SNSE: while the method $(\ref{DL7})$ has a smaller one-step error than $(\ref{NSM21}),$ it requires evaluation of spatial derivatives of $\bar{v}(t_{k},x)$.
Approximation of pressure\[secPres\]
------------------------------------
In the previous sections we constructed numerical methods for velocity $v(t,x),$ in this section we propose approximations for pressure $p(t,x)$.
Applying the projection operator $P^{\bot}$ to SNSE (\[NS1\])-(\[NS3\]), we get (see also (\[NS03\])): $$\nabla p(t,x)=-P^{\bot}\left[ (v(t,x),\nabla)v(t,x)\right] +P^{\bot}f(t,x).
\label{pre1}%$$ Based on (\[pre1\]), we complement the layer method (\[DL7\]) for the velocity by the approximation of pressure as follows: $$\nabla\bar{p}(t_{k+1},x)=-P^{\bot}\left[ (\bar{v}(t_{k+1},x),\nabla)\bar
{v}(t_{k+1},x)\right] +P^{\bot}f(t_{k+1},x). \label{DL7p}%$$
As a result, we obtain *the layer method* (\[DL7\]), (\[DL7p\]) for the solution of SNSE (\[NS1\])-(\[NS3\]).
It is clear that the numerical error $\nabla\bar{p}(t_{k+1},x)-\nabla p(t,x)$ is of the same order as the global errors of $\bar{v}(t_{k+1},x)$ and $\nabla\bar{v}(t_{k+1},x).$ We note that in (\[DL7p\]) to evaluate pressure at time $t_{k+1}$ we use velocity at time $t_{k+1},$ i.e., the updated velocity.
We observe that $\rho$ in $(\ref{DL30n})$ is such that $P^{\bot}\rho=0.$ Indeed, it follows from $(\ref{DL30n})$-$(\ref{DL3n})$ (with $t_{k+1}$ instead of $t_{k}$) that $$\nabla p(t_{k+1},x)=-P^{\bot}\left[ (v(t_{k+1},x),\nabla)v(t_{k+1},x)\right]
+P^{\bot}f(t_{k+1},x)+P^{\bot}\rho. \label{pre2}%$$ Comparing $(\ref{pre1})$ and $(\ref{pre2})$, we get $P^{\bot}\rho=0.$
Let us now return to the layer method (\[NSM21\]) for velocity. We have to complement it with an approximation of pressure. To this end, we approximate (see Remark \[Remfd2\] and (\[fd3\])) the term $(\bar{v}(t_{k+1}%
,x),\nabla)\allowbreak\bar{v}(t_{k+1},x)$ in (\[DL7p\]) by $\breve
{v}(t_{k+1},x)/\sigma\sqrt{h}$ with $\breve{v}(t_{k+1},x)$ from (\[NSM23\]) (with $t_{k+1}$ instead of $t_{k})$. We obtain $$\nabla\bar{p}(t_{k+1},x)=-\frac{1}{\sigma\sqrt{h}}P^{\bot}\breve{v}%
(t_{k+1},x)+P^{\bot}f(t_{k+1},x), \label{NSMp}%$$ where $\breve{v}(t_{k+1},x)$ is from (\[NSM23\]). Note that in the velocity approximation (\[NSM21\]) we use $\breve{v}(t_{k},x)$ while in the pressure approximation (\[NSMp\]) we use $\breve{v}(t_{k+1},x).$
As a result, we obtain *the layer method* (\[NSM21\])-(\[NSM23\]), (\[NSMp\]) for the solution of SNSE (\[NS1\])-(\[NS3\]).
We remark that the layer method (\[NS18\]) for velocity can be completed by approximating the pressure as it is either in (\[DL7p\]) with $\bar
{v}(t_{k+1},x)$ found due to (\[NS18\]) or in (\[NSMp\]) but with $\breve{v}(t_{k+1},x)$ from (\[NSM23\]) using $\bar{v}(t_{k+1},x)$ found due to (\[NS18\]).
To provide an example of an algorithm involving an approximation of pressure, let us return to the algorithm (\[alg2\]) for velocity. Based on (\[NSMp\]) (see also (\[N01\])), we obtain $$\bar{p}_{\mathbf{n}}(t_{k+1})=i\frac{L}{2\pi}\left( \frac{\breve
{v}_{\mathbf{n}}^{\top}(t_{k+1})\mathbf{n}}{\sigma\sqrt{h}|\mathbf{n}|^{2}%
}-\frac{f_{\mathbf{n}}^{\top}(t_{k+1})\mathbf{n}}{|\mathbf{n}|^{2}}\right)
,\ \ \mathbf{n\neq0,\ }\bar{p}_{\mathbf{0}}(t_{k+1})=0, \label{algp}%$$ where $\breve{v}_{\mathbf{n}}^{\top}(t_{k+1})$ are as in (\[alg3\]) with $t_{k+1}$ instead of $t_{k}$.
As a result, we obtain *the algorithm* (\[alg2\])-(\[alg3\]), (\[algp\]) for the solution of SNSE (\[NS1\])-(\[NS3\]) which corresponds to the layer method (\[NSM21\])-(\[NSM23\]), (\[NSMp\]).
Analogously, one can obtain algorithms corresponding to the other two layer methods considered in the paper.
Error analysis\[secER\]
=======================
In this section we provide theoretical support for the numerical methods from the previous section. For definiteness, we consider the layer method (\[NS18\]). Analogous results can be obtained for the other two layer methods proposed in Sections \[secSimL\] and \[secDiL\].
As before, $||u(\cdot)||=||u(x)||$ denotes the $\mathbf{L}^{2}$-norm of a function $u(x),$ $x\in Q.$ In this section we use the same letter $K$ for various deterministic constants and $C=C(\omega)$ for various positive random variables.
We start with analysis of the local mean-square error.
\[lemonest\]Let Assumptions 2.1 hold with $m_{0}>6$. The one-step error $$\rho(t_{k+1},x)=\hat{v}(t_{k+1},x)-v(t_{k+1},x) \label{onesterr}%$$ of the one-step approximation $(\ref{NSA1})$-$(\ref{NSA3})$ for the SNSE $(\ref{NS1})$-$(\ref{NS3})$ is estimated as $$||E(\rho(t_{k+1},x)|\mathcal{F}_{t_{k}}^{w})||\leq C(\omega)h^{2}%
,\ \label{lm1}%$$ and for $1\leq p<p_{0}$ $$\left( E||\rho(t_{k+1},\cdot)||^{2p}\right) ^{1/2p}\leq Kh^{3/2}%
,\ \label{lm22}%$$ where a random constant $C(\omega)>0$ with $EC^{2}<\infty$ does not depend on $h$ and $k,$ a deterministic constant $K>0$ does not depend on $h$ and $k$ but depends on $p,$ and $p_{0}=p_{0}(m_{0})>1$ is a positive number or $p_{0}=\infty.$
**Proof.** Using Assumptions 2.1, we expand the right-hand side of (\[NSA3\]), substitute the outcome in (\[NSA1\]), and obtain $$\begin{aligned}
\hat{v}(t_{k+1},x) & =v(t_{k},x)-hP\left[ (v(t_{k},x),\nabla)v(t_{k}%
,x)\right] +\frac{\sigma^{2}}{2}h\Delta v(t_{k},x)\label{lm3}\\
& +Pf(t_{k},x)h+\sum_{r=1}^{q}\gamma_{r}(t_{k},x)\Delta_{k}w_{r}+r_{1}%
(t_{k},x),\nonumber\end{aligned}$$ where the remainder $r_{1}(t_{k},x)$ has the form $$\begin{aligned}
r_{1}(t_{k},x) & =\frac{h^{2}}{2}\sum_{i,j=1}^{n}P\left[ v^{i}%
(t_{k},x)v^{j}(t_{k},x)\frac{\partial^{2}}{\partial x^{i}\partial x^{j}%
}v(t_{k},\Theta)\right] \\
& +\frac{\sigma^{2}h^{2}}{2}\sum_{i,j=1}^{n}P\left[ v^{j}(t_{k}%
,x)\frac{\partial^{2}}{\left( \partial x^{i}\right) ^{2}\partial x^{j}%
}v(t_{k},\tilde{\Theta})\right] \\
& +\frac{\sigma^{4}h^{2}}{24}2^{-n}\sum_{j=1}^{2^{n}}\sum_{i=1}^{n}P\left[
\frac{\partial^{4}}{\partial x^{i_{1}}\partial x^{i_{2}}\partial x^{i_{3}%
}\partial x^{i_{4}}}v(t_{k},\Xi_{j})\xi_{j}^{i_{1}}\xi_{j}^{i_{2}}\xi
_{j}^{i_{3}}\xi_{j}^{i_{4}}\right] ,\end{aligned}$$ and $\Theta$ and $\tilde{\Theta}$ are some intermediate points between $x$ and $x-v(t_{k},x)h,$ and $\Xi_{j}$ are some intermediate points between $x-v(t_{k},x)h$ and $x-v(t_{k},x)h+\sigma\sqrt{h}\xi_{j}$ (we note that $r_{1}$ is a vector and the intermediate points depend on the component of $r_{1}$ but we do not reflect this in the notation). It is not difficult to estimate that this remainder satisfies the inequalities $$||E\left( r_{1}(t_{k},x)|\mathcal{F}_{t_{k}}^{w}\right) ||\leq
C(\omega)h^{2},\ \ \left( E||r_{1}(t_{k},\cdot)||^{2p}\right) ^{1/2p}\leq
Kh^{2}. \label{lm4}%$$
We write the solution $v(s,x),\ s\geq t_{k},$ of (\[NS1\])-(\[NS3\]) as $$\begin{aligned}
v(s,x) & =v(t_{k},x)+\int_{t_{k}}^{s}\left[ \frac{\sigma^{2}}{2}\Delta
v(s^{\prime},x)-(v(s^{\prime},x),\nabla)v(s^{\prime},x)+f(s^{\prime
},x)\right] ds^{\prime}\label{lm5}\\
& -\int_{t_{k}}^{s}\nabla p(s^{\prime},x)ds^{\prime}+\sum_{r=1}^{q}%
\int_{t_{k}}^{s}\gamma_{r}(s^{\prime},x)dw_{r}(s^{\prime})\nonumber\end{aligned}$$ and, in particular,$$\begin{aligned}
v(t_{k+1},x) & =v(t_{k},x)+\int_{t_{k}}^{t_{k+1}}\left[ \frac{\sigma^{2}%
}{2}\Delta v(s,x)-(v(s,x),\nabla)v(s,x)+f(s,x)\right] ds\label{lm6}\\
& -\int_{t_{k}}^{t_{k+1}}\nabla p(s,x)ds+\sum_{r=1}^{q}\int_{t_{k}}^{t_{k+1}%
}\gamma_{r}(s,x)dw_{r}(s).\nonumber\end{aligned}$$ Substituting $v(s,x)$ from (\[lm5\]) in the integrand of the first integral in (\[lm6\]) and expanding $\gamma_{r}(s,x)$ at $(t_{k},x),$ we obtain $$\begin{aligned}
v(t_{k+1},x) & =v(t_{k},x)+h\frac{\sigma^{2}}{2}\Delta v(t_{k}%
,x)-h(v(t_{k},x),\nabla)v(t_{k},x)+hf(t_{k},x)\label{lm7}\\
& -\int_{t_{k}}^{t_{k+1}}\nabla p(s,x)ds+\sum_{r=1}^{q}\gamma_{r}%
(t_{k},x)\Delta_{k}w_{r}+r_{2}(t_{k},x),\nonumber\end{aligned}$$ where $$r_{2}(t_{k},x)=r_{2}^{(1)}(t_{k},x)+r_{2}^{(2)}(t_{k},x)$$ and $$\begin{aligned}
r_{2}^{(1)}(t_{k},x) & =\frac{\sigma^{2}}{2}\int_{t_{k}}^{t_{k+1}}\left[
\int_{t_{k}}^{s}\Delta\left( \frac{\sigma^{2}}{2}\Delta v(s^{\prime
},x)-(v(s^{\prime},x),\nabla)v(s^{\prime},x) \right. \right . \displaybreak[0]\\
& \bigg. \bigg. +f(s^{\prime},x)\bigg)ds^{\prime}\bigg] ds
-\frac{\sigma^{2}}{2}\int_{t_{k}}^{t_{k+1}}\int_{t_{k}}^{s}\Delta\nabla
p(s^{\prime},x)ds^{\prime}ds \displaybreak[0]\\
&-\int_{t_{k}}^{t_{k+1}}(v(s,x),\nabla)
\left[ \int_{t_{k}}^{s}\left(
\frac{\sigma^{2}}{2}\Delta v(s^{\prime},x)
-(v(s^{\prime},x),\nabla)v(s^{\prime},x)
\right. \right. \displaybreak[0]\\
&\bigg. \bigg. +f(s^{\prime},x)\bigg) ds^{\prime}\bigg] ds \displaybreak[0]\\
& +\int_{t_{k}}^{t_{k+1}}(v(s,x),\nabla)\int_{t_{k}}^{s}\nabla p(s^{\prime
},x)ds^{\prime}ds \displaybreak[0]\\
& -\int_{t_{k}}^{t_{k+1}}\left( \int_{t_{k}}^{s}\left( \frac{\sigma^{2}}%
{2}\Delta v(s^{\prime},x)-(v(s^{\prime},x),\nabla)v(s^{\prime},x) \right. \right. \displaybreak[0]\\
& \bigg. \bigg. +f(s^{\prime},x)\bigg) ds^{\prime},\nabla\bigg) v(s,x)ds \displaybreak[0]\\
&+\int_{t_{k}}^{t_{k+1}}\left( \int_{t_{k}}^{s}\nabla p(s^{\prime
},x)ds^{\prime},\nabla\right) v(s,x)ds \displaybreak[0]\\
& +\int_{t_{k}}^{t_{k+1}}(t_{k+1}-s)\frac{\partial}{\partial s}f(s,x)ds,\end{aligned}$$$$\begin{aligned}
r_{2}^{(2)}(t_{k},x) & =\frac{\sigma^{2}}{2}\sum_{r=1}^{q}\int_{t_{k}%
}^{t_{k+1}}\int_{t_{k}}^{s}\Delta\gamma_{r}(s^{\prime},x)dw_{r}(s^{\prime
})ds \displaybreak[0]\\
& -\sum_{r=1}^{q}\int_{t_{k}}^{t_{k+1}}\left[ (v(s,x),\nabla)\int_{t_{k}%
}^{s}\gamma_{r}(s^{\prime},x)dw_{r}(s^{\prime})\right] ds \displaybreak[0]\\
& -\sum_{r=1}^{q}\int_{t_{k}}^{t_{k+1}}\left( \int_{t_{k}}^{s}\gamma
_{r}(s^{\prime},x)dw_{r}(s^{\prime}),\nabla\right) v(s,x)ds \displaybreak[0]\\
& +\sum_{r=1}^{q}\int_{t_{k}}^{t_{k+1}}\left( w_{r}(t_{k+1})-w_{r}%
(s)\right) \frac{\partial}{\partial s}\gamma_{r}(s,x)ds.\end{aligned}$$ We see that the remainder $r_{2}(t_{k},x)$ consists of 1) $r_{2}^{(1)}%
(t_{k},x)$ with terms of mean-square order $h^{2}$ and 2) $r_{2}^{(2)}%
(t_{k},x)$ with terms containing $\mathcal{F}_{t_{k+1}}^{w}$-measurable Ito integrals of mean-square order $h^{3/2}$ which expectations with respect to $\mathcal{F}_{t_{k}}^{w}$ equal zero. Further, using Assumptions 2.1, one can show that $$|E\left( r_{2}(t_{k},x)|\mathcal{F}_{t_{k}}^{w}\right) |\leq C(\omega
)h^{2},\ \ \left( E\left\vert r_{2}(t_{k},x)\right\vert ^{2p}\right)
^{1/2p}\leq Kh^{3/2},\ \label{lm8}%$$ where $C(\omega)>0$ and $K>0$ do not depend on $k,$ $x,$ and $h.$ Based on the second inequality in (\[lm8\]), we obtain $$\begin{aligned}
E\left\vert |r_{2}(t_{k},\cdot)|\right\vert ^{2p} & =E\left( \int%
_{Q}\left[ r_{2}(t_{k},x)\right] ^{2}dx\right) ^{p}\leq KE\int%
_{Q}\left\vert r_{2}(t_{k},x)\right\vert ^{2p}dx\label{lm82}\\
& \leq K\int_{Q}E\left\vert r_{2}(t_{k},x)\right\vert ^{2p}dx\leq
Kh^{2p\times3/2}\ .\nonumber\end{aligned}$$
Applying the projector operator $P$ to the left- and right-hand sides of (\[lm7\]), we arrive at $$\begin{aligned}
v(t_{k+1},x) & =v(t_{k},x)+h\frac{\sigma^{2}}{2}\Delta v(t_{k}%
,x)-hP[(v(t_{k},x),\nabla)v(t_{k},x)]+hPf(t_{k},x)\label{lm9}\\
& +\sum_{r=1}^{q}\gamma_{r}(t_{k},x)\Delta_{k}w_{r}+r_{3}(t_{k}%
,x),\ \nonumber\end{aligned}$$ where the new remainder $r_{3}(t_{k},x)=Pr_{2}(t_{k},x).$ Using (\[lm82\]), we get $$E||r_{3}(t_{k},\cdot)||^{2p}=E||Pr_{2}(t_{k},\cdot)||^{2p}\leq E||r_{2}%
(t_{k},\cdot)||^{2p}\leq Kh^{2p\times3/2}. \label{lm100}%$$ Hence from here, (\[lm4\]) and (\[lm3\]), (\[lm9\]), we obtain (\[lm22\]).
Observing that expectation of projection $P$ of Ito integrals remains equal to zero, we get $E\left( Pr_{2}^{(2)}(t_{k},x)|\mathcal{F}_{t_{k}}^{w}\right)
=0.$ Since $r_{2}^{(1)}(t_{k},x)$ consists of terms of mean-square order $h^{2}$, we obtain$$\begin{aligned}
||E\left( r_{3}(t_{k},x)|\mathcal{F}_{t_{k}}^{w}\right) ||^{2} &
=||E\left( Pr_{2}^{(1)}(t_{k},x)|\mathcal{F}_{t_{k}}^{w}\right) ||^{2} \displaybreak[0]\\%
&=\int_{Q}\left[ E\left( Pr_{2}^{(1)}(t_{k},x)|\mathcal{F}_{t_{k}}%
^{w}\right) \right] ^{2}dx \displaybreak[0]\\
& \leq\int_{Q}E\left( \left[ Pr_{2}^{(1)}(t_{k},x)\right] ^{2}%
|\mathcal{F}_{t_{k}}^{w}\right) dx \displaybreak[0]\\
&=E\left( \int_{Q}\left[ Pr_{2}%
^{(1)}(t_{k},x)\right] ^{2}dx|\mathcal{F}_{t_{k}}^{w}\right) \displaybreak[0]\\
& \leq E\left( \int_{Q}\left[ r_{2}^{(1)}(t_{k},x)\right] ^{2}%
dx|\mathcal{F}_{t_{k}}^{w}\right) \leq C(\omega)h^{4}%\end{aligned}$$ whence $$||E\left( r_{3}(t_{k},x)|\mathcal{F}_{t_{k}}^{w}\right) ||\leq
C(\omega)h^{2}\ . \label{lm11}%$$ Then the estimate (\[lm1\]) follows from (\[lm4\]), (\[lm11\]) and (\[lm3\]), (\[lm9\]). $\square$
We recall that in Assumptions 2.1 we require existence of moments of order $m,$ $2\leq m<m_{0},$ of the solution and its spatial derivatives. The higher the $m_{0},$ the higher $p$, $1\leq p<p_{0},$ can be taken in $(\ref{lm22})$. In particular, to guarantee $(\ref{lm22})$ with $p=1,$ we need existence of moments of up to the order $m=6,$ while if the moments of any order $m$ (i.e., $m_{0}=\infty)$ are finite then $(\ref{lm22})$ is valid for any $p.$ We also note that the smoothness conditions on the SNSE solution (see Assumptions 2.1) required for proving Theorem \[lemonest\] are so that $v(t,x)$ should have continuous spatial derivatives up to order four and $p(t,x)$ – up to order three.
\[coras\]Let Assumptions 2.1 hold with the bounded moments of any order $m\geq2.$ Then for almost every trajectory $w(\cdot)$ and any $0<\varepsilon
<3/2$ there exists a constant $C(\omega)>0$ such that the one-step error from $(\ref{onesterr})$ is estimated as $$||\rho(t_{k+1},\cdot)||\leq C(\omega)h^{3/2-\varepsilon}, \label{coroe}%$$ i.e., the layer method $(\ref{NS18})$ has the one-step error of order $3/2-\varepsilon$ a.s. .
**Proof.** Here we follow the recipe used in [@Gyo98; @filter; @spde]. The Markov inequality together with (\[lm22\]) implies $$P(||\rho(t_{k+1},\cdot)||>h^{\gamma})\leq\frac{E||\rho(t_{k+1},\cdot)||^{2p}%
}{h^{2p\gamma}}\leq Kh^{2p(3/2-\gamma)}.$$ Then for any $\gamma=3/2-\varepsilon$ there is a sufficiently large $p\geq1$ such that (recall that $h=T/N)$ $$\sum_{N=1}^{\infty}P\left( ||\rho(t_{k+1},\cdot)||>\frac{T^{\gamma}%
}{N^{\gamma}}\right) \leq KT^{2p(3/2-\gamma)}\sum_{N=1}^{\infty}\frac
{1}{N^{2p(3/2-\gamma)}}<\infty.$$ Hence, due to the Borel-Cantelli lemma, the random variable $$\varsigma
:=\sup_{h>0}h^{-\gamma}||\rho(t_{k+1},\cdot)||$$ is a.s. finite which implies (\[coroe\]). $\square$
Since it is desirable for the order of the one-step error $||\rho
(t_{k+1},\cdot)||$ to be greater than one, we should impose the restriction on $\varepsilon$ in $(\ref{coroe})$ to be in $(0,0.5).$ If we restrict ourselves to fulfilment of the inequality $(\ref{coroe})$ with $\varepsilon
_{0}<\varepsilon<1/2,$ where $\varepsilon_{0}$ is some positive number, then the conditions of Corollary \[coras\] can be weakened since for such $\varepsilon$ it is sufficient to take $p_{0}=1/(2\varepsilon_{0}).\ $
The intuition built on numerics for ordinary stochastic differential equations (see, e.g. [@MT1]) and also based on layer methods for SPDEs [@filter; @spde] together with convergence results for layer methods for deterministic NSE [@BM; @NS5] suggests that the one-step error properties proved in Theorem \[lemonest\] should lead to mean-square convergence of the layer method (\[NS18\]) with order one, i.e., $$(E||\bar{v}(t_{k},\cdot)-v(t_{k},\cdot)||^{2p})^{1/2p}\leq Kh. \label{msqone}%$$ However, we have not succeeded in proving such a result. Below we prove an almost sure (a.s.) convergence of the method (\[NS18\]) with lower order of $1/2-\varepsilon$ for arbitrary $\varepsilon>0$ than the $1-\varepsilon$ a.s. order which should follow from (\[msqone\]) and the Borel-Cantelli-type of arguments (see, e.g. [@filter; @spde] and also the proof of Corollary \[coras\] above). In our numerical experiments (see Section \[secnum\]) we observed the first order (both mean-square and a.s.) convergence of a layer method on test examples.
Since we assumed in Assumptions 2.1 ** that the problem ** (\[NS1\])-(\[NS3\]) ** has a unique classical solution $v(t,x),$ $p(t,x)$ which has continuous derivatives in the space variable $x$ up to some order and since we are considering the periodic case, then $v(t,x),\ p(t,x)$ and their derivatives are a.s. finite on $[0,T]\times Q$.
To prove the below a.s. convergence Theorem \[tmhconuadd\], we make the following assumptions on the approximate solution $\bar{v}(t_{k},x)$ from (\[NS18\]).
**Assumptions 4.1.** *Let* $\bar{v}(t_{k},x),$ $k=0,\ldots,N,$ have continuous first-order spatial derivatives and $$\begin{aligned}
|\bar{v}(t_{k},x)| & \leq C(\omega),\ \label{NS201}\\
|\partial\bar{v}(t_{k},x)/\partial x^{i}| & \leq C(\omega),\ \ i=1,\ldots
,n,\nonumber\end{aligned}$$ *where* $C(\omega)>0$ *is an a.s. finite constant independent of* $x,\ h,\ k.$
The first inequality in (\[NS201\]) is necessary for a.s. convergence of the layer method (\[NS18\]). The second inequality is also necessary if one expects convergence of spatial derivatives of $\bar{v}(t,x).$ We note that even in the case of deterministic NSE [@BM; @NS5] it turns out to be problematic to derive the inequalities (\[NS201\]) for the approximate solutions. At the same time, verifying Assumptions 4.1 in numerical experiments is straightforward. We also note that in the case of Oseen-Stokes equations we succeeded in deriving such estimates for approximate solutions and their spatial derivatives.
\[tmhconuadd\]Let Assumptions 2.1 hold with the bounded moments of any order $m\geq2$ and Assumptions 4.1 also hold. For almost every trajectory $w(\cdot)$ and any $0<\varepsilon<1/2$ there exists a constant $C(\omega)>0$ such that $$||\bar{v}(t_{k},\cdot)-v(t_{k},\cdot)||\leq C(\omega)h^{1/2-\varepsilon},
\label{thm1}%$$ i.e., the layer method $(\ref{NS18})$ for the SNSE $(\ref{NS1})$-$(\ref{NS3})$ converges with order $1/2-\varepsilon$ a.s..
**Proof.** First, we note that it is easy to see that under Assumptions 2.1 and 4.1: $$\operatorname{div}\bar{v}(t_{k},x)=0. \label{bvdiv}%$$ Denote the error of the method (\[NS18\])-(\[NS19\]) on the $k$th layer by $$\varepsilon(t_{k},x)=\bar{v}(t_{k},x)-v(t_{k},x).$$ Due to (\[NS18\]) and (\[NS19\]), we obtain $$\begin{aligned}
\varepsilon(t_{k+1},x)+v(t_{k+1},x) & =\bar{v}(t_{k+1},x) \displaybreak[0]\\
& =2^{-n}\sum_{j=1}^{2^{n}}P\bar{v}(t_{k},x-\bar{v}(t_{k},x)h+\sigma\sqrt
{h}\xi_{j})+Pf(t_{k},x)h \displaybreak[0]\\
& +\sum_{r=1}^{q}\gamma_{r}(t_{k},x)\Delta_{k}w_{r}\\
& =2^{-n}\sum_{j=1}^{2^{n}}Pv(t_{k},x-\bar{v}(t_{k},x)h+\sigma\sqrt{h}\xi
_{j}) \displaybreak[0]\\
& +2^{-n}\sum_{j=1}^{2^{n}}P\varepsilon(t_{k},x-\bar{v}(t_{k},x)h+\sigma
\sqrt{h}\xi_{j}) \displaybreak[0]\\
& +Pf(t_{k},x)h+\sum_{r=1}^{q}\gamma_{r}(t_{k},x)\Delta_{k}w_{r}.\end{aligned}$$ Using Assumptions 2.1, we obtain $$v(t_{k},x-\bar{v}(t_{k},x)h+\sigma\sqrt{h}\xi_{j})=v(t_{k},x-v(t_{k}%
,x)h+\sigma\sqrt{h}\xi_{j})+r_{j}(t_{k},x), \label{thm2}%$$ where $$|r_{j}(t_{k},x)|\leq C(\omega)|\varepsilon(t_{k},x)|h \label{thm3}%$$ and $C(\omega)$ is an a.s. finite random variable. Hence $$\begin{aligned}
\varepsilon(t_{k+1},x)+v(t_{k+1},x) =2^{-n}\sum_{j=1}^{2^{n}}%
Pv(t_{k},x-v(t_{k},x)h+\sigma\sqrt{h}\xi_{j}) \displaybreak[0]\\
+2^{-n}\sum_{j=1}^{2^{n}} Pr_{j}(t_{k},x)
+2^{-n}\sum_{j=1}^{2^{n}}P\varepsilon(t_{k},x-\bar{v}(t_{k},x)h+\sigma
\sqrt{h}\xi_{j}) \displaybreak[0]\\
+Pf(t_{k},x)h+\sum_{r=1}^{q}\gamma_{r}(t_{k},x)\Delta_{k}w_{r}.\end{aligned}$$ Then we get $$\begin{aligned}
\varepsilon(t_{k+1},x)&=2^{-n}\sum_{j=1}^{2^{n}}P\varepsilon(t_{k},x-\bar
{v}(t_{k},x)h+\sigma\sqrt{h}\xi_{j})+2^{-n}\sum_{j=1}^{2^{n}}Pr_{j}%
(t_{k},x) \label{thm5} \displaybreak[0]\\
&+\rho(t_{k+1},x), \nonumber\end{aligned}$$ where $\rho(t_{k+1},x)$ is the error (see (\[onesterr\])) of the one-step approximation (\[NSA1\])-(\[NSA3\]) and this one-step error satisfies the inequality (\[coroe\]) from Corollary \[coras\]. It follows from (\[thm5\]), (\[thm3\]) and (\[coroe\]) that $$\begin{aligned}
||\varepsilon(t_{k+1},\cdot)|| \leq & 2^{-n}\sum_{j=1}^{2^{n}}||P\varepsilon
(t_{k},\cdot-\bar{v}(t_{k},\cdot)h+\sigma\sqrt{h}\xi_{j})||+2^{-n}\sum
_{j=1}^{2^{n}}||Pr_{j}(t_{k},\cdot)||\label{thm7} \displaybreak[0]\\
& +||\rho(t_{k+1},\cdot)||\nonumber \displaybreak[0]\\
\leq &2^{-n}\sum_{j=1}^{2^{n}}||\varepsilon(t_{k},\cdot-\bar{v}(t_{k}%
,\cdot)h+\sigma\sqrt{h}\xi_{j})||+2^{-n}\sum_{j=1}^{2^{n}}||r_{j}(t_{k}%
,\cdot)|| \nonumber \displaybreak[0] \\
&+||\rho(t_{k+1},\cdot)||\nonumber \displaybreak[0]\\
\leq & 2^{-n}\sum_{j=1}^{2^{n}}||\varepsilon(t_{k},\cdot-\bar{v}(t_{k}%
,\cdot)h+\sigma\sqrt{h}\xi_{j})||+C(\omega)||\varepsilon(t_{k},\cdot
)||h \nonumber \displaybreak[0] \\
&+C(\omega)h^{3/2-\varepsilon}.\nonumber\end{aligned}$$
Consider $\delta(x)=\varepsilon(t_{k},x-\bar{v}(t_{k},x)h+\sigma\sqrt{h}%
\xi_{j}).$ Due to Assumptions 4.1, the function $y(x)=x-\bar{v}(t_{k}%
,x)h+\sigma\sqrt{h}\xi_{j}$ is a differentiable function with continuous partial derivatives. Furthermore, using Assumptions 4.1, one can show that for sufficiently small $h>0$ the function $y(x)=x-\bar{v}(t_{k},x)h+\sigma\sqrt
{h}\xi_{j}$ is injective. Then, taking into account the $Q$-periodicity of $\bar{v}(t_{k},x)$ and $\varepsilon^{i}(t_{k},x),$ we obtain $$\begin{aligned}
||\delta(\cdot)||^{2} & =\int_{Q}\sum_{i=1}^{n}\left[ \varepsilon^{i}%
(t_{k},x-\bar{v}(t_{k},x)h+\sigma\sqrt{h}\xi_{j})\right] ^{2}dx\\
& =\int_{Q}\sum_{i=1}^{n}\left[ \varepsilon^{i}(t_{k},y)\right] ^{2}%
\frac{D(x^{1}\ldots x^{n})}{D(y^{1}\ldots y^{n})}dy.\end{aligned}$$ Due to Assumptions 4.1 and due to (\[bvdiv\]), we get $$\begin{aligned}
\frac{D(y^{1}\ldots y^{n})}{D(x^{1}\ldots x^{n})} & =\left\vert
\begin{array}
[c]{cccc}%
1-h\frac{\partial\bar{v}^{1}(t_{k},x)}{\partial x^{1}} & -h\frac{\partial
\bar{v}^{1}(t_{k},x)}{\partial x^{2}} & \cdots & -h\frac{\partial\bar{v}%
^{1}(t_{k},x)}{\partial x^{n}} \\
-h\frac{\partial\bar{v}^{2}(t_{k},x)}{\partial x^{1}} & 1-h\frac{\partial
\bar{v}^{2}(t_{k},x)}{\partial x^{2}} & \cdots & -h\frac{\partial\bar{v}%
^{2}(t_{k},x)}{\partial x^{n}} \\
\cdots & \cdots & \cdots & \cdots \\
-h\frac{\partial\bar{v}^{n}(t_{k},x)}{\partial x^{1}} & -h\frac{\partial
\bar{v}^{n}(t_{k},x)}{\partial x^{2}} & \cdots & 1-h\frac{\partial\bar{v}%
^{n}(t_{k},x)}{\partial x^{n}}%
\end{array}
\right\vert \label{thm8} \displaybreak[0]\\
& =1+C(\omega)h^{2},\nonumber\end{aligned}$$ where $C(\omega)$ is an a.s. finite random variable. Then, we also have$$\dfrac{D(x^{1}\ldots x^{n})}{D(y^{1}\ldots y^{n})}=1+C(\omega)h^{2}.$$
We obtain from (\[thm7\]) and (\[thm8\]): $$||\varepsilon(t_{k+1},\cdot)||\leq||\varepsilon(t_{k},\cdot)||+C(\omega
)||\varepsilon(t_{k},\cdot)||h+C(\omega)h^{3/2-\varepsilon}, \label{thm9}%$$ whence (\[thm1\]) follows.$\ \ \square$
We recall that we have proved in Theorem \[lemonest\] that the mean and mean-square one-step errors of the layer method $(\ref{NS18})$ (and analogously of the other two layer methods from Section \[secLayer\]) are of orders $O(h^{2})$ and $O(h^{3/2}),$ respectively. This has given us the basis to argue that the methods from Section \[secLayer\] are of global mean-square order one (see $(\ref{msqone})$). The same intuition implies that if we incorporate terms of mean-square order $O(h^{3/2})$ and of mean order $O(h^{2})$ in these first order methods (and thus make the mean-square one-step errors to be of order $O(h^{2})$ and the mean errors of order $O(h^{3}))$ then they become of global mean-square order $3/2.$ The required Ito integrals of mean-square order $O(h^{3/2})$ can be simulated in the constructive way (and hence these methods of order $3/2$ are constructive). In the case of deterministic NSE (i.e., when $\gamma_{r}=0)$ such a method of global mean-square order $3/2$ becomes of order two and coincides with the corresponding layer method derived in [@NS5].
Let us now consider the error of the approximations of pressure considered in Section \[secPres\]. In the next proposition we prove convergence of pressure evaluated by (\[DL7p\]), (\[NS18\]). Analogously, one can prove convergence of the other approximations of pressure derived in Section \[secPres\].
\[prp32\]Let assumptions of Theorem \[tmhconuadd\] hold. In addition assume that second-order spatial derivatives of the approximate solution are a.s. finite: $|\partial^{2}\bar{v}(t_{k},x)/\partial x^{i}\partial
x^{j}|\allowbreak\leq C(\omega).$ Then for almost every trajectory $w(\cdot)$ and any $0<\varepsilon<1/3$ there exists a constant $C(\omega)>0$ such that the approximate pressure $\bar{p}(t_{k},x)$ from $(\ref{DL7p})$, $(\ref{NS18})$ satisfies the following inequality$$\Vert\bar{p}(t_{k},\cdot)-p(t_{k},\cdot)\Vert\leq C(\omega)h^{1/3-\epsilon}.
\label{NS204}%$$
**Proof**. We have $$\begin{aligned}
\frac{\partial v^{i}}{\partial x^{j}}(t_{k},x) & =\frac{v^{i}(t_{k},x+\delta
e_{j})-v^{i}(t_{k},x-\delta e_{j})}{2\delta}+O(\delta^{2}),\label{NS25} \displaybreak[0]\\
\frac{\partial\bar{v}^{i}}{\partial x^{j}}(t_{k},x) & =\frac{\bar{v}%
^{i}(t_{k},x+\delta e_{j})-\bar{v}^{i}(t_{k},x-\delta e_{j})}{2\delta
}+O(\delta^{2}),\nonumber\end{aligned}$$ where $\delta$ is a positive sufficiently small number and $|O(\delta
^{2})|\leq C(\omega)\delta^{2}$. Due to Theorem \[tmhconuadd\], $$\begin{gathered}
\left\Vert \frac{v(t_{k},x+\delta e_{j})-v(t_{k},x-\delta e_{j})}{2\delta
}-\frac{\bar{v}(t_{k},x+\delta e_{j})-\bar{v}(t_{k},x-\delta e_{j})}{2\delta
}\right\Vert \label{NS255} \\
\leq C(\omega)\frac{h^{1/2-\epsilon/2}}{\delta}\ \text{a.s.} \nonumber\end{gathered}$$ Choosing $\delta=ch^{1/6+\epsilon/2}$ with some $c>0,$ we obtain from (\[NS25\]) and (\[NS255\]) that$$\left\Vert \frac{\partial v}{\partial x^{j}}(t_{k},\cdot)-\frac{\partial
\bar{v}}{\partial x^{j}}(t_{k},\cdot)\right\Vert \leq C(\omega)h^{1/3-\epsilon
}\ \ \ \text{a.s.\ .} \label{NS26}%$$ Subtracting (\[pre1\]) with $t=t_{k}$ from (\[DL7p\]) with $t_{k}$ instead of $t_{k+1}$, we get$$\begin{gathered}
\left\Vert \nabla\bar{p}(t_{k},\cdot)-\nabla p(t_{k},\cdot)\right\Vert
=\left\Vert P^{\bot}\left[ (v(t_{k},\cdot),\nabla)v(t_{k},\cdot)\right]
-P^{\bot}\left[ (\bar{v}(t_{k},\cdot),\nabla)\bar{v}(t_{k},\cdot)\right]
\right\Vert \label{NS27} \displaybreak[0]\\
\leq\left\Vert P^{\bot}\left[ (v(t_{k},\cdot),\nabla)(v(t_{k},\cdot)-\bar
{v}(t_{k},\cdot))\right] \right\Vert +\left\Vert P^{\bot}\left[
(v(t_{k},\cdot)-\bar{v}(t_{k},\cdot),\nabla)\bar{v}(t_{k},\cdot)\right]
\right\Vert \nonumber \displaybreak[0]\\
\leq\left\Vert (v(t_{k},\cdot),\nabla)(v(t_{k},\cdot)-\bar{v}(t_{k}%
,\cdot))\right\Vert +\left\Vert (v(t_{k},\cdot)-\bar{v}(t_{k},\cdot
),\nabla)\bar{v}(t_{k},\cdot)\right\Vert \ .\nonumber\end{gathered}$$ Due to Assumptions 2.1 and (\[NS26\]), $$\left\Vert (v(t_{k},\cdot),\nabla)(v(t_{k},\cdot)-\bar{v}(t_{k},\cdot
))\right\Vert \leq C(\omega)h^{1/3-\epsilon}\ \ \text{a.s.\ .} \label{NS28}%$$ Due to Assumptions 4.1 and Theorem \[tmhconuadd\], $$\left\Vert (v(t_{k},\cdot)-\bar{v}(t_{k},\cdot),\nabla)\bar{v}(t_{k}%
,\cdot)\right\Vert \leq C(\omega)h^{1/2-\epsilon}\ \ \text{a.s.\ .}
\label{NS29}%$$ Thus, (\[NS27\])-(\[NS29\]) imply (\[NS204\]). $\ \square$
To prove the estimate $$\left\Vert \frac{\partial v}{\partial x^{j}}(t_{k},x)-\frac{\bar{v}%
(t_{k},x+\delta e_{j})-\bar{v}(t_{k},x-\delta e_{j})}{2\delta}\right\Vert \leq
C(\omega)h^{1/3-\epsilon}\ \ \ \text{a.s.\ ,} \label{eremp1}%$$ we do not need in the assumption on boundedness of second-order spatial derivatives of the approximate solution. Then, under the conditions of Theorem \[tmhconuadd\] (without the additional assumption on second-order spatial derivatives of the approximate solution), we can analogously prove convergence with a.s. order $1/3-\epsilon$ of the approximate pressure $\bar{p}(t_{k},x)$ from $(\ref{NSMp})$ with $\breve{v}(t_{k+1},x)$ from $(\ref{NSM23})$ in which we substitute $\bar{v}(t_{k+1},x)$ found due to $(\ref{NS18})$.
\[remp2\]As we discussed earlier in this section, though we proved $1/2-\varepsilon$ a.s. convergence order for the velocity approximation in Theorem \[tmhconuadd\], we are expecting that the actual a.s. convergence order is $1-\varepsilon$ which was observed in our numerical experiments in Section \[secnum\]. Analogously, we expect that spatial derivatives of the approximate velocity converge with a.s. order $1-\varepsilon$ instead of $1/3-\epsilon$ shown in $(\ref{NS26})$. It is not difficult to see from the proof of Proposition \[prp32\] that a.s. convergence of both velocity and its first-order spatial derivatives with order $1-\varepsilon$ implies a.s. convergence of pressure with order $1-\varepsilon.$ In our numerical experiments (see Section \[secnum\]) we observed convergence (both mean-square and a.s.) of pressure with order one.
Numerical examples\[secnum\]
============================
In this section we test the numerical algorithm (\[alg2\]) from Section \[secSimL\] on two model problems. The experiments indicate that the algorithm has the first order mean-square convergence.
Model problems
--------------
We introduce two model examples of SNSE (\[NS1\])-(\[NS3\]) which solutions can be written in an analytic form. Both examples are generalizations of the deterministic model of laminar flow from [@Taylor] to the stochastic case.
**First model problem.** Let $$f(t,x)=0,\text{\ \ }\varphi(x)=0, \label{nl1}%$$$$\begin{aligned}
q & =1,\label{nl11} \displaybreak[0]\\
\gamma_{1}^{1}(t,x) & =A\sin\frac{2\pi\kappa\ x^{1}}{L}\cos\frac{2\pi
\kappa\ x^{2}}{L}\exp\left( -\sigma^{2}\left( \frac{2\pi\kappa}{L}\right)
^{2}t\right) \ ,\nonumber \displaybreak[0]\\
\gamma_{1}^{2}(t,x) & =-A\cos\frac{2\pi\kappa\ x^{1}}{L}\sin\frac{2\pi
\kappa\ x^{2}}{L}\exp\left( -\sigma^{2}\left( \frac{2\pi\kappa}{L}\right)
^{2}t\right) \ ,\ \kappa\in\mathbf{Z},\ \ A\in\mathbf{R},\nonumber\end{aligned}$$ then it is easy to check that the problem (\[NS1\])-(\[NS3\]), (\[nl1\])-(\[nl11\]) has the following solution $$\begin{aligned}
v^{1}(t,x) =&A\sin\frac{2\pi\kappa\ x^{1}}{L}\cos\frac{2\pi\kappa\ x^{2}%
}{L} \label{nl2} \displaybreak[0] \\
&\times \exp\left( -\sigma^{2}\left( \frac{2\pi\kappa}{L}\right) ^{2}t\right)
w(t)\ , \nonumber \displaybreak[0] \\
v^{2}(t,x) = &-A\cos\frac{2\pi\kappa\ x^{1}}{L}\sin\frac{2\pi\kappa\ x^{2}%
}{L} \nonumber \displaybreak[0] \\
&\times \exp\left( -\sigma^{2}\left( \frac{2\pi\kappa}{L}\right) ^{2}t\right)
w(t)\ ,\nonumber \displaybreak[0]\\
p(t,x) =&\frac{A^{2}}{4}\left( \cos\frac{4\pi\kappa\ x^{1}}{L}+\cos
\frac{4\pi\kappa\ x^{2}}{L}\right) \nonumber \displaybreak[0] \\
& \times \exp\left( -2\sigma^{2}\left( \frac
{2\pi\kappa}{L}\right) ^{2}t\right) (w(t))^{2}\ .\nonumber\end{aligned}$$
**Second model problem.** To construct this example, we recall the following proposition from [@HRoz07].
\[PropRoz\]Let $V(t,x),$ $P(t,x)$ be a solution of the deterministic NSE with zero forcing $($i.e., of $(\ref{NS1})$-$(\ref{NS3})$ with all $\gamma
_{r}=0$ and $f(t,x)=0)$ then the solution $v(t,x),$ $p(t,x)$ of $(\ref{NS1}%
)$-$(\ref{NS3})$ with constant $\gamma_{r}(t,x)=\gamma_{r}$ and $f(t,x)=0$ is equal to $$\begin{aligned}
v(t,x) & =V\left( t,x-\int_{0}^{t}\sum_{r=1}^{q}\gamma_{r}w_{r}%
(s)ds\right) +\sum_{r=1}^{q}\gamma_{r}w_{r}(t),\label{sgt}\\
p(t,x) & =P\left( t,x-\int_{0}^{t}\sum_{r=1}^{q}\gamma_{r}w_{r}%
(s)ds\right) . \label{sgt1}%\end{aligned}$$
Combining this proposition with the deterministic model of laminar flow from [@Taylor], we obtain that if $$\begin{gathered}
f(t,x)=0,\text{\ \ }\varphi(x)=\left( A\sin\frac{2\pi\kappa\ x^{1}}{L}%
\cos\frac{2\pi\kappa\ x^{2}}{L},-A\cos\frac{2\pi\kappa\ x^{1}}{L}\sin
\frac{2\pi\kappa\ x^{2}}{L}\right) ^{\top},\label{nl3} \displaybreak[0]\\
\kappa\in\mathbf{Z},\ \ A\in\mathbf{R},\nonumber\end{gathered}$$ and$$q=1,\ \gamma_{1}^{1}(t,x)=\gamma^{1},\ \ \gamma_{1}^{2}(t,x)=\gamma^{2}\ .
\label{nl31}%$$ then the problem (\[NS1\])-(\[NS3\]), (\[nl3\])-(\[nl31\]) has the following solution $$\begin{aligned}
v^{1}(t,x) & =A\sin\frac{2\pi\kappa\ \left( x^{1}-\gamma^{1}I(t)\right)
}{L}\cos\frac{2\pi\kappa\ \left( x^{2}-\gamma^{2}I(t)\right) }{L} \label{nl4} \displaybreak[0]\\
& \times \exp\left(
-\sigma^{2}\left( \frac{2\pi\kappa}{L}\right) ^{2}t\right)
+\gamma^{1}w(t),\nonumber \displaybreak[0]\\
v^{2}(t,x) & =-A\cos\frac{2\pi\kappa\ \left( x^{1}-\gamma^{1}I(t)\right)
}{L}\sin\frac{2\pi\kappa\ \left( x^{2}-\gamma^{2}I(t)\right) }{L} \nonumber \displaybreak[0]\\
& \times \exp\left(
-\sigma^{2}\left( \frac{2\pi\kappa}{L}\right) ^{2}t\right)
+\gamma^{2}w(t),\nonumber \displaybreak[0]\\
p(t,x) & =\frac{A^{2}}{4}\left( \cos\frac{4\pi\kappa\ \left( x^{1}%
-\gamma^{1}I(t)\right) }{L}+\cos\frac{4\pi\kappa\ \left( x^{2}-\gamma
^{2}I(t)\right) }{L}\right) \nonumber \displaybreak[0] \\
& \times \exp\left( -2\sigma^{2}\left( \frac{2\pi
\kappa}{L}\right) ^{2}t\right) ,\nonumber\end{aligned}$$ where $$I(t)=\int_{0}^{t}w(s)ds,\ w(s)=w_{1}(s).$$
Results of numerical experiments
--------------------------------
In our numerical experiments we test the algorithm (\[alg2\])-(\[alg3\]), (\[algp\]) which is a realization of the layer method (\[NSM21\])-(\[NSM23\]), (\[NSMp\]). This algorithm possesses the following properties.
\[Propcons\]**1.** The approximate solution of the problem $(\ref{NS1})$-$(\ref{NS3})$, $(\ref{nl1})$-$(\ref{nl11})$ obtained by the algorithm $(\ref{alg2})$-$(\ref{alg3})$, $(\ref{algp})$ contains only those modes which are present in the coefficient $\gamma_{1}(t,x)$ from $(\ref{nl11})$, i.e., which are present in the exact solution $(\ref{nl2})$.
**2.** The approximate solution of the problem $(\ref{NS1})$-$(\ref{NS3})$, $(\ref{nl3})$-$(\ref{nl31})$ obtained by the algorithm $(\ref{alg2})$-$(\ref{alg3})$, $(\ref{algp})$ contains only those modes which are present in the initial condition $\varphi(x)$ from $(\ref{nl3})$ and the zero mode, i.e., which are present in the exact solution $(\ref{nl4})$.
The proof of this proposition is analogous to the proof of a similar result in the deterministic case [@NS5] and it is omitted here.
We measure the numerical error in the experiments as follows. First, we consider the relative mean-square error defined as$$err_{msq}^{v}=\frac{\sqrt{E\sum_{\mathbf{n}}|\bar{v}_{\mathbf{n}%
}(T)-v_{\mathbf{n}}(T)|^{2}}}{\sqrt{E\sum_{\mathbf{n}}|v_{\mathbf{n}}(T)|^{2}%
}}\ ,\ \ \ err_{msq}^{p}=\frac{\sqrt{E\sum_{\mathbf{n}}|\bar{p}_{\mathbf{n}%
}(T)-p_{\mathbf{n}}(T)|^{2}}}{\sqrt{E\sum_{\mathbf{n}}|p_{\mathbf{n}}(T)|^{2}%
}}\ . \label{msqerr}%$$ Analysis of this error provides us with information about mean-square convergence of the numerical algorithm considered. To evaluate this error in the experiments, we use the Monte Carlo technique for finding the expectations in (\[msqerr\]) by running $K$ independent (with respect to realizations of the Wiener process $w(t))$ realizations of $\bar{v}_{\mathbf{n}}(T),$ $v_{\mathbf{n}}(T),$ $\bar{p}_{\mathbf{n}}(T),\ p_{\mathbf{n}}(T).$ Second, we consider the relative $L_{2}$-error for a fixed trajectory of $w(t):$ $$err^{v}=\frac{\sqrt{\sum_{\mathbf{n}}|\bar{v}_{\mathbf{n}}(T)-v_{\mathbf{n}%
}(T)|^{2}}}{\sqrt{\sum_{\mathbf{n}}|v_{\mathbf{n}}(T)|^{2}}}\ ,\ \ \ err^{p}%
=\frac{\sqrt{\sum_{\mathbf{n}}|\bar{p}_{\mathbf{n}}(T)-p_{\mathbf{n}}(T)|^{2}%
}}{\sqrt{\sum_{\mathbf{n}}|p_{\mathbf{n}}(T)|^{2}}}\ . \label{err}%$$ Analysis of this error provides us with information about a.s. convergence of the numerical algorithm. To evaluate this error in the tests, we fix a trajectory $w(t),$ $0\leq t\leq T,$ which is obtained with a small time step.
We note that in the case of the considered examples and the tested algorithm (see Proposition \[Propcons\]) $v_{\mathbf{n}}(T)$ are nonzero only for $|\mathbf{n}^{1}|=|\mathbf{n}^{2}|=|\kappa|$ and $p_{\mathbf{n}}(T)$ are nonzero only for $|\mathbf{n}^{1}|=2|\kappa|,$ $\mathbf{n}^{2}=0$ and $\mathbf{n}^{1}=0,$ $|\mathbf{n}^{2}|=2|\kappa|$. Hence, the sums in (\[msqerr\]) and (\[err\]) are finite here. This also implies that it is sufficient here to take the cut-off parameter $M$ in the algorithm (\[alg2\])-(\[alg3\]), (\[algp\]) to be equal to $2|\kappa|.$
The test results for the algorithm (\[alg2\])-(\[alg3\]), (\[algp\]) applied to the first model problem (\[NS1\])-(\[NS3\]), (\[nl1\])-(\[nl11\]) are presented in Tables$~$\[tab1\] and \[tab2\]. In Table$~$\[tab1\] the $\pm$ reflects the Monte Carlo errors in evaluating of $err_{msq}^{v}$ and $err_{msq}^{p}$, they give the confidence intervals for the corresponding values with probability $0.95$.
\[c\][lc]{}$h$ &
\[c\][ll]{}velocity & pressure
\
$0.2$ &\
$0.1$ &\
$0.05$ &\
$0.02$ &\
$0.01$ &\
\[c\][lc]{}$h$ &
\[c\][ll]{}velocity & pressure
\
$0.2$ &\
$0.1$ &\
$0.05$ &\
$0.02$ &\
$0.01$ &\
We can conclude from Table$~$\[tab1\] that both velocity and pressure found due to the algorithm (\[alg2\])-(\[alg3\]), (\[algp\]) demonstrate the mean-square convergence with order $1.$ We also see from Table$~$\[tab2\] that both velocity and pressure converge with order $1$ for a particular, fixed trajectory of $w(t).$ We note that we repeated the experiment for other realizations of $w(t)$ and observed the same behavior. The observed first order convergence of the algorithm is consistent with our prediction (see (\[msqone\]), the discussion after it, and Remark \[remp2\]).
The test results for the algorithm (\[alg2\])-(\[alg3\]), (\[algp\]) applied to the second model problem (\[NS1\])-(\[NS3\]), (\[nl3\])-(\[nl31\]) are presented in Table$~$\[tab3\]. In these tests we limit ourselves to simulation for a particular, fixed trajectory of $w(t)$ and observation of a.s. convergence. We note that evaluation of the exact solution (\[nl4\]) requires simulation of the integral $I(t).$ This was done in the following way. At each time step $k+1,$ $k=0,\ldots,N-1,$ we simulate a Wiener increment $\Delta_{k}w$ as i.i.d. Gaussian $\mathcal{N}(0,h)$ random variables (and we find $w(t_{k+1})=w(t_{k})+\Delta_{k}w)$ and i.i.d. Gaussian $\mathcal{N}(0,1)$ random variables $\eta_{k}.$ Then (see [@MT1 Chapter 1]): $$I(t_{k+1})=I(t_{k})+hw(t_{k})+\frac{h}{2}\Delta_{k}w+\frac{h^{3/2}}{\sqrt{12}%
}\eta_{k}\ .$$
\[c\][lc]{}$h$ &
\[c\][ll]{}velocity & pressure
\
$0.01$ &\
$0.005$ &\
$0.002$ &\
$0.001$ &\
$0.0005$ &\
Again, the observed first order convergence of the algorithm in Table \[tab3\] is consistent with our prediction (see the discussion after (\[msqone\]) and Remark \[remp2\]).
Acknowledgments {#acknowledgments .unnumbered}
===============
The work was partially supported by the Royal Society International Joint Project grant JP091142.
[99]{}
Ya. Belopolskaya, G.N. Milstein. An approximation method for Navier-Stokes equations based on probabilistic approach. *Statistics & Probability Letters* **64** (2003), 201–211.
Z. Brzeźniak, E. Carelli, A. Prohl. Finite element based discretizations of the incompressible Navier-Stokes equations with multiplicative random forcing. *IMA J. Num. Anal.* (2013), doi: 10.1093/imanum/drs032.
C. Canuto, M.T. Hussaini, A. Quarteroni, T.A. Zang. *Spectral Methods in Fluid Dynamics.* Springer, 1988.
E. Carelli, A. Prohl. Rates of convergence for discretizations of the stochastic incompressible Navier-Stokes equations. Available at http://na.uni-tuebingen.de/pub/prohl/papers/snse\_final\_cp.pdf
A.J. Chorin, J.E. Marsden. *A Mathematical Introduction to Fluid Mechanics*. Springer, 2000.
P.L. Chow. *SPDEs in Turbulence, in Probabilistic Analysis and Related Topics.* Vol. 1, ed. A.T. Bharucha-Reid, Academic Press, 1978, 1–43.
P.L. Chow. *Stochastic Partial Differential Equations*. Chapman and Hall/CRC, 2007.
P.L. Chow, R.Z. Khasminskii. Stationarity solutions of nonlinear stochastic evolution equations. *Stoch. Anal. Appl.* **15** (1997), 671–699.
G. Da Prato. *Kolmogorov Equations for Stochastic PDEs*. Birkhäuser, 2004.
F. Delarue, S. Menozzi. An interpolated stochastic algorithm for quasi-linear PDEs. *Math. Comp.* **77** (2008), 125–158.
P. Dörsek. Semigroup splitting and cubature approximations for the stochastic Navier-Stokes equations. *SIAM J. Num. Anal.* **50** (2012), 729-746.
F. Flandoli. An introduction to 3D stochastic fluid dynamics. In: SPDE in Hydrodynamic: Recent Progress and Prospects, *Lecture Notes in Mathematics* **1942**, Springer, 2008, 51–150.
I. Gyöngy. A note on Euler’s approximations. *Potential Analysis* **8** (1998), 205–216.
M. Hairer, J.C. Mattingly. Ergodicity of the 2D Navier-Stokes equations with degenerate stochastic forcing. *Annals of Mathematics* **164** (2006), 993–1032.
T.Y. Hou, W. Luo, B. Rozovskii, H.-M. Zhou. Wiener Chaos expansions and numerical solutions of randomly forced equations of fluid mechanics. *J. Comp. Phys.* **216** (2006), 687–706.
N.V. Krylov, B.L. Rozovskii. On the characteristics of degenerate second order parabolic Ito equations. *J. Soviet Math.* **32** (1986), 336–348.
H. Kunita. *Stochastic Flows and Stochastic Differential Equations*. Cambridge University Press, 1990.
J.C. Mattingly, E. Pardoux. Malliavin calculus and the randomly forced Navier-Stokes equation. *Comm. Pure Appl. Math.* **59** (2006), 1742–1790.
R. Mikulevicius. On the Cauchy problem for stochastic Stokes equation. *SIAM J. Math. Anal.* **34** (2002), 121–141.
R. Mikulevicius, B. Rozovskii. Stochastic Navier-Stokes equations for turbulent flows. *SIAM J. Math. Anal.* **35** (2004), 1250–1310.
R. Mikulevicius, B. Rozovskii. Global L2-solutions of stochastic Navier-Stokes equations. *Ann. Prob.* **33** (2005), 137–176.
G.N. Milstein. The probability approach to numerical solution of nonlinear parabolic equations. *Num. Meth. PDE* **18** (2002), 490–522.
G.N. Milstein, M.V. Tretyakov. *Stochastic Numerics for Mathematical Physics*. Springer, 2004.
G.N. Milstein, M.V. Tretyakov. Discretization of forward-backward stochastic differential equations and related quasilinear parabolic equations. *IMA J. Num. Anal.* **27** (2007), 24–44.
G.N. Milstein, M.V. Tretyakov. Monte Carlo algorithms for backward equations in nonlinear filtering. *Adv. Appl. Prob.* **41** (2009), 63–100.
G.N. Milstein, M.V. Tretyakov. Solving parabolic stochastic partial differential equations via averaging over characteristics. *Math. Comp.* **78** (2009), 2075–2106.
G.N. Milstein, M.V. Tretyakov. Probabilistic methods for the incompressible Navier-Stokes equations with space periodic conditions. *Adv. Appl. Prob.* **45** (2013), to appear.
G.N. Milstein, M.V. Tretyakov. Solving the Dirichlet problem for Navier-Stokes equations by probabilistic approach. *BIT Num. Math.* **52** (2012), 141–153.
E. Pardoux. Stochastic partial differential equations and filtering of diffusion processes. *Stochastics* **3** (1979), 127–167.
E. Pardoux, S. Peng. Backward doubly stochastic differential equations and systems of quasilinear SPDEs. Probab. Theory Relat. Fields **98** (1994), 209–227.
B.L. Rozovskii. *Stochastic Evolution Systems, Linear Theory and Application to Nonlinear Filtering*. Kluwer Academic Publishers, Dordrecht, 1991.
G.I. Taylor. *The Scientific Papers of Sir Geoffrey Ingram Taylor*. Edited by G.K. Batchelor. Vol. 2, Meteorology, oceanography and turbulent flow. The Decay of Eddies in a Fluid, paper written in 1923. Cambridge University Press, 1960, 190–192.
R. Temam. *Navier-Stokes Equations, Theory and Numerical Analysis.* AMS Chelsea Publishing, 2001.
R. Temam. *Navier-Stokes Equations and Nonlinear Functional Analysis*. SIAM, 1995.
[^1]: Ural State University, Lenin Str. 51, 620083 Ekaterinburg, Russia; email: [email protected]
[^2]: School of Mathematical Sciences, University of Nottingham, University Park, Nottingham, NG7 2RD, UK, email: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this article, we prove an “equivalence” between two higher even moments of primes in short intervals under Riemann Hypothesis. We also provide numerical evidence in support of these asymptotic formulas.'
author:
- Tsz Ho Chan
title: Higher moments of primes in short intervals I
---
Introduction
============
Recently, Montgomery and Soundararajan \[\[MS2\]\] studied the moments $$M_k(N;h) := \sum_{n=1}^{N} (\psi(n+h) - \psi(n) - h)^k$$ where $k$ is a positive integer, $\psi(x) = \sum_{n \leq x}
\Lambda(n)$ and $\Lambda(n)$ is von Mangoldt lambda function. They proved that, under a strong form of Hardy-Littlewood prime-$k$ tuple conjecture, for small $\epsilon > 0$, there is a $\delta >
0$, $$\label{1.1} M_k(N;h) = \mu_k h^{k/2} \int_{1}^{N}
(\log{\frac{x}{h}} + B)^{k/2} dx + O_k(h^{k/2} N^{1-\epsilon})$$ uniformly for $(\log N)^{15 k^2} \leq h \leq N^{1/k - \delta}$ where $\mu_k = 1 \cdot 3 \cdot \cdot \cdot (k-1)$ if $k$ is even, and $\mu_k = 0$ if $k$ is odd. Here $B = 1 - C_0 - \log 2\pi$ and $C_0$ denotes Euler’s constant. One further expects that (\[1.1\]) holds uniformly for $N^\delta \leq h \leq
N^{1-\delta}$. This implies that, for $0 \leq x \leq N$, the distribution of $\psi(x+h) -\psi(x)$ is approximately normal with mean $h$ and variance $h \log{N/h}$. It contradicts with the prediction of Cramér’s model of variance $h \log N$. In the last section, we will show numerical evidence in support of (\[1.1\]).
Now, $M_k(X;h)$ can be written as $$\label{1.2} \int_{1}^{X} (\psi(x+h) - \psi(x) - h)^k dx.$$ We also consider the following moments: $$\label{1.3} \widetilde{M}_k(X;\delta) := \int_{1}^{X} (\psi(x +
\delta x) - \psi(x) - \delta x)^k dx.$$ Goldston and Montgomery \[\[GM\]\] showed that, under Riemann Hypothesis (RH), the stronger form of the Pair Correlation Conjecture as formulated by Montgomery \[\[M\]\] is equivalent to an asymptotic formula for (\[1.2\]) in $X^\epsilon \leq h \leq
X^{1-\epsilon}$ or an asymptotic formula for (\[1.3\]) in $X^{-1+\epsilon} \leq \delta \leq X^{-\epsilon}$ when $k=2$. The author generalized these to include the second main terms in \[\[C\]\] (again only when $k=2$). So, the main purpose of this paper is to prove the “equivalence” between an asymptotic formula for (\[1.2\]) and an asymptotic formula for (\[1.3\]) in appropriate ranges of $h$ and $\delta$ for any positive even integer $k$. Roughly speaking, we have
\[theorem1.1\] Let $k$ be a positive even integer. Assuming RH, the following are equivalent: $$(i) \; \int_{1}^{X} (\psi(x+h) - \psi(x) - h)^k dx \sim \mu_k
h^{k/2+1} \int_{E}^{X/h} (\log{\frac{x}{E}})^{k/2} dx$$ holds uniformly for $X^\epsilon \leq h \leq X^{1-\epsilon}$. $$(ii) \; \int_{1}^{X} (\psi(x+\delta x) - \psi(x) - \delta x)^k
dx \sim \frac{\mu_k}{\frac{k}{2}+1} X^{k/2+1} \delta^{k/2}
\Bigl(\log{\frac{1}{E \delta}}\Bigr)^{k/2}$$ holds uniformly for $X^{-1+\epsilon} \leq \delta \leq X^{-\epsilon}$.
Here $E = 2\pi e^{C_0 - 1}$. Our method of proof replaces the brute-force calculations in \[\[C\]\]. We will assume RH throughout this paper and $k$ being a positive even integer unless stated otherwise.
This work is part of the author’s $2002$ PhD thesis with some improvements.
Some preparations
=================
First of all, $\psi(x) = x + O(x^{1/2} \log^2{x})$ by RH (see \[\[vK\]\]). One has the following: $$\label{ineq1} \int_{1}^{X} (\psi(x+\delta x) - \psi(x) - \delta x
)^k dx \ll X^{{k/2}+1} \log^{2k}{X}$$ for $0 \leq \delta \leq 1$, and $$\label{ineq2} \int_{1}^{X} (\psi(x+h) - \psi(x) - h)^k dx \ll
X^{{k/2}+1} \log^{2k}{X}$$ for $0 \leq h \leq X$. Also, estimating trivially, we have $$\label{ineq3} \int_{1}^{X} (\psi(x + \delta x) - \psi(x) - \delta
x)^k dx \ll \int_{1}^{X} (\delta x \log{X})^k dx \ll \delta^k
X^{k+1} \log^k{X}$$ for $0 \leq \delta \leq 1$, and $$\label{ineq4} \int_{1}^{X} (\psi(x + h) - \psi(x) - h)^k dx \ll
\int_{1}^{X} (h \log{X})^k dx \ll h^k X \log^k{X}$$ for $0 \leq h \leq X$. We also need some lemmas.
\[lemma2.1\] For any differentiable function $f(u)$ and any $0
\leq \eta \leq 1$, $$\int_{T}^{(1+\eta)T} f(u) du = \eta T f(T) + O\bigl(\eta^2 T^2 \mathop{max}_{T
\leq t \leq (1+\eta)T} |f'(t)|\bigr).$$
Proof: By mean-value theorem, we have $$\begin{aligned}
\int_{T}^{(1+\eta)T} f(u) du &=& \eta T f(T + \xi T) \mbox{ where
} 0 \leq \xi \leq \eta \\
&=& \eta T \bigl(f(T) + \xi T f'(T + \xi' T)\bigr) \mbox{ where }
0 \leq \xi' \leq \xi,\end{aligned}$$ and the lemma follows.
\[lemma2.2\] For any positive integer $k$, we have $$x^k-y^k=(x-y)P(x,y) + (x-y)^k$$ where $P(x,y)$ is some homogeneous polynomial of degree $k-1$.
Proof: By Factor Theorem, $z-1$ divides $z^k -1 -(z-1)^k$. So, $$(z-1)P(z) = z^k -1 -(z-1)^k$$ for some integer polynomial $P(z)$ of degree $k-1$. Set $z = {x
\over y}$ and multiply both sides by $y^k$, we get the desired result.
\[lemma2.3\] For any positive integer $k$, and any non-negative real numbers $\alpha$, $a_1,a_2,...,a_k$, we have $$(a_1+a_2+...+a_k)^{\alpha} {\ll}_{\alpha ,k} \; a_1^{\alpha} + a_2^{\alpha}
+...+ a_k^{\alpha}.$$ Here, ${\ll}_{\alpha ,k}$ means that the implicit constant may depend on $\alpha$ and $k$ but not on any $a_i$’s.
Proof: Without loss of generality, suppose that $a_1$ is the largest among the $a_i$’s. Then $$(a_1+a_2+...+a_k)^{\alpha} \leq (k a_1)^{\alpha} \leq k^{\alpha}
(a_1^{\alpha}+a_2^{\alpha}+ ... +a_k^{\alpha}).$$
(i) $\Rightarrow$ (ii)
======================
Throughout this and the next section, we think of $k$ as fixed.
\[theorem3.1\] Assume RH. If, for some small $\epsilon >
\epsilon_1 > 0$ (small in terms of $k$), $$\label{3.1} \int_{1}^{X} (\psi(x+h)-\psi(x)-h)^k dx = \mu_k h^{k/2
+ 1} \int_{E}^{X/h} \bigr(\log{\frac{x}{E}}\bigr)^{k/2} dx +
O_k(h^{k/2} X^{1-\epsilon_1})$$ holds uniformly for $X^{\epsilon} \leq h \leq X^{1-\epsilon}$, then $$\label{3.1result} \int_{1}^{X} (\psi(x + \delta x) - \psi(x) -
\delta x)^k dx = \frac{\mu_k}{\frac{k}{2}+1} X^{{k/2}+1}
\delta^{k/2} \Bigl( \log{\frac{1}{E\delta}} \Bigr) ^{k/2} +
O_k(\delta^{k/2} X^{{k/2}+1-\epsilon_2})$$ holds uniformly for $X^{-1 +2\epsilon + 2\epsilon_1} \leq \delta
\leq X^{-\epsilon}/2$ with some $\epsilon_2 > 0$.
Proof: Our method is that of Saffari and Vaughan \[\[SV\]\] employed in \[\[GM\]\] and \[\[C\]\]. Let $f(x,h)
=\psi(x+h)-\psi(x)-h$. Let $X^{-1+2\epsilon+\epsilon_1} \leq
\Delta \leq X^{-\epsilon}$. Say $\Delta = X^{-\mu}$ for some $\epsilon \leq \mu \leq 1 - 2\epsilon - \epsilon_1$. We want to calculate $$\label{3.2} \int_{V/2}^{V} \int_{0}^{\Delta} (\psi(x+\delta x) -
\psi(x) - \delta x)^k d\delta \; dx.$$ Substituting $h=\delta x$, (\[3.2\]) becomes $$\int_{\Delta V/2}^{\Delta V} \int_{h/\Delta}^{V} {f(x,h)^k \over x}
dx \; dh + \int_{0}^{\Delta V/2} \int_{V/2}^{V} {f(x,h)^k \over x}
dx \; dh$$ $$=\int_{\Delta V/2}^{\Delta V} \int_{h/\Delta}^{V} +
\int_{V^{\epsilon}}^{\Delta V/2} \int_{V/2}^{V} +
\int_{0}^{V^{\epsilon}} \int_{V/2}^{V} = I_1 + I_2 + I_3.$$ By integration by parts, we have from (\[3.1\]) that $$\begin{aligned}
& & \int_{U}^{V} {f(x,h)^k \over x} dx \\
&=& \Bigl[{1 \over x} \int_{1}^{x} f(u,h)^k du {\Bigr]}_{U}^{V} +
\int_{U}^{V}
\Bigl(\int_{1}^{x} f(u,h)^k du \Bigr){1 \over x^2} dx \\
&=& \mu_k h^{k/2+1} \Bigl[{1 \over V} \int_{E}^{V/h} \bigl(\log{x
\over E}\bigr)^
{k/2} dx-{1 \over U}\int_{E}^{U/h} \bigl(\log{x \over E}\bigr)^{k/2} dx\Bigr] \\
& &+ \mu_k h^{k/2+1}\int_{U}^{V} {1 \over x^2}
\int_{E}^{x/h}\bigl(\log{u \over E}
\bigr)^{k/2} du \; dx + O_{k}(U^{-\epsilon_1} h^{k/2}) \\
&=& T_1 + T_2 + O_{k}(U^{-\epsilon_1} h^{k/2})\end{aligned}$$ as long as $V^{\epsilon} \leq h \leq U^{1-\epsilon}$ with $U \leq
V \leq 2U$. $$\begin{aligned}
T_2 &=& \mu_k h^{k/2} \int_{U/h}^{V/h} \int_{E}^{y} \bigl(\log{u
\over E}\bigr)^
{k/2} du d({-1 \over y}) \\
&=& \mu_k h^{k/2} \Bigl[-{h \over V} \int_{E}^{V/h} \bigl(\log{u
\over E}\bigr)^
{k/2} du + {h \over U} \int_{E}^{U/h} \bigl(\log{u \over E}\bigr)^{k/2} du \\
& &+ \int_{U/h}^{V/h} {(\log{y/E})^{k/2} \over y} dy \Bigr].\end{aligned}$$ Therefore, $$\int_{U}^{V} {f(x,h)^k \over x} dx = {\mu_k \over {k \over 2}+1}
h^{k/2} \Bigl[\bigl(\log{V \over Eh}\bigr)^{k/2+1} -\bigl(\log{U
\over Eh}\bigr)^{k/2+1} \Bigr] + O_k(U^{-\epsilon_1} h^{k/2})$$ as long as $V^{\epsilon} \leq h \leq U^{1-\epsilon}$ with $U \leq
V \leq 2U$. Thus, for $I_1$ and $I_2$ to work, we need $$\label{3cond} V^{\epsilon} \leq {\Delta V \over 2} \leq \Bigl({V
\over 2} \Bigr)^{1-\epsilon}, \mbox{ and } V^{\epsilon} \leq h
\leq \Bigl({h \over \Delta }\Bigr)^{1-\epsilon} \mbox{ for }
{\Delta V \over 2} \leq h \leq \Delta V.$$ Since $\Delta = X^\mu$, for $X^{\mu + \epsilon} \leq V \leq X$, one can check that (\[3cond\]) are satisfied. Therefore, $$\begin{aligned}
I_1 &=& {\mu_k \over {k \over 2}+1} \Bigl[\int_{\Delta
V/2}^{\Delta V} h^{k/2} \bigl(\log{V \over Eh}\bigr)^{k/2+1} dh
-\int_{\Delta V/2}^{\Delta V} h^{k/2}
\bigl(\log{1 \over E\Delta} \bigr)^{k/2+1} dh \Bigr] \\
& &+ O_{k}({\Delta}^{k/2+1} V^{k/2+1-\epsilon_1}), \\
I_2 &=& {\mu_k \over {k \over 2}+1} \Bigl[\int_{0}^{\Delta V/2}
h^{k/2} \bigl(\log{V \over Eh}\bigr)^{k/2+1} dh -\int_{0}^{\Delta
V/2} h^{k/2} \bigl(
\log{V/2 \over Eh}\bigr)^{k/2+1} dh\Bigr] \\
& &+ O_{k}(V^{({k/2}+1)\epsilon} (\log{V})^{{k/2}+1}) +
O_{k}({\Delta}^{{k/2}
+1} V^{{k/2}+1-\epsilon_1}), \\
I_3 &\ll& V^{\epsilon} V^{k \epsilon} \log^k{V},\end{aligned}$$ Combining these, (\[3.2\]) equals $${\mu_k \over {k \over 2}+1} \biggl[ \int_{0}^{\Delta V} h^{k/2}\bigl(\log{V
\over Eh}\bigr)^{k/2+1} dh -\int_{0}^{\Delta V/2} h^{k/2}
\bigl(\log{V/2 \over Eh}\bigr)^{k/2+1} dh$$ $$- \bigl(\log{1 \over E\Delta}\bigr)^{k/2+1} \int_{\Delta V/2}^{\Delta V}
h^{k/2} dh \biggr] + O\bigl(V^{(k+1)\epsilon} \log^k {V} \bigr) +
O_{k} (\Delta^{k/2+1} V^{k/2+1-\epsilon_1})$$ for $X^{\mu +
\epsilon} \leq V \leq X$. Now, replacing $V$ by $X2^{-l}$ in the above, summing over $0 \leq l \leq M=[{ (1 - \mu - \epsilon)
\log{X} \over \log{2}}]$, $$\label{3.4}
\begin{split}
&\int_{X/2^M}^{X} \int_{0}^{\Delta} \bigl(\psi(x+\delta x)
-\psi(x) -\delta x \bigr)^k d\delta \; dx \\
=& {\mu_k \over {k \over 2}+1} \biggl[\int_{0}^{\Delta X} h^{k/2}
\bigl(\log{X \over Eh}\bigr)^{k/2+1} dh - {1 \over {k \over 2}+1}
\bigl(\log{1 \over E\Delta}
\bigr)^{k/2+1} (\Delta X)^{k/2+1} \biggr]\\
&+ O_k(X^{(k+1)\epsilon} \log^k {X}) + O_k({\Delta}^{k/2+1}
X^{k/2 + 1 - \epsilon_1}) \\
=& {\mu_k \over {k \over 2}+1} \int_{0}^{\Delta X} h^{k/2}
\bigl(\log{X \over Eh} \bigr)^{k/2} dh + O_k({\Delta}^{k/2+1}
X^{k/2+1-\epsilon_1})
\end{split}$$ by integration by parts, and as $\Delta = X^{-\mu} \geq X^{-1 +
2\epsilon + \epsilon_1}$, $$X^{(k+1)\epsilon} \log^k {X} \ll X^{(k+2)\epsilon + (k/2)
\epsilon_1} \ll \Delta^{k/2+1} X^{k/2+1-\epsilon_1}.$$ Using (\[ineq3\]), $$\label{3.5} \int_{1}^{X/2^M} \int_{0}^{\Delta} (\psi(x+\delta x) -
\psi(x)- \delta x)^k d\delta \; dx \ll_k \Delta^{k+1} (X^{\mu +
\epsilon})^{k+1} \log^k {X}.$$ But, since $\mu \leq 1 - 2\epsilon - \epsilon_1$, $$\label{3.6}
\begin{split}
\Delta^{k/2+1} X^{k/2+1-\epsilon_1} &= \Delta^{k+1} X^{k/2 + 1 +
(k/2) \mu - \epsilon_1} = \Delta^{k+1} X^{(k+1)\mu + (k/2+1)
(1-\mu) - \epsilon_1} \\
&\geq \Delta^{k+1} X^{(k+1)\mu + (k/2+1) (2\epsilon + \epsilon_1)}
\gg \Delta^{k+1} (X^{\mu + \epsilon})^{k+1} \log^k {X}.
\end{split}$$ Combining (\[3.4\]), (\[3.5\]) and (\[3.6\]), we have $$\label{3.7}
\begin{split}
& \int_{0}^{\Delta} \int_{1}^{X} (\psi(x+\delta x) - \psi(x)
-\delta x)^k dx \; d\delta \\
=& {\mu_k \over {k \over 2}+1} \int_{0}^{\Delta X} h^{k/2}
\bigl(\log{X \over Eh} \bigr)^{k/2} dh + O_k({\Delta}^{k/2+1}
X^{k/2+1-\epsilon_1})
\end{split}$$ for $X^{-1 + 2\epsilon + \epsilon_1} \leq \Delta \leq
X^{-\epsilon}$.
We now deduce (\[3.1result\]) from (\[3.7\]). Set $\eta =
X^{-2 \epsilon_1 / 3}$. By Lemma \[lemma2.1\], one has for $X^{-1 + 2\epsilon + 2\epsilon_1} \leq \Delta \leq
X^{-\epsilon}/2$, $$\label{3.8}
\begin{split}
& \int_{\Delta}^{(1+\eta)\Delta} \int_{1}^{X} (\psi(x+\delta
x) - \psi(x) -\delta x)^k dx \; d\delta \\
=& {\mu_k \over {k \over 2}+1} \int_{\Delta X}^{(1+\eta)\Delta X}
h^{k/2} \bigl( log{X \over Eh} \bigr)^{k/2} dh +
O_k(\Delta^{k/2+1} X^{k/2+1-\epsilon_1}) \\
=& {\mu_k \over {k \over 2}+1} (\Delta X)^{{k/2}+1} \Bigl(\log{1
\over E\Delta} \Bigr)^{k/2} \eta + O_{k}\Bigl(\eta^2 (\Delta
X)^{k/2+1}(\log{1 \over \Delta})^{k/2} \Bigr) \\
&+ O_{k}(\Delta^{k/2+1} X^{k/2+1-\epsilon_1}).
\end{split}$$ Let $g(x,\delta x)=f(x,\Delta x)$ for $\Delta \leq \delta \leq
(1+\eta)\Delta$. Then one can easily check that $f(x,\delta
x)-g(x,\delta x)=f\bigl((1+\Delta)x,(\delta-\Delta)x \bigr)$. So, $$\label{3.9}
\begin{split}
& \int_{\Delta}^{(1+\eta)\Delta} \int_{1}^{X} \bigl(f(x,\delta
x)-g(x, \delta x ) \bigr)^k dx d\delta = \int_{0}^{\eta \Delta
\over 1+ \Delta} \int_{1+\Delta}^
{(1+\Delta)X} f(x,\delta x)^k dx d\delta \\
&\ll_{k} (\eta X \Delta)^{{k/2}+1} (\log{1 \over \eta
\Delta})^{k/2}
\end{split}$$ by (\[3.7\]), the choice of $\eta$ and the range of $\Delta$. Thus, by Lemma \[lemma2.3\], (\[3.8\]) and (\[3.9\]), $$\label{3.10}
\begin{split}
\int_{\Delta}^{(1+\eta)\Delta} \int_{1}^{X} g(x,\delta x)^k dx \;
d\delta &\ll_k
\int \int |f(x,\delta x)|^k + \int \int |f(x,\delta x) - g(x,\delta x)|^k \\
\ll_{k}& \eta X^{{k/2}+1} \Delta^{{k/2}+1} (\log{1 \over
\Delta})^{k/2} + \Delta^{{k/2}+1} X^{{k/2}+1 -\epsilon_1}.
\end{split}$$ By Lemma \[lemma2.2\] and Holder’s inequality, $$\begin{aligned}
& &\int_{\Delta}^{(1+\eta)\Delta} \int_{1}^{X} f(x,\delta x)^k -
g(x,\delta x)^k dx \; d\delta \\
&=&\int_{\Delta}^{(1+\eta)\Delta} \int_{1}^{X} P(f,g) (f-g) +
\int_{\Delta}^{(1+\eta)\Delta} \int_{1}^{X} (f-g)^k \\
&\ll&\Bigl(\int \int |P(f,g)|^{k/(k-1)}\Bigr)^{(k-1)/k} \Bigl(\int
\int |f-g|^k
\Bigr)^{1/k} + \int \int |f-g|^k \\
&=& J_1^{(k-1)/k} J_2^{1/k} + J_2\end{aligned}$$ where $P(x,y)$ is a homogeneous polynomial of degree $k-1$. $$J_2 \ll_{k} \eta^{{k/2}+1} X^{{k/2}+1} \Delta^{{k/2}+1} (\log{1 \over \eta
\Delta})^{k/2}$$ by (\[3.9\]). And $$\begin{aligned}
J_1 &\ll_k& \int \int \bigl(\sum_{i+j=k-1} |f|^i |g|^j \bigr)^{k/(k-1)} \\
&\ll_k& \sum_{i+j=k-1} \int \int \bigl(|f| + |g|\bigr)^k \mbox{ by
binomial
theorem} \\
&\ll_k& \int \int f^k + g^k \mbox{ by Lemma \ref{lemma2.3}}\\
&\ll_k& \eta X^{{k/2}+1} \Delta^{{k/2}+1}(\log{1 \over \eta
\Delta})^{k/2} + \Delta^{{k/2}+1} X^{{k/2}+1 -\epsilon_1} \mbox{
by (\ref{3.8}) and (\ref{3.10})} \\
&\ll_k& \eta X^{{k/2}+1} \Delta^{{k/2}+1}(\log{1 \over \eta
\Delta})^{k/2} \mbox{ as } \eta=X^{-2\epsilon_1 /3}.\end{aligned}$$ Consequently, by Lemma \[lemma2.3\], $$\label{3.11} \int \int f^k-g^k \ll_k X^{{k/2}+1} \Delta^{{k/2}+1}
\eta^{3/2}(\log{1 \over \eta \Delta})^{k/2}.$$ Therefore, by (\[3.11\]) and (\[3.8\]), $$\begin{aligned}
& &\eta \Delta \int_{1}^{X} (\psi(x+\Delta x) - \psi(x) -
\Delta x )^k dx = \int \int g^k \\
&=&\int \int f^k + O\Bigl(X^{{k/2}+1} \Delta^{{k/2}+1}
\eta^{3/2}(\log{1 \over \eta \Delta})^{k/2}\Bigr) \\
&=&{\mu_k \over {k\over 2}+1} (X \Delta)^{{k/2}+1} \Bigl(\log{1
\over E\Delta} \Bigr)^{k/2} \eta
+ O_k(\Delta^{k/2+1} X^{k/2+1-\epsilon_1})\\
& &+ O_k\Bigl(X^{k/2+1} \Delta^{k/2+1} \eta^{3/2} (\log{1 \over
\eta \Delta})^ {k/2}\Bigr).\end{aligned}$$ Dividing through by $\eta \Delta$, we have $$\begin{split}
& \int_{1}^{X} (\psi(x+\Delta x) - \psi(x) - \Delta x)^k dx \\
=& {\mu_k \over {k\over 2}+1} X^{{k/2}+1} \Delta^{k/2}
\Bigl(\log{1 \over E \Delta}\Bigr)^{k/2} + O_k\Bigl({\Delta^{k/2}
X^{k/2 + 1 - \epsilon_1} \over \eta} \Bigr) \\
&+ O_k\Bigl(X^{k/2+1} \Delta^{k/2} \eta^{1/2} (\log{1 \over \eta
\Delta})^ {k/2}\Bigr).
\end{split}$$ Finally, recall $\eta = X^{-2\epsilon_1/3}$, one has the error terms $\ll X^{{k/2}+1-{\epsilon_1/4}} \Delta^{k/2}$. So, the theorem is true with $\epsilon_2 = {\epsilon_1 \over 4}$.
(ii) $\Rightarrow$ (i)
======================
\[theorem4.1\] Assume RH. If, for some small $\epsilon
> \epsilon_1 > 0$ (small in terms of $k$), $$\label{4.1} \int_{1}^{X} (\psi(x+\delta x) - \psi(x) - \delta x
)^k dx = {\mu_k \over {k\over 2}+1} X^{{k/2}+1} \delta^{k/2}
\Bigl( \log{1 \over E\delta}\Bigr)^{k/2} + O_k(\delta^{k/2}
X^{k/2+1-\epsilon_1})$$ holds uniformly for $X^{-1+\epsilon} \leq \delta \leq
X^{-\epsilon}$, then $$\label{4.2} \int_{1}^{X} (\psi(x+h)-\psi(x)-h)^k dx = \mu_k
h^{k/2+1} \int_{E}^ {X/h} \bigl(\log{x \over E}\bigr)^{k/2} dx +
O_k(h^{k/2} X^{1-\epsilon_2})$$ holds uniformly for $X^{2\epsilon + \epsilon_1} \leq h \leq
X^{1-(k/2+1) \epsilon - 2\epsilon_1} / 2$ with some $\epsilon_2 >
0$.
Proof: Let $f(x,h)=\psi(x+h)-\psi(x)-h$. Let $X^{2\epsilon} \leq H
\leq X^{1 - (k/2+1)\epsilon - 2\epsilon_1)}$. Say $H = X^\mu$ for some $2\epsilon \leq \mu \leq 1 - (k/2+1)\epsilon - 2\epsilon_1$. First, we calculate $$\label{4.3} \int_{V/2}^{V} \int_{0}^{H} (\psi(x+h) - \psi(x) - h
)^k dh \; dx.$$ Substituting $\delta = {h \over x}$, (\[4.3\]) becomes $$\int_{H/V}^{2H/V} \int_{V/2}^{H/\delta} f(x,\delta x)^k x dx \; d\delta +
\int_{0}^{H/V} \int_{V/2}^{V} f(x,\delta x)^k x dx \; d\delta$$ $$=\int_{H/V}^{2H/V} \int_{V/2}^{H/\delta} + \int_{({V \over 2})^{-1+\epsilon}}^
{H/V} \int_{V/2}^{V} + \int_{0}^{({V \over 2})^{-1+\epsilon}}
\int_{V/2}^{V} = I_1+I_2+I_3.$$ By integration by parts, we have from (\[4.1\]) that $$\begin{aligned}
& & \int_{U}^{V} f(x,\delta x)^k x dx \\
&=& {\mu_k \over {{k \over 2}+2}} \delta^{k/2} \Bigl(\log{1 \over
E\delta}\Bigr)^ {k/2} \Bigl[V^{{k/2}+2} - U^{{k/2}+2}\Bigr] +
O_k(\delta^{k/2} V^{{k/2} +2-\epsilon_1})\end{aligned}$$ as long as $U^{-1+\epsilon} \leq \delta \leq V^{-\epsilon}$ with $U \leq V \leq 2U$. In order for this to work for $I_1$ and $I_2$, we need $$\label{4cond} \Bigl({V \over 2}\Bigr)^{-1+\epsilon} \leq {H \over
V} \leq V^{-\epsilon}, \mbox{ and } \Bigl({V \over
2}\Bigr)^{-1+\epsilon} \leq \delta \leq \Bigl({H \over
\delta}\Bigr)^{-\epsilon} \mbox{ for } {H \over V} \leq \delta
\leq {2H \over V}.$$ Since $H = X^\mu$, one can check that (\[4cond\]) are satisfied for $X^{\mu + \epsilon} \leq V \leq X$. Thus, $$\begin{aligned}
I_1 &=& {\mu_k \over {k \over 2}+2} \int_{H/V}^{2H/V}
\Bigl[\bigl({H \over \delta}\bigr)^{{k/2}+2} - \bigl({V \over
2}\bigr)^{{k/2}+2} \Bigr]
\delta^{k/2} \Bigl(\log{1 \over E\delta}\Bigr)^{k/2} d\delta \\
& &+ O_k(H^{{k/2}+1} V^{1-\epsilon_1}), \\
I_2 &=& {\mu_k \over {k \over 2}+2} \Bigl[V^{{k/2}+2} - \bigl({V
\over 2}\bigr)^ {{k/2}+2} \Bigr] \int_{0}^{H/V} \delta^{k/2}
\Bigl(\log{1 \over E\delta}\Bigr)^{k/2} d\delta \\
& &+ O_k(V^{1+({k/2}+1) \epsilon} (\log{V})^{k/2}) +
O_k(H^{{k/2}+1} V^{1- \epsilon_1}), \\
I_3 &\ll& V^{-1+\epsilon} V^2 V^{k\epsilon} \log^k {V} =
V^{1+(k+1)\epsilon} \log^k {V}.\end{aligned}$$ Let $\nu_k = \mu_k / ({k \over 2}+2)$, then (\[4.3\]) $$\begin{aligned}
&=& \nu_k H^{{k/2}+2} \int_{H/V}^{2H/V} {1 \over \delta^2}
\Bigl(\log{1 \over E\delta}\Bigr)^{k/2} d\delta + \nu_k
V^{{k/2}+2} \int_{0}^{H/V}
\delta^{k/2} \Bigl(\log{1 \over E\delta}\Bigr)^{k/2} d\delta \\
& &- \nu_k \Bigl({V \over 2}\Bigr)^{{k/2}+2} \int_{0}^{2H/V}
\delta^{k/2} \Bigl(\log{1 \over E\delta}\Bigr)^{k/2} d\delta \\
& &+O_k(V^{1+(k+1)\epsilon} \log^k {V}) + O_k(H^{{k/2}+1}
V^{1-\epsilon_1})\end{aligned}$$ when $X^{\mu + \epsilon} \leq V \leq X$. Now, replacing $V$ by $X2^{-l}$ in the above, summing over $0 \leq l \leq M =[{(1 - \mu
- \epsilon) \log{X} \over \log{2}}]$, $$\label{4.4}
\begin{split}
& \int_{X/2^M}^{X} \int_{0}^{H} (\psi(x+h) - \psi(x) - h)^k dh \; dx \\
=& \nu_k H^{{k/2}+2} \int_{H/X}^{2^{M+1}H/X} {1 \over \delta^2}
\Bigl(\log{1 \over E\delta}\Bigr)^{k/2} d\delta + \nu_k
X^{{k/2}+2} \int_{0}^{H/X} \delta^{k/2} \Bigl(\log{1 \over E\delta}
\Bigr)^{k/2} d\delta \\
&- \nu_k \Bigl({X \over 2^{M+1}}\Bigr)^{{k/2}+2}
\int_{0}^{2^{M+1}H/X} \delta^{k/2} \Bigl(\log{1 \over
E\delta}\Bigr)^{k/2} d\delta + O_k(H^{{k/2}+1} X^{1-\epsilon_1}) \\
=& \nu_k H^{{k/2}+2} \int_{H/X}^{1/E} {1 \over \delta^2}
\Bigl(\log{1 \over E\delta}\Bigr)^{k/2} d\delta + \nu_k
X^{{k/2}+2} \int_{0}^{H/X} \delta^{k/2} \Bigl(\log{1 \over E\delta}
\Bigr)^{k/2} d\delta \\
&+ O_k(H^{{k/2}+1} X^{1-\epsilon_1}).
\end{split}$$ because, as $X^{2\epsilon} \leq H$, $X^{1+(k+1)\epsilon} \log^k
{X} \ll H^{k/2+1} X^{1-\epsilon_1}$. Also, the terms involving $2^{M+1}$ are absorbed into the error term as $\mu \leq 1 -
\epsilon - 2\epsilon_1$. Using (\[ineq2\]), $$\label{4.5} \int_{1}^{X/2^{M}} \int_{0}^{H} \bigl(\psi(x+h)
-\psi(x) -h \bigr)^k dh \; dx \ll H (X^{\mu + \epsilon})^{k/2+1}
\log^{2k} {X}.$$ But, since $H = X^\mu \leq X^{1 - (k/2+1)\epsilon - 2\epsilon_1}$, $$\label{4.6}
\begin{split}
H (X^{\mu + \epsilon})^{k/2+1} \log^{2k} {X} \leq & X^{1 -
(k/2+1)\epsilon - 2\epsilon_1} H^{k/2+1} X^{(k/2+1)\epsilon}
\log^{2k} X \\
\ll & H^{k/2+1} X^{1-\epsilon_1}.
\end{split}$$ Combining (\[4.4\]), (\[4.5\]) and (\[4.6\]), we have $$\label{4.7}
\begin{split}
& \int_{0}^{H} \int_{1}^{X} \bigl(\psi(x+h) - \psi(x) -h \bigr)^k dx \;
dh \\
=& \nu_k H^{{k/2}+2} \int_{H/X}^{1/E} {1 \over \delta^2}
\Bigl(\log{1 \over E \delta}\Bigr)^{k/2} d\delta + \nu_k
X^{{k/2}+2} \int_{0}^{H/X} \delta^{k/2}
\Bigl(\log{1 \over E\delta}\Bigr)^{k/2} d\delta \\
&+ O_k(H^{{k/2}+1} X^{1-\epsilon_1})
\end{split}$$ for $X^{2\epsilon} \leq H \leq X^{1-(k/2+1)\epsilon -
2\epsilon_1}$.
We now deduce (\[4.2\]) from (\[4.7\]). Set $\eta=X^{-2\epsilon_1/3}$. For $X^{2\epsilon + \epsilon_1} \leq H
\leq X^{1-(k/2+1)\epsilon - 2\epsilon_1}/2$, $$\begin{aligned}
& &\int_{H}^{(1+\eta)H} \int_{1}^{X} (\psi(x+h) - \psi(x) - h)^k
dx \; dh \\
&=& \nu_k ((1+\eta)H)^{k/2+2} \int_{(1+\eta)H/X}^{1/E} {1 \over
\delta^2}
\Bigl(\log{1 \over E\delta}\Bigr)^{k/2} d\delta \\
& & - \nu_k H^{k/2+2} \int_{H/X}^{1/E} {1 \over \delta^2}
\Bigl(\log{1 \over E \delta}\Bigr)^{k/2} d\delta \\
& &+ \nu_k X^{k/2+2} \int_{H/X}^{(1+\eta)H/X} \delta^{k/2}
\Bigl(\log{1 \over E \delta}\Bigr)^{k/2} d\delta
+ O_k(H^{{k/2}+1} X^{1-\epsilon_1}) \\
&=& -\nu_k H^{k/2+2} \int_{H/X}^{(1+\eta)H/X} {1 \over \delta^2}
\Bigl(\log{1 \over E\delta}\Bigr)^{k/2} d\delta + \mu_k \eta
H^{k/2+2} \int_{H/X}^{1/E} {1 \over \delta^2} \Bigl(\log{1 \over
E\delta}\Bigr)^{k/2} d\delta \\
& &+ \nu_k X^{k/2+2} \int_{H/X}^{(1+\eta)H/X} \delta^{k/2}
\Bigl(\log{1 \over E \delta} \Bigr)^{k/2} d\delta \\
& &+ O_k\Bigl(\eta^2 H^{k/2+1} X (\log{X \over H})^{k/2}\Bigr) +
O_k(H^{k/2+1} X^{1-\epsilon_1}) \\
&=& \mu_k \eta H^{k/2+2} \int_{H/X}^{1/E} {1 \over \delta^2}
\Bigl(\log{1 \over E\delta}\Bigr)^{k/2} d\delta \\
& &+ O_k\Bigl(\eta^2 H^{k/2+1} X (\log{X \over H})^{k/2}\Bigr) +
O_k(H^{k/2+1} X^{1-\epsilon_1})\end{aligned}$$ by Lemma \[lemma2.1\]. Therefore $$\label{4.8}
\begin{split}
& \int_{H}^{(1+\eta)H} \int_{1}^{X} \bigl(\psi(x+h) - \psi(x) -h
\bigr)^k dx \; dh \\
=& \mu_k \eta H^{k/2+2} \int_{E}^{X/H} \bigl(\log{u \over E}
\bigr)^{k/2} du \\
&+ O_k\Bigl(\eta^2 H^{k/2+1} X (\log{X \over H})^{k/2}\Bigr) +
O_k(H^{{k/2}+1} X^{1-\epsilon_1}).
\end{split}$$ Let $g(x,h)=f(x,H)$ for $H \leq h \leq (1+\eta)H$. Again, one can check that $f(x,h)-g(x,h)=f(x+H,h-H)$. So, $$\label{4.9}
\begin{split}
\int_{H}^{(1+\eta)H} \int_{1}^{X} \bigl(f(x,h)-g(x,h)\bigr)^k dx
dh &= \int_{0}^{\eta H} \int_{1+H}^{X+H} f(x,h)^k dxdh \\
&\ll_k \eta^{k/2+1} X H^{k/2+1} \Bigl(\log{X \over \eta
H}\Bigr)^{k/2}
\end{split}$$ by (\[4.7\]) as well as the choice of $\eta$ and the range of $H$. Thus, by Lemma \[lemma2.3\], (\[4.8\]) and (\[4.9\]), $$\label{4.10}
\begin{split}
\int_{H}^{(1+\eta)H} \int_{1}^{X} g(x,h)^k dx \; dh &\ll_k \int
\int |f(x,h)|^k + \int \int |f(x,h)-g(x,h)|^k \\
&\ll_k \eta X H^{k/2+1} (\log{X \over \eta H})^{k/2}+
H^{k/2+1}X^{1-\epsilon_1}.
\end{split}$$ By Lemma \[lemma2.2\] and Holder’s inequality, $$\label{4i1}
\begin{split}
&\int_{H}^{(1+\eta)H} \int_{1}^{X} f(x,h)^k - g(x,h)^k dx d\delta \\
=&\int_{H}^{(1+\eta)H} \int_{1}^{X} P(f,g)(f-g) +
\int_{H}^{(1+\eta)H} \int_{1}
^{X} (f-g)^k \\
\ll& \Bigl(\int \int |P(f,g)|^{k/(k-1)}\Bigr)^{(k-1)/k} \Bigl(\int
\int |f-g|^k\Bigr)^{1/k} + \int \int |f-g|^k \\
=&K_1^{(k-1)/k} K_2^{1/k} + K_2
\end{split}$$ where $P(x,y)$ is a homogeneous polynomial of degree $k-1$. From (\[4.9\]), $$\label{4i2} K_2 \ll_k \eta^{k/2+1} X H^{k/2+1} \Bigl(\log{X \over
\eta H}\Bigr)^{k/2}.$$ And similar to the proof in Theorem \[theorem3.1\], $$\label{4i3} K_1 \ll_k \int \int f^k+g^k \ll_k \eta X
H^{k/2+1}(\log{X \over \eta H})^{k/2}$$ by (\[4.8\]) and (\[4.10\]). Consequently, by (\[4i1\]), (\[4i2\]), (\[4i3\]) and Lemma \[lemma2.3\], $$\label{4.11} \int \int f^k -g^k \ll_k \eta^{3/2} X H^{k/2+1}
\Bigl(\log{X \over \eta H} \Bigr)^{k/2}.$$ Therefore, by (\[4.11\]) and (\[4.8\]), $$\begin{aligned}
& &\eta H \int_{1}^{X} \bigl(\psi(x+h) - \psi(x) -h \bigr)^k dx =
\int \int g^k \\
&=&\int \int f^k + O_k\Bigl(X H^{k/2+1} \eta^{3/2} (\log{X \over
\eta H})^{k/2}\Bigr) \\
&=& \mu_k \eta H^{k/2+2} \int_{E}^{X/H} \bigl(\log{u \over
E}\bigr)^{k/2} du + O_k(H^{{k/2}+1} X^{1-\epsilon_1}) \\
& &+ O_k\Bigl(X H^{k/2+1} \eta^{3/2} (\log{X \over \eta
H})^{k/2}\Bigr).\end{aligned}$$ Divide through by $\eta H$ and recall $\eta = X^{-2\epsilon_1/3}$, we get the theorem with $\epsilon_2 = {\epsilon_1 \over 4}$.
Numerical evidence
==================
In Montgomery and Soundararajan \[\[MS1\]\], they got some numerical data for the actual values of $M_k(X;h)$. One has the following table:
For $X = 10^{10}$ and $h = 10^5$.
[lll]{} k & M\_k(X;h) &\
2 & 9.066310\^[15]{} & 9.097810\^[15]{}\
4 & 2.499510\^[22]{} & 2.513110\^[22]{}\
6 & 1.157310\^[29]{} & 1.167510\^[29]{}
Using a C program, we get some numerical evidence in support of the truth of (ii) in Theorem \[theorem1.1\].
For $X=10^8$ and $\delta=10^{-4}$:
[lll]{} k & \_k(X;) &\
2 & 4.007510\^[12]{} & 3.897610\^[12]{}\
4 & 6.516110\^[17]{} & 6.076610\^[17]{}\
6 & 1.959210\^[23]{} & 1.776310\^[23]{}
For $X=10^{10}$ and $\delta=10^{-5}$:
[lll]{} k & \_k(X;) &\
2 & 5.052710\^[15]{} & 5.048510\^[15]{}\
4 & 1.021010\^[22]{} & 1.019510\^[22]{}\
6 & 3.864510\^[28]{} & 3.860210\^[28]{}
[9]{}
\[C\] T.H. Chan, [*More Precise Correlation of Zeros and Primes in Short Intervals*]{}, J. London Math. Soc. (2) [**68**]{}, 2003, no. 3, 579 - 598.
\[GM\] D.A. Goldston and H.L. Montgomery, [*On pair correlations of zeros and primes in short intervals*]{}, Analytic Number Theory and Diophantine Problems (Stillwater, OK, July 1984), Prog. Math. [**70**]{}, Birkauser, Boston, 1987, pp. 183-203.
\[M\] H.L. Montgomery, [*The pair correlation of zeros of the zeta function*]{}, Analytic Number Theory (St. Louis Univ., 1972), Proc. Sympos. Pure Math. [**24**]{}, Amer. Math. Soc., Providence, 1973, pp. 181-193.
\[MS1\] H.L. Montgomery and K. Soundararajan, [*Beyond pair correlation*]{}, Paul Erd[ö]{}s and his mathematics, I (Budapest, 1999), Bolyai Soc. Math. Stud. [**11**]{}, János Bolyai Math. Soc., Budapest, 2002, pp. 507-514.
\[MS2\] H.L. Montgomery and K. Soundararajan, [*Primes in Short Intervals*]{}, preprint.
\[SV\] B. Saffari and R.C. Vaughan, [*On the fractional parts of x/n and related sequences II*]{}, Ann. Inst. Fourier (Grenoble) (2) [**27**]{}, 1977, 1-30.
\[vK\] H. von Koch, [*Sur la distribution des nombres premiers*]{}, Acta Math. [**24**]{}, 1901, 159-182.
Tsz Ho Chan\
American Institute of Mathematics\
360 Portage Avenue\
Palo Alto, CA 94306\
USA\
[email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Complex motions for robots are frequently generated by switching among a collection of individual movement primitives. We use this approach to formulate robot motion plans as sequences of primitives to be executed one after the other. When dealing with *dynamical* movement primitives, besides accomplishing the high-level objective, planners must also reason about the effect of the plan’s execution on the safety of the platform. This task becomes more daunting in the presence of disturbances, such as external forces. To alleviate this issue, we present a framework that builds on rigorous control-theoretic tools to generate safely-executable motion plans for externally excited robotic systems. Our framework is illustrated on a 3D limit-cycle gait bipedal robot that adapts its walking pattern to persistent external forcing.'
author:
- 'Sushant Veer and Ioannis Poulakakis [^1] [^2]'
bibliography:
- 'bib\_lib.bib'
title: '**Safe Adaptive Switching among Dynamical Movement Primitives: Application to 3D Limit-Cycle Walkers**'
---
Introduction {#sec:intro}
============
Robots operating in the real world are expected to encounter a wide range of exogenous input signals due to contact or other types of interaction with a possibly time-varying, stochastic environment. Depending on the task, external signals may represent commands that need to be followed or disturbances that must be attenuated. A diverse collection of suitable primitive motions, and the capability to *switch* among them, can provide a sufficiently rich repertoire of behaviors for adapting to or compensating for such signals. For example, in a human-robot physical collaboration scenario [@motahar2015impedance; @motahar2017steering], a force exerted by a human with the intention of decelerating the robot can be followed by switching to a primitive of lower velocity. On the other hand, an undesirable force that pushes against a robot, which is tasked to maintain a constant desired velocity, can be compensated by switching to a primitive of higher velocity. Adopting a dynamical systems perspective, one approach to characterizing primitive motions is to represent them as attractors of dynamical systems, termed dynamical movement primitives[^3] (DMPs) in the relevant literature [@hogan2013dynamic; @ijspeert2013dynamical; @burridge1999sequential]. Adjusting the “landscape” of such attractors through coupling terms can realize both discrete and rhythmic motion patterns of high complexity, allowing a robot to perform challenging tasks in its workspace [@ijspeert2013dynamical]. An alternative way to generate sufficiently rich robot behaviors is to consider a discrete collection of suitably selected DMPs, and allow switching among them [@burridge1999sequential; @tedrake2010]. This paper focuses on the latter, and it leverages recent theoretical developments in [@veer2018ultimate; @veer2017poincare] to provide *explicitly computable* sufficient conditions that provably guarantee the safety of the robotic platform as it switches among different DMPs to accomplish a task and compensate for disturbances.
Safe switching among DMPs occupies a significant body of work in the literature of robotic motion planning. Inspired by early work in [@mason1985mechanics], a tree of positively-invariant Lyapunov funnels was constructed in [@burridge1999sequential] to generate controller switching policies that drive a robot to a goal while ensuring safety of the robotic platform. Later work in [@tedrake2010] provided a computationally tractable approach to estimate the Lyapunov funnels construction of [@burridge1999sequential] based on Sum-of-Squares (SoS) programming [@parrilo2000structured]. This method and its extensions have been experimentally successful for planning motions of a wheeled robot [@conner2011integrating], ball-bot [@nagarajan2013integrated], and a fixed-wing airplane [@majumdar2017funnel]. Other methods address composability of dynamic primitives, capturing the associated constraints in the form of maneuver [@frazzoli2005maneuver] or timed [@bouyer2017timed] automata. None of the aforementioned methods deals with persistent disturbances except [@majumdar2017funnel], which though requires knowledge of the disturbed dynamics to ensure safe operation; see [@majumdar2017funnel Section 4.3.1]. On the contrary, the conditions provided here for safe switching in the presence of persistent disturbances rely only on the zero-disturbance stability properties of the individual DMPs, and thus are *agnostic* to the disturbances. Our focus in this paper is on rhythmic motions of dynamically-stable robots; typical examples of such systems include walking or running machines [@raibert1985legged] and flapping-wing flying robots [@ramezani2017describing]. Mathematically, such motions can be idealized as attracting limit cycles that capture the fundamental oscillatory behavior of the underlying energy transformations. Restricting attention to legged robots, a variety of methods have been proposed to stabilize limit-cycle locomotion behaviors, including hybrid zero dynamics (HZD) [@westervelt2007feedback], geometric control reduction [@Gregg2010Geometric], and virtual holonomic constraints [@freidovich2009passive], just to name a few. These methods aim at stabilizing *individual* limit cycles that result in locomotion behaviors with certain desired attributes. However, to address the challenges faced by a legged robot moving in a time-varying environment, these individual limit cycles must be composed in response to external stimuli to form more complex motion patterns. Enabling such compositions is at the core of the proposed framework, the practical value of which lies on its ability to take such off-the-shelf robot controllers for individual primitive behaviors and switch among them in response to external signals, while affording rigorous safety guarantees. It should be noted that switching among limit-cycle gaits has been explored in the context of various applications, including navigation in environments cluttered by obstacles [@gregg-planning2012; @motahar2016composing; @veer2017driftless], speed adaptation [@veer2017continuum; @bhounsule2018switching], and robustness to disturbances [@saglam2013switching; @quan2016dynamic; @da2016first]. With the exception of [@saglam2013switching] which provides stochastic stability guarantees, the switching policies discussed in the aforementioned methods do not account for persistent external excitation.
In this paper, building upon our recent work [@veer2018ultimate], we present a general framework that formulates switching among externally excited DMPs as a switched system with multiple equilibria under disturbances. However, in [@veer2018ultimate], we assumed global stability properties for the individual systems and demanded knowledge of the external signal’s effect on them—either of which rarely hold true for robotic systems. The method we provide here, remedies these limitations by allowing us to furnish safety guarantees for switching under disturbances by studying an *unperturbed* switched system (Theorem \[thm:dist-agnostic\]). The framework is demonstrated for locomotion adaptation of a 3D bipedal walker that collaborates with a leader by switching among limit-cycle primitives.
![Adaptive planning block diagram. The high-level supervisor is shown in yellow and the low-level control loop is shown in blue.[]{data-label="fig:block-diagram"}](figures/block_diagram.pdf){width="0.95\columnwidth"}
Overview: Switching Framework {#sec:overview}
=============================
In this section we provide an overview of our motion planning framework; technical details are relegated for later sections. Our planning framework is organized in two hierarchical levels as shown in Fig. \[fig:block-diagram\]. On the high level, there is a supervisor that comprises a library of primitives and a switching logic that governs the choice of the primitive to be implemented on the robot. On the low level, there is a feedback control loop which executes the controller of the primitive chosen by the supervisor. It is the supervisor’s role to orchestrate switching among members of the primitive library in accordance with a prescribed high-level objective—such as navigate to a goal while avoiding obstacles [@motahar2016composing] or adapt to a collaborating leader [@veer2017supervisory]. However, due to the dynamics of the system and possible external excitation, the generated sequence of primitive switches may lead to instability, requiring the switching logic to reason about the dynamic limitations of the robot; to address these limitations, we consider the effect of switching among primitives on the robot dynamics.
We work with primitives that take the form of dynamical systems with equilibrium behaviors—equilibrium points or limit cycles. Hence, executing a motion plan naturally induces a switched system structure to the dynamics of the robot. In Section \[sec:switch-guarantees\], we analyze the switched system that arises and identify a class of switching signals that can be safely executed by the robot despite disturbances. This class of switching signals is characterized by a lower bound on the average time-gap between any two consecutive switches—formally known as average dwell-time. Informing the supervisor about this constraint allows the switching logic to generate motion plans that can be safely executed.
This framework can be used for adaptive planning in various scenarios, such as navigation of a robot in the presence of disturbances [@majumdar2017funnel], adapting to rough terrain [@quan2016dynamic], and teleoperation [@yang2018online], to name a few. In this paper, we apply the framework to a task that involves physical cooperation between a bipedal robot and a leading collaborator—a human or a robot. The success of such a task hinges on the ability of the biped to adapt its walking pattern in response to the leader’s intended trajectory, which is not explicitly available but is indirectly known through an interaction force applied by the leader. Hence, the interaction force serves as a command signal that the biped must adapt to. To achieve this, we supply the supervisor with a library of limit-cycle primitives—straight walking, turn right, and turn left—as shown in Fig. \[fig:block-diagram\]. The supervisor is provided with an appropriate input $\Phi_k$ that encodes the leader’s intention through the interaction force, and an average dwell-time constraint, which it uses to *safely* adapt the biped’s motion to the leader’s intended trajectory. Section \[sec:example\] of the paper will discuss this scenario in detail.
Safe Switching under Disturbances {#sec:switch-guarantees}
=================================
This section formalizes the concepts that underlie the framework briefly outlined above; see Fig \[fig:block-diagram\].
Library of Motion Primitives {#subsec:lib-prim}
----------------------------
In the proposed approach, motion primitives are characterized by point attractors of continuous- or discrete-time dynamical systems together with the vector fields capturing the corresponding dynamic behavior. Limit cycle primitives—such as those frequently employed in rhythmic behaviors—can be incorporated in our approach naturally, as they can be associated with point attractors of suitably constructed discrete-time systems via the method of Poincaré [@GUHO96]. Hence, in what follows we will develop the method in the context of discrete-time systems, noting that analogous results hold for continuous-time systems as well [@veer2018ultimate]. Let $\mathcal{P}$ be a finite index set and consider a collection of discrete-time nonlinear systems $$\begin{aligned}
\label{eq:discrete_system}
x_{k+1} = f_p(x_k) \enspace, & & p\in\mathcal{P} \enspace,\end{aligned}$$ where $x$ represents the state vector evolving in a space $\mathcal{X}_p\subseteq \mathbb{R}^n$, $f_p$ is a vector field describing the dynamics of the system and $k$ denotes discrete time. We can now formalize our notion of motion primitives, which are defined as two-tuples consisting of a vector field $f_p$ and a corresponding equilibrium (fixed) point $x_p^*$ satisfying $x_p^*=f_p(x_p^*)$; that is, $$\begin{aligned}
\label{eq:primitives}
\mathcal{G}_p:=\{f_p,x_p^*\} \enspace, & & p\in\mathcal{P} \enspace.\end{aligned}$$ The totality $\mathbb{G} := \{ \mathcal{G}_p ~|~ p \in \mathcal{P} \}$ of the motion primitives defines a library of motion primitives.
To ensure safe operation, $\mathbb{G}$ will only include motion primitives that correspond to exponentially stable fixed points. Safety certificates of this form can be obtained through the notion of a Lyapunov function [@khalil2002nonlinear]; i.e., a positive definite, radially unbounded, decrescent function $V_p$ that satisfies $$\begin{aligned}
V_p(f_p(x)) & \leq \lambda V_p(x) \label{eq:V-2-0-inp}\enspace,\end{aligned}$$ where[^4] $0<\lambda<1$. To implement switching policies among motion primitives that afford rigorous safety guarantees, we will also need an estimate of the basin-of-attraction (BoA) $\mathcal{B}_p$ associated with each $\mathcal{G}_p$ in $\mathbb{G}$. To conveniently characterize such estimates, we use sublevel sets of Lyapunov functions verified through sums-of-squares (SoS) programming [@parrilo2000structured]. The dimensional reduction afforded by the controllers employed in Section \[sec:example\] greatly improves computational efficiency in obtaining such estimates.
The motion primitives defined by can be employed directly to plan or adapt rhythmic behaviors to external signals; a concrete example is provided in Section \[sec:example\]. Indeed, such tasks can be facilitated via switching among limit cycle motion primitives $\{ \mathcal{O}_p ~|~ p \in \mathcal{P} \}$. By the method of Poincaré, however, each limit cycle $\mathcal{O}_p$ can be naturally associated with a fixed point $x_p^*$ of a discrete-time system with $f_p$ being the corresponding Poincaré map [@GUHO96].
Conditions for Safe Switching {#subsec:switch-system}
-----------------------------
A motion planner—or, a supervisor—is responsible for monitoring the state of the system as it interacts with its environment and deciding which motion primitive $\mathcal{G}_p$ out of the library $\mathbb{G}$ must be implemented at each time instant. This decision can be represented as a “descending” switching signal $\sigma:\mathbb{Z}_+\to\mathcal{P}$, which maps the current time $k$ to the index $p=\sigma(k)\in\mathcal{P}$ of the member $\mathcal{G}_{\sigma(k)}$ of $\mathbb{G}$ that must be executed at $k$. The process gives rise to a discrete switched system with multiple equilibria that has the form $$\label{eq:switched-system}
x_{k+1} = f_{\sigma(k)}(x_k) \enspace.$$ Quantifying safety for systems, the behavior of which is governed by a switched system like , can be challenging; yet, such systems emerge in a wide range of applications where a supervisor chooses among different controllers, each being suited for a particular mode of operation. Switching in effectively causes the system to “shift” to a different point attractor, and thus persistent switching in response to constantly varying environmental or task conditions causes the system to be in a “permanent” transient phase, *never* converging to any of the underlying equilibrium states. The resulting evolution can be highly irregular.
In this work, we will adopt the following set-characterization for safety. The switched system will be considered safe, if the following condition holds $$\begin{aligned}
\label{eq:safety_criterion}
\big\{ x^*_p \in \bigcap_{p\in\mathcal{P}} \mathcal{B}_p \text{~~and~~}
x_k \in \bigcap_{p\in\mathcal{P}} \mathcal{B}_p \text{~for all~} k \in \mathbb{Z}_+ \big\}\enspace,\end{aligned}$$ where $\mathcal{B}_p$ is an estimate of the BoA associated with the primitive $\mathcal{G}_p$. In words, this condition implies that the state of always remain trapped in a compact subset of the state space, which is explicitly characterized as the intersection of the estimates of the BoAs of all the motion primitives and includes all the equilibrium points. Our motivation for adopting this criterion is twofold. First, it implies that the state remains bounded. Second, it guarantees that if switching were to cease, the system would return to the equilibrium corresponding to the most recently implemented primitive.
The question we address next is to identify a class of switching signals $\sigma$ which guarantee that the condition is satisfied. Loosely speaking, we will require that switching is sufficiently slow on average. Intuitively, we require that the average time gap between any two consecutive switches is sufficiently long to ensure that the energy dissipated due to the exponentially stable nature of each primitive $\mathcal{G}_p$ dominates the possible energy gain due to a switch, resulting in an overall energy reduction. To make this precise, we will use the notion of the average dwell time, introduced in [@hespanha1999stability].
\[def:avg-dwell\] A switching signal $\sigma:\mathbb{Z}_+\to\mathcal{P}$ has average dwell-time $N_{\rm a}>0$ if the number of switches $N_\sigma(k,\underline{k})$ in any discrete interval $[\underline{k},k)\cap\mathbb{Z}_+$ satisfies $$\begin{aligned}
\label{eq:avg-dwell-time-def}
N_\sigma(k,\underline{k}) \leq N_0 + \frac{k-\underline{k}}{N_{\rm a}} \enspace, & & \forall k\geq \underline{k} \geq 0 \enspace,\end{aligned}$$ where $k,\underline{k}\in\mathbb{Z}_+$ and $N_0>0$ is a finite constant.
Below, we introduce a set construction that will allow us to obtain explicitly computable expressions for an average dwell time constraint, which guarantees that is satisfied; a rigorous justification can be found in [@veer2018ultimate]. We begin by choosing a $\kappa>0$ and defining the $\kappa$-sublevel set of $V_p$ $$\nonumber
\mathcal{M}_p(\kappa):=\{x\in\mathbb{R}^n~|~V_p(x)\leq \kappa\} \enspace;$$ see Fig. \[fig:set-construction\]. Let $\mathcal{M}(\kappa):=\bigcup_{p\in\mathcal{P}} \mathcal{M}_p(\kappa)$ denote the union of these subsets over $\mathcal{P}$, and define $$\label{eq:omega-def}
\omega(\kappa):=\max_{p\in\mathcal{P}}\max_{x\in\mathcal{M}(\kappa)} V_p(x) \enspace.$$ Then, it can be seen that $ \mathcal{M}(\kappa) \subseteq \bigcap_{p\in\mathcal{P}} \mathcal{M}_p(\omega(\kappa))$ as in Fig. \[fig:set-construction\]. Effectively, $\omega$ “inflates” the sets $\mathcal{M}_p(\kappa)$ to the sets $\mathcal{M}_p(\omega(\kappa))$, the intersection of which contains $\mathcal{M}(\kappa)$. To bound possible “energy" gain due to switching, let[^5] $$\label{eq:mu-def}
\mu(\kappa):=\max_{p,q\in\mathcal{P}} \max_{x\in \mathcal{B}_p\setminus \accentset{\circ}{\mathcal{M}}_p(\kappa)} \frac{V_q(x)}{V_p(x)} \enspace,$$ which captures the ratio of all Lyapunov functions and is well-defined since $\mathcal{P}$ is finite and $\mathcal{B}_p\setminus \accentset{\circ}{\mathcal{M}}_p(\kappa)$ is compact; the exclusion of $\accentset{\circ}{\mathcal{M}}_p(\kappa)$ is to prevent the denominator of from approaching 0, as $V_p(x_p^*)=0$. Note also that the interchangeability of $p$ and $q$ implies that $\mu(\kappa)\geq 1$.
With $\omega(\kappa)$ and $\mu(\kappa)$ as in and and $\lambda$ as in , a lower bound on the average dwell-time can be computed as $$\label{eq:avg-dwell}
\overline{N}_{\rm a}= \frac{\ln{\mu(\kappa)}}{\ln{(\epsilon/\lambda)}} \enspace,$$ where $\epsilon$ is an arbitrary constant in $(\lambda,1)$. Switching signals with $N_\mathrm{a} \geq \overline{N}_{\rm a}$ guarantee that the evolution of remains bounded. Intuitively, a large $\mu$ implies that the “energy” gain due to switching may be large, causing $\overline{N}_{\rm a}$ to increase, so that choosing $N_{\rm a} \geq \overline{N}_{\rm a}$ in will result to slower switching on average. Similarly, a slow rate of convergence $\lambda$ during the interval between switches will have a similar effect.
Satisfying $N_\mathrm{a} \geq \overline{N}_{\rm a}$ ensures boundedness of the state, however, according to we also need to ensure that the corresponding bounded “trapping” set lies within $\bigcap_{p\in\mathcal{P}}\mathcal{B}_p$. It turns out that, to comply with this requirement, the chosen $\kappa$ must be such that the condition $$\label{eq:feasible-0-inp}
\overline{\mathcal{M}}:=
\mathcal{M}(\mu(\kappa)^{\overline{N}_0}\omega(\kappa)) \subset \bigcap_{p\in\mathcal{P}} \accentset{\circ}{\mathcal{B}}_p \enspace.$$ is verified for some $\overline{N}_0\geq 1$. Then, for any $\sigma$ with $$N_0 \leq \overline{N}_0 \text{~~and~~} N_{\rm a}\geq \overline{N}_{\rm a} \enspace,$$ the solution of starting from any $x_0 \in \bigcap_{p\in\mathcal{P}} \mathcal{M}_p(\omega(\kappa))$ will remain within $\bigcap_{p\in\mathcal{P}}\mathcal{B}_p$ for all $k\in\mathbb{Z}_+$. Note also that since $\mathcal{M}$ represents the union of sublevel sets, it includes all the equilibrium points, thereby implies that the first part of the condition is also satisfied.
{width="0.65\columnwidth"}
\[fig:set-construction\]
The rigorous justification of this result—which is of “local” nature—is due to Theorem \[thm:dist-agnostic\] below, which in fact establishes that the bounds $(\overline{N}_0, \overline{N}_{\rm a})$ remain valid even in the case where the system is perturbed by external disturbances, as long as they are sufficiently small. The proof of this theorem, which is made available in the appendix, also provides a clear motivation for the set constructions described above. For the purpose of implementation though, the procedure of obtaining $(\overline{N}_0, \overline{N}_{\rm a})$ can be summarized as follows. First, select a $\kappa>0$ and compute $\omega(\kappa)$ and $\mu(\kappa)$ by and , respectively. If holds for some $\overline{N}_0 \geq 1$, then safe switching is achieved by providing the supervisor with the numbers $(\overline{N}_0, \overline{N}_{\rm a})$ where $\overline{N}_{\rm a}$ is obtained by . Otherwise, choose a new $\kappa$ and repeat the procedure.
Note that, for a given $\kappa>0$, obtaining $\omega(\kappa)$ and $\mu(\kappa)$ numerically may be computationally challenging, particularly for high-dimensional systems. However, for quadratic Lyapunov functions—as is frequently the case in practical applications—an upper bound for $\omega$ and $\mu$ can be *analytically* computed using [@veer2018ultimate Proposition 1].
Conditions for Safe Switching: The case of disturbances {#subsec:switch-system-disturb}
-------------------------------------------------------
In this section, we state a theorem that rigorously justifies the assertions informally made in the previous section. With a slight abuse of notation, consider the perturbed version $$\begin{aligned}
\label{eq:switch-system-disturb}
x_{k+1} = f_{\sigma(k)}(x_k,d_k) \enspace,\end{aligned}$$ of the switched system . In the disturbance signal is represented as a sequence $d:=\{d_k\}_{k\in\mathbb{Z}_+}$ with values in the set of admissible disturbances[^6] $\mathcal{D}$. Let $\|\cdot\|_{\mathcal{D}}$ be the norm on $\mathcal{D}$ and define $\|d\|_\infty:=\sup_{k\in\mathbb{Z}_+} \|d_k\|_{\mathcal{D}}$. We further assume that the vector fields $f_p:\mathcal{X}_p\times\mathcal{D}\to\mathcal{X}_p$ are locally Lipschitz.
Then, Theorem \[thm:dist-agnostic\] below shows that the switching conditions stated in the absence of disturbances in Section \[subsec:switch-system\] guarantee that the switched system will be safe according to despite the presence of disturbances.
\[thm:dist-agnostic\] Consider where $\sigma$ is a switching signal and $\sigma(k) = p \in \mathcal{P}$. Assume that
(i) for each $p\in\mathcal{P}$, $x^*_p$ is an exponentially stable equilibrium of $f_p$ in the absence of disturbances, let $V_p$ be a locally Lipschitz Lyapunov function and $\mathcal{B}_p$ an estimate of the BoA based on $V_p$;
(ii) there exists a $\kappa>0$, $\overline{N}_0>0$ such that holds.
Then, there exists a $\delta>0$ such that for any disturbance $\{d_k\}_{k\in\mathbb{Z}_+}$ with $\|d\|_\infty<\delta$, and for any switching signal $\sigma$ that satisfies Definition \[def:avg-dwell\] with $N_0 \leq \overline{N}_0$ and $N_{\rm a}\geq \overline{N}_{\rm a}$ defined in , the solution of satisfies $$\begin{aligned}
\label{eq:target-set}
x_0 \in \bigcap_{p\in\mathcal{P}} \mathcal{M}_p(\omega(\kappa)) \Rightarrow x_k \in \mathcal{M}(\bar{\omega}(\|d\|_\infty)) \subset \bigcap_{p\in\mathcal{P}} \mathcal{B}_p \enspace,\end{aligned}$$ for all $k\in\mathbb{Z}_+$, where $\bar{\omega}(\|d\|_\infty):=\mu(\kappa)^{N_0}\omega(\kappa) + \alpha(\|d\|_\infty)$ and[^7] $\alpha\in\mathcal{K}_\infty$.
The proof of Theorem \[thm:dist-agnostic\] is detailed in the appendix.
Adaptation of a 3D Limit-Cycle Biped {#sec:example}
====================================
In this section we apply the proposed switching framework to adapt the walking gaits of a 3D limit-cycle biped to an externally applied force.
Task and Robot Model
--------------------
Our motivation stems from collaborative object transportation tasks, in which a leading co-worker—either a robot or a human—physically interacts with a biped to direct its motion. It is assumed that the leader’s intention can be represented as a sufficiently smooth trajectory $p_{\rm L}(t)$. Although the biped does not know $p_{\rm L}(t)$ explicitly, it experiences an interaction force $F_{\rm e}(t)$ applied by the collaborator, which carries information about the intended trajectory. Following [@rahman2002investigation; @tsuji2004analysis], to model this interaction the leader’s intention $p_{\rm L}(t)$ is translated to the force $F_{\rm e}(t)$ using an impedance model $$\nonumber
F_{\rm e}(t) = K_{\rm L}(p_{\rm L}(t)-p_{\rm E}(t)) + N_{\rm L}(\dot{p}_{\rm L}(t)-\dot{p}_{\rm E}(t)) \enspace,$$ where $p_{\rm E}$ is the point at which the force is applied and $K_{\rm L}$ and $N_{\rm L}$ are suitable stiffness and damping matrices, respectively; see [@motahar2015impedance; @motahar2017steering; @veer2017supervisory] for more details. The bipedal robot model employed here is shown in Fig. \[fig:model\] and is similar to models that have appeared in the literature [@3D-robotica2012; @motahar2017steering]; thus, the exposition below will be terse. The model features nine degrees of freedom, and its configuration can be described by the coordinates $q:=(q_1,q_2,...,q_9)$ as in Fig. \[fig:model\]. It is assumed that all degrees of freedom other than yaw $q_1$ and pitch $q_2$ are actuated. The walking cycle is composed of alternating sequences of single and double support phases. As in [@3D-robotica2012], we assume that the double support phases are instantaneous and can be modeled as an impact event based on the hypotheses listed in [@westervelt2007feedback Chapter 3]. Defining $\hat{x}:=(q,\dot{q})$ as the state of the robot, walking can be represented as a system with impulse effects $$\begin{aligned}
\Sigma:
\begin{cases}
~~\dot{\hat{x}} = f(\hat{x}) + g(\hat{x})u + g_{\rm e}(\hat{x}) F_{\rm e}, & \mathrm{if}~\hat{x}\not\in\mathcal{S} \\
~~\hat{x}^+ = \Delta(\hat{x}^-), & \mathrm{if}~\hat{x}^-\in\mathcal{S}
\end{cases} \enspace, \label{eq:SIE-robot}\end{aligned}$$ where $u$ are the inputs, $(f, g, g_{\rm e})$ describe the swing phase dynamics in the presence of the external force $F_{\rm e}$. In , $\mathcal{S}$ represents the ground surface and $\Delta$ is a mapping taking the states $\hat{x}^-$ prior to impact to the states $\hat{x}^+$ right after impact. More details on the model can be found in [@3D-robotica2012] with the difference due to the existence of the external force, which is taken into account as explained in [@motahar2017steering].
{width="0.55\columnwidth"}
\[fig:model\]
Family of Controllers {#subsec:control-fam}
---------------------
We work with a finite family of feedback controllers $u=\Gamma_p(x, F_{\rm e})$, where $x$ includes all components of $\hat{x}$ except $q_1$, i.e. $\hat{x}:=(q_1,x)$. Each controller results in a limit cycle $\mathcal{O}_p$ that corresponds to a straight-line or a turning walking motion. The controllers are designed within the HZD framework as in [@motahar2017steering; @3D-robotica2012], and they guarantee exponential stability of $\mathcal{O}_p$ in the absence of the external force. In the interest of space, we will only discuss some important properties of the controllers $\Gamma_p$; details can be found in [@motahar2017steering]. Associated to each $\Gamma_p$ is a zero dynamics surface $\mathcal{Z}_p$ which, for each $p\in\mathcal{P}$, has the following properties:
(i) $\mathcal{Z}_p$ is invariant under the biped’s closed-loop dynamics despite the presence of $F_{\rm e}$;
(ii) for each $\mathcal{Z}_p$, we have $\mathcal{S}\cap\mathcal{Z}_p = \mathcal{S}\cap\mathcal{Z}_q$ for all $p,q\in\mathcal{P}$.
A point of clarification is due here. The properties listed above are specific to the controllers used in this example; any other control method that generates exponentially stable walking limit cycles could be used without changing the procedure of Section \[sec:switch-guarantees\] for safe switching. However, these properties vastly simplify computation of the Lyapunov functions and the corresponding estimates of the BoAs involved in establishing the conditions of Section \[sec:switch-guarantees\] due to the associated dimensional reduction.
Limit-Cycle Gait Primitives and Switching Among Them
----------------------------------------------------
The behavior of the limit cycle walking motions $\mathcal{O}_p$ can be studied via the corresponding *forced* Poincaré map [@veer2017poincare] $$\label{eq:discrete-dyn-xhat}
\hat{x}_{k+1} = \hat{P}_p(\hat{x}_k, F_{{\rm e},k}) \enspace,$$ where $F_{{\rm e}, k}$ is the externally applied force over one stride. Although the forced Poincaré map is 17-dimensional, thus making the estimation of the BoA very challenging, the controller properties listed in Section \[subsec:control-fam\] allow us to work with a drastically lower-dimensional system. Due to the invariance of the associated zero dynamics surfaces $\mathcal{Z}_p$, the state always returns to $\mathcal{S}\cap\mathcal{Z}_p\subset\mathbb{R}^3$ at the end of each step. Taking into account that rotations around the yaw axis do not alter the dynamics—i.e., the system is equivariant to $q_1$ as shown in [@3D-robotica2012; @motahar2016composing]—allows us to work with the two-dimensional restricted forced Poincaré map $$\begin{aligned}
z_{k+1} & = \rho_p(z_k,F_{{\rm e},k}) \enspace, \label{eq:discrete-dyn-z}\end{aligned}$$ where $z$ are suitable coordinates on $\mathcal{S}\cap\mathcal{Z}_p$. As a result, the limit cycle walking motions $\mathcal{O}_p$ can be represented as motion primitives of the form $\mathcal{G}_p:=\{\rho_p,z_p^*\}$ where $z_p^*$ is a fixed point of $\rho_p$. Switching among these primitives according to a signal $\sigma(k)$ can be described by the switched system $$\label{eq:switch-system-zd}
z_{k+1} = \rho_{\sigma(k)}(z_k,F_{{\rm e},k}) \enspace.$$ It is worth reminding that the dimensional reduction afforded by HZD resulted in a 2-dimensional switched systems instead of the original 17-dimensional system.
Simulation Results {#sec:sim-results}
------------------
Working within the family of controllers discussed in Section \[subsec:control-fam\], we generate three gait primitives: $\mathcal{G}_0$ to turn clockwise by $30^{\circ}$, $\mathcal{G}_1$ to walk straight, and $\mathcal{G}_2$ to turn counter-clockwise by $30^{\circ}$. Using SoS programming, $\mathcal{B}_p$’s are obtained and plotted in Fig. \[fig:BoA\] as dashed ellipses; exact details of the SoS program can be found in [@motahar2016composing]. We choose $\kappa=0.002$ and compute upper bounds for $\mu(\kappa)$ and $\omega(\kappa)$ using [@veer2018ultimate Proposition 1]. Using these we obtain $\overline{\mathcal{M}}$ defined in and it can be noted from Fig. \[fig:BoA\], that this set lies within $\accentset{\circ}{\mathcal{B}}_0\cap\accentset{\circ}{\mathcal{B}}_1 \cap\accentset{\circ}{\mathcal{B}}_2$ for $\overline{N}_0=2$; further, computing gives $\overline{N}_{\rm a}=0.99$. With this choice of $\overline{N}_0$ and $\overline{N}_{\rm a}$, arbitrary switching signals satisfy . Hence, by Theorem \[thm:dist-agnostic\], the biped can switch to a different primitive each stride without compromising its safety, despite the external force.
![Estimates of the BoA for the primitives in $\mathbb{G}$ and verification of . The BoA estimates $\mathcal{B}_0$, $\mathcal{B}_1$, and $\mathcal{B}_2$ are the dashed red, green, and blue ellipses, respectively. The grey region is $\overline{\mathcal{M}}$ defined in for $\kappa=0.002$, $\overline{N}_0=2$. Black crosses are the solution of for the simulation in Fig. \[fig:leader\_follow\]. ](figures/BoA_sim_data.pdf){width="0.9\columnwidth"}
\[fig:BoA\]
{width="0.78\columnwidth"}
\[fig:biped\_lead\_straight\]
Now we turn our attention towards the switching policy that adapts the biped’s gait to the leader’s intended trajectory $p_{\rm L}(t)$, which is not directly available to the biped, requiring the planner to harness the external force feedback as a cue for adaptation. Our switching policy estimates the “average" heading direction $\Phi_k$ that the force is pointing towards over a stride, and then chooses the primitive that turns the biped towards this estimated heading. To compute $\Phi_k$, we integrate the force along the $X$ and $Y$ directions over a stride; see Fig. \[fig:model\] for the global coordinate frame. Let $t_0=0$ be the initial time and $t_k$ be the time at the end of the $k$-th stride. Then, over the $(k+1)$-th stride, the integral of the force components are $$\begin{aligned}
\nonumber
F_k^X:=\int_{t_k}^{t_{k+1}} F_{\rm e}^X(t)~dt, & & F_k^Y:=\int_{t_k}^{t_{k+1}} F_{\rm e}^Y(t)~dt \enspace,\end{aligned}$$ which are used to compute the “average" heading as $\Phi_k=\arctan(F_k^Y/F_k^X)$. The switching policy is chosen to be $$\begin{aligned}
\sigma(k+1)= \mathrm{sign}(\Phi_k)+1 \enspace, \label{eq:policy}\end{aligned}$$ where the $\mathrm{sign}$ function returns -1, 0, 1 for negative, 0, and positive $\Phi_k$, respectively. It can be observed from the switching policy that there is a one-step time delay in response to the force, i.e., $\Phi_k$ is used to obtain $\sigma(k+1)$.
We first test our planning framework for walking straight. It can be noted from Fig. \[fig:biped\_lead\_straight\] that with a single primitive $\mathcal{G}_0$ the biped drifts away from $p_{\rm L}(t)$. However, the switching framework is able to adapt the biped’s gait to keep it within the vicinity of the leader’s intended trajectory. Next, we simulate a more complex scenario shown in Fig. \[fig:leader\_follow\] where $p_{\rm L}(t)$ is represented by the red line, along which the leader intends to move at a constant speed of 0.65 m/s. Following the switching signal generated by the supervisor, the biped is able to adapt to the leader’s intended trajectory while maintaining its safety, as verfied by Fig. \[fig:BoA\] where the solution of the switched system , denoted by black crosses, lies within $\bigcap_{p\in\mathcal{P}} \mathcal{B}_p$, satisfying . Lastly, note that the biped naturally adapts its speed to the external force without requiring the planner’s intervention; see [@veer2015adaptation; @motahar2017steering] for details.
{width="0.95\columnwidth"}
\[fig:leader\_follow\]
Conclusion {#sec:conclusion}
==========
In this paper we proposed a general planning framework that facilitates adaptive planning in the face of uncertainty and external signals. Our framework views switching among externally excited dynamic primitives as a switched system with multiple equilibria under disturbances. We analyze this switched system to extract an average dwell-time bound that succinctly captures the dynamic limitations of the robot, resulting in motion plans consistent with the dynamics. The planning framework is particularized to a biped-leader collaborative task, where the biped’s gait is adapted to follow the leader’s intended trajectory, harnessing the interaction force as a command.
For convenience, denote the space of uniformly bounded sequences in $\mathcal{D}$ by $l_\infty^{\mathcal{D}}$; and $\mathcal{B}_p$ as the $\overline{\kappa}_p>0$ sublevel set of $V_p$, i.e., $\mathcal{B}_p=\{x\in\mathbb{R}^n~|~V_p(x)\leq \overline{\kappa}_p\}$.
Before proceeding to the proof of Theorem \[thm:dist-agnostic\], we first establish the following lemma, which shows that a Lyapunov function on a compact set of the state space is also an ISS-Lyapunov function on the same set, provided, that the disturbances are sufficiently small.
\[lem:ISS-Lyap\] Let $V_p:\mathcal{X}_p \to \mathbb{R}_+$ be a Lyapunov function for $f_p(\cdot,0)$ that satisfies . Then, there exist a[^8] $\delta>0$ and $\alpha_p\in\mathcal{K}_\infty$, such that $V_p$ is an ISS-Lyapunov function, that satisfies for all $x\in\mathcal{B}_p$ and $d:=\{d_k\}_{k\in\mathbb{Z}_+}\in l_\infty^{\mathcal{D}}$ with $\|d\|_\infty<\delta$, $$\nonumber
V_p(f_p(x,d_k)) \leq \lambda V_p(x) + \alpha_p(\|d\|_\infty)$$ where $0<\lambda<1$ is the same as in .
This proof builds on [@veer2017poincare Theorem 2], from which it follows that a Lyapunov function $V_p$ is also an ISS-Lyapunov function in a set where $g_p:=V_p\circ f_p:\mathcal{X}_p\times\mathcal{D}\to\mathbb{R}_+$ has a uniform Lipschitz constant. Hence, we first establish this property in our region of interest.
*Claim 1:* There exists a $\delta>0$ such that $g_p$ is Lipschitz for all $x\in\mathcal{B}_p$ and[^9] $d\in B_{\delta}(0)$.\
The proof for Claim 1 is detailed after the proof of Theorem \[thm:dist-agnostic\]. It is worth noting that if $\mathcal{D}$ were a finite-dimensional Banach space, $\mathcal{B}_p\times\overline{B}_\delta(0)$ would be a compact set by Heine-Borel theorem, and then Claim 1 would follow from the fact that locally Lipschitz functions are Lipschitz on compact sets. However, as we allow for an infinite-dimensional Banach space, more technical care is required.
With Claim 1, we can merely repeat the proof of [@veer2017poincare Theorem 2] to complete the proof of Lemma \[lem:ISS-Lyap\].
Now we are ready to present the proof of Theorem \[thm:dist-agnostic\].
*\[Theorem \[thm:dist-agnostic\]\]* The proof of this theorem utilizes [@veer2018ultimate Corollary 2] which shows that for switched systems with *global* ISS subsystems, if the switching signal satisfies the average dwell-time bound , then, there exists a non-empty compact set $\mathcal{C}\subset\mathcal{M}(\mu(\kappa)^{\overline{N}_0}\omega(\kappa))$ such that for all $k\in\mathbb{Z}_+$, $$\label{eq:cor-use}
x_0\in\mathcal{C} \implies x_k\in \mathcal{M}(\bar{\omega}(\|d\|_\infty)) \enspace.$$ The proof of [@veer2018ultimate Corollary 2] explicitly characterizes $\mathcal{C}$ as $\mathcal{C}=\bigcap_{p\in\mathcal{P}}\mathcal{M}_p(\omega(\kappa))$. With this, and , we can obtain . However, [@veer2018ultimate Corollary 2] requires global ISS-Lyapunov functions for each subsystem, but the $V_p$ from Lemma \[lem:ISS-Lyap\] are ISS-Lyapunov functions only for $x\in\mathcal{B}_p$ and $d\in\mathcal{D}$ with $\|d\|_\infty<\delta$. Therefore, to use this corollary we restrict the solutions of to evolve in the domain $\bigcap_{p\in\mathcal{P}}\mathcal{B}_p$ for $d\in\mathcal{D}$ with $\|d\|_\infty<\delta$, where each ISS-Lyapunov function is valid.
To ensure that the solutions stay within $\bigcap_{p\in\mathcal{P}}\mathcal{B}_p$, we claim that it is sufficient to show that for some $\delta>0$, $$\label{eq:desired-relation}
\mathcal{M}(\bar{\omega}(\delta)) \subset \bigcap_{p\in\mathcal{P}} \accentset{\circ}{\mathcal{B}}_p \enspace.$$ By the monotonicity of the sublevel sets, if $\|d\|_\infty<\delta$, $\mathcal{M}(\bar{\omega}(\|d\|_\infty))\subset \mathcal{M}(\bar{\omega}(\delta))$, which on using in gives $\mathcal{M}(\bar{\omega}(\|d\|_\infty)) \subset \bigcap_{p\in\mathcal{P}} \accentset{\circ}{\mathcal{B}}_p$. Hence, by the solutions would be restricted to $\bigcap_{p\in\mathcal{P}} \accentset{\circ}{\mathcal{B}}_p$ for all $k\in\mathbb{Z}_+$ and we can obtain . In what follows, we will show that there exists a $\delta>0$, such that for disturbances smaller than $\delta$, holds.
For the sake of convenience, let $\underline{\mathcal{B}}:=\bigcup_{p\in\mathcal{P}} \mathcal{B}_p\setminus \bigcap_{p\in\mathcal{P}} \accentset{\circ}{\mathcal{B}}_p$ which does not contain any $x\in\mathcal{M}(\mu(\kappa)^{\overline{N}_0}\omega(\kappa))$ as $\mathcal{M}(\mu(\kappa)^{\overline{N}_0}\omega(\kappa))$ is in the interior of $\bigcap_{p\in\mathcal{P}} \mathcal{B}_p$. Therefore, $$\begin{aligned}
\label{eq:boundary-Vp}
\forall p\in\mathcal{P}, & ~~~ V_p(x)>\mu(\kappa)^{\overline{N}_0}\omega(\kappa), & \forall x\in \underline{\mathcal{B}} \enspace.\end{aligned}$$ Let $\overline{\kappa}$ be defined as $$\label{eq:kappa-min-def}
\overline{\kappa}:=\min_{p\in\mathcal{P}} \min_{x\in \underline{\mathcal{B}}} V_p(x) \enspace,$$ which is well-defined because $ \underline{\mathcal{B}}$ is compact and $\mathcal{P}$ is finite. From , it follows that $\overline{\kappa}>\mu(\kappa)^{\overline{N}_0}\omega(\kappa)$. Let $0< c < \overline{\kappa} - \mu(\kappa)^{\overline{N}_0}\omega(\kappa)$, and shrink $\delta$ if necessary to ensure $0<\delta < \alpha^{-1}(c)$. Then, for any $p\in\mathcal{P}$, and $x\in\mathcal{M}_p(\bar{\omega}(\delta))$, $$\begin{aligned}
V_p(x) & \leq \mu(\kappa)^{\overline{N}_0}\omega(\kappa) + \alpha(\delta) \nonumber\\
& < \mu(\kappa)^{\overline{N}_0}\omega(\kappa) + c < \overline{\kappa} \enspace. \label{eq:Vp<kappa-boundary}\end{aligned}$$ Furthermore, with this choice of $\delta$, $$\label{eq:Mp-in-Bp}
\mathcal{M}_p(\bar{\omega}(\delta))\subset \mathcal{B}_p \enspace.$$ To see this, let $x\in\mathcal{M}_p(\bar{\omega}(\delta))$, then[^10] $V_p(x)<\overline{\kappa}\leq \overline{\kappa}_p$ by and $\overline{\kappa}\leq \min_{p\in\mathcal{P}}\overline{\kappa}_p$. It can be observed that $\overline{\kappa}\leq \min_{p\in\mathcal{P}}\overline{\kappa}_p$ because if not, then by the definition of $\overline{\kappa}$ in , for each $p\in\mathcal{P}$ and $x\in\underline{\mathcal{B}}$, $V_p(x)\geq \overline{\kappa} >\min_{p\in\mathcal{P}}\overline{\kappa}_p$, implying that every point in $\underline{\mathcal{B}}$ is strictly “outside" some $\mathcal{B}_p$, resulting in a contradiction with the definition of $\underline{\mathcal{B}}$ that contains elements of each $\mathcal{B}_p$. This follows from the fact that $\mathcal{B}_p$ is closed, hence its boundary $\partial\mathcal{B}_p$ is a subset of $\mathcal{B}_p$, but by the definition of the boundary of a set, $\partial \mathcal{B}_p$ is not in $\accentset{\circ}{\mathcal{B}}_p$, hence $\underline{\mathcal{B}}$ includes each $\partial \mathcal{B}_p\subset \mathcal{B}_p$.
Finally, we claim that with the given choice of $\delta$ as above, holds. To check this claim, assume that does not hold for this choice of $\delta$. Then, there must exist a $\hat{x}\in \mathcal{M}(\bar{\omega}(\delta))$ which lies in the complement of $\bigcap_{p\in\mathcal{P}} \accentset{\circ}{\mathcal{B}}_p$. There exists a $p\in\mathcal{P}$ for which $\hat{x}\in\mathcal{M}_p(\bar{\omega}(\delta))$, which by gives $\hat{x}\in\mathcal{B}_p$ but $\hat{x}\not\in\bigcap_{p\in\mathcal{P}} \accentset{\circ}{\mathcal{B}}_p$, further implying that $\hat{x}\in\underline{\mathcal{B}}$. Hence, by the definition of $\overline{\kappa}$ in , $V_p(\hat{x})\geq \overline{\kappa}$. On the other hand, as $\hat{x}\in\mathcal{M}_p(\bar{\omega}(\delta))$, by it follows that $V_p(\hat{x})< \overline{\kappa}$, leading to a contradiction with $V_p(\hat{x})\geq \overline{\kappa}$. Hence, there exists a $\delta>0$ for which holds, completing the proof of Theorem \[thm:dist-agnostic\].
*\[Claim 1 in Lemma \[lem:ISS-Lyap\]\]* As $V_p$ and $f_p$ are locally Lipschitz in their arguments, their composition $g_p:=V_p\circ f_p$ is locally Lipschitz as well. Hence, for any $(x,0)\in \mathcal{B}_p \times \mathcal{D}$, there exists a $\delta_x>0$ and $L_x>0$ such that $\|g_p(x_1,d_1)-g_p(x_2,d_2)\|\leq L_x \| (x_1-x_2,d_1-d_2)\| $, for any $x_1,x_2 \in B_{\delta_x}(x)$ and $d_1,d_2 \in B_{\delta_x}(0)$. Construct an open cover $\bigcup_{x\in\mathcal{B}_p} B_{\delta_x/2}(x)$ of $\mathcal{B}_p$ which is compact, hence there exists $\hat{x}_1,~\hat{x}_2,\cdots,\hat{x}_N$ such that $\mathcal{B}_p \subset \bigcup_{i=1}^n B_{\delta_i/2}(\hat{x}_i)$ where $\delta_i:=\delta_{\hat{x}_i}$ and define $\delta:=\min\{\delta_1/2,\cdots,\delta_N/2\}$.
Consider $x_1,x_2\in\mathcal{B}_p$ and $d_1,d_2\in B_\delta(0)$. Then, the following two cases arise.
*Case (a):* There exists an $i\in\{1,\cdots,N\}$ such that $x_1,x_2\in B_{\delta_i}(\hat{x}_i)$.\
As $d_1,d_2\in B_\delta(0)\subset B_{\delta_i/2}(0)$, and by the assumption of this case $x_1,x_2\in B_{\delta_i}(\hat{x}_i)$, we can use the Lipschitz continuity of $g_p$ in the $\delta_i$ neighborhood of $(\hat{x}_i,0)$ to obtain $\|g_p(x_1,d_1)-g_p(x_2,d_2)\|\leq L_i \| (x_1-x_2,d_1-d_2)\| $ where $L_i:=L_{\hat{x}_i}$. Define $\hat{L}:=\max\{L_1,\cdots,L_N \}$, then we can express the Lipschitz bound as $$\label{eq:lipschitz-case-a}
\|g_p(x_1,d_1)-g_p(x_2,d_2)\|\leq \hat{L} \| (x_1-x_2,d_1-d_2)\| \enspace.$$ *Case (b):* There does not exist any $i\in\{1,\cdots,N\}$ such that $x_1,x_2\in B_{\delta_i}(\hat{x}_i)$.\
To obtain the Lipschitz bound in this case we first need to establish uniform boundedness of $g_p$ over $\mathcal{B}_p\times B_\delta(0)\subset\mathcal{X}_p\times\mathcal{D}$. Note that $g_p(\cdot,0):\mathcal{X}_p\to\mathcal{X}_p$ is Lipchitz on the compact set $\mathcal{B}_p$ as it is locally Lipshcitz in its arguments. Hence, there exists a $\tilde{L}>0$ such that $\|g_p(y_1,0)-g_p(y_2,0)\|\leq \tilde{L} \|y_1-y_2\|$ for any $y_1,y_2\in \mathcal{B}_p$. Further, using the boundedness (compactness) of $\mathcal{B}_p \subset \mathbb{R}^n$, there exists a $r>0$ such that $\|y_1-y_2\|\leq r$ for any $y_1,y_2\in\mathcal{B}_p$. As $\mathcal{B}_p \subset \bigcup_{i=1}^n B_{\delta_i/2}(\hat{x}_i)$, there exist $\hat{x}_m$ and $\hat{x}_n$ such that $\|x_1-\hat{x}_n\|<\delta_n/2$ and $\|x_2-\hat{x}_m\|<\delta_m/2$. Then, $$\begin{aligned}
& \|g_p(x_1,d_1)- g_p(x_2,d_2)\| \nonumber \\
& = \|g_p(x_1,d_1)-g_p(\hat{x}_n,0) + g_p(\hat{x}_n,0) - g_p(\hat{x}_m,0) \nonumber \\
& ~~~ + g_p(\hat{x}_m,0) - g_p(x_2,d_2) \| \nonumber \\
& \leq \|g_p(x_1,d_1)-g_p(\hat{x}_n,0)\| + \|g_p(\hat{x}_n,0) - g_p(\hat{x}_m,0)\| \nonumber \\
& ~~~ + \|g_p(\hat{x}_m,0) - g_p(x_2,d_2) \| \nonumber \\
& \leq L_n \big( \|x_1-\hat{x}_n\| + \|d_1\| \big) + \tilde{L} \|\hat{x}_n-\hat{x}_m\| \nonumber \\
& ~~~ + L_m \big( \|x_2-\hat{x}_m\| + \|d_2\| \big) \nonumber \\
& \leq 2\hat{L} \big( r+\delta \big) + \tilde{L}r =: M \enspace. \label{eq:P-uniform-bound}\end{aligned}$$ Also, it can be noted that $\|x_1-x_2\|\geq \delta$ which can be shown by the way of contradiction. Suppose $\|x_1-x_2\| < \delta$. Let $\hat{x}_n$ be such that $\|x_1-\hat{x}_n\|<\delta_n/2$ which exists because $\mathcal{B}_p \subset \bigcup_{i=1}^n B_{\delta_i/2}(\hat{x}_i)$. Then, adding and subtracting this $\hat{x}_n$ in $\|x_1-x_2\|$, and using reverse triangle inequality gives $$\begin{aligned}
\|x_2-\hat{x}_n\|-\|x_1-\hat{x}_n\| & \leq\|x_1-\hat{x}_n+\hat{x}_n-x_2\| <\delta \nonumber
$$ which leads to $\|x_2-\hat{x}_n\| < \delta + \|x_1-\hat{x}_n\| < \delta_n/2 + \delta_n/2=\delta_n$ implying that $x_2\in B_{\delta_n}(\hat{x}_n)$, which along with the fact that $x_1\in B_{\delta_n}(\hat{x}_n)$ leads to a contradiction with the assumption of Case (b). Hence, $\|x_1-x_2\|\geq \delta$ which is used in to obtain $$\begin{aligned}
& \|g_p(x_1,d_1)- g_p(x_2,d_2)\| \nonumber \\
& \leq M \leq \frac{M}{\delta}\|x_1-x_2\| \leq \frac{M}{\delta} \|(x_1-x_2,d_1-d_2)\| \enspace. \label{eq:lipschitz-case-b}\end{aligned}$$ With the bounds and in Case (a) and (b), respectively, let $L:=\max\{\hat{L},M/\delta\}$ to obtain $$\nonumber
\|g_p(x_1,d_1)- g_p(x_2,d_2)\| \leq L \|(x_1-x_2,d_1-d_2)\| \enspace,$$ for any $x_1,x_2\in\mathcal{B}_p$ and $d_1,d_2\in B_\delta(0)$.
[^1]: S. Veer and I. Poulakakis are with the Department of Mechanical Engineering, University of Delaware, Newark, DE 19716, USA; e-mail: [{veer, poulakas}@udel.edu.]{}
[^2]: This work is supported in part by NSF CAREER Award IIS-1350721 and by NRI-1327614.
[^3]: The terms dynamical movement primitives and dynamical motion primitives are interchangeably used in the paper.
[^4]: For simplicity, we assume that the convergence rate $\lambda$ is the same for all primitives; if this is not the case, we can always choose $\lambda:=\max_{p\in\mathcal{P}} \lambda_p$.
[^5]: Notation: If $S$ is a set, then $\accentset{\circ}{S}$ denotes its interior.
[^6]: The disturbance $d_k$ can be a continuous function of time rendering $\mathcal{D}$ the mathematical structure of a Banach space and $f_p$ a nonlinear functional [@royden2010real]; indeed this will be the case in Section \[sec:example\].
[^7]: Terminology: A function $\alpha:\mathbb{R}_+ \to \mathbb{R}_+$ belongs to $\mathcal{K}_\infty$ if it is continuous, strictly increasing, $\alpha(0)=0$, and $\lim_{s\to\infty} \alpha(s)=\infty$.
[^8]: We assume that $\delta$ is the same for each $p\in\mathcal{P}$; if this is not the case we can choose $\delta:=\min_{p\in\mathcal{P}} \delta_p$ without loss of generality.
[^9]: Notation: We use $B_\delta(a)$ to denote an open-ball of radius $\delta$ centered at $a$. This notation can be used for open-balls in $\mathbb{R}^n$, as well as $\mathcal{D}$. It will be clear from context the space to which the ball belongs.
[^10]: Reminder: $\mathcal{B}_p$ is the $\overline{\kappa}_p$ sublevel set of $V_p$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present 74 MHz radio continuum observations of the Galactic center region. These measurements show nonthermal radio emission arising from molecular clouds that is unaffected by free-free absorption along the line of sight. We focus on one cloud, G0.13–0.13, representative of the population of molecular clouds that are spatially correlated with steep spectrum ($\alpha^{74\rm MHz}_{\rm 327MHz}=1.3\pm0.3$) nonthermal emission from the Galactic center region. This cloud lies adjacent to the nonthermal radio filaments of the Arc near l$\sim0.2^{\circ}$ and is a strong source of 74 MHz continuum, SiO (2-1) and FeI K$\alpha$ 6.4 keV line emission. This three-way correlation provides the most compelling evidence yet that relativistic electrons, here traced by 74 MHz emission, are physically associated with the G0.13–0.13 molecular cloud and that low energy cosmic ray electrons are responsible for the FeI K$\alpha$ line emission. The high cosmic ray ionization rate $\sim10^{-13}$ s$^{-1}$ H$^{-1}$ is responsible for heating the molecular gas to high temperatures and allows the disturbed gas to maintain a high velocity dispersion. LVG modeling of multi-transition SiO observations of this cloud implies H$_2$ densities $\sim 10^{4-5}$ cm$^{-3}$ and high temperatures. The lower limit to the temperature of G0.13-0.13 is $\sim100$K, whereas the upper limit is as high as 1000K. Lastly, we used a time-dependent chemical model in which cosmic rays drive the chemistry of the gas to investigate for molecular line diagnostics of cosmic ray heating. When the cloud reaches chemical equilibrium, the abundance ratios of HCN/HNC and N$_2$H$^+$/HCO$^{+}$ are consistent with measured values. In addition, significant abundance of SiO is predicted in the cosmic ray dominated region of the Galactic center. We discuss different possibilities to account for the origin of widespread SiO emission detected from Galactic center molecular clouds.'
author:
- 'F. Yusef-Zadeh$^{1*}$, M. Wardle$^2$, D. Lis$^3$, S. Viti$^4$, C. Brogan$^5$, E. Chambers$^6$, M. Pound$^7$ & M. Rickert$^1$'
title: '74 MHz Nonthermal Emission from Molecular Clouds: Evidence for a Cosmic Ray Dominated Region at the Galactic Center'
---
[km s$^{-1}$]{} \#1 [10\^[\#1]{}]{} \#1 \#2 [ \^[\#2]{}]{} u\#1 [ ]{}
[$^{\circ}$]{} i v
Introduction
============
The interstellar medium (ISM) of the nucleus of our Galaxy has unique characteristics. Gas clouds are subject to the strong tidal field of the nucleus and are required to have higher density than in the Galactic disk in order to become gravitationally unstable and form stars. This region of the Galaxy hosts two components in its central few hundred parsec. One is the reservoir of molecular gas showing enhanced molecular emission with higher velocity dispersion ($\sim20-30$) and gas temperature (50-200K) compared to elsewhere in the rest of the Galaxy$^{1-7}$. The other is the prevalence of a mixture of thermal emission from ionized gas at a temperature of few thousand degrees and nonthermal radio synchrotrone emission from relativistic electrons (e.g.; Nord et al. 2004; Law et al. 2008). The synchrotron emission is best viewed at low radio frequencies ($< 1$GHz) whereas thermal free-free emission is better detected at high frequencies ($> 1$GHz). Large-scale surveys of this region suggest that relativistic electrons and molecular gas co-exist and possibly interact with each other. However, the apparent correlation could result from chance coincidence along the long line of sight towards the Galactic center. We present the first evidence for low frequency nonthermal emission closely tracing molecular gas, establishing that cosmic ray electrons are physically associated with individual molecular clouds. We study one molecular cloud G0.13–0.13 in detail and postpone the discussion of other Galactic center clouds to elsewhere.
The interaction between cosmic-ray particles and molecular gas has several far reaching consequences. For one, cosmic rays play an important role in star formation processes as they are the primary source of ionization in molecular clouds that are self-shielded from UV radiation field. The interaction of cosmic ray electrons heats the gas to a higher temperature, which increases the Jeans mass and causes the initial mass function (IMF) to become top-heavy$^{8-10}$. The higher ionization fraction due to the impact of these electrons reduces the damping of MHD waves and helps to maintains the high velocity dispersion of molecular gas in the nuclear disk. In addition, this interaction strengthens the coupling of gas to the magnetic field slowing star formation by increasing the time scale for ambipolar diffusion before the onset of gravitational collapse.
Recent measurements indicate a vast amount of H$_3^+$ and H$_3$O$^+$ $^{11-14}$ as well as high temperature molecular gas distributed in the Galactic center. The inferred minimum cosmic ray ionization rate, $\zeta\sim 10^{-15}$ s$^{-1}$ H$^{-1}$, is one to two orders of magnitude higher in the Galactic center region than in the Galactic disk$^{10,13}$. The elevated cosmic ray ionization rate will increase the ionization fraction of electrons in molecular gas, and drive ion-neutral chemistry. Another consequence is enhanced fluorescent FeI K$\alpha$ 6.4 keV emission$^{15,16}$. This nonthermal X-ray emission is uniquely detected from Galactic center clouds and results from filling the inner K-shell vacancies of neutral iron created by the impact of low energy cosmic ray electrons or by irradiation by X-ray photons$^{17-20}$. A widely accepted model argues that the 6.4 keV line is a result of irradiation by a hypothetical transient source associated with Sgr A\*, which was active about 400 and again 100 years ago, and that we are seeing this transient’s light echo in the 6.4 keV emission$^{21}$. We discuss another scenario in which low energy cosmic rays can contribute significantly in production of the FeI K$\alpha$ line emission from Galactic center clouds.
The Galactic center population of molecular clouds have physical conditions similar to hot cores, but with the exception of Sgr B2, there is no strong evidence for ongoing massive star formation throughout the inner few hundred parsecs of the Galaxy, a region known as the central molecular zone (or CMZ)$^{22}$. Line emission from a variety of molecular species has been detected throughout the CMZ including those that are produced in the gas phase (e.g.; HCN, HCO$^+$, HNC, N$_2$H$^+$) as well as those processed on grain surfaces, e.g. SiO $^{6,7}$. With the exception of the strong radiation field near Sgr A\* and the Arches cluster, PDRs can not be very important in the dense self-shielded clouds. In addition, the X-ray flux from the Galactic center is too weak to qualify for the application of XDR models. Global heating by cosmic rays can explain why the gas temperature is significantly higher than the dust temperature in a large fraction of the gas and dust clouds in the nuclear disk. In the cosmic-ray dominated region of the Galactic center, cosmic-ray heating should be more significant than ambipolar and turbulent heating$^{23}$. Thus, it is natural to examine the chemistry of the gas in the cosmic-ray dominated region and attempt to identify molecular line diagnostics of cosmic ray heating.
To this end, we present high resolution 74 MHz observations of the Galactic center and show the spatial correlation between molecular gas and 100 MeV cosmic ray electrons. Enhanced 74 MHz emission appears to coincide with a large concentration of molecular gas in the inner 5$^{\circ}$ of the Galactic center$^{24}$. Here, we focus on G0.13–0.13, a representative molecular cloud, and compare its 74 MHz emission with that of excited rotational transitions of SiO and CS molecules. This molecular cloud lies along the nonthermal filaments of the radio Arc near l$\sim0.2^{\circ}$, the most prominent network of magnetized filaments in the Galactic center. Molecular line emission from CO, CS, SiO, H$^{13}$CO$^+$ and CS suggests that this cloud has a high column density and gas temperature$^{25-27}$. The kinematics of CS line emission from G0.13–0.13 suggests an expansion of molecular gas into the nonthermal filaments$^{25}$. CO observations imply gas temperature T$\ge70$K and column density N(H$_2$)=6–7$\times10^{23}$ cm$^{-2}$. Multiple transitions of NH$_3$ line emission from this cloud have measured two temperature components giving a range T$\sim$25K and T$\sim125-200$K$^{28}$. The low temperature component of molecular gas is similar to the measured 18–22K dust temperature in clouds in the inner 2$^{\circ}\times1^{\circ}$ of the Galaxy$^{29,30}$.
The structure of this paper is as follows. We first show 74 MHz emission from Galactic center molecular cloud G0.13–0.13 and estimate the variation of spectral index of nonthermal emission between 327 and 74 MHz. We then show the association of the nonthermal radio emission from G0.13–0.13 with FeI K$\alpha$ line emission at 6.4 keV and estimate the cosmic ray ionization rate needed to produce the FeI K$\alpha$ line emission. We will also present new molecular line observations of four rotational transitions of SiO to determine the temperature and density of molecular gas in G0.13–0.13 more accurately. We consider the interaction of relativistic electrons with the gas and compute the cosmic ray ionization rate, implied by the 74 MHz emission, as a function of assumed magnetic field strength. We present the dependence of the total cooling rate of the gas on gas temperature for high molecular gas densities. Lastly, we study the chemical consequences of the interaction of cosmic ray electrons with the molecular gas by modeling the abundance ratios of several molecular species as a function of time. A time-dependent gas-grain chemical model$^{31}$ is used to explore how abundance ratios of five representative molecular species vary with gas density and cosmic-ray ionization rate.
Our study of G0.13–0.13 is not only relevant to understanding the molecular component of the nuclear disk of our Galaxy$^{8}$ but also to external galaxies and ultraluminous infrared galaxies (ULIRGs) where cosmic rays are thought to be the driving mechanism for star formation$^{9}$.
Observations and Data Reduction
===============================
We imaged the Galactic center at 74 MHz and mapped molecular line emission from several molecular clouds in the Galactic center using Mopra, CSO and CARMA. We only present observations of G0.13–0.13 here and will describe details of observations of other Galactic center clouds elsewhere.
VLA
---
The 74 MHz radio continuum observations of the Galactic center were described previously$^{24}$. Briefly, these measurements were taken in multiple configurations of the Very Large Array (VLA) of the National Radio Astronomy Observatory[^1] (NRAO), with a resolution of 114$''\times60''$ (PA$=-5^{\circ}$) and rms sensitivity of $\sim 0.12$ Jy. These observations are unprecedented in spatial resolution and sensitivity at these low frequencies. Given the strong frequency dependence ($\nu^{-2.1}$) of free-free absorption, foreground and embedded thermal sources are seen in absorption against the strong nonthermal emission from the Galactic center. The suppression of thermal emission due to high opacity of ionized gas at 74 MHz allows us to readily identify nonthermal sources that emit at this frequency.
CSO
---
Observations of the SiO (5–4) and (6–5) rotational transitions at 217.1 and 260.5 GHz, respectively, were carried out in 2010 September using the wideband 230 GHz “Z-Rx” receiver of the Caltech Submillimeter Observatory (CSO) on Mauna Kea, Hawaii. The weather conditions were good, characterized by a 225 GHz zenith opacity of $\sim 0.05-0.07$, equivalent to 1–1.5 mm of precipitable water vapor. Typical single sideband system temperatures at the relatively low elevation of the Galactic center were 250 K. The CSO FWHM beam size at the two frequencies is $\sim 34^{\prime\prime}$ and $28^{\prime\prime}$, respectively, and the main beam efficiency, as determined from total power observations of Jupiter, was $\sim 70\%$. Spectra were taken in the “on-the-fly mapping” mode, on a $\sim 15^{\prime\prime}$ grid, using a designated off position at $(\alpha, \delta)_{2000.0} =$, out of the Galactic plane. Multiple maps scanned in orthogonal directions were averaged together to avoid striping. The final per pixel integration time (ON) was 14–17 sec for SiO 5–4 and 25–31 sec for SiO 6–5. We used the high-resolution CSO facility FFTS spectrometer with 8192 channels and a total bandwidth of 1 GHz.
Mopra
-----
We observed G0.13–0.13 on 2010, June 28 using Mopra[^2] in the on-the-fly mapping mode. The map size is 5$'$ by 5$'$, and is centered at Galactic longitude and latitude (l, b)=($0.1092^{\circ}, -0.1000^{\circ}$). We used a latitude scan direction (orthogonal to the Galactic plane), and a row spacing of $\sim15''$. We used the MOPS spectrometer in Wideband mode, with a central frequency of 89.690 GHz. This results in a frequency coverage of $\sim~85.6 - 93.8$GHz across four sub-bands. Each sub-band has 8192 channels, a channel width of 269.5 kHz ($\sim0.9$ km s$^{-1}$), and two linear polarizations. The final map is the average of the two polarizations. Regular pointing checks done throughout the observations, and were within $\sim$ 4$''$ (the beam size is $\sim$ 38$''$ at these frequencies). The intensity is given in T$^*_A$ which can be divided by the main beam efficiency 0.49 to convert the intensity to T$_{mb}$.
We also made deep pointed observations toward four positions (A-D) within G0.13–0.13 on 2010, June 30 (see Table 1). These spectra were obtained in an on-off position switching mode for a total on-source time of 288 s. The MOPS spectrometer was configured in wideband mode, identically to the setup of the map described above. A total of 16 molecular lines CH$_3$CN, C$^{13}$CN, $^{13}$CS, N$_2$H$^+$, HNC, HCCN, HNCO, HCN, HCO$^+$, $^{28}$SiO, HC$^{15}$N, SO, H$^{13}$CN, H$^{13}$CO, SiO, NH$^{15}$C and C$_2$H are detected from four positions in G0.13–0.13 (see Fig. 7).
BIMA
----
We mapped G0.13–0.13 in CS(2-1) at 97.980968 GHz and HCO$^+$(1-0) at 89.188518 GHz with the Berkeley-Illinois-Maryland Association (BIMA) interferometer[^3] during the 2002-2003 observing season. The BIMA array was a mm-wavelength interferometer located in Hat Creek, California$^{32}$, consisting of ten 6.1 m antennas. These antennas have since been combined with the Owens Valley Radio Observatory 10 m antennas into the CARMA[^4] millimeter array at Cedar Flat, California.
The cloud was observed with 20 mosaicked pointings on a hexagonal grid corresponding to Nyquist sampling of the primary beam and covering a total area of about 5 $\times$ 6. The pointing center was $(\alpha, \delta)_{2000.0} =$ with a velocity of 30 with respect to the local standard of rest. The HCO$^+$ and CS data were obtained simultaneously in opposite sidebands. Each of the spectral lines were observed with a 50.0 MHz correlator window with 256 channels. Along with the spectral data, continuum windows with a total coverage of about 800 MHz were also observed. We observed in the C and B configuration of the array, sampling spatial frequencies from 1 to 78 k$\lambda$. Absolute flux calibration was derived from observations of Mars or Uranus immediately before or after the source track. The quasar 1733-130 was used as the phase calibrator and secondary flux calibrator. To fill the inner [*uv*]{} spacings, we also obtained fully-sampled maps covering the region imaged with the array in the HCO$^+$, and CS using the Five Colleges Radio Astronomy Observatory 14 m telescope, as well as the previously published CS(2-1) maps made with Nobeyama 45-m radio telescope (NRO)$^{25}$. A more detailed high resolution kinematic study of G0.13–0.13 including the results of HCO$^+$ observations will be given elsewhere.
Results
=======
Figure 1a,b shows the distribution of 74 MHz continuum and the CO (3-2) line from the inner 1.5$^{\circ}\times12'$ of the Galactic center. A layer of molecular gas in the CMZ runs parallel to the Galactic plane and is seen in the distribution of CO (3-2) line emission$^{33}$. This layer includes some of the most prominent molecular clouds in the CMZ such as Sgr B2 and Sgr B1, the 40, 50, –30 and 20 Sgr A and Sgr C clouds. Some of the prominent clouds are labeled on Figures 1a,b. At positive longitudes, the Sgr B2 complex, G0.36-0.09 and G0.13–0.13 have 74 MHz counterparts and at negative longitudes the layer of molecular gas that runs parallel to the Galactic plane G359.75-0.23 shows a 74 MHz counterpart. Another presentation of the CO (3-2) and 74 MHz emission is shown in Figure 1c. The distribution of 74 MHz emission appears to follow the CO (3-2) line emission except in two types of regions. One type is associated with holes created by free-free absorption suppressing background 74 MHz emission and the other is where nonthermal continuum sources have no molecular counterparts. Overall, the 74 MHz emission shows similar morphology to that of CO (3-2) line emission apart from absorption features manifesting as holes in the 74 MHz distribution. To clarify the correlation between absorption features and free-free emission, Figure 1d presents a continuum map at 8.5 GHz based on GBT observations$^{34}$. Prominent continuum sources are labeled on this figure. Holes in the 74 MHz continuum map coincide with strong continuum emission at 8.5 GHz, consistent with the free-free absorption coefficient increasing significantly at low frequencies. Thus, thermal sources become optically thick and are seen in absorption against strong nonthermal background emission$^{35}$ whereas nonthermal sources that are not contaminated by significant foreground thermal emission are clearly identified at 74 MHz. The morphological comparison of these figures suggest that nonthermal continuum emission at 74 MHz arises from Galactic center molecular clouds, implying that cosmic-ray electrons are interacting with molecular gas in this unique region of the Galaxy.
We now focus on the G0.13–0.13 cloud. Figure 2a shows contours of CS (1-0) molecular line emission from this cloud superimposed on a grayscale 5 GHz continuum image. A network of linearly polarized filaments of the radio Arc running almost perpendicular to the Galactic plane at a PA$\sim-170^{\circ}$ lies at the eastern edge of the cloud$^{37}$. The morphology of molecular line and radio continuum emission from G0.13–0.13 suggests that the molecular gas is surrounded by, and is interacting with, nonthermal radio filaments. The kinematics of the molecular gas suggests that expansion of the cloud is responsible for a dynamical interaction between the eastern limb of the G0.13–0.13 and the magnetized nonthermal filaments$^{25}$. Figure 2b shows high resolution CS (2-1) line emission from the G0.13–0.13 molecular cloud (red) superimposed on a 1.4 GHz continuum image (green). The molecular gas distribution is clumpy with the morphology of a “boot", and is edge-brightened parallel to the nonthermal radio filaments of the Arc near l$\sim0.2^{\circ}$ on the eastern edge of the cloud. We note that the western edge of the cloud lies parallel to another magnetized nonthermal filament G0.087-0.087$^{38}$. The molecular line images show a ridge of emission that deviated from a straight line perpendicular to the Galactic plane (labelled as “Meanderning feature” on Figure 2a) near l$\sim8'$, b$\sim-7'$. This ridge of emission appears to curve around a nonthermal filament near l$=8'\, 4''$, b=$-6'\, 55''$. This high resolution map of molecular line emission provides the strongest morphological support for the interaction of the nonthermal filaments and the eastern edge of G0.13–0.13.
Figure 3a,b show contours of 74 and 327 MHz emission, respectively, superimposed on a grayscale image at 5 GHz. The 74 MHz emission is produced mainly by nonthermal processes whereas 5 GHz and 327 MHz emission can trace emission produced by both thermal and nonthermal processes. The distribution of 74 MHz emission differs remarkably from the 327 MHz and 5 GHz continuum. The 74 MHz emission arises from the vertical filaments as well as the region where the G0.13–0.13 molecular cloud lies. There is no evidence of thermal emission from G0.13-0.13 at 327 MHz or at higher frequencies, thus, the spectrum derived here is not contaminated by thermal emission. The vertical filaments seen in the grayscale image at 5 GHz have counterparts at 74 MHz except in the thermal region to the north of the filaments near G0.18-0.04. This is because G0.18-0.04 is an HII region dominated by free-free emission at high frequencies, thus the lack of 74 MHz emission is due to significant free-free absorption$^{35}$.
What is remarkable about images shown in Figure 3a,b is the discovery that enhanced nonthermal emission at 74 MHz arises from the molecular cloud G0.13–0.13. The spectrum of emission from the center of the cloud must be steep given that there is no strong 327 MHz emission from G0.13–0.13. Figure 3c shows the spectral index distribution (where the flux density F$_{\nu} \propto \nu^{-\alpha}$) between 327 and 74 MHz. The spectrum is inverted ($\alpha < 0$) at the position of the filaments and is steep ($\alpha\sim1.3$) where the molecular line and 74 MHz emission peak. There is no evidence of thermal emission from G0.13-0.13 at 327 MHz or at higher frequencies, thus, the spectrum derived here is not contaminated by thermal emission.
Another version of the spectral index distribution is made by comparing the background subtracted profiles of the emission at 74 and 327 MHz. Figures 3d,e show two different cross cuts made at constant longitude l=$-6'$ and constant latitude b=$-8'$, respectively. This technique follows earlier work by Law et al. (2008). The spectral index distribution, as shown in Figure 3d, is consistent with a steep spectrum ($\alpha=1.1\pm0.2$) becoming flatter or inverted at the nonthermal vertical filaments. Figure 3e shows $\alpha\approx-0.2\pm0.05$ between 74 and 327 MHz at the position of nonthermal filaments. The electron energy spectrum of the center of G0.13–0.13 is remarkably steep, p$\approx$3.2 (where p=$2\alpha$+1 corresponding to energy spectrum E$^{-p}$). In contrast, the spectral index of the vertical filaments is remarkably flat p$\approx$0.6. The variation of the spectral index of nonthermal emission from the filaments and G0.13–0.13, is $\Delta\alpha\sim 1.5$.
G0.13–0.13 is one of the Galactic center molecular clouds showing bright FeI K$\alpha$ line emission at 6.4 keV$^{39}$. Figure 4 shows contours of fluorescent FeI K$\alpha$ line emission based on Chandra observations convolved to a resolution of 30$''$ and superimposed on a grayscale distribution of integrated CS (1-0) line emission. High resolution Chandra and CARMA images of this cloud show a clumpy distribution of K$\alpha$ line emission$^{20}$ tracing CS (1-0) line emission. The overall distribution of FeI K$\alpha$ line emission arising from G0.13–0.13 is similar to that of low frequency radio emission molecular line CS and SiO emission. The distribution of the K$\alpha$ line emission is clumpy and shows a ridge of emission adjacent to a nonthermal feature at l=8$'\, 10''$ b=$-6'\, 55''\,$. Remarkably, this X-ray ridge feature is similar to that of CS emission (cf. Meandering feature in Fig. 2b). Both the X-ray and CS ridges curve around a nonthermal filament. The morphology of 6.4 keV line emission supports the idea that nonthermal radio filaments are interacting with the edge brightened FeI K$\alpha$ emitting molecular cloud G0.13–0.13 at 6.4 keV. These images suggest a three-way correlation between cosmic ray electrons, molecular gas and FeI K$\alpha$ line emission.
In order to estimate the gas density and temperature in G0.13–0.13, we measured the line intensities of four transitions of SiO (6-5), (5-4), (2-1) and (1-0). We used the SiO (1-0) and (2-1) images taken from Handa et al. (2006) and convolved them to the same resolution and velocity coverage, between -10 to 50 as those of SiO (6-5) and (5-4) transitions. Figure 5a shows contours of velocity integrated SiO (5-4) line emission mapped over a limited area of G0.13–0.13 superimposed on a 5 GHz continuum image. We applied a Large Velocity Gradient (LVG) model to the two E and NW rectangular boxes, as shown schematically, in order to derive the physical parameters of molecular gas in G0.13–0.13. The LVG grid was computed for a constant SiO column density of 1.6$\times10^{13}$ cm$^{-2}$ and a line width of 30 . This approximately reproduces the observed line intensities for the range of densities and temperatures applicable to the gas in the Galactic center region. With these assumptions, the emission is optically thin (typical optical depths ranging between 0.1 and 10$^{-3}$, depending on the transition). Thus, optical depth effects do not influence the LVG model results as long as the beam filling factor is close to unity. A smaller beam filling factor will change the optically thin assumption we have made. Figure 5b,c show the density and kinetic temperature constraints estimated from three SiO line ratios of the E and NW boxes in three different colors, respectively. The three line ratios give a wider range of gas temperature between $\approx100-1000$K for box E and and $\approx400-1000$K for box NW. A lower range of gas temperatures $>100$K is found from the region near the nonthermal filaments of the Arc, box E as shown in Figure 5a, than that of the peak emission in the NW box having T$>$400K. The implied kinetic temperatures are consistent with earlier measurements based on multiple transitions of NH$_3$ $^{28}$. Additional constraints have also been made by estimating molecular gas pressure from the NW box of Figure 5a. The top three panels of Figure 5d show the plots of line ratios for all models in the grid as function of logarithm of pressure (nT). The bottom panel shows a sum of \[(model-observed)/observed\]$^2$ for the three line ratios and the the best fit is obtained for log(nT) in the range of 6.4–6.9. The line ratios give values of density and temperature that are consistent with log(n)$\sim4.4$ and log(T)$\sim2.3$. for minimum $\chi^2$ values. Similar values were obtained for the E box of Figure 5a.
Tsuboi et al. (2011) used the SiO (2-1)/(1-0) line ratios to estimate a molecular gas density of $\sim10^4$ cm$^{-3}$ at a kinetic temperature of 60K. On the other hand, Oka et al. (2001) used the CO (2-1)/(1-0) intensity ratios and argued that G0.13–0.13 has a low density $\sim10^2$ cm$^{-3}$, high kinetic temperature $\ge70$K component. Overall, our model fits which uses multiple transitions of SiO are suited to measure gas density and temperature, though with systematic uncertainties from using different telescopes. The application of the LVG excitation code constrains the gas density of hydrogen nuclei n$_{\rm H}\sim (1-3)\times10^4$ cm$^{-3}$ and temperature T$\sim (1-2)\times100$K but we can not rule out gas temperature as much as 1000K and lower densities n$\sim10^3$ cm$^{-3}$.
Two studies have made discrepant estimates of the column density of the gas in G0.13–0.13 $^{26,40}$ Using kinetic temperature T$_k$=70K, Handa et al. (2006) used H$^{13}$CO$^+$ (J=1-0) line emission which is expected to be optically thin with abundance of 10$^{-10}$ to derive an H$_2$ column density (6-7)$\times10^{23}$ cm$^{-3}$. On the other hand, Amo-Baladron et al. (2009) determined the physical condition by applying an LVG model to the SiO (2-1) and (3-2) transitions and then estimated the column of H$^{13}$CO$^+$ (J=1-0). Using the same abundance as that of Handa et al. (2006), the derived column density of molecular hydrogen from this study is (2-8)$\times10^{22}$ cm$^{-3}$. To resolve this discrepancy, we estimated the column density of molecular gas by using submillimeter emission from dust grains in G0.13–0.13. We used the data taken from Pierce-Price et al. (2000) who derived 513 Jy$^{-1}$ beam$^{-1}$ at 850$\mu$m with the assumption that dust temperature is 20K, metallicity twice solar and opacity index $\beta$=2. Using background subtracted flux densities of 0.5 and 1.5 Jy from the NW and E boxes at 850$\mu$m and a beam of 15$''$, we find the column density of molecular hydrogen ranges between $\sim3.5\times10^{22}$ and $\sim10^{23}$, cm$^{-2}$ respectively. There is uncertainty in accurately determining the background emission. Given this uncertainty, our estimate of the column density falls between previous measurements of Handa et al. (2006) and Amo-Baladron et al. (2009). Herschel maps presented by Molinari et al. (2011) give column density values that are consistent with our estimates.
Discussion
==========
Our interest in studying Galactic center molecular clouds stems from three earlier studies, all of which had suggested the interaction of cosmic ray electrons with molecular gas and that cosmic ray electrons are enhanced in this region. First, fluorescent 6.4 keV K$\alpha$ iron line emission from several molecular clouds in the Galactic center may be the consequence of the interaction of low energy cosmic ray electrons with molecular gas$^{20,41}$, or alternatively, the irradiation of molecular clouds by a hypothetical X-ray flash associated with past activity of Sgr A\*$^{21}$. Second, nonthermal bremsstrahlung from this electron population can also explain the diffuse $\gamma$-ray emission from the central 2$^{\circ}\times1^{\circ}$ of the Galactic center$^{42}$. Third, the strong H$_3^+$ absorption along several lines of sight towards the Galactic center imply that the ionization rate is at least an order of magnitude higher than elsewhere in the Galaxy$^{10}$. This is consistent with the low energy cosmic ray electron production of the FeI K$\alpha$ line emission. Finally, low frequency diffuse radio emission$^{43}$ combined with spectral index measurements$^{23}$ suggest strong nonthermal continuum emission from the Galactic center. Our detailed study provides the most compelling evidence yet that relativistic electrons, here traced by 74 MHz emission, are physically associated with the G0.13–0.13 molecular cloud. This cloud is one of the strongest sources of FeI K$\alpha$ line emission in the Galactic center. We complete our analysis by studying the complex morphology of G0.13-0.13, estimating the cosmic ray ionization rate, the expected FeI K$\alpha$ line emission at 6.4 keV, and modeling chemical signature of the interaction.
Morphology & Kinematics
-----------------------
G0.13–0.13 is a complex molecular cloud with kinematics that suggest expansion into the vertical filaments of the radio Arc$^{25}$ Morphologically, this cloud lies at the center of a circular-shaped structure, known as the radio Arc bubble, apparent in mid-IR images taken with the Spitzer and MSX observatories$^{44-47}$. Figure 6 shows contours of SiO (2-1) line emission, obtained with Mopra, superimposed on a three-color Spitzer IRAC image. The Arc bubble is noted exterior to the south of G0.13–0.13. The origin of the bubble and its relation to the radio Arc and the G0.13–0.13 cloud is not well understood. Past studies have suggested that the radio Arc bubble lies in the vicinity of the Quintuplet star cluster and is produced by stellar winds or supernova explosions sweeping up interstellar material$^{47}$. The nonthermal radio emission at 74 MHz and expansion of G0.13–0.13 point to the possibility that the Arc bubble may be produced by the same event. It is possible that an energetic event is driving a shock into G0.13–0.13. The shock reaches the edge of the cloud and before it encounters much lower density of the gas exterior to G0.13–0.13, then sweeps through the low density gas and creates a edge-brightened bubble. In this picture, the expansion of G0.13–0.13 into the nonthermal vertical filaments accelerates particles along the filaments$^{25,48}$, thus generating a young population of electrons running along the magnetized filaments with an unusually flat energy spectrum (p$\sim$0.6). The puzzle, however, is the origin of the steep energy spectrum of particles with p$>3$ in G0.13–0.13. This is because energy losses are increasingly severe for lower energy electrons, tending to flatten their energy spectrum at low energies.
Cosmic Ray Ionization Rate
--------------------------
Using equation (3) of Yusef-Zadeh et al. (2012), the cosmic ray ionization per second per hydrogen nucleus, $\zeta$, depends on the observed synchrotron intensity I$_{\nu}$, the magnetic field B, the depth of the source of emission along the line of sight, L, and the energy spectrum of the electrons with the index p, assumed to be power-law E$^{-p}$ between 0.1 MeV and 1 GeV and $\alpha = (p-1)/2$ is the synchrotron spectral index and $I_{\nu}\propto\nu^{-\alpha}$.
$$\zeta \approx \frac{3.1\times10^{-14}}{p-1}\,
\frac{I_\nu}{\u Jy \ut arcmin -2 }\, \left(\frac{\nu}{\u GHz }\right)^{\!\alpha} \, \left( \frac{L}{30\u pc
}\right)^{-1}\, \left(\frac{B}{100\,\mu\mathrm{G} }\right)^{\!\!-(1+\alpha)}\, \ut s -1 \ut H -1$$
Figure 7a shows the inferred cosmic ray ionization rate as a function of the magnetic field strength for different values of $\alpha$. We assumed that L$\sim$2.5 pc (1$'$ corresponds to 2.4pc at the Galactic center distance 8.5 kpc) and the observed surface brightness of 0.71 Jy arcmin$^{-2}$ at 74 MHz averaged over the inner $5'\times5'$ of G0.13–0.13. The black dot on the curve gives the equipartition magnetic field for the electron population producing diffuse synchrotron emission in the dense molecular cloud G0.13–0.13. The ionization rate due to electrons increases for steeper spectral index values for a fixed emissivity at the observed frequency, as there is a successively larger population of lower-energy electrons radiating at lower frequencies. The equipartition magnetic field ranges between 30 $\mu$G and 0.3 mG for $\alpha$ values between 0.25 to 1.5. The corresponding $\zeta$ is estimated to be $\sim4\times10^{-15}$ and 10$^{-11}$ s$^{-1}$, respectively. A cosmic ray ionization rate $\ge10^{-13}$ s$^{-1}$ H$^{-1}$ is sufficient to fully dissociate the gas on a time scale of $\leq10^6$ years. To avoid this, the magnetic field has to be larger than 1mG for $\alpha=1.25$.
The 6.4 keV Neutral Iron Line Emission
--------------------------------------
We apply the cosmic ray model to the G0.13–0.13 molecular cloud by estimating the K$\alpha$ line emission from the interaction of the low energy nonthermal electrons responsible for synchrotron radio emission detected at 74 MHz. Using equation (6) of Yusef-Zadeh et al. (2012), $$I_{\mathrm{K}\alpha} \approx 8.7\ee -8 \left(\frac{\zeta}{10^{-14}\ut s -1 }\right)
\left(\frac{N_\mathrm{H}}{10^{23}\ut cm -2 }\right) {\rm \ ph \rm \ s^{-1} \ cm^{-2} \ arcmin^{-2}}\,,$$ for N(H$_2$)=$(1-3)\times10^{23}$ cm$^{-2}$, then the cosmic ray ionization rate needed to give the background subtracted K$\alpha$ line flux of $3\times10^{-6} \u ph\, \rm cm^{-2}\, \rm s^{-1}$ arcmin$^{-2}$ derived from Suzaku observations$^{39}$ is $\zeta\sim(1-3)\times10^{-13}$ s$^{-1}$, respectively. A value of cosmic ray ionization rate $\zeta\sim10^{-13}$ s$^{-1}$ is indicated by a dashed line on Figure 7a. Using the spectral index values between $\alpha=1$ and 1.25 and the gas density between few times $10^{4-6}$ cm$^{-3}$, the magnetic field strength is constrained to values between $\sim0.5$ and 1 mG, respectively.
A widely accepted alternative model argues that the 6.4 keV FeI K$\alpha$ line emission results from irradiation of molecular clouds near the Galactic center by a hypothetical transient source associated with Sgr A\*, which was active about 400 and again 100 years ago$^{21}$. The variability of 6.4 keV line emission on a short time scale provided a strong evidence in support of this hypothesis. However, the evidence for nonthermal emission from molecular clouds, as presented here, also predicts time variability in the context of the cosmic ray picture$^{42}$. A supernova event expanding inside the cloud could change not only the ionization characteristics of the molecular gas in G0.13–0.13 but also causes the observed 6.4 keV time variability. Thus, the time variability of K$\alpha$ line emission need not be clear support for the X-ray irradiation model.
Cosmic Ray Heating of Molecular Gas
-----------------------------------
To estimate the temperature of molecular gas subject to high levels of cosmic-ray ionization, we extended our previous estimates of the cooling rate$^{42}$ to higher densities, using the fitting formula for cooling by rotational transitions of CO, H$_2$, and H$_2$O provided by Neufeld & Kaufman (1993) and Neufeld, Lepp & Melnick (1995). We also included an LVG calculation of the contribution by OI fine-structure lines, using the atomic data and collision rates in the Leiden Atomic and Molecular Database$^{49}$. We adopted an abundance of CO relative to H$_2$ of $2.8\times10^{-4}$, twice the values adopted by Neufeld, Lepp & Melnick (1995) to reflect the higher metallicity of the gas in the CMZ. The abundance of OI and H$_2$O is temperature-dependent because of the conversion of OI and O$_2$ to H$_2$O by neutral-neutral reactions at temperatures above 200 K, so we assume twice the chemical equilibrium abundances of OI and H$_2$O found by Neufeld, Lepp & Melnick (1995): OI$/$H$_{2} = 2\times10^{-4} / (1+(T/220\,\mathrm{K})^{14})$, and H$_2$O$/$H$_2 = \mathrm{dex}( -6.222 + 2.831 / ( 1 + (245 \mathrm{K}/T)^{14} ) )$. We adopted a cloud column of $2\times
10^{23}$H$_2$cm$^{-2}$ and a line FWHM 20kms$^{-1}$.
The resulting total cooling rate per H$_2$ molecule is plotted in Figure 7b for representative H$_2$ densities of $10^3$, $10^4$ and $10^5$cm$^{-3}$. The cooling rates on the left hand axis are mapped to the right-hand axis showing the cosmic ray ionization rate that would supply heat at the same rate, assuming that each ionization is associated with the deposition of 12.4eV of heat$^{42,52}$. The dip in the cooling rate at $T\approx250$K is due to the increasing importance of OI and H$_2$O cooling at higher densities and the switch from OI to H$_2$O at this temperature. We conclude that an ionization rate of $10^{-14}$ to 10$^{-13}$s$^{-1}$ would yield gas temperatures in the range 50–200 K, the upper range consistent with the temperatures estimated using LVG shown in Figure 5b,c.
High Velocity Dispersion of Molecular Clouds
--------------------------------------------
High cosmic-ray fluxes in molecular clouds affect star formation by heating the gas and increasing its ionization fraction. Higher cloud temperatures increase the Jeans mass, potentially changing the IMF, while high ionization increases magnetic coupling to the cloud material, reducing ambipolar diffusion and increasing the time taken for gravitationally unstable cores to contract to the point that they overwhelm their magnetic support. Another consequence of increased ionization is the therefore reduced damping of MHD waves, contributing to sustaining Alfvenic velocity fields within the clouds, which may assist in explaining the observed high velocity dispersion of molecular clouds in the nuclear disk (e.g. Oka et al. 1998; Martin et al. 2004). In a weakly ionized medium, waves with frequencies $\omega \sim kv_A$ below the collision frequency of neutral particles with ions, $\nu_{ni} = n_i <\sigma v>$, are damped on a time scale $2 \nu_{ni} / \omega^2 $ $^{53,54}$ directly proportional to $n_i$ which is increased by a factor of $\sim100$ compared to the Galactic disk for $\zeta\sim10^{-13}$ s$^{-1}$. The power input required to maintain wave motions on a given scale is reduced by the same factor.
Molecular Line Spectra toward G0.13–0.13
----------------------------------------
Figure 8a-d show line emission from 16 molecular lines toward four positions in G0.13–0.13 corresponding to A to D in Table 1, respectively. The spectra of these pointed observations show a rich chemistry similar to that seen throughout the CMZ. We focus only on five molecular lines HCN (1–0), HCO$^+$(1–0), HNC(1–0), N$_2$H$^+$(1–0) and SiO (2–1), as Table 2 show the peak velocities and peak intensities in T$^*_A$ of all five spectral lines. Positions A and D fall in the E and NW boxes (see Fig. 5a). As a demonstration, the spectrum of these five molecular species toward position A is shown in Figure 9. The strong velocity component associated with G0.13-0.13 is centered around 60 (see Table 2). We also notice an additional weak velocity component between -50 and 0 .
Using RADEX[^5] (van der Tak et al. 2007)$^{55}$, we derived the column density of each molecular species to match the observed intensity with the assumption that molecular gas temperature and gas density are T$=200$ K and 10$^4$ cm$^{-3}$, respectively and the velocity dispersion is 30 . Table 3 shows the log of column density ratio of HCN/HNC, HCO$^+$/N$_2$H$^+$ and SiO/N$_2$H$^+$ in columns 2-4 and the corresponding antenna temperature ratios in column 5-8 based on deep Mopra observations at the four positions A-D given column 1. Calibration uncertainties of 25% have not been included in tables 2 and 3. We note that position C lies at the center of the boot-shaped structure, cf. Fig. 2b, where the 74 MHz emission peaks, implying that the cosmic ray ionization rate is higher than the other three positions.
Chemical Modelling
------------------
To examine the effect of high cosmic-ray ionization rates on chemistry, we make use of a time-dependent gas-grain chemical model, UCL\_CHEM$^{31}$ to qualitatively investigate the behavior of the observed species. The UCL\_CHEM is a gas-grain time-dependent model; for the purpose of this paper it is used as a two-stage model: [Phase I follows the free fall collapse of a diffuse (10$^2$ cm$^{-3}$) gas to a denser state (where the final density is a free parameter), while Phase II follows the chemistry as the gas and dust warm up]{} due to either the increase of temperature and/or an enhanced cosmic ray ionization rate (Phase II). Apart from the very initial diffuse state (where the temperature is about 100K), during the collapse in [Phase I]{} the (coupled) gas and dust temperature remains constant at 10 K and atoms and molecules collide with, and freeze on to, grain surfaces. The advantage of this approach is that the ice composition is [***[not]{}***]{} assumed but it is derived by a time-dependent computation of the chemical evolution of the gas–dust interaction process which, in turns, depends on the density of the gas. Hydrogenation occurs rapidly on these surfaces, so that, for example, some percentage of carbon atoms accreting will rapidly become frozen out methane, CH$_4$. The justification for a constant 10 K during most of Phase I is two-fold: first, in order for a diffuse gas to collapse under gravity and form a dense core, the temperature must remain low. Second, it is believed that methanol must form on icy mantles (and this implies an efficient freeze out of CO which can only occur for dust temperatures less than 20K) since experiments show that it can not form in the gas phase$^{56}$ (e.g., Geppert et al. 2006). In Phase II [(which effectively represents the observed phase)]{} the gas and dust temperature are decoupled; the temperature of the gas is varied from 50 to 200 K, while the dust temperature was kept to 20-30K. We employ the reaction rate data from the UMIST astrochemical database, augmenting it with grain-surface (hydrogenation) reactions$^{31,57}$ In both Phases nonthermal desorption is also considered$^{58}$.
Figure 10 shows the abundance ratios of selected species as a function of time for a subset of our model grid where the final density was varied from 10$^4$ cm$^{-3}$ to 10$^6$ cm$^{-3}$ and the cosmic ray ionization rate from 10$^{-15}$ s$^{-1}$ to 10$^{-13}$ s$^{-1}$. The gas temperature in Phase II was assumed to be $\sim200$K. Note, however, that while molecular ratios are in principle powerful tools in constraining the physical and chemical characteristics of a cloud, the assumption that all the selected species arise from the same region may well be incorrect. Nevertheless, we assume that the selected species are co-spatial and consider a comparison between the theoretical column density ratios with those derived from our observations (see Table 3). The \[HCN/HNC\] (black) and \[HCO$^+$/N$_2$H$^+$\] (green) values, as drawn on the top right panel in dashed lines, agree with models of low density ($10^4$ cm$^{-3}$) and high cosmic ray ionization rate $\zeta=10^{-13}$ s$^{-1}$ H$^{-1}$. We note an increase in the gas density and or a decrease in the cosmic ray ionization rate does not change the abundance ratio \[HCN/HNC\] but increases the HCO$^+$/N$_2$H$^+$ (green) ratio to values well above the observed range of 0.005 to 0.19 of Table 3. The cosmic ray ionization rate, inferred independently from this model, is remarkably similar to that estimated from FeI K$\alpha$ line measurements (see Fig. 7a). In addition, the gas density predicted from this chemical model is consistent within the range of values inferred from LVG models, as shown in Figure 5b,c.
It is known that Galactic center molecular clouds have high abundance of SiO$^{59,60,61}$. In one estimate, Minh et al. (1992) find the SiO abundance relative to molecular hydrogen to be between $\sim10^{-7}$ and $10^{-8}$. In other studies, Martin Pintado et al. (1997) also find a high abundance of SiO $\sim10^{-9}$ toward Galactic center molecular clouds. The observed \[SiO/N$_2$H$^+$\] abundance ratio ranges between 0.27–0.44 and can not be matched by any model at late enough times that chemical equilibrium has been reached. The models predict even higher fractional abundance of \[SiO/N$_2$H$^+$\]. In our models, we used an initial atomic abundance for Si depleted by a factor of 100 with respect to solar, as this represents the lower limit of measured initial abundances in a large sample of molecular clouds$^{62}$. However, all initial abudances$^{63}$ are then scaled to match a metallicity twice the solar value. Note that the abundance of SiO derived by our models is consistent with the fact that with high cosmic ray ionization rates the icy mantles are released and therefore Si is released in the gas phase. It is therefore puzzling to observe such a low fractional abundance in a gas where we expect cosmic ray ionization rates to be enhanced. There are three possibilities that could account for the low fractional abundance of \[SiO/N$_2$H$^+$\] compared to that predicted by chemical model. One is that the medium is quite turbulent and hence chemical equilibrium is never reached and the gas is constantly being recycled; then the best solution must be found at earlier times ($< 10^4$ years). In this scenario all ratios can be qualitatively matched for the range of densities 10$^{4-5}$ cm$^{-3}$ and the whole range of cosmic ray ionization rates investigated and no further constraints can be given by the models. The other is that the initial atomic abundance of Si is low. Although we already employ, in models shown in Figure 10, an abundance of the initial Si 1.64$\times10^{-7}$ that is up to a factor of 100 depleted with respect to solar$^{64}$. It could be that in these environments even more than a factor of 100 is depleted. Lastly, it is possible that our derivation of the column density from SiO (2-1) intensity is not accurate, e.g. optically thick. Thus, the intensity of high rotational transitions are needed to determine accurately the SiO column density. Additional detailed work is needed to sort out the discrepancy between the observed and modeled \[SiO/N$_2$H$^+$\] abundance ratio. What is clear from our chemical modeling, as in shocks and and PDRs, a high SiO can also be generated by cosmic rays.
Summary
-------
We showed that the molecular emission from a Galactic center cloud G0.13-0.13, a representattive molecular cloud in the CMZ, coincides with 74 MHz and FeI K$\alpha$ line emission at 6.4 keV. This three-way correlation provides compelling evidence that relativistic electrons are physically associated with the G0.13–0.13 molecular cloud. The high cosmic ray electron ionization rate $\sim10^{-13}$ s$^{-1}$ H$^{-1}$ is responsible for the FeI K$\alpha$ line emission, heating the gas clouds to high temperatures and maintaining the velocity dispersion of clouds by increasing the wave damping time scale. One of the important conclusions of this study is that LVG modeling of multi-transition SiO observations gave molecular gas densities $\sim 10^{4-5}$ cm$^{-3}$ and temperature $\sim100-1000$K. The variation of the spectral index of nonthermal emission from G0.13-0.13 was shown to be unusually large $\Delta\alpha\sim 1.5$. We suggested that the expansion of the G0.13–0.13 molecular cloud into nonthermal vertical filaments accelerates particles along the filaments$^{25,48}$, thus producing a young population of electrons with an unusually flat energy spectrum.
Evidence for FeI K$\alpha$ line and 74 MHz synchrotron emission from warm molecular gas supports the view that the Galactic center is a cosmic ray dominated region. We explored the chemistry in a region subject to high levels of cosmic ray ionization and showed that enhanced SiO emission, widespread throughout Galactic center molecular clouds, can be produced in the context of cosmic ray interaction with molecular gas. The model to explain enhanced SiO, as well as NH$_3$ and CH$_3$OH, emission by cosmic ray driven chemistry is an alternative to the shock-driven chemistry throughout Galactic center molecular clouds.
This research is supported in part by grants from the $Fermi$ Guest Investigator Program as well as the grant AST-0807400 from the NSF the National Science Foundation. The Caltech Submillimeter Observatory is operated by the California Institute of Technology under cooperative agreement with the National Science Foundation (AST-0838261). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.
[99]{}
\(1) Bally, J.; Stark, A.A.; Wilson, R.W.; Henkel, C. Galactic Center Molecular Clouds. II - Distribution and Kinematics. *Astrophys. J.* **1988**, *324*, 223-247
\(2) Dahmen, G.; Huttemeister, S.; Wilson, T. L.; Mauersberger, R.; Linhart, A.; Bronfman, L.; Tieftrunk, A. R.; Meyer, K.; Wiedenhoever, W.; Dame, T. M. et al. Molecular gas in the Galactic Center Region. I. Data from a Large Scale C$^{18}$O(J = 1–0) Survey. *Astron. Astrophys., Suppl. Ser.* **1997**, *126*, 197-236
\(3) Oka, T.; Hasegawa, T.; Sato, F.; Tsuboi, M.; Miyazaki, A. A Large-Scale CO Survey of the Galactic Center. *Astrophys. J., Suppl. Ser.* **1998**, *118*, 455-515
\(4) Martin, C. L.; Walsh, W. M.; Xia, K.; Lane, A. P.; Walker, C. K.; Stark, A. A. The AST/RO Survey of the Galactic Center Region. I. The Inner 3 Degrees. *Astrophys. J., Suppl. Ser.* **2004**, *150*, 239-262
\(5) Tsuboi, M.; Handa, T.; Ukita, N. Dense Molecular Clouds in the Galactic Center Region. I. Observations and Data. *Astrophys. J., Suppl. Ser.* **1999**, *120*, 1-39
\(6) Riquelme, D.; Bronfman, L.; Mauersberger, R.; May, J.; Wilson, T. L. A Survey of the Galactic Center Region in HCO$^{+}$, H$^{13}$CO$^{+}$, and SiO. *Astron. Astrophys.* **2010**, *523*, A45
\(7) Jones, P. A.; Burton, M. G.; Cunningham, M. R.; Requena-Torres, M. A.; Menten, K. M.; Schilke, P.; Belloche, A.; Leurini, S.; Martin-Pintado, J.; Ott, J.; et al. Spectral Imaging of the Central Molecular Zone in Multiple 3-mm Molecular Lines. *Mon. Not. R. Astron. Soc.* **2012**, *419*, 2961-2986
\(8) Yusef-Zadeh, F.; Wardle, M.; Roy, S. Cosmic-Ray Heating of Molecular Gas in the Nuclear Disk: Low Star Formation Efficiency. *Astrophys. J.* **2007**, *665*, L123
\(9) Papadopoulos, P. P. A Cosmic-ray-dominated Interstellar Medium in Ultra Luminous Infrared Galaxies: New Initial Conditions for Star Formation. *Astrophys. J.* **2010**, *720*, 226-232
\(10) Oka, T.; Geballe, T. R.; Goto, M.; Usuda, T.; McCall, B. J. Hot and Diffuse Clouds near the Galactic Center Probed by Metastable H$^{+}$$_{3}$1. *Astrophys. J.* **2005**, *632* 882-893
\(11) Goto, M.; Usuda, T.; Nagata, T.; Geballe, T. R.; McCall, B. J.; Indriolo, N.; Suto, H.; Henning, T.; Morong, C. P.; Oka, T. Absorption Line Survey of H$^{+}$$_{3}$ toward the Galactic Center Sources. II. Eight Infrared Sources within 30 pc of the Galactic Center. *Astrophys. J.* **2008**, *688*, 306-319
\(12) Goto, M.; Usuda, T.; Geballe, T. R.; Indriolo, N.; McCall, B. J.; Henning, T.; Oka, T. Absorption-Line Survey of H$_{3}$$^{+}$ toward the Galactic Center Sources. III. Extent of Warm and Diffuse Clouds. *Publ. Astron. Soc. Jpn.* **2011**, *63*, L13-L17
\(13) van der Tak, F.F.S.; Belloche, A.; Schilke, P.; Güsten, R.; Philipp, S.; Comito, C.; Bergman, P.; Nyman, L.-A. APEX Mapping of H$_{3}$O$^{+}$ in the Sgr B2 Region. *Astron. Astrophys.* **2006**, *454*, L99-L102
\(14) Lis, D. C.; Carlstrom, J. E. Submillimeter Continuum Survey of the Galactic Center. *Astrophys. J.* **1994**, *424*, 189-199
\(15) Koyama, K.; Maeda, Y.; Sonobe, T.; Takeshima, T.; Tanaka, Y.; Yamauchi, S. ASCA View of Our Galactic Center: Remains of Past Activities in X-Rays?. *Publ. Astron. Soc. Jpn.* **1996**, *48*, 249-255
\(16) Murakami, H.; Koyama, K.; Maeda, Y. Chandra Observations of Diffuse X-Rays from the Sagittarius B2 Cloud. *Astrophys. J.* **2001**, *558*, 687-692
\(17) Sunyaev, R.; Churazov, E. Equivalent Width, Shape and Proper Motion of the Iron Fluorescent Line Emission from Molecular Clouds as an Indicator of the Illuminating Source X-ray Flux History. *Mon. Not. R. Astron. Soc.* **1998**, *297*, 1279-1291
\(18) Koyama, K.; Inui, T.; Matsumoto, H.; Tsuru, T. G. A Time-Variable X-Ray Echo: Indications of a Past Flare of the Galactic-Center Black Hole. *Publ. Astron. Soc. Jpn.* **2008**, *60*, S201-S206
\(19) Valinia, A.; Tatischeff, V.; Arnauld, K.; Ebisawa, K.; Ramaty, R. On the Origin of the Iron K Line in the Spectrum of The Galactic X-Ray Background. *Astrophys. J.* **2000**, *543*, 733-739
\(20) Yusef-Zadeh, F.; Law, C.; Wardle, M. The Origin of X-Ray Emission from a Galactic Center Molecular Cloud: Low-Energy Cosmic-Ray Electrons. *Astrophys. J.* **2002**, *568*, L121-L125
\(21) Ponti, G.; Terrier, R.; Goldwurm, A.; Belanger, G.; Trap, G. Discovery of a Superluminal Fe K Echo at the Galactic Center: The Glorious Past of Sgr A\* Preserved by Molecular Clouds. *Astrophys. J.* **2010**, *714*, 732-747
\(22) Morris, M.; Serabyn, E. The Galactic Center Environment. *Annu. Rev. Astron. Astrophys.* **1996**, *34*, 645-702
\(23) Li, P. S.; Myers, A.; McKee, C. Ambipolar Diffusion Heating in Turbulent Systems. *Astrophys. J.* **2012**, *760*, 33
\(24) Brogan, C. L.; Nord, M.; Kassim, N.; Lazio, J.; Anantharamaiah, K. Spatially Resolved Very Large Array 74 MHz Observations Toward the Galactic Center. *Astron. Nachr. Suppl.*, **2003**, *324*, 17-24
\(25) Tsuboi, M.; Ukita, N.; Handa, T. An Expanding Shell-like Molecular Cloud near the Galactic Center Arc. *Astrophys. J.* **1997**, *481*, 263
\(26) Handa, T.; Sakano, M.; Naito, S.; Hiramatsu, M.; Tsuboi, M. Thermal SiO and H$^{13}$CO$^{+}$ Line Observations of the Dense Molecular Cloud G0.11-0.11 in the Galactic Center Region. *Astrophys. J.* **2006**, *636*, 261-266
\(27) Tsuboi, M.; Tadaki, K.; Miyazaki, A.; Handa, T. Sagittarius A Molecular Cloud Complex in H$^{13}$CO$^{+}$ and Thermal SiO Emission Lines. *Publ. Astron. Soc. Jpn.* **2011**, *63*, 763-794
\(28) Hüttemeister, S.; Wilson, T.L.; Banina, T.M.; Martin-Pintado, J. Kinetic Temperatures in Galactic Center Molecular Clouds. *Astron. Astrophys.* **1993**, *280*, 255-267
\(29) Pierce-Price, D.; Richer, J. S.; Greaves, J. S.; Holland, W. S.; Jenness, T.; Lasenby, A. N.; White, G. J.; Matthews, H. E.; Ward-Thompson, D.; Dent, W. R. F.; et al. A Deep Submillimeter Survey of the Galactic Center. *Astrophys. J.* **2000**, *545*, L121-L125
\(30) Molinari, S.; Bally, J.; Noriega-Crespo, A.; Compiegne, M.; Bernard, J. P.; Paradis, D.; Martin, P.; Testi, L.; Barlow, M.; Moore, T.; et al. A 100 pc Elliptical and Twisted Ring of Cold and Dense Molecular Clouds Revealed by Herschel Around the Galactic Center. *Astrophys. J.* **2011**, *735*, L33
\(31) Viti, S.; Collings, M. P.; Dever, J. W.; McCoustra, M. R. S.; Williams, D. A. Evaporation of Ices Near Massive Stars: Models Based on Laboratory Temperature Programmed Desorption Data. *Mon. Not. R. Astron. Soc.* **2004**, *354*, 1141-1145
\(32) Welch, W. J.; Thornton, D. D.; Plambeck, R. L.; Wright, M. C. H.; Lugten, J.; Urry, L.; Fleming, M.; Hoffman, W.; Hudson, J.; Lum, W. T.; et al. The Berkeley-Illinois-Maryland-Association Millimeter Array. *Publ. Astron. Soc. Pac.* **1996**, *108*, 93
\(33) Oka, T.; Nagai, M.; Kamegai, K.; Tanaka, K.; Kuboi, N. A CO J = 3-2 Survey of the Galactic Center. *Publ. Astron. Soc. Jpn.* **2007**, *59*, 15-23
\(34) Law, C. J.; Yusef-Zadeh, F.; Cotton, W. D.; Maddalena, R. J. Green Bank Telescope Multiwavelength Survey of the Galactic Center Region. *Astrophys. J., Suppl. Ser.* **2008**, *177*, 255-274
\(35) Nord, M.E.; Lazio, T.J.W.; Kassim, N.E.; Hyman, S.J.; LaRosa, T.N.; Brogan, C.L.; Duric, N. High-Resolution, Wide-Field Imaging of the Galactic Center Region at 330 MHz. *Astron. J.* **2004**, *128*, 1646-1670
\(36) Camilo, F.; Ransom, S. M.; Gaensler, B. M.; Lorimer, D. R.
Discovery of the Energetic Pulsar J1747-2809 in the Supernova Remnant G0.9+0.1. *Astrophys. J.* **2009**, *700*, L34-L38
\(37) Yusef-Zadeh, F.; Morris, M. G0.18-0.04 - Interaction of Thermal and Nonthermal Radio Structures in the Arc Near the Galactic Center. *Astron. J.* **1987**, *94*, 1178-1184
\(38) Reich, W. G0.087-0.087, A Highly Polarized Flat Spectrum Filament Near the Galactic Centre Arc. *Astron. Astrophys.* **2003**, *401*, 1023-1026
\(39) Uchiyama, H.; Nobukawa, M.; Tsuru, T.; Koyama, K.; Matsumoto, H. Global Distribution of Fe K[$\alpha$]{} Lines in the Galactic Center Region Observed with the Suzaku Satellite. *Publ. Astron. Soc. Jpn.* **2011**, *63*, S903-S911
\(40) Amo-Baladron M. A.; Martin-Pintado, J.; Morris, M. R.; Muno, M. P.; Rodriguez-Fernandez, N. J. SiO Emission as a Tracer of X-Ray Dominated Chemistry in the Galactic Center. *Astrophys. J.* **2009**, *694*, 943-950
\(41) Yusef-Zadeh, F.; Muno, M.; Wardle. M.; Lis, D.C. The Origin of Diffuse X-Ray and [$\gamma$]{}-Ray Emission from the Galactic Center Region: Cosmic-Ray Particles. *Astrophys. J.* **2007**, *656*, 847-869
\(42) Yusef-Zadeh, F.; Hewitt, J. W.; Wardle, M.; Tatischeff, V.; Roberts, D.; Cotton, W.; Uchiyama, H.; Nobukawa, M.; Tsuru, T. G.; Heinke, C. et al. Interacting Cosmic Rays with Molecular Clouds: A Bremsstrahlung Origin of Diffuse High Energy Emission from the Inner 2deg by 1deg of the Galactic Center. *Astrophys. J.* **2013**, 762, 1-33
\(43) LaRosa, T. N.; Brogan, C. L.; Shore, S. N.; Lazio, T. J.; Kassim, N. E.; Nord, M. E. Evidence of a Weak Galactic Center Magnetic Field from Diffuse Low-Frequency Nonthermal Radio Emission. *Astrophys. J.* **2005**, *626*, L23-L27
\(44) Egan, M. P.; Shipman, R. F.; Price, S. D.; Carey, S. J.; Clark, F. O.; Cohen, M. A Population of Cold Cores in the Galactic Plane. *Astrophys. J.* **1998**, *494*, L199
\(45) Price, S. D.; Egan, M. P.; Carey, S. J.; Mizuno, D. R.; Kuchar, T. A. Midcourse Space Experiment Survey of the Galactic Plane. *Astron. J.* **2001**, *121*, 2819-2842
\(46) Rodriguez-Fernandez, N. J.; Martin-Pintado, J.; Fuente, A.; de Vicente, P.; Wilson, T.L.; Hüttemeister, S. Warm H$_{2}$ in the Galactic Center Region. *Astron. Astrophys.* **2001**, *365*, 174-185
\(47) Simpson, J. P.; Colgan, S. W. J.; Cotera, A. S.; Erickson, E. F.; Hollenbach, D. J. Warm H$_{2}$ in the Galactic Center Region. *Astrophys. J.* **2007**, *670*, 1115-1131
\(48) Oka, T.; Hasegawa, T.; Sato, F.; Tsuboi, M.; Miyazaki, A. A Molecular Cloud and an Expanding Cavity Adjacent to the Nonthermal Filaments of the Galactic Center Radio Arc. *Publ. Astron. Soc. Jpn.* **2001**, *53*, 779-786
\(49) Schöier, F.L.; van der Tak, F.F.S.; van Dishoeck E.F.; Black, J.H. An Atomic and Molecular Database for Analysis of Submillimetre Line Observations. *Astron. Astrophys.* **2005**, *432*, 369-379
\(50) Neufeld, D. A.; Kaufman, M. J. Radiative Cooling of Warm Molecular Gas. *Astrophys. J.* **1993**, *418*, 263
\(51) Neufeld, D. A.; Lepp, S.; Melnick, G. J. Thermal Balance in Dense Molecular Clouds: Radiative Cooling Rates and Emission-Line Luminosities. *Astrophys. J. Supp* **1995**, *100*, 132
\(52) Glassgold, A.; Galli, D.; Padovani, M. Cosmic-Ray and X-Ray Heating of Interstellar Clouds and Protoplanetary Disks. *Astrophys. J.* **2012**, *756*, 157
\(53) Kulsrud, R.; Pearce, W. P. The Effect of Wave-Particle Interactions on the Propagation of Cosmic Rays. *Astrophys. J.* **1969**, *156* 445
\(54) Zweibel, E. G.; Josafatsson, K. Hydromagnetic Wave Dissipation in Molecular Clouds. *Astrophys. J.* **1983**, *270*, 511-518
\(55) van der Tak, F. F. S.; Black, J. H.; Schöier, F. L.; Jansen, D. J.; van Dishoeck, E. F. A Computer Program for Fast Non-LTE Analysis of Interstellar Line Spectra. With Diagnostic Plots to Interpret Observed Line Intensity Ratios. *Astron. Astrophys.* **2007**, *468*, 627-635
\(56) Geppert, W. D.; Hamberg, M.; Thomas, R. D.; Osterdahl, F.; Hellberg, F.; Zhaunerchyk, V.; Ehlerding, A.; Millar, T. J.; Roberts, H.; Semaniak, J. et al. Dissociative recombination of protonated methanol *Chemical Evolution of the Universe, Faraday Discussions* **2006**, *133*, 177-190
\(57) Viti, S.; Jimenez-Serra, I.; Yates, J. A.; Codella, C.; Vasta, M.; Caselli, P.; Lefloch, B.; Ceccarelli, Water and Ammonia as Diagnostics of Shock Temperature. *Astrophys. J.* **2011**, *740*, L3-5
\(58) Roberts, J. F.; Rawlings, J. M. C.; Viti, S. & Williams, D. A. Desorption from Interstellar Ices. *Mon. Not. R. Astron. Soc.* **2007**, *382*, 733-742
\(59) Minh, Y. C.; Irvine, W. M. & Friberg, P. Molecular Abundances in the Sagittarius A Molecular Cloud. *Astron. & Astrophys.* **1992**, *258*, 489-494
\(60) Martin-Pintado, J.; de Vicente, P.; Fuente, A. & Planesas, P. SiO Emission from the Galactic Center Molecular Clouds. *Astrophys. J.* **1997**, *482*, L45-L48
\(61) Martin-Pintado, J.; Rizzo, J. R. de Vicente, P.; Rodriguez-Fernandez, N. J. & Fuente, A. & Planesas, P. Large-Scale Grain Mantle Disruption in the Galactic Center. *Astrophys. J.* **2001**, *548*, L65-L68
\(62) Kimura, Hiroshi; Mann, Ingrid; Jessberger, Elmar K. Elemental Abundances and Mass Densities of Dust and Gas in the Local Interstellar Cloud *Astrophys. J.* **2003**, *582*, 846-858
\(63) Sofia, U. J.; Meyer, D. M. Interstellar Abundance Standards Revisited. *Astrophys. J.* **2001**, *554*, L221-L224
\(64) Sofia, U. J.; Cardelli, J. A.; Savage, B. D. The Abundant Elements in Interstellar Dust. *Astrophys. J.* **1994**, *430*, 650-666
![ [*(Top to Bottom)*]{} [*(a)*]{} A grayscale 74 MHz distribution of the inner 1.5$^{\circ}\times12'$ (Galactic coordinates $l\times b$) convolved to a resolution of $2'\times2'$ based on VLA observations. The 74 MHz flux range is between -1 and 4 Jy. [*(b)*]{} Similar to [*(a)*]{} except the distribution of CO (3-2) integrated over $\pm$200 at $34''$ resolution$^{48}$ with flux range 100 and 2500 K. [*(c)*]{} Contours of the CO (3-2) line emission shown in (b) are superimposed on the 74 MHz image shown in [*(a)*]{}. [*(d)*]{} Similar to [*(a)*]{} except observed with the GBT at 8.5 GHz with a resolution of 88$''$ $^{34}$. The flux range is between -5e-3 and 0.4 Jy. ](fig1_new.pdf)
![ \[fig:20cm\] [*(a - Left)*]{} Contours of CS (1-0) line emission integrated between 0 and 50 from the G0.13–0.13 molecular cloud with a resolution of 45$''^{25}$ are superimposed on a 5 GHz continuum image with a spatial resolution of 10.8$''\times5.5''$ (PA=-2.6$^{\circ}$). Contour levels are set at 2 to 10 K (T$^*_A$) with 1 K interval. [*(b - Right)*]{} The distribution of CS (2-1) line emission (red) from G0.13–0.13 integrated over velocities between –12.1 $< v_{LSR}<76.4$ with a resolution of of $3.1''\times2.6''$ (PA=20.5$^{\circ}$) is based on BIMA observations. The extended nonthermal filaments of the Arc at 1.4 GHz with a resolution of $10.7''\times10.1''$ (PA=28$^{\circ}$) are shown in green. ](fig2a_new.pdf "fig:") ![ \[fig:20cm\] [*(a - Left)*]{} Contours of CS (1-0) line emission integrated between 0 and 50 from the G0.13–0.13 molecular cloud with a resolution of 45$''^{25}$ are superimposed on a 5 GHz continuum image with a spatial resolution of 10.8$''\times5.5''$ (PA=-2.6$^{\circ}$). Contour levels are set at 2 to 10 K (T$^*_A$) with 1 K interval. [*(b - Right)*]{} The distribution of CS (2-1) line emission (red) from G0.13–0.13 integrated over velocities between –12.1 $< v_{LSR}<76.4$ with a resolution of of $3.1''\times2.6''$ (PA=20.5$^{\circ}$) is based on BIMA observations. The extended nonthermal filaments of the Arc at 1.4 GHz with a resolution of $10.7''\times10.1''$ (PA=28$^{\circ}$) are shown in green. ](fig2b.pdf "fig:")
![ [*(a - Top Left)*]{} Contours of 74 MHz emission with a resolution 122$''\times64''$ (PA=-$5^{\circ}$) with levels set at 0.8, 0.9,...2.3 Jy beam$^{-1}$ are superimposed on a grayscale image at 5 GHz. [*(b - Top Right)*]{} Contours of 327 MHz emission with levels (0.8, 0.9, ...2.3)$\times2$ are shown with the same resolution as in (a) and are superimposed on the 5 GHz image. The white dashed lines show the location of cross cuts (see (d) below). [*(c - Left)*]{} The spectral index distribution $\alpha$ between 74 and 327 MHz emission. The 74 and 327 MHz are convolved to the same resolution and are background subtracted by 2 and 0.5 Jy, respectively. The bar at the bottom of the figure shows the color scale version of spectral index values ranging between -0.5 and 2 (see text). [*(d-e - Right)*]{} Top two panels show background subtracted intensity profiles at 327 and 74 MHz (top) and the corresponding spectral index distribution between these two frequencies (below) made from vertical latitude cross cut at constant longitude l=$-6'$, as drawn on (b) with a vertical dashed line. Bottom two panels are the same as the top two panels except that the cross cuts made horizontally at constant b=$-8'$. Cross cuts are made on the 74 and 327 MHz images. ](fig3a_new.pdf "fig:") ![ [*(a - Top Left)*]{} Contours of 74 MHz emission with a resolution 122$''\times64''$ (PA=-$5^{\circ}$) with levels set at 0.8, 0.9,...2.3 Jy beam$^{-1}$ are superimposed on a grayscale image at 5 GHz. [*(b - Top Right)*]{} Contours of 327 MHz emission with levels (0.8, 0.9, ...2.3)$\times2$ are shown with the same resolution as in (a) and are superimposed on the 5 GHz image. The white dashed lines show the location of cross cuts (see (d) below). [*(c - Left)*]{} The spectral index distribution $\alpha$ between 74 and 327 MHz emission. The 74 and 327 MHz are convolved to the same resolution and are background subtracted by 2 and 0.5 Jy, respectively. The bar at the bottom of the figure shows the color scale version of spectral index values ranging between -0.5 and 2 (see text). [*(d-e - Right)*]{} Top two panels show background subtracted intensity profiles at 327 and 74 MHz (top) and the corresponding spectral index distribution between these two frequencies (below) made from vertical latitude cross cut at constant longitude l=$-6'$, as drawn on (b) with a vertical dashed line. Bottom two panels are the same as the top two panels except that the cross cuts made horizontally at constant b=$-8'$. Cross cuts are made on the 74 and 327 MHz images. ](fig3b_new.pdf "fig:")\
![continued[]{data-label="fig:example"}](fig3c.pdf "fig:") ![continued[]{data-label="fig:example"}](fig3de.pdf "fig:")
![ [*(a)*]{} Contours of SiO (5-4) line emission integrated between -20 and 100 with levels at 8, 10, 12, 14, 16, 18, 20, 23 K (T$_{\rm mb}$ velocity) are superimposed on a 5 GHz grayscale image. Two rectangular boxes NW and E are drawn for LVG analysis. [*(b)*]{} The parameters of density n$_{\rm H}$ cm$^{-3}$ and temperature T (K) using LVG model are presented. SiO (5-4), (6-5), (2-1) line ratios are displayed in the bottom left corner in red, green and black, respectively. The dashed lines show the corresponding calibration uncertainties of the derived parameters at a level of 25%. The SiO line intensity ratios are extracted from the region presented by the E rectangular box, as drawn schematically on (a). [*(c)*]{} Similar to [*(b)*]{} except that data are from the NW box shown in (a). [*(d)*]{} Top three panels show the plots of line ratios as function of logarithm of pressure (nT) based on a grid of models. The observed line ratio values with their corresponding errors are drawn horizontally (green) in solid and dashed lines, respectively. The bottom panel shows the best $\chi^2$ fit to the grid of models. The vertical lines (green) are centered near minimum $\chi^2$ giving log(nT) values in the range of 6.4–6.9. ](fig5a.pdf "fig:")\
![continued[]{data-label="fig:example"}](fig5b.pdf "fig:") ![continued[]{data-label="fig:example"}](fig5c.pdf "fig:")\
![continued[]{data-label="fig:example"}](fig5d_press_lvg.pdf "fig:")
![Contours of SiO (2-1) line emission based on Mopra observations with levels 2, 3.5, 5, 6.5, 8 K are superimposed on a three color Spitzer/IRAC image combining Channels 4 (8$\mu$m) is red), 2 (4.5$\mu$m) is green and 1 (3.6$\mu$m) is blue. The temperature scale is in T$^*_A$. The straight line in white shows the orientation of the nonthermal filament of the radio arc whereas the dashed circle marks roughly the size of the arc bubble. The positions where molecular abundances are measured are drawn as A, B, C and D. The box indicated the region mapped by Mopra. \[fig:20cm\] ](fig6_new.pdf)
![ \[fig:20cm\] [*(a - Top)*]{} The variation of cosmic ray ionization rate as a function of magnetic field strength for different values of the spectral index of the radiation $\alpha$. The black dot on each curve gives the value at which the magnetic field and particle energies are in equipartition. [*(b - Bottom)*]{} Solid curves show the total cooling rate for diffuse molecular gas for H$_2$ densities of $10^3$, $10^4$ and $10^5$cm$^{-3}$. Rotational transitions of H$_2$, CO and H$_2$O, and by fine-structure transitions of OI have been included; their individual contributions for $10^3$cm$^{-3}$ are indicated by the dashed curves. The right hand axis shows the ionization rate by cosmic-ray electrons needed to supply the corresponding heating rate (see text). The vertical axis in both figures are logarithmic. ](fig7a+b.pdf)
![ [*(a - Top Left 4)*]{} Deep pointed observations showing the spectra of 16 molecules toward position A in G0.13–0.13 [*(b - Top Right 4)*]{} Similar to [*(a)*]{} except for position B. [*(c - Top Left 4)*]{} Similar to [*(a)*]{} except for position C. [*(d - Top Right 4)*]{} Similar to [*(a)*]{} except for position D. ](fig8a.pdf "fig:") ![ [*(a - Top Left 4)*]{} Deep pointed observations showing the spectra of 16 molecules toward position A in G0.13–0.13 [*(b - Top Right 4)*]{} Similar to [*(a)*]{} except for position B. [*(c - Top Left 4)*]{} Similar to [*(a)*]{} except for position C. [*(d - Top Right 4)*]{} Similar to [*(a)*]{} except for position D. ](fig8b.pdf "fig:")\
![continued[]{data-label="fig:example"}](fig8c.pdf "fig:") ![continued[]{data-label="fig:example"}](fig8d.pdf "fig:")
[c c c]{} A & 0.1007 & $-$0.0792\
B & 0.0632 & $-$0.0709\
C & 0.1132 & $-$0.1084\
D & 0.1299 & $-$0.1292\
[c c c c c c c c c c c]{} A & 60.65$\pm0.9$ & 5.02$\pm0.03$ & 57.91$\pm0.90$ & 3.49$\pm0.03$ & 54.92$\pm0.90$ & 2.12$\pm0.02$ & 58.13$\pm0.90$ & 2.49$\pm0.02$ & 54.29$\pm0.90$ & 1.09$\pm0.03$\
B & 57.88$\pm0.90$ & 3.96$\pm0.03$ & 51.53$\pm0.90$ & 3.10$\pm0.03$ & 50.42$\pm0.90$ & 2.59$\pm0.02$ & 48.54$\pm0.90$ & 2.64$\pm0.03$ & 49.60$\pm0.90$ & 0.96$\pm0.03$\
C & 34.19$\pm0.90$ & 2.65$\pm0.03$ & 34.33$\pm0.90$ & 2.33$\pm0.03$ & 28.15$\pm0.90$ & 1.12$\pm0.03$ & 32.08$\pm0.90$ & 1.19$\pm0.02$ & 26.36$\pm0.90$ & 0.32$\pm0.02$\
D & 32.35$\pm0.90$ & 1.98$\pm0.03$ & 41.56$\pm0.90$ & 1.98$\pm0.03$ & 30.81$\pm0.90$ & 0.97$\pm0.03$ & 33.80$\pm0.90$ & 1.00$\pm0.03$ & 42.1$\pm0.90$ & 0.38$\pm0.03$\
[c c c c c c c ]{} A & 0.33$\pm0.01$ & 0.13$\pm0.01$ & -0.10$\pm0.01$ & 2.37$\pm0.03$ & 1.40$\pm0.02$ & 0.44$\pm0.01$\
B & 0.01$\pm0.001$ & 0.005$\pm0.001$ & -0.20$\pm0.02$ & 1.53$\pm0.02$ & 1.17$\pm0.02$ & 0.36$\pm0.01$\
C & 0.35$\pm0.005$ & 0.19$\pm0.01$ & -0.41$\pm0.03$ & 2.37$\pm0.05$ & 1.96$\pm0.04$ & 0.27$\pm0.02$\
D & 0.26$\pm0.01$ & 0.16$\pm0.01$ & -0.24$\pm0.03$ & 2.04$\pm0.06$ & 1.98$\pm0.07$ & 0.38$\pm0.03$\
[^1]: The National Radio Astronomy Observatory is a facility of the National Science Foundation, operated under a cooperative agreement by Associated Universities, Inc.
[^2]: The data was obtained using the Mopra radio telescope, a part of the Australia Telescope National Facility which is funded by the Commonwealth of Australia for operation as a National Facility managed by CSIRO. The University of New South Wales (UNSW) digital filter bank (the UNSW-MOPS) used for the observations with Mopra was provided with support from the Australian Research Council (ARC), UNSW, Sydney and Monash Universities, as well as the CSIRO.
[^3]: The BIMA interferometer was operated under a joint agreement between the University of California, Berkeley, the University of Illinois, and the University of Maryland with support from the National Science Foundation.
[^4]: Combined Array for Research in Millimeter-wave Astronomy; http://www.mmarray.org
[^5]: http://www.sron.rug.nl/ vdtak/radex/index.shtml
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Enhanced quantization offers a different classical/quantum connection than that of canonical quantization in which $\hbar>0$ throughout. This result arises when the only allowed Hilbert space vectors allowed in the quantum action functional are coherent states, which leads to the classical action functional augmented by additional terms of order $\hbar$. Canonical coherent states are defined by unitary transformations of a fixed, fiducial vector. While Gaussian vectors are commonly used as fiducial vectors, they cannot be used for all systems. We focus on choosing fiducial vectors for several systems including bosons, fermions, and anyons.'
author:
- 'T. C. Adorno[^1]'
- 'J. R. Klauder[^2]'
title: |
**Examples of Enhanced Quantization:\
Bosons, Fermions, and Anyons**
---
Introduction
============
Canonical quantization would seem to be, as its name suggests, a closed subject. However, some attempts at canonical quantization lead to unnatural results—most notably the “Triviality of $\phi^4_n$ for $n\ge5$”, and possibly for $n=4$ as well. Rather than accept such an outcome, a natural result may perhaps be obtained with an alternative quantization procedure, which, for the $\phi^4_n$ example, has been presented elsewhere [@cc]. The basis for an alternative quantization procedure [@EQ] is readily summarized here; for some early motivation, see [@Kla63I; @Kla63II; @Kla63III]. For the simple example of a single particle, the classical action functional is given by A\_C=\_0\^Tdt, while the corresponding quantum action functional is given by A\_Q=\_0\^T(t)||(t)dt.\[e2\] A stationary variation of the paths $p(t)$ and $q(t)$ leads to Hamilton’s equations of motion, while a stationary variation of normalized Hilbert state vectors $\<\psi(t)|$ (and $|\psi(t)\>$) leads to Schrödinger’s equation of motion (and its adjoint). The connection between these two theories—known as “canonical quantization”—comes when we promote $p\ra P$ and $q\ra Q$ as Hermitian operators, and $\H(P,Q)$ is chosen as $H_c(P,Q)$ apart from possible ${\cal O}(\hbar)$ corrections. This prescription works best when $(p,q)$ are Cartesian coordinates [@dirac]—despite the fact that phase space carries no metric with which to determine Cartesian coordinates!
There are many other proposed connections between classical and quantum expressions. We consider just two examples: (1) The Wigner [@Wigner] phase-space representation of a quantum state leads to a distribution function, but it generally suffers from not being a positive distribution; (2) The Husimi [@Husimi] phase-space representation provides a positive representation of an associated phase-space distribution function, but its partners in forming expectations are often singular distributions [@Sudar]. Further examples are presented in most textbooks dealing with quantum theory, e.g., [@BerShu91].
Let us examine an alternative quantization procedure which has been called “Enhanced Quantization” [@EQ]. The derivation of Schrödinger’s equation of motion assumes that sufficiently many vectors $\<\psi(t)|$ can be varied in (\[e2\]), but suppose that is not possible. A macroscopic observer can only change a few features of a quantum system, such as its position or velocity, which thanks to Galilean covariance, can be realized by moving the observer rather than disturbing the microscopic system. Thus, choosing a suitable, normalized fiducial vector $|\eta\>$, and with two self-adjoint operators $P$ and $Q$, $[Q,P]=i\hbar$, and initially choosing $|\eta\>$ as $|0\>$, an harmonic oscillator ground state determined by $(Q+iP)\s|0\>=0$, we generate two, independent unitary transformations of $|0\>$ leading to |p,qe\^[-iqP/]{}e\^[ipQ/]{}|0. \[coherstates\]We choose this set of states for all $(p,q)\in{\mathbb R}^2$ as states that can be varied by a macroscopic observer. This set is also well known as a set of canonical coherent states with $|0\>$ serving as its fiducial vector [@jkbs]. Using these states, the restricted quantum action functional becomes A\_[Q(R)]{}-1.3em&&=\_0\^Tp(t),q(t)||p(t),q(t)dt\
&&=\_0\^Tdt. Thus a stationary variation of the restricted action leads to Hamilton’s equations with $H(p,q)$ serving as the enhanced classical Hamiltonian since $\hbar>0$ still. Specifically, H(p,q)-1.3em&&p,q|(P,Q)|p,q\
&&=0|(P+p,Q+q)|0\
&&=(p,q)+[O]{}(; p,q; |0).\[kla5\]
It is important to note that this classical/quantum relation is exactly what is meant by the relation between the classical and quantum Hamiltonians when $(p,q)$ are the favored Cartesian coordinates. Although the phase space cannot determine Cartesian coordinates, the Hilbert space provides a metric. Since the overall phase of a Hilbert space vector has no physical significance in quantum theory, we use a suitably scaled version of the distance determined by $D_{ray}(|\psi\>;|\phi\>)^2\equiv(2\hbar)\min_\a \||\psi\>-e^{i\a}\s|\phi\>\|^2$, which, when applied to two infinitesimally close coherent states, leads to d(p,q)\^2(2)\[d|p,q\^2-|p,q|d|p,q|\^2\]= dp\^2+dq\^2. Thus, it is the Hilbert space that determines that the chosen coordinates are Cartesian, and, if one so desired, that metric could be added to the phase space as well. In that sense the alternative classical/quantum connection used in enhanced quantization has nevertheless yielded the very same result as canonical quantization since the Cartesian coordinates determined by the Hilbert space metric are linked with the very same choice of the Hamiltonian operator as related to the classical Hamiltonian as is chosen by canonical quantization.
There are additional features of enhanced quantization that are covered elsewhere, e.g., [@EQ; @Moscow; @Ben]. They also include affine quantization, which although not discussed in this article, could also be related to our further discussions. In fact, the main role of this article is to feature the choice of the fiducial vector $|\eta\>$. So far we have chosen $|0\>$ as the fiducial vector, as this is a common choice whenever discussing coherent states. On the other hand, the choice of $|0\>$ is not always optimal or even possible, and in the rest of this paper we discuss the features and requirements of various fiducial vectors $|\eta\>$.
Choosing the Fiducial Vector
============================
General properties
------------------
Coherent states are often defined with the aid of a group, and in so doing the action of the group acting on a fixed, fiducial vector defines the coherent states. As an example consider the set of canonical coherent states given by |p,q;e\^[-iqP/]{}e\^[ipQ/]{}|. Normally, the choice of the normalized fiducial vector $|\eta\>$ is left implicit, but on this occasion, since we are deciding on how to choose a “good”—or the “best”—fiducial vector, we include it explicitly in the previous equation. The transformation of the fiducial vector to make coherent states involves unitary transformations by the given expressions, which means that both P and Q must be self-adjoint operators so they may generate unitary operators. The real variables $p$ and $q$ each generate one-parameter groups expressed in so-called canonical group coordinates of the second kind [@Cohen]. Since such unitary operators are strongly continuous in their parameters, e.g., $\|\s[\s e^{ipQ/\hbar}-\one\s]\s|\psi\>\s\|\ra0$ as $p\ra0$ for all $|\psi\>\in{\mathfrak{ H}}$, it follows that the coherent states are strongly continuous in their parameters for any $|\eta\>$. This property ensures the continuity of the coherent state representation for any abstract vector $|\psi\>$ given by $\psi(p,q;\eta)\equiv\<p,q;\eta\s|\psi\>$ for any $|\eta\>$.
Besides continuity, the other basic feature of coherent states is a resolution of unity by an integral over the entire phase space of coherent state projection operators that involves an absolutely continuous measure with a suitable positive weighting, which, for the example under consideration, is given by =|p,q;p,q;|d(p,q),3em d(p,q)dpdq/(2), a relation that holds weakly (as well as strongly) for any choice of the normalized fiducial vector $|\eta\>$. \[Remark: For other sets of putative coherent states, it sometimes happens that the resolution of unity fails; in this case we deal with so-called “weak coherent states” [@jkbs]. When that happens, it is useful to let the inner product of weak coherent states serve as a reproducing kernel and to generate a reproducing kernel Hilbert space [@repro].\]
First look at choosing the fiducial vector
------------------------------------------
In enhanced quantization, such as discussed in Sec. 1 of this paper, the restricted quantum action functional is given by A\_[Q(R)]{}-1.3em&&=\_0\^Tp,q;||p,q;dt\
&& =\_0\^T\[p(t)[q]{}(t)-H(p(t),q(t))\]dt. and thus it is necessary that the the coherent states are in the domain of the Hamiltonian operator $\H$, which, for an unbounded Hamiltonian operator, will already induce a certain restriction on $|\eta\>$. Additionally, in giving physical meaning to the variables $p$ and $q$, it is useful to impose “physical centering”, i,e, $\<P\>=\<Q\>=0$, wherein we have introduced the shorthand that $\<(\cdot)\>\equiv\<\eta|(\cdot)|\eta\>$. The virtue of physical centering becomes clear when we note that it leads to p,q;|P|p,q;=p,3em p,q;|Q|p,q;=q, yielding a natural physical interpretation of the parameters $p$ and $q$.
The enhanced classical Hamiltonian $H(p,q)$ (with $\hbar>0$) differs from the classical Hamiltonian $H_c(p,q)=\lim_{\hbar\ra0} H(p,q)$ by terms of order $\hbar$. As an example, let $\H(P,Q)=P^2+Q^2+Q^4$. in which case H(p,q)=p\^2+q\^2+q\^4+6q\^2Q\^2 +,\[11K\] assuming for simplicity that odd expectations vanish. Apart from a constant, there is the term $6q^2\<Q^2\>$ which will modify the usual equations of motion and their solution. If we choose $\<x|\eta\>\propto \exp(-\omega\s x^2/2\hbar)$, it follows that $\<Q^2\>\propto \hbar$, meaning that the correction is important only when $p$ and $q$ are “quantum-sized” themselves. That limitation makes good sense since then the enhanced classical description is effectively unchanged for macroscopic motion, only showing quantum “uncertainty”, i.e., dependence on $\omega$, for quantum-sized motion. Alternatively, we could in principle choose $\<x|\eta\>\propto \exp(-a\s x^2/2)$, where, say, $a=10^{-137}\s m^{-2}$ and independent of $\hbar$, which means that the additional term would modify the quadratic term by a potentially huge amount even for macroscopic motions. Such a choice is mathematically possible, just as studying a simple harmonic oscillator with displacements on a planetary scale or energies equivalent to the mass energy of the Earth are mathematically possible. However, they are unphysical applications of the mathematical description of a simple harmonic oscillator. In a similar story, although it is mathematically possible to choose fiducial vectors so that $\<Q^2\>$ leads to macroscopic modifications, such a fiducial vector could never be physically realized. Thus we conclude that it is logical to choose the fiducial vector supported largely on a “quantum-sized” region. However, that still leaves open many possibilities. Indeed, Troung [@TTT] has showed that choosing the ground state of a quartic Hamiltonian as the fiducial vector can recast Schrödinger’s equation into a new form that offers novel analysis options.
Second look at choosing the fiducial vector
-------------------------------------------
It is popular to choose the fiducial vector to be a “quantum-sized” Gaussian; indeed, we have done so in Sec. 1. This choice often leads to fairly simple analytic expressions, and sometimes plausible arguments can be advanced that even help choose the variance parameter for such a vector [@gazeau]. However, it is important to understand that such a choice is [*not*]{} suitable in all cases. Let us reexamine the old discussion about [*“The rest of the universe”*]{} [@fey]. A single system seldom exists in isolation; instead, it is surrounded by other systems. If we can imagine one specific system, then it is possible to imagine $N$ independent, identical systems, and even infinitely many such systems, i.e., $N=\infty$. A system may involve several degrees of freedom; however, for clarity our basic system has a single degree of freedom.
For example, consider the Hamiltonian operator $\H_{(N)}\equiv\Sigma_{n=1}^N\,\H(P_n,Q_n)$. which represents $N$ independent, identical copies of the “original” Hamiltonian $\H(P_1,Q_1)$ involving independent operator pairs $(P_n,Q_n)$. The first system (for $P_1$ and $Q_1$) is chosen as the “physical” one, while the other sub-systems are “spectator” systems. The coherent states we choose are for the physical system only, that is only for the [*first*]{} operator pair; specifically |p,q;=e\^[-iqP\_1/]{}e\^[ipQ\_1/]{}|,3em |\_[n=1]{}\^N|\_n. To preserve the equivalence of all subsystems we choose each $|\eta_n\>$ to be identical to one another, i.e., the same single-system fiducial vector. In this case it follows that A\_[Q(R)]{}=\_0\^T\[p(t)[q]{}(t)-H\_[(N)]{}(p(t),q(t)\]dt, where, in the present case, H\_[(N)]{}(p,q)=H(p,q)+\_[n=2]{}\^NH(P\_n,Q\_n), \[e3\] which differs by a constant from the single system story. So long as $N<\infty$, that constant is finite and not important. But, as $N\ra\infty$, and we eventually deal with an infinite number of spectator systems, it becomes important that that constant must be zero. At this point in the argument we restrict attention to quantum systems that have a non-negative spectrum and a unique ground state, which we choose as $|\eta\>$, with an energy eigenvalue adjusted to be zero. If that is the case, then the added constant in (\[e3\]) vanishes for all $N$ including $N=\infty$. It may be argued that we could choose $|\eta\>$, say, as the first excited state for each subsystem and subtract that energy to obtain a zero. However, that would imply, for $N=\infty$, that there were infinitely many energy levels with $-\infty$ for their energy level, which is clearly an unphysical situation. The remedy for that situation is to insist that the fiducial vector be chosen as the unique ground state of each system with an energy level adjusted to vanish. This choice applies to the physical system, and to all of the spectator systems as well. Thus we can imagine any one of the $N$ identical systems being the physical one and the remaining $N-1$ systems as spectators.
Moreover, we can imagine there are many other spectator systems different from the physical one we have chosen. For example, suppose there is another multiple-system type, with a Hamiltonian $\tH_{(\widetilde{N})}=\Sigma_{\wn=1}^{\widetilde{N}}\tH(P_\wn,Q_\wn)$, that is also present. This operator, too, is assumed to have a unique ground state with zero energy eigenvalue. Thus this new Hamiltonian could be present in the overall Hamiltonian but it would contribute nothing to the restricted quantum action functional because coherent states for it have not been “turned on”. Indeed there could be many such systems, even infinitely many such new (sub)systems. This argument can be carried to yet new families of spectator systems all of which are there, just “resting”, or “hibernating”, in their own ground state, and contributing nothing to the restricted quantum action functional. In this fashion we have found how to include “the rest of the universe” in such a way that it makes no contribution whatsoever, just as if we had ignored it altogether at the beginning of the story.
This desirable property requires that we choose the fiducial vector as the [*unique ground state of the system under consideration adjusted to have zero energy*]{}, which for this esoteric exercise of dealing with the surroundings proved extremely convenient if not absolutely necessary. This choice of fiducial vector also eliminates any nonsense regarding intrusion of the micro world into the macro world as we argued above. Of course, concerns about the intrusion of the surroundings are not always necessary, and thus it is acceptable to consider other fiducial vectors that are “close” to the ground state in some unspecified way, if one so desires.
Finally, we need to comment on other model systems that have Hamiltonians, which (i), near a lower bound, have a continuous spectrum, or (ii) instead have a spectrum that is unbounded below. These are interesting mathematical models, but it it is difficult to find any real physical systems that have such features.
Having shown how we can, if necessary, deal with the rest of the universe, we revert to simple systems without concerning ourselves with such big issues. Thus, in what follows, we allow ourselves to consider a variety of useful fiducial vectors, particularly those where Gaussian form are not appropriate to describe certain physical systems, as in the case of fermions and anyons.
Bosons, Fermions, and Anyons
============================
In this section we confine ourselves to $2+1$ spacetime dimensions so that anyons can be treated as well as bosons and fermions, although the results and calculations for bosons and fermions are not limited to $2+1$ dimensions; generalizations to arbitrary spacetime dimensions for them are straightfoward.
Charged particles orbiting around a magnetic flux tube and interacting with it have fractional statistics in $2+1$ dimensions [@Wilczek]. Such composites, known as anyons [@Wilczek], are basically characterized by their peculiar statistics: under a half rotation centered on a magnetic flux, leading to a particle permutation, the state of the system changes by a complex phase factor,
$$\begin{aligned}
& \psi (\mathbf{r}_{2},\mathbf{r}_{1})=e^{i\alpha }\psi (\mathbf{r}_{1},%
\mathbf{r}_{2})\,,\,\,(0\leq \alpha <2\pi )\,, \label{any1} \\
& \mathbf{r}_{1}=(x_{1},y_{1})\,,\,\,\mathbf{r}_{2}=(x_{2},y_{2})\,, \notag\end{aligned}$$
rather than the standard bosons ($\alpha =0$) or fermions ($\alpha =\pi $) statistics. Here and throughout in this paper $\mathbf{r}_{\sigma }$ designates the two-dimensional cartesian coordinates of the first ($\sigma
=1 $) particle and the second ($\sigma =2$) particle, respectively.
As has been noticed originally by Wilczek [@Wilczek], this consideration has a very clear mathematical explanation based on the fact that in two space dimensions the rotation group $SO(2)$ is isomorphic with the Abelian unitary group $U(1)$ whose representations are labeled by real numbers. Due to the latter fact and the spin-statistics connection, the wave function of these particles may admit an arbitrary phase under rotations in $2+1$ spacetime. More precisely, circling a magnetic flux in the same direction, by two half turns, the state of the system may not necessarily need to return to its original state, meaning that now the system can be described by a multivalued wave function.
Anyons have attracted much attention due to their own richness, both in theoretical treatment of fundamental concepts, as well as in their physical implications. Particles with fractional statistics were considered in the $%
O(3)\,\sigma$ model due to the existence of solitons [@WilZee83], it also has been shown that anyons can be described as ordinary particles interacting with a Chern-Simons field [@IenLec90; @IenLec92]. A relativistic wave equation for anyons was formulated in [@JacNai91] and more recently in [@HorPlyVal10]. Regarding the physical implications, anyons have a central role in the explanation of the quantum Hall effect [@Laughlin88], in high-$T_{c}$ superconductivity [@FetHanLau89], and more recently in topological quantum computation [@Sarma05; @Collins06; @Stern08]. For a complete review underlying the fundamental theoretical descriptions and applications, we recommend the reviews [@Stern08; @IenLec92; @Wilczek90].
All this interest motivates us to study the subject. Particularly, we are interested in the enhanced classical theory of anyons in $2+1$ dimensions. In this respect it should be noted that classical theories for anyons have been considered before [@Pluyshchay90; @Ghosh9495]. In these papers the authors follow a canonical quantization procedure [dirac,CanQuant]{} to derive the corresponding quantum theory for anyons. Here, we adopt a different construction to quantize the theory of particles with fractional statistics by the application of the coherent state quantization and derive, for the first time, the corresponding enhanced classical theory. More precisely we consider a rotationally invariant Hamiltonian operator with a quartic interaction as a basic example.
Beneath the physical consideration behind anyons in $2+1$ dimensions, a coherent state representation [@Kla63I; @Kla63II] of the quantum theory for anyons must exhibit the same property as (\[any1\]). A direct consequence of such a requirement is the statistical invariance of the corresponding representation since now the fiducial vectors also obey,
$$\left\langle \mathbf{r}_{2},\mathbf{r}_{1}|\eta \right\rangle =\eta \left(
\mathbf{r}_{2},\mathbf{r}_{1}\right) =e^{i\alpha }\eta \left( \mathbf{r}_{1},%
\mathbf{r}_{2}\right) \,. \label{any2}$$
In choosing a coherent state representation we must, as a starting point, provide a suitable fiducial vector consistent with the property (\[any2\]). One possible way is to assume the fiducial vector is composed by two parts $\eta \left( \mathbf{r}_{1},\mathbf{r}_{2}\right) =A\left( \mathbf{r}%
_{1},\mathbf{r}_{2}\right) S\left( \mathbf{r}_{1},\mathbf{r}_{2}\right) $ where $S\left( \mathbf{r}_{1},\mathbf{r}_{2}\right) $ is a symmetrical function and $A\left( \mathbf{r}_{1},\mathbf{r}_{2}\right) $ has a nonsymmetrical form. In particular,$$A\left( \mathbf{r}_{2},\mathbf{r}_{1}\right) =\left( -1\right) ^{\gamma
}A\left( \mathbf{r}_{1},\mathbf{r}_{2}\right) \,,$$and the condition (\[any2\]) is recovered for $\gamma =\alpha /\pi $. This construction leads to a proper description of nonrelativistic fermions for $%
\gamma =1$, and it is generalized for anyons assuming $0<\gamma <2\,\,
(\gamma\neq 1) $.
The wave functions for two or more anyon systems have been discussed before. One of the first proposals was a multivalued function with a complex nonsymmetrical part and a Gaussian-like symetrical part [Halperin84,Wu84]{}. Generalizations of wave functions of this form have been considered in several works (e.g. [Chou91,MurLawBraBha91,IenLec92,Poly91]{}) and in particular represents a bound state for two-anyons system in a $\left( 2+1\right) $-dim. interaction to an external harmonic potential like $\left( 1/2\right) \Omega ^{2}\mathbf{r}%
_{i}^{2}$ [@Poly91]. Following these constructions, we study a family of fiducial vectors, parametrized by $\lambda $, suitable to discuss bosons, fermions, and anyons in the same framework,$$\begin{aligned}
&&\eta _{1}\left( \mathbf{r}_{1},\mathbf{r}_{2}\right) =N_{\gamma }\left(
z_{1}-z_{2}\right) ^{\gamma }e^{-\frac{\lambda }{2}\left( \mathbf{r}_{1}^{2}+%
\mathbf{r}_{2}^{2}\right) }\,,\ \ 0\leq \gamma <2\,,\ \ \gamma =\frac{\alpha
}{\pi }\,, \label{any3} \\
&&z_{1}=x_{1}+iy_{1}\,,\ \ z_{2}=x_{2}+iy_{2}\,,\ \ \mathbf{r}_{\sigma
}^{2}=x_{\sigma }^{2}+y_{\sigma }^{2}\,,\ \ \lambda =\frac{\Omega }{\hbar }%
\,,\ \ \Omega =\mathrm{const.}\,, \notag\end{aligned}$$where $z_{1}$,$\ z_{2}$ are position of particles in the complex notation, $%
\Omega $ is an arbitrary real constant and $N_{\gamma }$ are the corresponding normalization constants. It is clear that interchanging $%
z_{1}$ and $z_{2}$ is equivalent to changing the corresponding positions of the particles, and, as a result, one obtains the desired phase $\eta
\left( \mathbf{r}_{2},\mathbf{r}_{1}\right) =\left( -1\right) ^{\gamma }\eta
\left( \mathbf{r}_{1},\mathbf{r}_{2}\right) $. We initially study bosons and fermions in the next subsection. Exact solutions and results for anyons are presented afterwards.
General calculations for bosons and fermions
--------------------------------------------
In this subsection we examinate, in detail, the choice of fiducial vectors (\[any3\]) for the particular cases of bosons $\left( \gamma =0\right) $ and fermions $\left( \gamma =1\right) $. More precisely, we are interested to obtain the enhanced classical description for these particles with bosons and fermions under an influence of a quartic interaction as (\[11K\]). To do it one has to evaluate expectation values of the self-adjoint operators $Q$, $P$, $Q^{2}$, $P^{2}$ and $Q^{4}$, and the subsequent subsections are reserved to present that. Detailed derivations are placed in the Appendix.
### Bosons $\left( \protect\gamma =0\right) $
From (\[any3\]) the fiducial vectors for bosons are pure Gaussian functions,$$\eta _{1}\left( \mathbf{r}_{1},\mathbf{r}_{2}\right) =N_{0}e^{-\frac{\lambda
}{2}\left( \mathbf{r}_{1}^{2}+\mathbf{r}_{2}^{2}\right) }\,,\ \ N_{0}=\frac{%
\lambda }{\pi }\,. \label{any4}$$
One of the useful conditions underlying usual coherent state representations lies in the fact that expectation values for the position and momentum operators are zero. Due to parity properties one can see that $\left\langle
Q_{x_{i}}\right\rangle =0$ for the functions (\[any4\]). Moreover it is straightfoward to see that $\left\langle Q_{x_{i}}^{2n+1}\right\rangle =0$ for $n\in \mathbb{N}$. The expectation value of the momentum operator is equivalently zero since the function (\[any4\]) is symmetric in these variables. Therefore,$$\left\langle Q_{x_{i}}\right\rangle =0=\left\langle P_{x_{i}}\right\rangle
\,. \label{any6}$$Intrinsic to the calculation of the enhanced Hamiltonian are the expectation values of $Q_{x_{i}}^{2}$, $P_{x_{i}}^{2}$ and $Q_{x_{i}}^{4}$. For example, the expectation value of $Q_{x_{1}}^{2}$ is,$$\left\langle Q_{x_{1}}^{2}\right\rangle =N_{0}^{2}\left( \frac{\pi }{\lambda
}\right) \int_{0}^{2\pi }d\vartheta \cos ^{2}\vartheta \int_{0}^{\infty
}dr_{1}r_{1}^{3}e^{-\lambda r_{1}^{2}}=\frac{1}{2\lambda }=\frac{\hbar }{%
2\Omega }\,, \label{any7.1}$$where $\vartheta $ is the polar angle. \[Remark: Here and in what follows we adopt the convention that one integration symbol over a bold letter means a double integration over two coordinates, for example, $\int_{c}^{c^{\prime }}d\mathbf{r}f\left( x,y\right)\equiv\int_{c}^{c^{\prime }}dx\int_{c}^{c^{\prime }}dyf\left( x,y\right) $\]. In addition the expectation value for $P_{x_{1}}^{2}$ is $$\left\langle P_{x_{1}}^{2}\right\rangle =-\hbar ^{2}\lambda
N_{0}^{2}\int_{-\infty }^{\infty }d\mathbf{r}_{1}\int_{-\infty }^{\infty }d%
\mathbf{r}_{2}\left( \lambda x_{1}^{2}-1\right) e^{-\lambda \left( \mathbf{r}%
_{1}^{2}+\mathbf{r}_{2}^{2}\right) }=\frac{\hbar ^{2}\lambda }{2}=\frac{%
\hbar \Omega }{2}\,, \label{any7.2}$$and the remaining expectation values have the same result.
In order to discuss the enhanced classical theory for bosons with a quartic interaction one has to evaluate several expectation values like $%
\left\langle Q_{x_{\sigma }}^{4}\right\rangle $, $\langle Q_{x_{\sigma
}}^{2}Q_{x_{\sigma ^{\prime }}}^{2}\rangle $ and $\langle
Q_{x_{\sigma }}^{2}Q_{y_{\sigma ^{\prime }}}^{2}\rangle $. However, due to the symmetry of the Gaussian fiducial vectors (\[any4\]), only $%
\left\langle Q_{x_{1}}^{4}\right\rangle $ is independent,$$\left\langle Q_{x_{1}}^{4}\right\rangle =N_{0}^{2}\left( \frac{\pi }{\lambda
}\right) \int_{0}^{2\pi }d\vartheta \cos ^{4}\vartheta \int_{0}^{\infty
}dr_{1}r_{1}^{5}e^{-\lambda r_{1}^{2}}=\frac{3}{4\lambda ^{2}}=\frac{3\hbar
^{2}}{4\Omega ^{2}}\,. \label{any7.3}$$To show that consider, for example, $\left\langle
Q_{x_{1}}^{2}Q_{x_{2}}^{2}\right\rangle $ (which is equivalent to $%
\left\langle Q_{x_{1}}^{2}Q_{y_{1}}^{2}\right\rangle $ in the present case). From its form it is straightfoward to conclude that$$\left\langle Q_{x_{1}}^{2}Q_{x_{2}}^{2}\right\rangle =N_{0}^{2}\left( \frac{%
\pi }{\lambda }\right) \int_{-\infty }^{\infty }dx_{1}x_{1}^{2}e^{-\lambda
x_{1}^{2}}\int_{-\infty }^{\infty }dx_{2}x_{2}^{2}e^{-\lambda x_{2}^{2}}=%
\frac{\left\langle Q_{x_{1}}^{4}\right\rangle }{3}\,. \label{any7.4}$$
### Fermions $\left( \protect\gamma =1\right)$
For fermions the fiducial vectors (\[any3\]) take the form$$\eta _{1}\left( \mathbf{r}_{1},\mathbf{r}_{2}\right) =N_{1}\left(
z_{1}-z_{2}\right) e^{-\frac{\lambda }{2}\left( \mathbf{r}_{1}^{2}+\mathbf{r}%
_{2}^{2}\right) }\,, \label{fer1.1}$$and the corresponding normalization constant is obtained as usual,$$\left\vert N_{1}\right\vert =\left( \int_{-\infty }^{\infty }d\mathbf{r}%
_{1}\int_{-\infty }^{\infty }d\mathbf{r}_{2}\left\vert \mathbf{r}_{1}-%
\mathbf{r}_{2}\right\vert ^{2}e^{-\lambda \left( \mathbf{r}_{1}^{2}+\mathbf{r%
}_{2}^{2}\right) }\right) ^{-\frac{1}{2}}=\sqrt{\frac{\lambda ^{3}}{2\pi ^{2}%
}}\,, \label{fer2.1}$$The expectation value of the coordinate operators, $Q_{x_{1}},...$ etc, are zero for these states. For example, $\left\langle Q_{x_{1}}\right\rangle $ has the form$$\left\langle Q_{x_{1}}\right\rangle =N_{1}^{2}\int_{0}^{2\pi }d\vartheta
\cos \vartheta \int_{0}^{\infty }dr_{1}r_{1}^{2}e^{-\lambda
r_{1}^{2}}\int_{-\infty }^{\infty }d\mathbf{r}_{2}\left\vert \mathbf{r}_{1}-%
\mathbf{r}_{2}\right\vert ^{2}e^{-\lambda \mathbf{r}_{2}^{2}}=0\,,
\label{fer3}$$In addition the expectation values for the momentum operator $\left\langle
P_{x_{1}}\right\rangle $ reads,$$\left\langle P_{x_{1}}\right\rangle =-i\hbar N_{1}^{2}\int_{-\infty
}^{\infty }d\mathbf{r}_{1}\int_{-\infty }^{\infty }d\mathbf{r}_{2}\left\{
\left( z_{1}^{\ast }-z_{2}^{\ast }\right) -\lambda x_{1}\left\vert \mathbf{r}%
_{1}-\mathbf{r}_{2}\right\vert ^{2}\right\} e^{-\lambda \left( \mathbf{r}%
_{1}^{2}+\mathbf{r}_{2}^{2}\right) }=0\,, \label{fer4}$$and the remaining linear expectation values are also zero. The expectation value for $Q_{x_{1}}^{2}$ are$$\left\langle Q_{x_{1}}^{2}\right\rangle =N_{1}^{2}\int_{-\infty }^{\infty }d%
\mathbf{r}_{1}\int_{-\infty }^{\infty }d\mathbf{r}_{2}\left( \mathbf{r}%
_{1}^{2}+\mathbf{r}_{2}^{2}\right) x_{1}^{2}e^{-\lambda \left( \mathbf{r}%
_{1}^{2}+\mathbf{r}_{2}^{2}\right) }=\frac{3}{4\lambda }=\frac{3\hbar }{%
4\Omega }\,, \label{fer5.0}$$and for $P_{x_{1}}^{2}$ are$$\begin{aligned}
\left\langle P_{x_{1}}^{2}\right\rangle &=&-\hbar ^{2}\lambda
N_{1}^{2}\int_{-\infty }^{\infty }d\mathbf{r}_{1}\int_{-\infty }^{\infty }d%
\mathbf{r}_{2}\left\{ \left\vert \mathbf{r}_{1}-\mathbf{r}_{2}\right\vert
^{2}\left( \lambda x_{1}^{2}-1\right) -x_{1}^{2}\right\} e^{-\lambda \left(
\mathbf{r}_{1}^{2}+\mathbf{r}_{2}^{2}\right) } \label{fer6.0} \\
&=&\frac{3\hbar ^{2}\lambda }{4}=\frac{3\hbar \Omega }{4}\,, \notag\end{aligned}$$
In order to discuss rotationally invariant quartic interactions we have to evaluate, for example, expectation values as $\left\langle Q_{x_{\sigma
}}^{4}\right\rangle $, $\langle Q_{x_{\sigma }}^{2}Q_{x_{\sigma
^{\prime }}}^{2}\rangle $ and $\langle Q_{x_{\sigma
}}^{2}Q_{y_{\sigma ^{\prime }}}^{2}\rangle $. The first one has the form$$\left\langle Q_{x_{1}}^{4}\right\rangle =N_{1}^{2}\int_{-\infty }^{\infty }d%
\mathbf{r}_{1}\int_{-\infty }^{\infty }d\mathbf{r}_{2}\left\vert \mathbf{r}%
_{1}-\mathbf{r}_{2}\right\vert ^{2}x_{1}^{4}e^{-\lambda \left( \mathbf{r}%
_{1}^{2}+\mathbf{r}_{2}^{2}\right) }=\frac{3}{2\lambda ^{2}}=\frac{3\hbar
^{2}}{2\Omega ^{2}}\,, \label{fer7}$$and the same result can be obtained for $\left\langle
Q_{y_{1}}^{4}\right\rangle $ ... etc. Next it can easely seen that expectation values like $\langle Q_{x_{\sigma }}^{2}Q_{x_{\sigma
^{\prime }}}^{2}\rangle $ and $\langle Q_{x_{\sigma
}}^{2}Q_{y_{\sigma ^{\prime }}}^{2}\rangle $ enjoy the same property as (\[any7.4\]). At last, there exists one more interesting expectation value $\left\langle Q_{x_{1}}Q_{x_{2}}\right\rangle $ which vanishes for bosons but is nonzero for fermions,$$\left\langle Q_{x_{1}}Q_{x_{2}}\right\rangle =-2N_{1}^{2}\int_{-\infty
}^{\infty }d\mathbf{r}_{1}\int_{-\infty }^{\infty }d\mathbf{r}%
_{2}x_{1}x_{2}\left( \mathbf{r}_{1}\cdot \mathbf{r}_{2}\right) e^{-\lambda
\left( \mathbf{r}_{1}^{2}+\mathbf{r}_{2}^{2}\right) }=-\frac{1}{4\lambda }=-%
\frac{\hbar }{4\Omega }\,. \label{fer8}$$which again, due to the symmetry of the problem, this result coincides with $%
\left\langle Q_{y_{1}}Q_{y_{2}}\right\rangle $.
General calculations for anyons
-------------------------------
In this section we generalize the previous results for arbitrary $\gamma $ within the range $\left[ 0,2\right) $, with the $\left( 0,1\right) \cup
(1,2) $ used for anyons. More precisely, the choice of fiducial vector ([any3]{}) admits an exact solution for arbitrary $\gamma $ and the results here are an extension to those particular cases with $\gamma =0$ and $1$. The first step is to evaluate the norm of the corresponding fiducial vectors,$$\left\Vert \left\vert \eta \right\rangle \right\Vert ^{2}=\left\vert
N_{\gamma }\right\vert ^{2}\int_{-\infty }^{\infty }d\mathbf{r}%
_{1}\int_{-\infty }^{\infty }d\mathbf{r}_{2}\left\vert \mathbf{r}_{1}-%
\mathbf{r}_{2}\right\vert ^{2\gamma }e^{-\lambda \left( \mathbf{r}_{1}^{2}+%
\mathbf{r}_{2}^{2}\right) }\,. \label{any8}$$where again the normalization constant is obtained as usual$$\left\vert N_{\gamma }\right\vert =\left( \int_{-\infty }^{\infty }d\mathbf{r%
}_{1}e^{-\lambda \mathbf{r}_{1}^{2}}\int_{-\infty }^{\infty }d\mathbf{r}%
_{2}\left\vert \mathbf{r}_{1}-\mathbf{r}_{2}\right\vert ^{2\gamma
}e^{-\lambda \mathbf{r}_{2}^{2}}\right) ^{-\frac{1}{2}}\,, \label{any8.1}$$This rotationally-invariant integral can simplified by choosing a convenient coordinate system. For example, in the case where $\mathbf{r}_{2}$ is aligned along the axis $y_{1}$ (\[any8.1\]) has the form,$$\begin{aligned}
&&\left\vert N_{\gamma }\right\vert ^{2}=\left( \int_{-\infty }^{\infty }d%
\mathbf{r}_{2}e^{-\lambda \mathbf{r}_{2}^{2}}\rho _{\gamma }^{\left(
-\right) }\left( r_{2}\right) \right) ^{-1}\,,\ \ \rho _{\gamma }^{\left(
-\right) }\left( r_{2}\right) \equiv \int_{0}^{\infty
}dr_{1}r_{1}e^{-\lambda r_{1}^{2}}\mathcal{I}_{\vartheta }^{\left( -\right)
}\left( r_{1},r_{2};\gamma \right) \,, \notag \\
&&\mathcal{I}_{\vartheta }^{\left( -\right) }\left( r_{1},r_{2};\gamma
\right) \equiv \int_{0}^{2\pi }d\vartheta \left(
r_{1}^{2}+r_{2}^{2}-2r_{1}r_{2}\sin \vartheta \right) ^{\gamma }\,,
\label{any9}\end{aligned}$$where $\vartheta $ is the angle between the axis $x_{1}$ and $\mathbf{r}_{1}$ such that $\mathbf{r}_{1}\cdot \mathbf{r}_{2}=r_{1}r_{2}\cos \left( \pi
/2-\vartheta \right) =r_{1}r_{2}\sin \vartheta $. Detailed solution of $\rho _{\gamma }^{\left( -\right)
}\left( r_{2}\right) $ and subsequent calculations can be found at the Appendix (\[ap1\])-(\[ap5.1\]). Using those results the norm (\[any8\]) has the final form$$\left\Vert \left\vert \eta \right\rangle \right\Vert ^{2}=\left\vert
N_{\gamma }\right\vert ^{2}\frac{\pi ^{2}\sqrt{\pi }}{2\lambda ^{\gamma +2}}%
\frac{\Gamma \left( 2+\gamma \right) }{\Gamma \left( \frac{3}{2}\right) }%
\left. _{2}F_{1}\left( \frac{1-\gamma }{2},-\frac{\gamma }{2};\frac{3}{2}%
;1\right) \right. \,, \label{any14}$$from which follows that$$\left\vert N_{\gamma }\right\vert =\left[ \frac{\pi ^{2}\sqrt{\pi }}{%
2\lambda ^{\gamma +2}}\frac{\Gamma \left( 2+\gamma \right) }{\Gamma \left(
\frac{3}{2}\right) }\left. _{2}F_{1}\left( \frac{1-\gamma }{2},-\frac{\gamma
}{2};\frac{3}{2};1\right) \right. \right] ^{-1/2}\,, \label{any15}$$ where $\left. _{2}F_{1}\left(\alpha ,\beta ;\delta ;z\right) \right. $ is the hypergeometric function [@Gradshtein]. The expectation value $Q_{x_{1}}^{2}$ is conveniently evaluated in the same reference system, whose form is$$\left\langle Q_{x_{1}}^{2}\right\rangle =\left\vert N_{\gamma }\right\vert
^{2}\int_{-\infty }^{\infty }d\mathbf{r}_{1}e^{-\lambda \mathbf{r}%
_{1}^{2}}x_{1}^{2}\rho _{\gamma }^{\left( -\right) }\left( r_{1}\right) \,,
\label{any15.1}$$where $\rho _{\gamma }^{\left( -\right) }\left( r_{1}\right) $ is defined in (\[ap1\]). Using its solution (\[ap4\]) this expectation value can be written as[^3],$$\left\langle Q_{x_{1}}^{2}\right\rangle =\frac{\pi ^{2}}{2\lambda ^{2}}%
\left\vert N_{\gamma }\right\vert ^{2}\sum_{s=0}^{\infty }\Lambda
_{s}^{\left( -\right) }R_{1}^{\left( -\right) }\left( \gamma ,s\right)
=\left( \frac{\pi ^{2}\sqrt{\pi }}{8\lambda ^{3+\gamma }}\right) \left\vert
N_{\gamma }\right\vert ^{2}\frac{\Gamma \left( 3+\gamma \right) }{\Gamma
\left( \frac{3}{2}\right) }\left. _{2}F_{1}\left( \frac{1-\gamma }{2},-\frac{%
\gamma }{2};\frac{3}{2};1\right) \right. \,, \label{any16}$$and substituting (\[any15\]) we get a remarkably simple result:$$\left\langle Q_{x_{1}}^{2}\right\rangle =\frac{1}{4\lambda }\frac{\Gamma
\left( 3+\gamma \right) }{\Gamma \left( 2+\gamma \right) }=\frac{\left(
\gamma +2\right) }{4\lambda }=\frac{\hbar \left( \gamma +2\right) }{4\Omega }%
\,. \label{any17}$$
By the same symmetry arguments, which has been discussed in the section on bosons and fermions, one can see that $\left\langle Q_{x_{1}}\right\rangle =0$ (again, the integral with respect to $x_{1}$ has an odd integrand within a symmetric interval). This is also true for the remaining coordinates expectation values. By virtue of this fact it follows that the expectation values of the momentum operators are also zero, for example,$$\begin{aligned}
\left\langle P_{x_{1}}\right\rangle &=&\left\vert N_{\gamma }\right\vert
^{2}\int_{-\infty }^{\infty }d\mathbf{r}_{1}\int_{-\infty }^{\infty }d%
\mathbf{r}_{2}\left\{ \left( z_{1}^{\ast }-z_{2}^{\ast }\right) \left[
\gamma -\lambda x_{1}\left( z_{1}-z_{2}\right) \right] \right. \notag \\
&\times &\left. \left( \mathbf{r}_{1}^{2}+\mathbf{r}_{2}^{2}-2\mathbf{r}%
_{1}\cdot \mathbf{r}_{2}\right) ^{\gamma -1}e^{-\lambda \left( \mathbf{r}%
_{2}^{2}+\mathbf{r}_{1}^{2}\right) }\right\} =0\,, \label{any18}\end{aligned}$$since it only has odd contributions.
The full expression of $\left\langle P_{x_{1}}^{2}\right\rangle $ is,$$\begin{aligned}
\left\langle P_{x_{1}}^{2}\right\rangle &=&-\hbar ^{2}\left\vert N_{\gamma
}\right\vert ^{2}\int_{-\infty }^{\infty }d\mathbf{r}_{1}\int_{-\infty
}^{\infty }d\mathbf{r}_{2}e^{-\lambda \left( \mathbf{r}_{2}^{2}+\mathbf{r}%
_{1}^{2}\right) } \{\left\vert \mathbf{r}_{1}-\mathbf{r}_{2}\right\vert
^{2\left( \gamma -2\right) } \\
&\times &\left[ \gamma \left( \gamma -1\right) \left( z_{1}^{\ast
}-z_{2}^{\ast }\right) ^{2}-2\lambda \gamma x_{1}\left( z_{1}^{\ast
}-z_{2}^{\ast }\right) \left\vert \mathbf{r}_{1}-\mathbf{r}_{2}\right\vert
^{2}\right. \\
&+&\left. \left. \lambda \left( \lambda x_{1}^{2}-1\right) \left\vert
\mathbf{r}_{1}-\mathbf{r}_{2}\right\vert ^{4}\right] \right\} \,.\end{aligned}$$Due to the appearance of odd coefficients some terms in this integral vanishe. Using previous results (\[any14\]), (\[any16\]) and the definitions (\[ap1\]), (\[ap3\]), (\[ap5.1\]), this expectation value can be written as (replacing $\lambda $ by $\Omega /\hbar $),$$\begin{aligned}
\left\langle P_{x_{1}}^{2}\right\rangle &=&-\hbar ^{2}\left\vert N_{\gamma
}\right\vert ^{2}\sum_{s=0}^{\infty }\Lambda _{s}^{\left( -\right)
}R_{1}^{\left( -\right) }\left( \gamma -1,s\right) +\hbar ^{2}\lambda \left(
1-\lambda \left\langle Q_{x_{1}}^{2}\right\rangle \right) \notag \\
&=&\frac{\Omega \hbar }{2}\left[ 1+\gamma \left( \frac{\left.
_{2}F_{1}\left( \frac{2-\gamma }{2},\frac{1-\gamma }{2};\frac{3}{2};1\right)
\right. }{\left. _{2}F_{1}\left( \frac{1-\gamma }{2},-\frac{\gamma }{2};%
\frac{3}{2};1\right) \right. }\right) -\frac{\gamma }{2}\right] \,.
\label{any21}\end{aligned}$$It should be noted that both results (\[any17\]) and (\[any21\]) coincide with the particular considerations (\[any7.1\]), (\[any7.2\]) and (\[fer5.0\]), (\[fer6.0\]), as expected.
Among all possible rotationally-invariant potentials within the problem under consideration are quartic interactions. There is a particular interest in rotational invariant models, which justifies our study. We consider the rotationally invariant potential $V$ given by,$$V\equiv \left( \mathbf{Q}_{1}^{2}+\mathbf{Q}_{2}^{2}\right) ^{2}\,,\ \
\mathbf{Q}_{1}^{2}\equiv Q_{x_{1}}^{2}+Q_{y_{1}}^{2}\,,\ \ \mathbf{Q}%
_{2}^{2}\equiv Q_{x_{2}}^{2}+Q_{y_{2}}^{2}\,, \label{any22.1}$$whose expectation value can separated into a homogenous part $V_{\mathrm{H}}$ and in a nonhomogeneous part $V_{\mathrm{N}}$,$$\left\langle V\right\rangle =\left\langle V_{\mathrm{H}}\right\rangle
+\left\langle V_{\mathrm{N}}\right\rangle \,,\ \ \left\langle V_{\mathrm{H}%
}\right\rangle \equiv\left\langle \mathbf{Q}_{1}^{4}+\mathbf{Q}%
_{2}^{4}\right\rangle \,,\ \ \left\langle V_{\mathrm{N}}\right\rangle
\equiv 2\left\langle \mathbf{Q}_{1}^{2}\mathbf{Q}_{2}^{2}\right\rangle \,.
\label{any22.2}$$From its form it can be seen that the corresponding expectation value contains terms like $\left\langle Q_{x_{1}}^{4}\right\rangle $ (the same for all other coordinates operators), as well as $\left\langle
Q_{x_{1}}^{2}Q_{y_{1}}^{2}\right\rangle $, and correlation functions such as $%
\left\langle Q_{x_{1}}^{2}Q_{x_{2}}^{2}\right\rangle $. For example, the expectation value of $Q_{x_{1}}^{4}$ is$$\left\langle Q_{x_{1}}^{4}\right\rangle =\left\vert N_{\gamma }\right\vert
^{2}\int_{-\infty }^{\infty }d\mathbf{r}_{1}e^{-\lambda \mathbf{r}%
_{1}^{2}}x_{1}^{4}\rho _{\gamma }^{\left( -\right) }\left( r_{1}\right) =%
\frac{3\pi }{8\lambda ^{3}}\left\vert N_{\gamma }\right\vert
^{2}\sum_{s=0}^{\infty }\Lambda _{s}^{\left( -\right) }R_{2}^{\left(
-\right) }\left( \gamma ,s\right) \,, \label{any23}$$which, after a few simplifications, can be split into two parts$$\begin{aligned}
&&\left\langle Q_{x_{1}}^{4}\right\rangle =\left\vert N_{\gamma }^{\left(
1\right) }\right\vert ^{2}\left( \frac{3\pi ^{2}\sqrt{\pi }}{64\lambda
^{\gamma +4}}\right) \frac{\Gamma \left( \gamma +4\right) }{\Gamma \left(
5/2\right) }\left( \mathcal{F}_{1}+\mathcal{F}_{2}\right) \,, \\
&&\mathcal{F}_{1}=\sum_{s=1}^{\infty }\frac{1}{\left( s-1\right) !}\left(
\frac{1-\gamma }{2}\right) _{s}\left( -\frac{\gamma }{2}\right) _{s}\left(
\frac{5}{2}\right) _{s}^{-1}\,,\ \ \mathcal{F}_{2}=2\left. _{2}F_{1}\left(
\frac{1-\gamma }{2},-\frac{\gamma }{2};\frac{5}{2};1\right) \right. \,.\end{aligned}$$Using the standard relation $\left( a\right) _{s+1}=a\left( a+1\right) _{s}$ the first series $\mathcal{F}_{1}$ can be simplified and be identified as a new hypergeometric function,$$\mathcal{F}_{1}=\frac{\gamma \left( \gamma -1\right) }{10}\left.
_{2}F_{1}\left( \frac{3-\gamma }{2},\frac{2-\gamma }{2};\frac{7}{2};1\right)
\right. \,,$$and consequently the expectation value of $\left\langle
Q_{x_{1}}^{4}\right\rangle $ has its final form (replacing $\lambda $ by $%
\Omega /\hbar $)$$\begin{aligned}
\left\langle Q_{x_{1}}^{4}\right\rangle &=&\frac{\hbar ^{2}\left( \gamma
+3\right) \left( \gamma +2\right) }{16\Omega ^{2}\left. _{2}F_{1}\left(
\frac{1-\gamma }{2},-\frac{\gamma }{2};\frac{3}{2};1\right) \right. }\left[
2\left. _{2}F_{1}\left( \frac{1-\gamma }{2},-\frac{\gamma }{2};\frac{5}{2}%
;1\right) \right. \right. \notag \\
&+&\left. \frac{\gamma \left( \gamma -1\right) }{10}\left. _{2}F_{1}\left(
\frac{3-\gamma }{2},\frac{2-\gamma }{2};\frac{7}{2};1\right) \right. \right]
\,. \label{any24}\end{aligned}$$Although (\[any24\]) has an apparently complicated form, one can see that it reduces to (\[any7.3\]) and (\[fer5.0\]) when $\gamma =0$, $\gamma =1$, respectively. Since $\left. _{2}F_{1}\left( a,b;c;1\right) \right. =1$ for $a=0$ or $b=0$, one can readily check that$$\left. \left\langle Q_{x_{1}}^{4}\right\rangle \right\vert _{\gamma =0}=%
\frac{12\hbar ^{2}}{16\Omega ^{2}}=\frac{3\hbar ^{2}}{4\Omega ^{2}}\,,\ \
\left. \left\langle Q_{x_{1}}^{4}\right\rangle \right\vert _{\gamma =1}=%
\frac{24\hbar ^{2}}{16\Omega ^{2}}=\frac{3\hbar ^{2}}{2\Omega ^{2}}\,,$$as expected.
The second kind of interaction from (\[any22.1\]) has the form $%
\left\langle Q_{x_{1}}^{2}Q_{y_{1}}^{2}\right\rangle $. It can be seen that for this interaction$$\left\langle Q_{x_{1}}^{2}Q_{y_{1}}^{2}\right\rangle =\left\vert N_{\gamma
}\right\vert ^{2}\int_{-\infty }^{\infty }d\mathbf{r}_{1}e^{-\lambda \mathbf{%
r}_{1}^{2}}x_{1}^{2}y_{1}^{2}\rho _{\gamma }^{\left( -\right) }\left(
r_{1}\right) \,, \label{any25.1}$$differs from (\[any23\]) only in the first integral which, after changing to polar coordinates, one obtains the simple relation$$\left\langle Q_{x_{1}}^{2}Q_{y_{1}}^{2}\right\rangle =\frac{\left\langle
Q_{x_{1}}^{4}\right\rangle }{3}\,. \label{any25.2}$$It is worth noting that this result obeys the same standard property of pure Gaussian states, i. e., (\[any25.2\]) is true for Gaussian states and here we see that this same property holds true. The reason behind it is due to the fact that the non-Gaussian part of these expectation values is a rotational invariant function which, as a matter of fact, does not spoil such a property. More specifically, the property$$\int_{-\infty }^{\infty }dx\int_{-\infty }^{\infty }dy\,e^{-\left(
x^{2}+y^{2}\right) }x^{2}y^{2}f\left( r\right) =\frac{1}{3}\int_{-\infty
}^{\infty }dx\int_{-\infty }^{\infty }dy\,e^{-\left( x^{2}+y^{2}\right)
}x^{4}f\left( r\right) \,,$$is true for any rotational invarian function $f\left( r\right) $.
At last, the remaining interactions are mixed and nonhomogeneous like $%
\left\langle Q_{x_{1}}^{2}Q_{x_{2}}^{2}\right\rangle $. These are slightly different from the previous cases (\[any23\]), (\[any25.1\]) as one can see from its form,$$\begin{aligned}
&&\left\langle Q_{x_{1}}^{2}Q_{x_{2}}^{2}\right\rangle =\left\vert N_{\gamma
}\right\vert ^{2}\int_{-\infty }^{\infty }d\mathbf{r}_{1}x_{1}^{2}e^{-%
\lambda \mathbf{r}_{1}^{2}}\rho _{\gamma }^{\left( +\right) }\left(
r_{1}\right) \,,\ \ \rho _{\gamma }^{\left( +\right) }\left( u_{1}\right)
=\int_{0}^{\infty }dr_{2}r_{2}^{3}e^{-\lambda r_{2}^{2}}\mathcal{I}%
_{\vartheta }^{\left( +\right) }\left( r_{1},r_{2};\gamma \right) \,, \notag
\\
&&\mathcal{I}_{\vartheta }^{\left( +\right) }\left( r_{1},r_{2};\gamma
\right) =\int_{0}^{2\pi }d\vartheta \cos ^{2}\vartheta \left(
r_{1}^{2}+r_{2}^{2}-2r_{1}r_{2}\sin \vartheta \right) ^{\gamma }\,,
\label{any26.1}\end{aligned}$$and the solution of $\mathcal{I}_{\vartheta }^{\left( +\right) }\left(
r_{1},r_{2};\gamma \right) $ can be found at the appendix (\[ap2\]). Using the series representation of the hypergeometric function as (\[ap2.1\]) the integral $\rho _{\gamma }^{\left( +\right) }\left( u_{1}\right) $ can be written as (\[ap3\])$$\begin{aligned}
\rho _{\gamma }^{\left( +\right) }\left( u_{1}\right) &=&\sum_{s=0}^{\infty
}\Lambda _{s}^{\left( +\right) }\varrho _{\gamma ,s}^{\left( +\right)
}\left( u_{1}\right) u_{1}^{s}\,,\ \ \Lambda _{s}^{\left( +\right) }=\frac{%
\pi }{\lambda ^{\gamma +2}}\frac{4^{s}}{s!}\left( \frac{1-\gamma }{2}\right)
_{s}\left( -\frac{\gamma }{2}\right) _{s}\left( 2\right) _{s}^{-1} \notag \\
\varrho _{\gamma ,s}^{\left( +\right) }\left( u_{1}\right) &=&\Gamma \left(
s+2\right) u_{1}^{\frac{\gamma -s+1}{2}}e^{\frac{u_{1}}{2}}W_{\left( \gamma
-3s-1\right) /2,\left( s-\gamma -2\right) /2}\left( u_{1}\right) \,,
\label{any26.2}\end{aligned}$$where the formula (\[ap4\]) has been used above. After some minor algebraic manipulations and using (\[ap5.1\]) we get the final result (replacing $\lambda $ by $\Omega /\hbar $),$$\left\langle Q_{x_{1}}^{2}Q_{x_{2}}^{2}\right\rangle =\left( \frac{\hbar
^{2}\left( \gamma +3\right) \left( \gamma +2\right) }{24\Omega ^{2}}\right)\frac{
\left. _{2}F_{1}\left( \frac{1-\gamma }{2},-\frac{\gamma }{2};\frac{5%
}{2};1\right) \right.}{ \left. _{2}F_{1}\left( \frac{1-\gamma }{2},-%
\frac{\gamma }{2};\frac{3}{2};1\right) \right.} \,. \label{any26.3}$$
Finally, taking into account the symmetries of the several homogeneous and nonhomogeneous expectation values under the changes $x_{1}\leftrightarrow
y_{1}$, $x_{2}\leftrightarrow y_{2}$, $x_{1}\leftrightarrow x_{2}$, $%
y_{1}\leftrightarrow y_{2}$ the full expectation value of the potential ([any22.1]{}) has the form (replacing $\lambda $ by $\Omega /\hbar $),$$\begin{aligned}
\left\langle V\right\rangle &=&\frac{16}{3}\left\langle
Q_{x_{1}}^{4}\right\rangle +4\left\langle
Q_{x_{1}}^{2}Q_{x_{2}}^{2}\right\rangle =\frac{\hbar ^{2}\left( \gamma
+3\right) \left( \gamma +2\right) }{6\Omega ^{2}}\left( \left.
_{2}F_{1}\left( \frac{1-\gamma }{2},-\frac{\gamma }{2};\frac{3}{2};1\right)
\right. \right) ^{-1} \notag \\
&\times &\left\{ 5\left. _{2}F_{1}\left( \frac{1-\gamma }{2},-\frac{\gamma }{%
2};\frac{5}{2};1\right) \right. +\frac{\gamma \left( \gamma -1\right) }{5}%
\left. _{2}F_{1}\left( \frac{3-\gamma }{2},\frac{2-\gamma }{2};\frac{7}{2}%
;1\right) \right. \right\} \,. \label{any27}\end{aligned}$$
Selected Hamiltonian operators
==============================
Traditionally the Hamiltonian operator for particles with fractional statistics is represented by the standard nonrelativistic Hamiltonian for a charged particle interacting with a magnetic flux tube through the minimum coupling with a vector potential [@Wilczek; @Wu84]. The latter potential may conveniently be removed by a gauge transformation which, in particular, gives the multivalued character to the corresponding wave function [Halperin84,Wu84]{}[^4]. Once we are working with a multivalued wave function (more preciselly with multivalued coherent states), the nonrelativistic momentum operator does have such potential vector. Moreover we add to the previous standard descriptions [@Wu84] a quartic potential of the form (\[any22.1\]). For us this potential has a great importance due to its close relation with rotational invariant models.
The nonrelativistic Hamiltonian operator in consideration has the form$$\begin{aligned}
&&\mathcal{H}\left( \mathbf{P},\mathbf{Q}\right) =\sum_{\sigma =1}^{2}\left[
\frac{\mathbf{P}_{\sigma }^{2}}{2m}+\frac{m\varpi ^{2}}{2}\mathbf{Q}_{\sigma
}^{2}\right] +gV\,,\ \ V=\left( \mathbf{Q}_{1}^{2}+\mathbf{Q}_{2}^{2}\right)
^{2}\,, \label{cl1} \\
&&\mathbf{P}_{\sigma }=\left( P_{x_{\sigma }},P_{y_{\sigma }}\right) \,,\ \
\mathbf{Q}_{\sigma }=\left( Q_{x_{\sigma }},Q_{y_{\sigma }}\right) \,,
\notag\end{aligned}$$where $m$ represent the mass of the particles, $\varpi $ is the harmonic potential frequency, $g$ is the quartic interaction coupling constant. In all cases the enhanced classical hamiltonian is identified with the expectation value of the Hamiltonian operator (\[cl1\]) with respect to the coherent states,$$\begin{aligned}
H\left( \mathbf{p},\mathbf{q}\right) &=&\left\langle \mathbf{p},\mathbf{q}%
\right\vert \mathcal{H}\left( \mathbf{P},\mathbf{Q}\right) \left\vert
\mathbf{p},\mathbf{q}\right\rangle \notag \\
&=&\sum_{\sigma =1}^{2}\left[ \frac{\mathbf{p}_{\sigma }^{2}}{2m}+\frac{%
m\varpi ^{2}}{2}\mathbf{q}_{\sigma }^{2}+\frac{\left\langle \mathbf{P}%
_{\sigma }^{2}\right\rangle }{2m}+\frac{m\varpi ^{2}}{2}\left\langle \mathbf{%
Q}_{\sigma }^{2}\right\rangle\right] +g\left\langle {\bf{p}},{\bf{q}}\vert V\vert {\bf{p}},{\bf{q}}\right \rangle \,,
\label{cl2}\end{aligned}$$where the coherent states are defined, as in (\[coherstates\]), by$$\left\vert \mathbf{p},\mathbf{q}\right\rangle =\prod_{\sigma =1}^{2}U\left(
\mathbf{q}_{\sigma }\right) V\left( \mathbf{p}_{\sigma }\right) \left\vert
\eta \right\rangle \,,\ \ U\left( \mathbf{q}_{\sigma }\right) =e^{-i\mathbf{q%
}_{\sigma }\cdot \mathbf{P}_{\sigma }/\hbar }\,,\ \ V\left( \mathbf{p}%
_{\sigma }\right) =e^{i\mathbf{p}_{\sigma }\cdot \mathbf{Q}_{\sigma }/\hbar
}\,. \label{cl0}$$As one can see from (\[cl2\]) there is no fundamental[^5] difference for the enhanced classical Hamiltonian between bosons, fermions or anyons. Nevertheless there does exist quantum differences between them, or more precisely, the numerical values of the coefficients proportional to $\hbar $ have a different value for each considered particle. By virtue of (\[kla5\]) and symmetries behind the several expectation values of (\[any22.2\]), as discussed in the subsection above, the expectation value of the potential $V$ with respect to the coherent states (\[cl0\]) reads$$\begin{aligned}
\left\langle \mathbf{p},\mathbf{q}\right\vert V\left\vert \mathbf{p},\mathbf{%
q}\right\rangle &=&\sum_{\sigma =1}^{2}\left[ \left( \mathbf{q}_{x_{\sigma
}}^{2}+\mathbf{q}_{y_{\sigma }}^{2}\right) ^{2}+10\,\mathbf{q}_{\sigma
}^{2}\left\langle Q_{x_{1}}^{2}\right\rangle \right] \notag \\
&+&2\left( \mathbf{q}_{1}\cdot \mathbf{q}_{2}\right) \left( \left\langle
Q_{x_{1}}Q_{x_{2}}\right\rangle +\left\langle
Q_{y_{1}}Q_{y_{2}}\right\rangle \right) +\frac{16}{3}\left\langle
Q_{x_{1}}^{4}\right\rangle +4\left\langle
Q_{x_{1}}^{2}Q_{x_{2}}^{2}\right\rangle \,. \label{cl2.2}\end{aligned}$$
Labeling $H_{k}\left( \mathbf{p},\mathbf{q}\right) $ as the enhanced Hamiltonian for bosons $\left( k=b\right) $, fermions $\left( k=f\right) $, $%
\left( k=\gamma \right) $ for anyons and using the results of the previous subsections we list below the enhanced hamiltonian for bosons and fermions: $$\begin{aligned}
H_{b}\left( \mathbf{p},\mathbf{q}\right) &=&H_{c}\left( \mathbf{p},\mathbf{q%
}\right) +\hbar \sum_{\sigma =1}^{2}\left( \frac{3g}{\Omega }\right) \mathbf{%
q}_{\sigma }^{2}+\hbar \left( \frac{\Omega }{m}+\frac{m\varpi ^{2}}{\Omega }%
\right) +\hbar ^{2}\left( \frac{3g}{\Omega ^{2}}\right) \,, \label{cl3} \\
H_{f}\left( \mathbf{p},\mathbf{q}\right) &=&H_{c}\left( \mathbf{p},\mathbf{q%
}\right) +6\,\hbar \sum_{\sigma =1}^{2}\left( \frac{3g}{\Omega }\right)
\mathbf{q}_{\sigma }^{2}+\hbar \left( \frac{3\Omega }{2m}+\frac{3m\varpi ^{2}%
}{2\Omega }\right) +2\,\hbar ^{2}\left( \frac{3g}{\Omega ^{2}}\right) \,,
\label{cl4}\end{aligned}$$ and for anyons we have$$\begin{aligned}
H_{\gamma }\left( \mathbf{p},\mathbf{q}\right) &=&H_{c}\left( \mathbf{p},%
\mathbf{q}\right) +\hbar \left[ \frac{\Omega }{m}\left( 1+\gamma \frac{%
\left. _{2}F_{1}\left( \frac{2-\gamma }{2},\frac{1-\gamma }{2};\frac{3}{2}%
;1\right) \right. }{\left. _{2}F_{1}\left( \frac{1-\gamma }{2},-\frac{\gamma
}{2};\frac{3}{2};1\right) \right. }-\frac{\gamma }{2}\right) +\frac{m\varpi
^{2}}{2\Omega }\left( 2+\gamma \right) \right] \notag \\
&+&\frac{\hbar g\left( \gamma +2\right) }{\Omega }\left[ \sum_{\sigma =1}^{2}%
\frac{5\mathbf{q}_{\sigma }^{2}}{2}-\left( \mathbf{q}_{1}\cdot \mathbf{q}%
_{2}\right) \left( \frac{\left. _{2}F_{1}\left( -\frac{1+\gamma }{2},-\frac{%
\gamma }{2};\frac{3}{2};1\right) \right. }{\left. _{2}F_{1}\left( \frac{%
1-\gamma }{2},-\frac{\gamma }{2};\frac{3}{2};1\right) \right. }-1\right) %
\right] \notag \\
&+&\frac{\hbar ^{2}\left( \gamma +3\right) \left( \gamma +2\right) }{6\Omega
^{2}}\left( \left. _{2}F_{1}\left( \frac{1-\gamma }{2},-\frac{\gamma }{2};%
\frac{3}{2};1\right) \right. \right) ^{-1} \notag \\
&\times &\left\{ 5\left. _{2}F_{1}\left( \frac{1-\gamma }{2},-\frac{\gamma }{%
2};\frac{5}{2};1\right) \right. +\frac{\gamma \left( \gamma -1\right) }{5}%
\left. _{2}F_{1}\left( \frac{3-\gamma }{2},\frac{2-\gamma }{2};\frac{7}{2}%
;1\right) \right. \right\} \,. \label{cl5}\end{aligned}$$where the *classical Hamiltonian* $H_{c}\left( \mathbf{p},\mathbf{q}%
\right) $, in which $\hbar \rightarrow 0$, has the same form for all cases,$$H_{c}\left( \mathbf{p},\mathbf{q}\right) =\lim_{\hbar \rightarrow
0}H_{k}\left( \mathbf{p},\mathbf{q}\right) =\sum_{\sigma =1}^{2}\left( \frac{%
\mathbf{p}_{\sigma }^{2}}{2m}+\frac{m\varpi ^{2}}{2}\mathbf{q}_{\sigma
}^{2}+g\left( \mathbf{q}_{x_{\sigma }}^{2}+\mathbf{q}_{y_{\sigma
}}^{2}\right) ^{2}\right) \,. \label{cl7}$$
Conclusion
==========
In this paper we have focussed on a central question in enhanced quantization using canonical coherent states, namely, the choice of the fiducial vector and the issues that choice involves. Initially, it was argued that a good choice is largely dictated by the explicit form of the Hamiltonian operator under consideration, and, in many cases the choice of the unique ground state as the fiducial vector has several virtues. However, that choice can also be relaxed to consider other fiducial vectors, and we can illustrate that choice by focussing attention on a Hamiltonian operator with a quartic interaction. One reason behind this choice is basically due to the fact that it is, effectively, the simplest example in which ${\cal O}(\hbar)$ coefficients of dynamical terms are involved that modify the classical description. Secondly, this choice can exhibit a problem with symmetry, e.g., rotational invariance, and similar properties can be extended to other models. In particular, fiducial vectors based on a Gaussian form are appropriate for bosons to deal with these Hamiltonian operators and $\hbar$-dynamical coefficients can be consistently treated with them. Generally, such coefficients can also be reduced by choosing Gaussian fiducial vectors, although this is not required according to the principles of enhanced canonical quantization [@EQ; @Kla63I; @Kla63II; @Kla63III]. In such cases, the corresponding enhanced Hamiltonian is a symbol [@Husimi] of the respective Hamiltonian operator where $\hbar$-dependence is included.
Although commonly used, it is a fact that Gaussian fiducial vectors are not always suitable to consistently describe certain physical systems. In this respect, we have chosen two examples where this form clearly fails: the enhanced quantization of fermions and anyons. In these cases the fiducial vectors cannot be independent Gaussians, and, instead, they must involve cross correlations for fermions or snyons. This latter property has a necessary physical consequence, namely, ensuring that permutations of the variables of the coherent state representation of Hilbert space vectors involve the required change of phase.
In section 3 non-Gaussian fiducial vectors have been used in the coherent states constructed for fermions and anyons. We have calculated several expectation values for both systems and, in this regard, the exact calculations for anyons also include the corresponding results for fermions simply by choosing $\gamma =1$. Using these results, we have calculated the enhanced classical Hamiltonian and after taking the limit in which $\hbar\ra0$, we have shown that bosons, fermions, and anyons all have the same classical Hamiltonian, despite the fundamental differences between their properties when $\hbar>0$.
Appendix {#appendix .unnumbered}
========
In dealing with expectation values of coordinate and momentum operators for the anyon case we frequently encounter integrals of the form \[recall the convention: $\tint_a^b\s(\cdot)\s d{\bf r}=\tint_a^b(\cdot)\s dx\,\tint_a^b(\cdot)\s dy$\] $$\int_{-\infty }^{\infty }d\mathbf{r}_{\sigma ^{\prime }}\left\vert \mathbf{r}%
_{\sigma ^{\prime }}-\mathbf{r}_{\sigma }\right\vert ^{2\gamma }e^{-\lambda
\mathbf{r}_{\sigma ^{\prime }}^{2}}\,, \label{ap1}$$where the indexes $\sigma ,\sigma ^{\prime }$ label particles. Due to rotational symmetry, one can choose a particular reference system to simplify its solution. Choosing, for example, a reference frame where the vector $\mathbf{r}_{\sigma }$ is aligned along the axis $y_{\sigma ^{\prime
}}$ we have $\mathbf{r}_{\sigma }\cdot \mathbf{r}_{\sigma ^{\prime
}}=r_{\sigma }r_{\sigma ^{\prime }}\cos \left( \pi /2-\vartheta \right)
=r_{\sigma }r_{\sigma ^{\prime }}\sin \vartheta $. Here $\vartheta $ is the angle between the axis $x_{\sigma ^{\prime }}$ and $\mathbf{r}_{\sigma
^{\prime }}$ such that (\[ap1\]) admits the form [@Gradshtein],$$\begin{aligned}
\rho _{\gamma }^{\left( \pm \right) }\left( r_{\sigma }\right)
&=&\int_{0}^{\infty }dr_{\sigma ^{\prime }}r_{\sigma ^{\prime }}^{2\pm
1}e^{-\lambda r_{\sigma ^{\prime }}^{2}}\mathcal{I}_{\vartheta }^{\left( \pm
\right) }\left( r_{\sigma },r_{\sigma ^{\prime }};\gamma \right) \,, \notag
\\
\mathcal{I}_{\vartheta }^{\left( \pm \right) }\left( r_{\sigma },r_{\sigma
^{\prime }};\gamma \right) &=&\int_{0}^{2\pi }d\vartheta \left( \cos
\vartheta \right) ^{1\pm 1}\left( r_{\sigma }^{2}+r_{\sigma ^{\prime
}}^{2}-2r_{\sigma }r_{\sigma ^{\prime }}\sin \vartheta \right) ^{\gamma }
\notag \\
&=&\left( \frac{3\mp 1}{2}\right) \pi \left( r_{\sigma }^{2}+r_{\sigma
^{\prime }}^{2}\right) ^{\gamma }\left. _{2}F_{1}\left( \frac{1-\gamma }{2},-%
\frac{\gamma }{2};\frac{3\pm 1}{2};\frac{4r_{\sigma }^{2}r_{\sigma ^{\prime
}}^{2}}{\left( r_{\sigma }^{2}+r_{\sigma ^{\prime }}^{2}\right) ^{2}}\right)
\right. \,. \label{ap2}\end{aligned}$$To evaluate $\rho _{\gamma }^{\left( \pm \right) }\left( r_{\sigma }\right) $ it is convenient to make use the series representation of the hypergeometric functions [@Gradshtein],$$\left. _{2}F_{1}\left( \frac{1-\gamma }{2},-\frac{\gamma }{2};\frac{3\pm 1}{2%
};\frac{4r_{\sigma }^{2}r_{\sigma ^{\prime }}^{2}}{\left( r_{\sigma
}^{2}+r_{\sigma ^{\prime }}^{2}\right) ^{2}}\right) \right.
=\sum_{s=0}^{\infty }\frac{1}{s!}\left( \frac{1-\gamma }{2}\right)
_{s}\left( -\frac{\gamma }{2}\right) _{s}\left( \frac{3\pm 1}{2}\right)
_{s}^{-1}\frac{\left( 4r_{\sigma }r_{\sigma ^{\prime }}\right) ^{2s}}{\left(
r_{\sigma }^{2}+r_{\sigma ^{\prime }}^{2}\right) ^{2s}}\,, \label{ap2.1}$$which, after the change of variables $u_{\sigma }=\lambda r_{\sigma }^{2}$ and $u_{\sigma ^{\prime }}=\lambda r_{\sigma ^{\prime }}^{2}$, can be brought into the form$$\begin{aligned}
&&\rho _{\gamma }^{\left( \pm \right) }\left( u_{\sigma }\right)
=\sum_{s=0}^{\infty }\Lambda _{s}^{\left( \pm \right) }\varrho _{\gamma
,s}^{\left( \pm \right) }\left( u_{\sigma }\right) u_{\sigma }^{s}\,, \notag
\\
&&\varrho _{\gamma ,s}^{\left( \pm \right) }\left( u_{\sigma }\right) \equiv
\int_{0}^{\infty }du_{\sigma ^{\prime }}e^{-u_{\sigma ^{\prime }}}u_{\sigma
^{\prime }}^{s+\frac{1\pm 1}{2}}\left( u_{\sigma ^{\prime }}+u_{\sigma
}\right) ^{\gamma -2s}\,, \notag \\
&&\Lambda _{s}^{\left( \pm \right) }=\frac{\pi }{\lambda ^{\gamma +\frac{%
3\pm 1}{2}}}\frac{4^{s}}{s!}\left( \frac{1-\gamma }{2}\right) _{s}\left( -%
\frac{\gamma }{2}\right) _{s}\left( \frac{3\pm 1}{2}\right) _{s}^{-1}\,,
\label{ap3}\end{aligned}$$where $\left( n\right) _{s}=\Gamma \left( n+s\right) /\Gamma \left( s\right)
$ is the Pochammer symbol [@Gradshtein]. The integral above can be analytically solved,$$\begin{aligned}
&&\varrho _{\gamma ,s}^{\left( \pm \right) }\left( u_{\sigma }\right)
=\Gamma \left( s+\frac{3\pm 1}{2}\right) u_{\sigma }^{\frac{2\left( \gamma
-s\right) +1\pm 1}{4}}e^{\frac{u_{\sigma }}{2}}W_{\mu ,\nu }\left( u_{\sigma
}\right) \,, \notag \\
&&\mu =\frac{2\left( \gamma -3s\right) -1\mp 1}{4}\,,\ \ \nu =\frac{2\left(
s-\gamma \right) -3\mp 1}{4}\,, \label{ap4}\end{aligned}$$with $W_{\mu ,\nu }\left( u_{\sigma }\right) $ being a Wittaker function [@Gradshtein].
To complete the description it is also needed to solve one more integral of the expressions above, namely an integral over the variables $u_{\sigma }$ [@Gradshtein],$$\begin{aligned}
R_{n}^{\left( \pm \right) }\left( \gamma ,s\right) &\equiv \int_{0}^{\infty
}du_{\sigma }e^{-\frac{u_{\sigma }}{2}}u_{\sigma }^{\frac{\gamma +s}{2}%
+n}W_{\mu ,\nu }\left( u_{\sigma }\right)\notag \\
&=\frac{\Gamma \left( n+s+\frac{%
3\mp 1}{4}\right) \Gamma \left( n+\gamma +\frac{9\pm 1}{4}\right) }{\Gamma
\left( 2s+n+\frac{9\pm 1}{4}\right) }\,,\ \ n>\frac{-3\pm 1}{4}\,.
\label{ap5.1}\end{aligned}$$
Acknowledgements {#acknowledgements .unnumbered}
================
T. C. Adorno acknowledges support of FAPESP under the contracts 2013/00840-9 and 2013/16592-4. He is also thankful to the Dept. of Physics of the University of Florida for its kind hospitality.
[99]{} J. R. Klauder, Divergences in Scalar Quantum Field Theory: The Cause and the Cure, Mod. Phys. Lett. A **27**, 1250117 (9pp) (2012); arXiv:1112.0803.
J. R. Klauder, Enhanced quantization: a primer, J. Phys. A: Math. Theor. **45**, 285304 (8pp) (2012); arXiv:1204.2870; J. R. Klauder, “Revisiting Canonical Quantization”, arXiv:1211.7351.
J. R. Klauder, “Continuous-Representation Theory I. Postulates of Continuous Representation Theory", J. Math. Phys. [**4**]{}, 1055-1058 (1963).
J. R. Klauder, “Continuous-Representation Theory II. Generalized Relation Between Quantum and Classical Dynamics", J. Math. Phys. [**4**]{}, 1058-1073 (1963).
J. R. Klauder, “Continuous-Representation Theory III. On Functional Quantization of Classical Systems", J. Math. Phys. [**5**]{}, 177-187 (1964).
P. A. M. Dirac, *The Principles of Quantum Mechanics*, (Clarendon Press, Oxford, 1947), page 114.
E. Wigner, “On the Quantum Correction For Thermodynamic Equilibrium”, Phys. Rev. [**40**]{}, 749 (1932).
K. Husimi, “Some Formal Properties of the Density Matrix” Proc. Phys. Math. Soc. Japan [**22**]{}, 264 (1940).
E.C.J. Sudarshan, “Equivalence of Semiclassical and Quantum Mechanical Descriptions of Statistical Light Beams”, Phys. Rev. Lett. 10 (1963 ) 277.
F. A. Berezin and M. A. Shubin, *“The Schrödinger Equation”*, (Kluwer Academic Publishers, Drodrecht 1991).
J. R. Klauder and B.-S. Skagerstam, “*Coherent States: Applications to Physics and Mathematical Physics*", editor plus original introduction (World Scientific, Singapore, 1985).
J. R. Klauder, Enhanced Quantum Procedures that Resolve Difficult Problems; arXiv:1206.4017
J. Ben Geloun and J. R. Klauder, Enhanced Quantization on a Circle, Phys. Scr. 87 035006 (5pp) (2013) ; arXiv:1206.1180
P. M. Cohn, [*Lie Groups*]{}, (Cambridge University Press, London, 1961)).
N. Aronszajn, Proc. Cambridge Phil. Soc. [**39**]{}, 133 (1943); Trans. Am. Math. Soc. [**68**]{}, 337 (1950); H. Meschkowski, [*Hilbertsche Räume mit Kernfunktion*]{}, (Springer Verlag, Berlin, 1962). T T Truong, “Weyl quantization of anharmonic oscillator”, J. Math. Phys. [**16**]{}, 1034-1042 (1975).
H. Bergeron, J.-P. Gazeau, and A. Youssef, “Are the Weyl and coherent state descriptions physically equivalent?”, Phys. Lett. A [**377**]{}, 598 (2013); arXiv:1102.3556.
R. P. Feynman, [*Statistical Mechanics*]{}, (Benjamin/Cummings: Reading, Mas- sachusetts, 1972).
F. Wilczek, “Magnetic Flux, Angular Momentum, and Statistics”, Phys. Rev. Lett **48**, 1144 (1982); “Quantum Mechanics of Fractional-Spin Particles”, **49**, 957 (1982).
F. Wilczek and A. Zee, “Linking Numbers, Spin, and Statistics of Solitons”, Phys. Rev. Lett. **51**, 2250 (1983).
R. B. Laughlin, “Superconducting Ground State of Noninteracting Particles Obeying Fractional Statistics”, Phys. Rev. Lett. **60**, 2677 (1988).
A. L. Fetter, C. B. Hanna and R. B. Laughlin, “Random phase approximation in the fractional statistics gas”, Phys. Rev. B **39**, 9697 (1989).
S. Das Sarma, M. Freedman and C. Nayak, “Topologically Protected Qubits from a Possible Non-Abelian Fractional Quantum Hall State”, Phys. Rev. Lett. **94**, 166802 (2005).
G. Collins, “Computing with Quantum Knots”, Sci. Am. (Int. Ed.) **294**, 57 (2006).
C. Nayak, S. H. Simon, A. Stern, M. Freedman and S. Das Sarma, “Non-Abelian anyons and topological quantum computation”, Rev. Mod. Phys. **80**, 1083 (2008).
B. I. Halperin, “Statistics of Quasiparticles and the Hierarchy of Fractional Quantized Hall States”, Phys. Rev. Lett. **52**, 1583 (1984).
Yong-Shi Wu, “Multiparticle Quantum Mechanics Obeying Fractional Statistics” Phys. Rev. Lett. **53**, 111 (1984).
R. Iengo and K. Lechner, “Quantum Mechanics of Anyons on a Torus”, Nucl. Phys. B **346**, 551 (1990).
C. Chou, “Multi-anyon quantum mechanics and fractional statistics”, Phys. Lett A **155**, 245 (1991).
M. V. N. Murthy, J. Law, M. Brack and R. K. Bhaduri, “Quantum Spectrum of Three Anyons in an Oscillator Potential”, Phys. Rev. Lett. **67**, 1817 (1991).
R. Iengo and K. Lechner, “Anyon Quantum Mechanics and Chern-Simons Theory”, Phys. Rep. **213**, 179 (1992).
F. Wilczek, *Fractional Statistics and Anyon Superconductivity* (World Scientifc, Singapore 1990).
R. Jackiw and V. P. Nair, “Relativistic wave equation for anyons”, Phys. Rev D **43**, 1933 (1991).
P. Horváthy, M. Plyushchay and M. Valenzuela, “Bosons, fermions and anyons in the plane, and supersymmetry”, Ann. Phys. **325**, 1931 (2010).
M. S. Plyushchay, “Relativistic model of the anyon”, Phys. Rev. B **248**, 107 (1990).
S. Ghosh, “Spinning Particles in 2+1 dimensions”, Phys. Lett. B **338**, 235 (1994); “Anyons in an electromagnetic field and the Bargmann-Michel-Telegdi equation”, Phys. Rev. D **51**, 5827 (1995).
D. M. Gitman and I. V. Tyutin, *Quantization of Fields with Constraints* (Springer-Verlag, Berlin, 1990); M. Henneaux and C. Teitelboim, *Quantization of Gauge Systems* (Princeton University Press, Princeton New Jersey 1994); K. Sundermeyer, *Constrained Dynamics with applications to Yang-Mills Theory, General Relativity, Classical Spin, Dual String Model*, Lecture Notes in Physics 169 (Springer-Verlag, Berlin Heidelberg New York 1982).
A. Polychronakos, “Exact anyonic states for a general quadratic hamiltonian”, Phys. Lett. B **264**, 362 (1991).
I. S. Gradshteyn and I. M. Ryzhik, *Tables of Integrals, Series and Products*, (70$^{th}$ edition, Academic Press 2007).
[^1]: [email protected], [email protected]
[^2]: [email protected]
[^3]: For the explicit definition of $\Lambda _{s}^{\left( -\right) }$ and $%
R_{1}^{\left( -\right) }\left( \gamma ,s\right) $ see the Appendix (\[ap2\]), (\[ap5.1\]).
[^4]: Regarding the gauge transformation there is a frequently used terminology associated with that [@IenLec90; @IenLec92]. The corresponding *singular* gauge transformation defines the so-called *anyons gauge* in which the wave function is a multivalued function and the Hamiltonian does not have the potential vector. The other possibility is the so-called *CS-gauge* where the wave function obeys the standard statistics but the Lagrangian contains a *Chern-Simons* term. In this paper we do not follow the latter approach.
[^5]: More precisely speaking, the functional form of the enhanced Hamiltonian for bosons, fermions and anyons are the same. The difference appears only in the numerical values of the expectation values which, as a matter of fact, depend on the choice of fiducial vectors under consideration.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Synthetic power systems that imitate functional and statistical characteristics of the actual grid have been developed to promote researchers’ access to public system models. Developing time series to represent different operating conditions of these synthetic systems will expand the potential of synthetic power systems applications. This paper proposes a methodology to create synthetic time series of bus-level load using publicly available data. Comprehensive validation metrics are provided to assure that the quality of synthetic time series data is sufficiently realistic. This paper also includes an example application in which the methodology is used to construct load scenarios for a 10,000-bus synthetic case.'
author:
- |
Hanyue Li, Ju Hee Yeo, \
Ashly L.Bornsheuer, Thomas J. Overbye, [^1]
bibliography:
- 'citation.bib'
title: |
The Creation and Validation of Load Time\
Series for Synthetic Electric Power Systems
---
Synthetic time series, residential commercial and industrial load, synthetic power systems
Introduction
============
access to real power system data is limited due to confidentiality concerns. Synthetic power system models and data are created to be functionally and statistically similar to real power systems. Synthetic systems are synthesized using public data of the actual grid, and they don’t represent the actual system located on the same geographic footprint, or contain any confidential information about the actual grid.
Many efforts have been made on the creation of synthetic power system base cases, which contains systems topology and many of them have AC or DC power flow solutions. Early work of [@Zwang1]-[@Zwang2] came up with an approach to create transmission grid topologies based on the small world graph network. A methodology for generating large scale synthetic transmission systems with AC power flow solutions on the footprint of United States was proposed in [@Adam1]-[@Adam2], and several synthetic systems of different sizes and footprints were created. The work of [@Soltan1]-[@Soltan2] investigated the geographic and structural properties of North American and Mexican transmission grids, and created large electric systems with synthetic nodes and node connections. European synthetic transmission grids with DC power flow solutions were also developed based on public information from utilities and regulatory agencies in the synchronous grid of Continental Europe (UCTE) [@Bialek].
Since synthetic power system base cases only reflect a one-time snapshot of the system, there is a natural need to expand the work and develop time series to represent changing system operating conditions over time. The combined data set of synthetic grid models and power system time series can be used as a benchmark for system scenario studies, a test bed for algorithms such as time sequence power flow, optimal power flow, and unit commitment.
Power system time series consist of data relevant to system status in a time sequence manner. It can span many aspects of the system such as the load, transmission line status, generator dispatch and electricity price. Real time series power system data are generally more publicly accessible compared to the data of actual grid topologies and models. For example, the Open Power System Data, for example, is a data depository that has load, wind, solar and price data in hourly resolution of 37 European countries [@OpenPowerSystemData]. In North America, for the transparent operation of electricity market, Independent System Operators (ISOs) often have hourly resolution load and price time series data publicly available as well [@ISO-NE]-[@IESO].
Besides actual system-level data, the creation of synthetic time series for household-level load is also a well-researched topic. The work of [@household1] developed a probabilistic mathematical model for residential load simulation. A top-down approach using domestic load patterns for household profile adaptation was implemented in [@household2]. A machine learning method to generate synthetic residential building load time series from smart meter data was proposed in [@household3]. For generation, the synthetic creation of solar data using a Markov Chain based approach was discussed in [@syntheticsolar1]. Methods and initial results of creating synthetic high frequency solar simulations was demonstrated using low-frequency solar variability [@syntheticsolar2]. Using Weibull distribution and ARMA time series model, synthetic wind speed and power output forecasts were presented [@syntheticwind2]. However, as most of the power system studies require load data resolution at the bus-level, either real system-level or synthetic household-level time series data can be directly applied to the synthetic grid.
This paper presents a methodology for synthesizing bus-level time series load data, building on the results of [@Scenarios]. The maximum value of load time series aims to match the corresponding load bus size determined in the above mentioned base case. The unique variation of each bus-level time series is a result of the heuristic aggregation of prototypical building and facility load time series. To ensure the quality and realism of the synthetic load time series, comprehensive validation metrics are established from the actual system-level load time series. An example application using the created time series for synthetic electric grid scenario study is also given.
The creation and validation of load time series for 2,000 and 10,000- bus synthetic grids (i.e., the ACTIVSg2000 and ACTIVSg10K grids from [@tamurepository]) are used as examples. The ACTIVSg2000 synthetic system shares the same footprint as the Electric Reliability Council of Texas (ERCOT), and the ACTIVSg10K has the same service region as the Western Interconnection (WI). However, the method of creating and validating synthetic time series is general enough to be applied to any system.
Creation of Bus-level Load Time Series
======================================
Electric load time series reflect electricity consumption patterns and provides insight on the absolute level and changing rate of load at different times. Having access to bus-level load time series is essential for the unit dispatch and commitment in power system operations since generation always needs to follow the time-varying system load. The load time series in synthetic power systems has hourly resolution with a duration of a year, and is created on the bus level so that every bus in the synthetic grid model has a unique profile. Each bus-level load time series is created using an iterative aggregation approach, where prototypical building load profiles are aggregated based on the size and composition of load buses.
Location and size of bus-level electric load
--------------------------------------------
The location and size of the electric loads are determined during the creation of the synthetic base case discussed in [@Adam1]-[@Adam2]. The load buses are located based on the clustering of geographic coordinates associated with postal codes that are obtained from the public U.S. census database. The size of each load bus is then scaled according to the population of the corresponding postal code and the per-capita MW consumption. A fixed power factor is assigned to each load as an assumption.
The base case is used as a reference to create load time series. The size of load buses in the base case are considered to be the peak value of each bus-level time series, and the geographic coordinates assigned to each load bus are then used to determine the unique location-dependent load features such as load composition ratio and prototypical building load time series.
Load bus composition ratio
--------------------------
The assignment of a composition ratio of residential, commercial and industrial load on each bus is helpful to realistically represent the uniqueness of load. It establishes the geographic and demographic dependence of electric load similar to reality.
U.S. utility companies’ service territories as well as their residential, commercial and industrial megawatt-hours sales values from the Annual Electric Power Industry Report are used to determine the bus load composition ratios [@AnnualElectricPowerIndustryReport]. Each load bus is assigned to one utility company based off its geographic coordinates, and the company’s sales ratio of the three load types is used as the average bus load composition ratio.
Figure \[fig:2k\_contour\] contour the dominated load type of each bus in ACTIVSg2000 and ACTIVSg10K synthetic power systems. In the contour maps, locations with high residential and commercial load concentration are also the major U.S. metropolitan and urbanized areas. Many industrial dominated areas are located close to major mining and manufacturing sites.
Prototypical building- and facility-level load time series
----------------------------------------------------------
To bridge between the bus load composition ratio, and a unique hourly profile, prototypical end user level load time series under residential, commercial and industrial load types are synthesized from public data. Building- and facility-level time series gives the desired bus load a good base to incorporate both individual user load patterns and the aggregation effect. Different categories of buildings and facilities and their prototypical load time series are realistic approximations to represent the most common and important load features.
### Prototypical residential and commercial building load
The prototypical building load time series synthesized in this paper are the same as the ones developed in [@Scenarios], where open source data of simulated hourly residential and commercial building energy consumption are used [@OpenEI].
The residential data contains buildings’ hourly electricity usage value from space heating/cooling, High Voltage Alternating Current (HVAC) fan, interior/exterior lighting, as well as appliances and miscellaneous loads. Each data file covers one typical meteorological year 3 (TMY3) location in the United States, which represents geographic locations with different meteorology[@OpenEIRes]. For commercial load, under each TMY3 location, 16 building electric load profiles are simulated using the Department of Energy (DOE) commercial reference building models[@OpenEICom], and the contents in each time series data are like that of residential data set.
Under the United States footprint, 1020 residential and 16,320 commercial building time series are calculated as a summation of all electricity consumption categories under each building type. They are created to exhibit unique profiles of residential and commercial buildings, and electricity consuming variations over time and geographic location.
Figure \[fig:res\_3tmy\], for example, shows prototypical residential building load time series in a winter week and a summer week. The load shapes in two seasons are distinguishable, where winter profiles tend to have two peaks in one day due to winter heating, while summer profiles only have one peak per day. The magnitude of load can also be very different in each season, depending on geographic locations. In winter, regions with colder climate such as Helena, Montana, would have higher average load. While in summer, load within hot and arid climate zones, such as Phoenix,Arizona, will have much more electricity consumption.
Similarly, figure \[fig:com\_3tmy\] shows prototypical commercial load time series for the large office building type. The load shape of a specific building type is generally consistent regardless of the location, and the load level is slightly higher in summer season compared to that in winter. In figure \[fig:com\_3type\], weekly load profiles of three commercial building types (full-service restaurant, small office, and strip mall) in Los Angeles, California are shown. The load shape and size under each building type is unique. Small offices have steady load during weekdays and low load during weekends. For full-service restaurants and trip malls, load levels are constant through out the week, while full-service restaurants observe two peaks near lunch and dinner time, strip malls only peak once every day.
### Prototypical industrial facility load
Prototypical industrial facility load time series are created based on publicly available per-unit industrial load curves from Oak Ridge National Laboratory (ORNL) [@ind] and the industrial assessment data in Industrial Assessment Centers (IAC) Database [@IAC].
The ORNL per-unit curves provide daily profiles of different industrial sectors, presented by different Standard Industrial Classification (SIC) codes with their unique load factor. The IAC Database contains information on the industry SIC code, total electricity usage and yearly operating hours of over 14,000 facilities in the United States, which are used to modify the ORNL curves into facility-specific load time series for a year.
For each industrial facility, the yearly operating hour is first used to determine the total number of operating days. The ORNL daily curves of the same SIC code is then expanded to a yearly load curve, with small white noise imposed and a random selection of starting day of the year. The synthesized yearly curve is then scaled so that the integral value of the curve matches the total electricity usage.
Figure \[fig:ind\_4type\] presents prototypical industrial load time series for four facilities from food, petroleum and refining, primary metal, as well as electronic and electrical equipment industries. As those load curves are adopted from the ORNL per-unit daily curves, they have similar daily variations and weekly shapes, with different load levels and load factors.
Aggregation of load
-------------------
The bus-level load time series is created by iteratively aggregating prototypical building and facility load time series of each load type. This aggregation process has three main aspects: integrating realistic amounts of end users under each load type, selecting representative prototypical time series, and mimicking the effect of load aggregation described in [@AggregationEffect].
A flow chart of this aggregation process is shown in figure \[fig:heuristic\]. The reference peak values of residential, commercial and industrial bus load type of each bus are first determined by the multiplication of bus load size and the load composition ratio. This is used to integrate realistic amount of end users under each load type, where the peak component values are the stopping criterion for the iterative aggregation process.
A pool of representative prototypical building load time series are then selected for each load bus. For residential and commercial load, the selection used the top five shortest distances between the load bus geographic coordinates and TMY3 locations. All industrial facility load time series that have smaller maximum value than the calculated peak industrial load component are included in this pool since industrial loads are less correlated with geographic locations.
Under each load type, within one iteration, only one building or facility load time series is picked based on a heuristic weighted distribution as equation (\[eq:1\]), where time series with a lower average load is more likely to be chosen.
$$\label{eq:1}
P(TS_i) = \frac{\frac{1}{\sqrt{\overline{TS_i}}}}{\sum_{i=1}^{N} \frac{1}{\sqrt{\overline{TS_i}}}}$$
where:
[$P(TS_i)$]{} probability of load time series i being chosen\
[$N$]{} number of time series selected in one load type\
[$\overline{TS_i}$]{} yearly average value of load time series i\
This selected building- or facility-level load time series is then processed through three types of transformations: time shift, time permutation, and small white noise insertion. Those transformations diversify the load profiles of end users, so that the smoothing effect for load aggregation can be produced.
The original prototypical load time series can be shifted for an integer hour following a normal distribution in equation (\[eq:2\]) , where the expectation is 0, and the standard deviation is determined based off the load type.
$$\label{eq:2}
P(time \: shift = x) = \int_{x}^{x+1} \frac{1}{{\sigma \sqrt {2\pi } }}e^{{{ - \left( {t} \right)^2 } \mathord{\left/ {\vphantom {{ - \left( {t} \right)^2 } {2\sigma ^2 }}} \right. \kern-\nulldelimiterspace} {2\sigma ^2 }}} dt$$
where:
[$\sigma = 0.4$]{} prototypical load time series is residential\
[$\sigma = 0.3$]{} prototypical load time series is commercial\
[$\sigma = 0.1$]{} prototypical load time series is industrial\
To imitate the random surges or drops of load for individual customers, certain hours (100, 100 and 50 hour pairs for residential, commercial and industrial respectively) are randomly chosen within the year to be permutated. To avoid bus load time series being overly conforming due to the use of similar prototypical building or facility time series, a small white Gaussian noise is also imposed. This transformed time series is then added to the corresponding type of load component, and the iteration would stop once after the load component maximum value calculated in previous step has been reached.
To mimic the effect of increasing load factor as load aggregates to a higher level [@AggregationEffect], public feeder load time series managed by National Renewable Energy Laboratory (NREL) [@TaxonomyFeeders] is used to adjust the bus-level load factors to a realistic range. This distribution feeder load time series is populated from the taxonomy distribution feeders of different geographic regions.
The geographic coordinates of each load bus in the synthetic system are used to randomly select a subset of taxonomy feeders from the same geographic region, so that the summation of feeder load time series is on the same scale as the bus load. The load factor of the aggregated feeder load time series is calculated to be the reference value. A constant component is added to the created bus-level load time series to adjust its load factor to a realistic value according to equation (\[eq:3\]).
$$\label{eq:3}
\frac{Constant + Average\:Load}{Constant + Max\:Load} = Reference\:Load\:Factor$$
Example Results
===============
The load time series created for ACTIVSg2000 and ACTIVSg10K synthetic power systems are discussed in this paper as example results. There are 1125 and 4170 load buses in those two cases, with 71 and 132 GW of system peak load respectively.
The plots in figure \[fig:2k\_contour\] show the dominant bus load type using the method discussed in section II.B, where the load type with highest composition percentage is considered to be the dominant type of its load bus. In ACTIVSg2000 system, 67.4% of the buses are dominated by residential load, 18.8% are primary composed of commercial load, and 13.8% for industrial load. For the ACTIVSg10K system, 45.4%, 33.7% and 20.9% of buses have residential, commercial and industrial load as primary composition respectively.
On the bus-level, each load time series is unique based off the location and load composition ratio of the load bus. Average bus-level load time series of different dominant load types are shown in figure \[fig:bus\_load\_result\]. Residentially-dominated bus load time series exhibit noticeable seasonal differences, where the electricity consumption in summer and winter seasons tend to have higher average values as well as higher variations. Commercially-dominated bus load time series have distinct daily patterns, while the electricity consumption base line stays relatively constant throughout the year. Industrially-dominated bus loads usually have the lowest variation and highest load factor. The average size of industrially-dominated buses are larger than the other two types.
The system-level synthetic load time series and the actual system load from their footprint regions are shown in figure \[fig:2k\_system\] and figure \[fig:10k\_system\]. Although duplicating system-level load time series is not the desired outcome, synthetic load time series on the system level should exhibit similar general load shapes and trends compared to the actual system.
Figure \[fig:2k\_system\] and figure \[fig:10k\_system\] show that the synthetic loads share similar size with the load of the actual system in the corresponding service territory. ACTIVSg2000 synthetic system has 71.1 GW of peak load, and 48.7 GW of average load, and the actual load of ERCOT system has 71.2 GW of maximum load and 41.0 GW of average load. For ACTIVSg10K system, there is 132.5 GW of peak load and 88.1 GW of average load. The corresponding actual system, the United States Western Interconnection, has 136.2 GW of peak load and 83.8 GW of average load.
Daily and weekly patterns can be seen from the ACTIVSg2000 and ACTIVSg10K synthetic load time series. It is also observed that the synthetic system has similar seasonal trends compared to the actual system. ACTIVSg2000 and ERCOT systems both experience peak load in summer, and also have some high load days weeks in winter. ACTIVSg10K and WI system also peak in summer, while their profile in the winter season is flatter.
Creation of Renewable Energy Time Series
========================================
The fast growing renewable generation in power system is changing the way that transmission grid is utilized and operated [@RenewableImpact]. According to the Energy Information Administration (EIA), renewable resources in the United States, including wind, solar, and hydro power, account for 20 percent of the total electrical generation during the first half of 2018 [@generatormonth]. With the assumption that renewable energies in the synthetic system are integrated into the transmission grid without curtailment, generation time series at each renewable unit is an important input to define system operating conditions.
The renewable generation time series in synthetic power systems should realistically reflect the relative size and level of variation of each renewable energy source. It has hourly resolution with duration of a year and is established on the bus level. The location and size of generators are determined from the synthetic base case [@Adam1]-[@Adam2], where publicly available data for all electric generators operating in the U.S. grid is used.
Synthetic wind generation time series
-------------------------------------
The hourly real power generations of wind units are synthesized from Wind Integration National Data Set (WIND) developed by NREL [@WINDTOOLKIT]. It contains 5-minute resolution wind turbine output and locations of more than 126,000 sites in the continental United States.
The geographic coordinates of wind units in the synthetic system is used to identify a closest wind site in the WIND Data Set. The 5-minute resolution data from this selected site is then up-sampled to hourly resolution time series, where the average value of 12 time points in the same hour is used. The hourly time series is also scaled so that the maximum value is consistent with the unit size defined in the synthetic base case.
Synthetic solar generation time series
--------------------------------------
The hourly real power output time series of each solar plant in the synthetic system is created based off the Solar Resource Data Set (SOLAR) from NREL [@Solarresourcedata], using equation \[eq:PVoutput\].
The solar panel yield represents the utilization efficiency, where its average value usually fluctuates between 15% and 18% [@PVoutputcalculation]. The 30-minute resolution global horizontal irradiance (GHI) record from SOLAR is up-sampled to hourly interval. PR is a coefficient for loss, which ranges between 0.5 and 0.9 [@PVoutputcalculation].
$$\label{eq:PVoutput}
Solar \, Output = S \times r \times GHI \times PR$$
where:
$S$ &total solar panel area\
$r $ & solar panel yield\
$GHI $ &hourly global horizontal irradiance\
$PR $ &performance ratio\
Each solar unit in the synthetic system is assigned with one solar output time series based on geographic distance, this time series is also scaled so that its peak value is consistent with the unit size from the base case.
Synthetic hydro generation time series
--------------------------------------
The hourly synthetic hydro generation time series is calculated by the multiplication of hydro unit maximum output, and a variation factor. The hydro generator size from the base case is used as its maximum value. The variation factor is calculated heuristically as the weighted combination of three components using equation \[eq:hydro\].
$$\begin{aligned}
\label{eq:hydro}
Variation \: Factor &= A\times Seasonal\: Component \\
&+ B\times Trend\: Component \\
&+ C\times Residual\: Component
\end{aligned}$$
The seasonal component is obtained from the California ISO hourly hydro break down data [@caisohydro]. The trend component reflects the monthly net U.S. hydroelectric generation from EIA’s Electric Power Monthly Report [@eiaepmr]. White noise is imposed as the residual component to account for irregular behavior patterns. This factor then is normalized between 0 and 1 every hour, where 1 stands for the maximum generation.
Results
=======
In ACTIVSg2000 system, 87 wind farms, 22 solar sites and 25 hydro generations are integrated into the synthetic transmission grid. For ACTIVSg10K systems, the number of renewable units is 243, 391 and 715 for wind, solar, and hydro respectively. The total amount of wind generation in ACTIVSg2000 system covers 23.4% of demand, and the total hydro generation in ACTIVSg10K system supplies 41.8% of its load. It is a realistic representation of the current renewable integration status of the footprint regions.
Wind generation pattern can vary considerably across regions, mostly based on local atmospheric and geographic conditions. Figure \[fig:wind\_seasonal\_location\] also shows that a specific location can have a unique wind generation pattern. For example, Roscoe in Texas has the world’s largest wind farm in actual grid. Also, Texas’s wind plant performance is at its highest in the early and later months of the year, not the summer. Therefore, realistically, Figure \[fig:wind\_seasonal\_location\] illustrates that Roscoe supplies prominently huge amount of wind power compared to two other cities, and the total amount of wind generation is higher during the winter than it is in the summer. In California, however, wind generation surges through June and fall. Its seasonal wind performance pattern is due to the strong, cold Pacific current on the West Coast and the land-sea breeze effect working with the normal west-to-east winds due the earth’s rotation [@windseasonalpattern]. This pattern also can be seen from our target city, Birds Landing, California. Glenns Ferry, which is located in the Northwest, does not generate substantial wind power annually. Hence, the amount of wind power generation is relatively smaller than other cities.
Figure \[fig:wind\_solar\_hydro\] shows the created synthetic generation time series of wind, solar, and hydro from three representative sites in the two synthetic power systems, of a week in winter and summer seasons. The generation from a hydro unit have the highest magnitude among the three renewable resources. It has similar pattern in different seasons since hydro power is less affected by seasonal variation of weather and temperature. Unlike hydro power, wind and solar power show different pattern depending on the season. For wind generation, more steady overall output is observed during the summer than it is in the winter, while the maximum output stays constant. Wind generation also peaks during the night in the synthetic generation time series, as wind is normally stronger at night [@dailysolarwindercot]. For solar generation, much higher amount of power is generated during the summer due to the longer exposure of the sun. Solar output also follows the daily cycle of sunlight, which peaks during the day and returns to zero at night.
Validation of Synthetic Time Series
===================================
since synthetic time series are fictitious, the validation of the created data against the actual data is critical to determine the quality and realism of the time series. A statistical based validation approach is implemented in this paper, as synthetic time series aim to realistically represent behaviors of load over time, instead of being an exact duplicate or forecast of the actual system time series.
A comprehensive set of validation metrics enables researchers to use synthetic time series with ease, but at the same time to be aware of the underlying assumptions.
Load time series of synthetic ACTIVSg10k and ACTIVSg2000 systems are first compared to the EIA-930 U.S. electric system operating data [@ElectricSystemOperationData], as they share the same geographic footprint.
Shown in Table \[table:validation\_table\], basic statistics of load time series of synthetic load time series and actual system that share the same footprint are compared. The actual load time series of United States Western Interconnection is obtained as the sum of all the balancing authorities in the region.
----------- ---------- ---------- --------- ---------
ERCOT
\[0.2cm\] 82.8 GW 88.1 GW 40.0 GW 48.8 GW
\[0.1cm\] 136.2 GW 132.4 GW 71.0 GW 71.2 GW
\[0.1cm\] 58.9 GW 56.5 GW 25.1 GW 36.0 GW
\[0.1cm\] 0.160 0.178 0.235 0.133
\[0.1cm\]
----------- ---------- ---------- --------- ---------
: Validation of Synthetic ACTIVSg System Load
\[table:validation\_table\]
Validation metrics that are generic and independent from geographic locations are summarized using statistical characteristics found in public load data of 37 European countries [@OpenPowerSystemData] and 66 United States Balancing Authorities [@ElectricSystemOperationData], so that synthetic load time series without a geographic footprint or have no availability of actual load time series can also be validated.
Load factors
------------
Load factor is defined as the ratio of average and peak value of a load time series. It is one effective metric to quantitatively validate the overall shape of the synthetic load profile. For profiles with relatively constant load level, such as regions with a high industrial composition, load factors are usually higher; while heavily residential areas tend to have lower load factors due to light occupation during the day [@LoadFactor].
The range of load factors of each month is summarized from public load data mentioned above and shown in figure \[fig:BA\_lf\] as the green shaded region. It is observed that as a general trend, the value of load factors are slightly higher in summer months, due to the increase of base electricity consumption from spacing cooling. There is also a consistent difference between the lowest and highest load factors of actual load every month, where systems with smaller size and less industrial load often have lower load factors.
The load factors of ACTIVSg10K and ACTIVSg2000 load time series lie inside the range observed from actual load time series, and also follow the same monthly trend.
Load distribution curves
------------------------
Load distribution curves show the percentage of time that load is at different levels. The load time series is normalized based off its mean value, where load levels exceeding yearly average would have per unit values larger than one.
The green shaded band in Figure \[fig:load\_disrtibution\] shows the range of load distribution curves found in real load time series, where load levels are scattered in between 0.4 and 1.8 per unit, with a denser distribution in the range from 0.8 to 1.2. The distribution of ACTIVSg10K and ACTIVSg2000 load time series follow the same general trend, load at most of the time points are within 0.8 to 1.2 times its yearly average.
Autocorrelations
----------------
Autocorrelation exhibits the relationship between time points of load time series that are certain time lags apart. It provides a validation perspective in time sequence order, instead of observing time series values as if they are independent recordings.
Figure \[fig:load\_autocorr\] shows the autocorrelations of actual and ACTIVSg synthetic load time series for time lags up to 48 steps. According to the real load time series data, the autocorrelation plot appears to be a sinusoidal-like wave of 24-hour cycle, with its magnitude slightly decreasing every cycle. The plot of ACTIVSg2000 synthetic load’s autocorrelation exhibits a similar trend as the actual system, where autocorrelation drops from 1 to around 0 when time lag increases from 0 to 12. Also, it increases back from 0 to 1 when time lag future increase from 12 to 24. The ACTIVSg10K synthetic load’s autocorrelation also shares the same sinusoidal trend as real load time series’s, while the lowest autocorrelation value would become negative around 12 hour time lag. This discrepancy is potentially introduced from the time shifting of prototypical building load time series during the load aggregation step.
Validation of renewable generation time series
----------------------------------------------
Similar to load time series, the created synthetic renewable generation time series is compared to publicly available renewable generation statistics and the characteristics found in actual system-level renewable time series for validation.
Due to the correlation of renewable energy output and the climate at the site location, the capacity factors of synthetic renewable time series are only compared to the statistics found on the same footprint. The auto-correlation is compared to actual system time series.
### Capacity factors
The capacity factor is defined as the ratio of utilized capacity to the installation capacity.
The actual wind generation capacity factor is obtained from EIA [@windseasonalpattern]. It sections the United States into 7 regions with unique wind generation profiles. The geographic footprints of ACTIVSg2000 and ACTIVSg10K synthetic system lie within the Northwest, California, Lower Plains, and “the rest of United States” regions, thus the median capacity factors of synthetic wind generation time series are compared with the reference capacity factor curve in the corresponding region, shown in figure \[fig:wind\_cf2\].
These seasonal capacity factor patterns differ by regions since the wind patterns range across location. As a general trend, wind capacity factor usually increases or stays flat from January to April, fall through August or September, and then increases again the remainder of the year [@windseasonalpattern]. However, wind capacity factor in California exhibits an opposite pattern [@windseasonalpattern]. The capacity factors of synthetic wind time series are observed to have similar trends in all footprint regions.
Yearly solar capacity factors of U.S. States are used for the validation of synthetic solar time series [@eiaepmr]. States with more than 10 solar plants are assorted individually, those with less than 10 plants are categorized as “other states”. According to this, solar time series in ACTIVSg2000 and ACTIVSg10K synthetic systems are classified as Texas, California, Nevada, Arizona, New Mexico, and “other states”. The yearly capacity factors from actual statistics and synthetic solar data exhibit similar trends by states as illustrated in Figure \[fig:solar\_cf\].
The difference in system-level wind and solar capacity is potentially the cause of the capacity factor discrepancy between the real and synthetic time series.
### Autocorrelations
The hourly renewable generation time series from 37 European countries are utilized to find reference autocorrelation curves for the validation of synthetic wind, solar and hydro data [@OpenPowerSystemData]. The autocorrelations of actual and ACTIVSg synthetic renewable time series were shown as Figure \[fig:generation\_autocorr\].
The magnitude of the actual wind autocorrelation gradually decreases from 1 to around 0.3. The ACTIVSg2000 and ACTIVSg10k synthetic wind autocorrelation curves are decreasing in a sinusoidal shape but they coincide at near 0.3.
The plots of the actual solar and hydro autocorrelations seem to be sinusoidal-like wave of 24-hour cycle. Their magnitudes fall when time lag increases from 0 to 12 and rise when time lag increases back from 12 to 24. The ACTIVSg2000 and ACTIVSg10k synthetic solar and hydro also have the same sinusoidal trend as their actual counterparts. The curves are also decreasing until 12 and increasing from 12 to 24, which is the same pattern as the actual system.
Synthetic Time Series Applications Example
==========================================
The creation of synthetic time series enables the potential of scenario development, time sequence simulations, and wide-area visualizations of large-scale power systems. Those time series data reflect the typical behavior of electric load in a hourly manner for the whole year, which is fundamental for power system steady-state analysis desired at different times and of different durations. The buses each load time series locates also cover a wide range of North America regions and have longitudes and latitudes. This allows non-uniform alterations of the time series to construct realistic power system scenarios. The alterations can be made considering the coupling of power system and location- based factors such as weather and major events. Along with transmission system models, generator cost functions, and other system data, synthetic load time series can facilitate power flow, optimal power flow analysis, as well as unit commitment of various scenarios.
As an example, this paper utilizes the bus-level load time series in the ACTIVSg10K system as the benchmark and develops a high behind-the-meter (BTM) solar scenario in the Western United States region. The BTM solar energies are solar generating units on the consumer’s side of the retail meter that serve all or part of the customer’s retail load with electric energy [@BTMG]. They are often treated as “negative loads” to be subtracted from the total load at the demand side. The increasing capacity of BTM solar installation changes the shape of daily net load on the system-level, when solar generation peaks at midday, the net load is low and when solar generation trails off at the end of the day, the total demand ramps quickly upward [@DuckCurve]. This new load shape is often referred as a “duck curve”. It has raised concerns on a conventional power system’s ability to accommodate the ramp rate and range needed to effectively supply the load and fully utilize the renewable energy [@NRELDuckChart].
The bus-level load time series in the synthetic power system is utilized to construct rare and extreme scenarios that can be useful to study this impact over time. Based off the composition of bus load type and the location-based solar potentials, the benchmark load time series at each bus is altered so that a system-level “duck curve” is created for the ACTIVSg10K system. This “duck curve” scenario can be used as the input of power flow analysis and unit commitment to provide analytical insights on transmission line loadings, generator dispatches schedules, system costs, and other system conditions.
This example uses a 24-hour time period in late spring from the benchmark hourly time series to develop a daily duck curve since such scenario usually occurs during the spring and summer seasons [@DuckCurve]. The BTM solar generation capacity is set to be 30,000 MW in the ACTIVSg10K system, and is distributed among load buses weighting their load sizes and the documented average solar resource outputs. Since most BTM solar installations are in the non-industrial sectors, only the size of residential and commercial load on each bus are considered to calculate the weights indicating bus solar potential, where buses with higher combined load are assigned with a higher peak BTM solar generation. On the other hand, as the solar potential is also dependent on solar radiance that varies with geographic locations, a solar resource data set from National Renewable Energy Laboratory (NREL) is also utilized to determine the weight to distribute the system BTM solar capacity [@NRELGIS]. This data set provides the monthly average solar output $(kWh/m^2/day)$ at each zip code location.
The solar potential that determines the BTM solar capacity at each bus is calculated as the weighted summation of normalized load size and normalized solar resource output in equation (\[eq:4\]) - (\[eq:5\]). To construct the 24-hour “duck curve” scenario in the ACTIVSg10k synthetic system, a BTM solar output time series is created and subtracted from the original load time series of each bus. The bus-level BTM solar capacity is considered as the peak value of each BTM solar output time series that would occur at a random time step in between 1 pm and 3 pm. The starting and ending time step of solar output are randomly chosen from 6 to 8 am, and 6 to 8 pm respectively. Before the starting and after the ending time point, the BTM solar output is zero.
$$\label{eq:4}
\begin{split}
BTM \: solar \: potential(i) & = X_1 \frac{load\:size(i)}{max(load\:size)} \\
& + X_2 \frac{solar\:resource(i)}{max(solar\:resource)}
\end{split}$$
$$\label{eq:5}
\begin{split}
BTM \: solar \: capacity(i) & = system \: BTM \: solar \: capacity \times \\
& \frac{BTM \:solar \: potential(i)}{\sum_{i = 1}^{Nload}BTM \:solar \: potential(i)}
\end{split}$$
where:
[$X_1,X_2$]{} weights on bus load size and solar resource\
The load size and solar resources data, as well as the BTM solar capacity determined for buses in ACTIVSg10K system are shown in the contour plots Figure \[fig:contour\]. It can be observed that most of the buses with higher BTM solar capacity are locations with high solar resources. Besides, major metropolitan areas with dense residential and commercial demands also have higher BTM solar capacities. The benchmark and duck curve load time series on the system-level are shown in Figure \[fig:duckcurve\].
Conclusions
===========
This paper proposed a methodology to synthesize and validate bus-level load time series in the existing synthetic power systems. The creation of time series uses an iterative bottom-up approach. Based on the geographic location and load type composition of each bus, prototypical building and facility level time series are integrated to construct a bus-level time series with unique profiles. Each time series has hourly resolution, and spans for a year. To validate and improve the realism and quality of synthetic load time series, actual load time series obtained from electric systems of different sizes are analyzed statically so that representative and comprehensive set of validation metrics can be developed.
Since the data set utilized in the synthesizing process is publicly available, the created time series can be accessed and distributed freely without any confidentiality concerns. The wide geographic coverage, time resolution and duration of bus-level time series enable its versatile applications in system scenario development and studies. As an example, this paper demonstrated the construction of a “duck curve” scenario in ACTIVSg10K system using the bus-level load time series as the benchmark.
Acknowledgment {#acknowledgment .unnumbered}
==============
This work was supported in part by the U.D. Department of Energy Advanced Research Projects Agency-Energy (ARPA-E) under the GRID DATA project.
[Hanyue Li]{}(S‘14) received the B.Sc. degree in electrical engineering from Illinois Institute of Technology, Chicago, IL, USA, in 2016, and the M.Sc. degree in electrical and computer engineering in Carnegie Mellon University, Pittsburgh, PA, USA, in 2017. She is currently working toward the Ph.D. degree in electrical engineering at Texas A&M University, College Station, TX, USA.
[Ju Hee Yeo]{} (S‘17) received the B.Sc. degree in electrical engineering from Sangmyung University, Seoul, South Korea, in 2017. She is currently working toward the Ph.D. degree in electrical engineering at Texas A&M University, College Station, TX, USA.
[Ashly L. Bornsheuer]{} (S‘17) received the B.Sc. degree in electrical engineering from Texas A&M University, College Station, TX, USA in 2018.
[Thomas J. Overbye]{} (S‘87-M‘92-SM‘96-F‘05) received the B.Sc., M.Sc., and Ph.D. degrees in electrical engineering from the University of Wisconsin-Madison, Madison, WI, USA. He is currently a TEES distinguished research Professor in the Department of Electrical and Computer Engineering at Texas A&M University, College Station, TX, USA.
[^1]: The authors are with the Department of Electrical and Computer Engineering, Texas A&M University, College Station, TX, 77840 USA (e-mail:[email protected], [email protected], [email protected], [email protected]).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present results from a near/mid IR search for submillimetre galaxies over a region of 6230 sq deg. of the southern sky. We used a cross-correlation of the VISTA Hemispheric Survey (VHS) and the WISE database to identify bright galaxies (K$_s\leq$ 18.2) with near/mid IR colours similar to those of the high redshift lensed sub-mm galaxy SMM J2135-0102. We find 7 galaxies which fulfill all five adopted near/mid IR colour (NMIRQC) criteria and resemble the SED of the reference galaxy at these wavelengths. For these galaxies, which are broadly distributed in the sky, we determined photometric redshifts in the range z=1.6-3.2. We searched the VHS for clusters of galaxies, which may be acting as gravitational lenses, and found that 6 out of the 7 galaxies are located within 3.5 arcmin of a cluster/group of galaxies. Using the J-K$_s$ vs J sequences we determine photometric redshifts for these clusters/groups in the range z=0.2-0.9. We propose the newly identified sources are ultra-bright high redshift lensed SMG candidates. Follow-up observations in the sub-mm and mm are key to determine the ultimate nature of these objects.'
author:
- |
S. Iglesias-Groth$^{1,2}$[^1], A. Díaz–Sánchez$^{3}$, R. Rebolo$^{1,2}$ and H. Dannerbauer$^{1,2}$[^2]\
$^{1}$ Instituto de Astrofísica de Canarias, La Laguna 38200, Spain\
$^{2}$Departamento de Astrofísica de la Universiad de La Laguna, Avda. Francisco Sánchez, La Laguna, 38200, Spain\
$^{3}$Departamento Física Aplicada, Universidad Politécnica de Cartagena, Campus Muralla del Mar, 30202 Cartagena, Murcia, Spain
date: 2016
title: 'A near/mid infrared search for ultra-bright submillimetre galaxies: Searching for Cosmic Eyelash Analogues'
---
\[firstpage\]
galaxies – submillimetre: galaxies.
Introduction
============
Since their discovery (Smail et al. 1997; Hughes et al. 1998) the so-called submillimetre galaxies (SMGs) are a vital galaxy population in order to understand the formation and evolution of massive galaxies in the distant universe (see for a review Casey et al. 2014). These systems have extreme star formation rates of several hundred to thousand solar masses per year (e.g., Magnelli et al. 2012), are molecular gas-rich with M$_{mol-gas}$= $\sim$ few times 10$^{10}$ M$_{\odot}$ (e.g., Greve et al. 2005; Bothwell et al. 2013) and the redshift distribution peaks $\approx$ $z=2.2-3.0$ (Chapman et al. 2005; Simpson et al. 2014; Miettinen et al. 2015; Strandet et al. 2016) depending on the selected wavelength in the (sub)mm window. Typical sizes of these dusty starbursts range between R$_{e}=0.6-2.0$ kpc (Ikarashi et al. 2015; Simpson et al. 2015; Hodge et al. 2016). Mergers and cold gas infall are potential agents of such star formation, however the available studies have not yet established the main cause of the high star-formation rates (e.g., Swinbank et al. 2008; Hayward et al. 2013; Michalowski et al. 2012; Narayanan et al. 2015).
Despite intrinsic luminosities of L$_{bol}$$\geq$ few times 10$^{12}$ to 10$^{13}$ L$_{\odot}$ (Magnelli et al. 2012; Ivison et al. 2013), these dusty galaxies are very challenging objects for observational studies in the near IR (e.g., Dannerbauer et al. 2002, 2008; Younger et al. 2007) and mid infrared (e.g. Pope et al. 2008). The brightest unlensed SMGs have observed flux densities of up to $\sim$ 10 mJy at 850 $\mu$m (e.g., Karim et al. 2013). Gravitational lensing via massive galaxy clusters can enhance the apparent brightness of SMGs without altering their colours (Smail et al. 1997) and cluster surveys (e.g. Smail et al. 2002; Johansson et al. 2011) indeed led to the detection of SMGs with high amplification factors (30-40) like that of the outstanding galaxy SMM J2135$-$0102 at $z=2.3259$ (the Cosmic Eyelash, hereafter SMM J2135, Swinbank et al. 2010; Ivison et al. 2010; Danielson et al. 2011). The South Pole Telescope and the Herschel and Planck space missions have unveiled similarly bright lensed SMGs (Negrello et al. 2010; Vieira et al. 2010, 2013; Weiss et al. 2013; Cañameras et al. 2015; Harrington et al. 2016; Strandet et al. 2016). In spite of the amplification factors provided by lensing, these SMGs are still rather faint in the optical and near-IR as a consequence of internal dust obscuration, and their properties in this spectral range are not well understood yet (e.g. Dannerbauer et al. 2002, 2004, 2008; Dunlop et al. 2004; Younger et al. 2007; Walter et al. 2012; Simpson et al. 2014; Hodge et al. 2015; Dye et al. 2015). The spectral energy distribution of bright high redshift sources like SMM J2135$-$0102 (hereafter SMM J2135) shows a very steep increase in flux as we move from the optical to the near-IR and mid-IR which could potentially be used to identify other similar galaxies.
We took advantage of the VISTA Hemisphere Survey (VHS)[^3] which has already covered an area of more than 8000 sq. deg. of the southern sky at near-IR bands and the Wide-field Infrared Survey Explorer (WISE) mission archive database in the mid-IR in order to find brighter analogues of SMM J2135. We adopted as a reference the near and mid-IR colours of this galaxy. Given the limiting magnitudes of the VHS/WISE surveys we expect the selected sample of similar colour galaxies will consist mainly of gravitationally lensed sources (Blain 1996, Negrello et al. 2007). Lensing of SMGs is strongly favored due to the steep number counts of this source population (Blain et al. 1996; Perrotta et al. 2002, 2003; Negrello et al. 2007; Negrello et al. 2010). Thus, if we conduct searches for sources with similar SEDs as known dusty starbursts in the near/mid IR wavelength regime, the above mentioned lensing bias should guarantee that the majority of the selected sources are massive galaxies with strong FIR/submm emission triggered by on-going intense star formation. Thus, the contamination by massive, early type galaxies should be rather negligible.
Our goal is to find the brightest high redshift (z$\sim$2) analogues of SMM J2135 in the sky such that follow-up studies can be performed in the optical, near/mid IR with moderate investment of observing time at the largest telescopes. Finding new examples of this important class of strongly lensed star-forming galaxies should help to determine their extinction properties in the visible/IR, perform extensive characterisation in the millimetre/submillimetre range and obtain a better understanding of how and where stars are formed in these galaxies. Brighter analogues of the SMG SMM J2135 in the redshift range z=2-3 will also allow us to investigate star formation on 100 pc scales as a function of redshift using ALMA (Dye et al. 2015, Swinbank et al. 2015).
In the following sections we describe our search for bright SMGs in a large fraction of (approx. 30 %) of the southern sky and the resulting candidate galaxies. In Section 2 we present the infrared data from VISTA/VHS and WISE (Vega mag-system) used in the search, the photometry of the reference galaxy SMM J2135 and the adopted colour criteria to identify potential analogues of this galaxy. Section 3 reports the results of the search and the identification of clusters of galaxies close to the new candidate SMGs.
$^1$ From the VHS J-band image.
$^2$ Transformed from the Spitzer \[3.6\] band using the template SED of SMM J2135.
$^3$ Transformed from the Spitzer \[4.5\] band using the template SED of SMM J2135.
$^4$ Transformed from the Spitzer \[8\] band using the template SED of SMM J2135.
{width="18cm" height="4cm"}
{width="8cm"}
The Search
==========
We will take advantage of the large sky coverage provided by the near-IR VISTA Hemispheric Survey (VHS) and the mid-IR WISE mission to conduct a search for bright analogues of the SMG SMM J2135.
The VISTA Hemispheric Survey: VHS
---------------------------------
The VHS is a near-infrared ESO public survey designed to map the entire Southern hemisphere in the J and K$_s$ broad band filters with average 5$\sigma$ depths of J = 19.5 $\pm$ 0.3 mag and K$_s$ = 18.5 $\pm$ 0.3 mag. In some particular areas also Y and H band observations are performed. The 4-m VISTA telescope (Emerson 2001; Emerson et al. 2004) operates since 2009 at ESO’s Cerro Paranal Observatory in Chile and has so far covered a sky area of about 8000 deg$^2$. It is equipped with a wide-field infrared camera VIRCAM (Dalton et al. 2006) composed of 16 Raytheon detectors 2048-2048 pixel array each, with a mean plate scale of 0.34 arcsec, giving a field of view of 1.65 degrees in diameter.
The VHS images are processed and calibrated automatically by a dedicated science pipeline implemented by the Cambridge Astronomical Survey Unit (CASU). Standard reduction and processing steps include dark and sky subtraction, flat field correction, linearity correction, destripe and jitter stacking. For a detailed description we refer to the CASU webpage http://casu.ast.cam.ac.uk/surveys-projects/vista as well as to Irwin et al. (2004) and Lewis et al. (2010). The photometry provided in the VHS catalogue is calibrated using the magnitudes of colour-selected 2MASS stars converted onto the VISTA system using colour equations including terms accounting for interstellar reddening. We used the catalogue J and K$_s$-band 1 arcsec aperture corrected (aperMag3) magnitudes for the selection of targets.
The Wide-field Infrared Survey Explorer
---------------------------------------
The NASA mission WISE completed an all-sky survey in four mid-IR bands (Wright et al. 2010) centred at 3.4, 4.6, 12, and 22 $\mu$m (referred as W1, W2, W3 and W4, respectively) with an angular resolution of 6.1$Ó$, 6.4$Ó$, 6.5$Ó$, and 12.0$Ó$ and 5$\sigma$ point source sensitivities of 0.07, 0.1, 0.86 and 5.4 mJy, respectively. The WISE photometric pipeline provides a number of quality flags useful in selecting point sources and extended sources. We use the AllWISE Source Catalog PSF profile-fit measurements that are made by chi-squared minimization on the “stack” of all Single-exposure frames in all bands covering a deep source detection [^4]. Only measurements with S/N$>$3.0 will be considered.
{width="8cm"}
{width="14cm"}
![Colour-colour diagrams for our 7 candidates, SMM J2135, three SPT (South Pole Telescope) sources with redshift 2.23, 2.51 and 2.78, and a z=2.79 SMG lensed by a galaxy cluster given in Gonzalez et al. (2011). The grey points are field sources. The colour cuts in Table 3 are given by point-dashed lines and typical error bars are shown in each diagram. We have used continuous lines to show the position of the SMM J2135 GRASIL SED and Arp220 SED as a function of redshift (indicated by a number, crosses are drawn in the curves for each 0.5 increment of redshift).[]{data-label="fig:f5"}](fig5a.pdf){width="6.5cm"}
![Colour-colour diagrams for our 7 candidates, SMM J2135, three SPT (South Pole Telescope) sources with redshift 2.23, 2.51 and 2.78, and a z=2.79 SMG lensed by a galaxy cluster given in Gonzalez et al. (2011). The grey points are field sources. The colour cuts in Table 3 are given by point-dashed lines and typical error bars are shown in each diagram. We have used continuous lines to show the position of the SMM J2135 GRASIL SED and Arp220 SED as a function of redshift (indicated by a number, crosses are drawn in the curves for each 0.5 increment of redshift).[]{data-label="fig:f5"}](fig5b.pdf "fig:"){width="6.5cm"} ![Colour-colour diagrams for our 7 candidates, SMM J2135, three SPT (South Pole Telescope) sources with redshift 2.23, 2.51 and 2.78, and a z=2.79 SMG lensed by a galaxy cluster given in Gonzalez et al. (2011). The grey points are field sources. The colour cuts in Table 3 are given by point-dashed lines and typical error bars are shown in each diagram. We have used continuous lines to show the position of the SMM J2135 GRASIL SED and Arp220 SED as a function of redshift (indicated by a number, crosses are drawn in the curves for each 0.5 increment of redshift).[]{data-label="fig:f5"}](fig5c.pdf "fig:"){width="6.5cm"}
Reference galaxy
----------------
As a reference for our search we adopted the near and mid-IR colours of SMM J2135. This extensively studied galaxy is lensed by a cluster of galaxies with no evidence for any intervening galaxy in the line of sight which could be affecting its near/mid-IR colours. SMM J2135 has an intrinsic rest-frame 8-1000 $\mu$m luminosity, L$_{bol}$, of 2.3 x 10$^{12}$ L$_{\odot}$ and a star formation rate estimated at $\sim$ 400 M$_{\odot}$/yr (Swinbank et al. 2010, Ivison et al. 2010). In the far-IR the SED can be described by dust components with temperatures of 30 and 60 K. A comparison with local ultra-luminous infrared galaxies like Arp 220 (Ivison et al. 2010) reveals that SMM J2135 is slightly less luminous in the rest-frame optical to mid-IR (possibly due to stronger dust extinction) while rather similar in the mm/sub-mm range.
Fig. 1 shows VHS images in J, H, K$_s$ and WISE images in the W1, W2, W3 and W4 bands of SMM J2135. No detection of this source was reported in the VHS and AllWISE catalogues, however after close inspection of the relevant images, SMM J2135 can be detected in the K$_s$-band and in the W3 and W4 bands. Subsequently, we measured its magnitudes using aperture photometry and obtained $K_s=18\pm 0.3$, $W3=12.6\pm 0.3$ and $W4=8.8 \pm 0.3$ (Vega system). An upper limit on the J-band magnitude was also derived from the corresponding VHS image. The photometry was calibrated using nearby sources in the catalogue free of contaminants. In the H, W1 and W2 bands we achieve a marginal detection of SMM J2135 using the Laplacian of Gaussian filter, often called ’mexican hat’, which helped to improve the detectability of sources in these filters. However, no reliable magnitudes could be determined from these images and the W1 and W2 magnitudes given in Table 1 were finally estimated converting the fluxes measured by Ivison et al in the \[3.6\] and \[4.5\] Spitzer bands, respectively. The final adopted fluxes and magnitudes of SMM J2135 are listed in Table 1.
The fluxes of SMM J2135 from the optical to centimetre wavelengths are plotted in Fig. 2 (black dots) in comparison with its SED as modeled by GRASIL from Lapi et al. (2011) and the SED of Arp 220 redshifted to z=2.32 with appropriate normalisation. The blue squares indicate our measurements in the K$_s$, W3 and W4 bands.
Search method
-------------
In order to find bright analogues of SMM J2135 we have correlated the AllWISE Source Catalogue with about 8000 $deg^2$ available from the VHS catalogue. The search is restricted to galactic latitudes $|b|\geq$20$^{o}$, reducing the surveyed area to 6230 sq. deg., we retain galaxies for which we have detections in all the WISE bands and identify counterparts in the VHS catalogue within 1 arcsec of the AllWISE targets. To avoid bad candidates due to Galactic extinction, the VHS and AllWISE magnitudes are corrected for Galactic extinction with the use of the Schlegel et al. (1998) maps and the $A_{\lambda} /E(B-V)$ coefficients taken from Cardelli et al. (1989). Then, we selected those galaxies with colours consistent with the colour ranges listed in Table 2, i.e. with colours within $\pm$ 0.5 mag of the SMM J2135 colours. We also restricted the selection to objects with S/N$\geq$3 in all the WISE bands. A total of 7 galaxies verified all these conditions.
Very recently, colour selection procedures have been refined successfully to find SMGs (see Chen et al. 2016 and references therein). These authors used the so-called Optical-IR Triple Colour (OIRTC) selection procedure. We can also adopt a similar approach, imposing just one limit for each single colour in order to select SMGs, i.e take away one of the limits in each of the colours listed in Table 2. When we adopt the colour limits of Table 3 we find the same candidates. Therefore we adopt for simplicity the colour criteria of Table 3 which we could call the NIR/MIR Quintuple Colour (NMIRQC) procedure.
In Table 4 we list photometry for the 7 resulting SMG candidates. Their sky positions are shown in figure 3. Images of these candidates extracted from the surveys VHS and WISE are plotted in Fig. 4.
All our candidate sub-mm galaxies have well measured positions in the NIR images (accuracy better than 0.3$^{\prime\prime}$) and are the only likely counterpart of the WISE sources. Only candidate 3 appears to be double in the J and K$_s$ images. We note that this NIR/MIR selection could be advantageous with respect to other techniques in the sub-mm range which usually require follow-up interferometry to determine subarcsec position of the candidate galaxies.
Results and Discussion
======================
The 7 candidates that fulfill the colour criteria of the lensed sub-mm galaxy SMM J2135 present $J$-band magnitudes in the range 18.8-20.4 and $K_s$-band magnitudes in the range 16.8-18.2. The candidates are well identified in all bands. In spite of the different spatial resolution of the VHS and WISE, there is good agreement in the positions of the identified sources.
In Figure 5 we plot colour-colour diagrams for the 7 candidates, SMM J2135, three SPT (South Pole Telescope) sources we have detected in VHS/WISE and one additional SMG lensed by a galaxy cluster reported by Gonzalez et al. 2011 (in this last case we have transformed Spitzer magnitudes to WISE magnitudes and there is a lower limit to the K$_s$ magnitude). The grey points are field sources. The colour cuts in Table 3 are given by point-dashed lines and typical error bars are shown in each diagram. We have also represented with continuous curves the SMM J2135 GRASIL SED and the Arp220 SED as a function of redshift (redshift is indicated by a number and crosses are drawn in the curves for each 0.5 increment of redshift). The three SPT sources have redshift 2.23, 2.51 and 2.78, they are lensed by foreground galaxies. While their WISE colours fall into the colour ranges we have adopted for SMGs, in the VHS bands there is contamination by the intervening lensing galaxy and the K$_s$-W1 colours lie out of the adopted range. The colours of SMM J2135 (z=2.32) and the SMG identified by Gonzalez et al. 2011 (z=2.79) do probably represent the true colours of the original lensed galaxy because they are lensed by galaxy clusters and therefore photometric contamination in the near-IR bands is not expected. From these diagrams we can see that the redshift of our selected sources are likely between z$\sim$ 1.6 and z$\sim$ 3.2. Also, it is shown that the contamination of our sample by field sources is not large given the distant location of the adopted colour cuts with respect the bulk of the field sources. The SMM J2135 GRASIL SED does not match well the W3 band measurement, as we can see in figure 2, and colour differences between the SED and SMM J2135 come from this. Nevertheless the extreme values in colours W3-W4 and W2-W3 are obtained for redshift z$\sim$ 2.5, rather far from the colour-colour regions where most field sources are located.
In order to discuss the potential contamination of galaxies other than SMGs we have used the SWIRE template library (Polletta et al. 2007 ) which contains 25 templates including 3 ellipticals, 7 spirals, 6 starbursts (SB), 7 AGNs (3 type 1 AGNs, 4 type 2 AGNs), and 2 composite (starburst+AGN). The 3 ellipticals are 2, 5 and 13 Gyr old and the 7 spirals range from early to late types (S0-Sdm). Templates of moderately luminous AGN, representing Seyfert 1.8 and Seyfert 2 galaxies, were obtained by combining models, and spectra of a random sample of 28 Seyfert galaxies. The other six AGN templates include three templates representing optically-selected QSOs with different values of infrared/optical flux ratios and two type 2 QSOs. The composite (AGN+SB) templates are empirical templates, these objects contain a powerful starburst component, mainly responsible for their large infrared luminosities and an AGN component that contributes to the mid-IR luminosities.
In Figure 6 we can see that only spiral galaxies are located in the selection region in the $W3-W4$ vs. $W2-W3$ colour-colour diagram at redshift $z \sim 2$ (top panel) but in the $W3-W2$ vs. $K_s-W1$ diagram (central panel) these galaxies only go into the selection region for $z\sim 4$. So they do not fulfill simultaneously both colour conditions at the same redshift and therefore are not a likely source of contamination. The bottom panel of figure 6 shows the $W3-W4$ vs. $W2-W3$ colour-colour diagram for the 7 AGNs and the 2 composite SB+AGN galaxies, none of them are in the selection region. The SED of SB galaxies is not too different from that of Arp220 and SMM J2135 and indeed these galaxies may enter the selection region for $1.7 \lesssim z \lesssim 3$.
In Tables 4 and 5 we complement the photometry of several SMG candidates with VHS measurements available in the H band and in the optical bands using CFHTLS. We have used redshifted GRASIL SEDs of SMM J2135 to determine the photometric redshift of our candidates, from J,H,K$_s$,W1,W2,W3,W4-band measurements (H-band when available) using a standard $\chi^2$ minimization procedure. The results are given in Table 4. All derived redshifts lie in the range z=1.6-3.2 as expected. We find 2 galaxies with K$_s$ $\leq$ 16.8, 3 galaxies in the range K$_s$=17-17.2 and 2 galaxies with K$_s$=17.5-18.2 Only 1 galaxy has photometric redshifts higher than z=3 in our sample. In general our selected candidates are about 1-2 magnitudes brighter than typical lensed SMGs selected from Herschel surveys (e.g. Ma et al. 2015a; Calanog et al. 2014) or the SPT survey (Ma et al. 2015b). However, one source, HATLAS J142935.3-002836 alias G15v2.19, at z=1.027 has similar NIR properties as our brightest candidates (Messias et al. 2014; Calanog et al. 2014; Ma et al. 2014).
In Fig. 7 we plot the SEDs of the candidate SMGs in comparison with the GRASIL SED of SMM J2135 shifted to the photometric redshift determined for each candidate. We note that in most cases the model fits very well all the near-IR and mid-IR bands. In the case of candidate 6 (VHSJ0902-0448), however, the fluxes deviate from the model at the shorter wavelength bands, probably because of the presence of an intervening galaxy in the line of sight which however is not detected in the available J and K$_s$ images. From these fits we can estimate peak sub-mm fluxes above $\sim$500 micron and a flux density at 1.4 mm above $\sim$ 15 mJy for all of our targets. Very few sources are expected to be intrinsically bright enough to exceed these fluxes (Karim et al. 2013) and therefore we may expect that our sample is conformed by gravitationally lensed galaxies.
Weiss et al. (2013) measured the redshift distribution of 26 strongly lensed sources selected from the 2500 square-degrees of the South Pole Telescope (SPT) survey with fluxes at 1.4mm above 20 mJy. They found a redshift distribution of the SPT sample with a mean of z=2.0, only 1 out of their 26 targets had redshift in the range z=2.5-3.2. Our targets have redshifts consistent with the SPT distribution, but only for 5 targets we estimate 1.4 mm fluxes above 20 mJy from the fits plotted in Fig. 7. The surface density we infer for SMG sources selected in the near/mid IR to match 1.4 mm fluxes above 20 mJy would be 5/6230 deg$^{-1}$ i.e. approximately 10 times less surface density that that found by Weiss et al. (2013) for potentially equivalent objects. We argue that this difference can be due to our selection criteria mostly identifying galaxies which are lensed by clusters of galaxies. Galaxies amplified by an intervening galaxy in the line of sight are likely to have photometric magnitudes in the J and K$_s$ bands affected by the lens and may not verify our selection criteria. Only a few cases in SPT are cluster lensing galaxies. Spilker et al. (2016) find that 4 of 47 SPT sources are lensed by clusters. The NMIRQC procedure, proposed here, could be a very effective tool to identify strongly lensed SMGs by clusters of galaxies. SMGs detected by Herschel have 500 $\mu$m fluxes in the range 100-350 mJy (Bussmann et al. 2013). For our candidates, we estimated 500 $\mu$m fluxes from the fitted SEDs, the results range from 160 to 980 mJy and are listed in Table 4. Follow-up sub-mm/mm observations of the candidates will demonstrate if indeed we are dealing with strongly lensed SMGs.
From the same SEDs we obtained 60 and 100 $\mu$m fluxes which used with the FIR/radio relation by Condon et al. (1992) allowed us to predict radio emission at 1.4 GHz in the range 0.09-0.36 mJy. To our knowledge no detection of radio emission has been reported in the literature for our targets.
Clusters of galaxies and sub-mm galaxy candidates
-------------------------------------------------
Using the red sequence method (Gladders et al. 2000) and the VHS survey we have searched for clusters of galaxies near our SMG candidates and identified several galaxy clusters in the range $0.2 < z < 0.9$.
We searched for local over densities of galaxies using the $J-K_s$ versus $J$ colour-magnitude diagram. For these redshifts, the slope of the red-sequence in the diagram is nearly $-0.03$ (Cohn et al., 2007). We took $J-K_s$ colour slices of amplitude $0.2$ magnitudes and searched for local overdensities in circular areas around the brightest cluster galaxy (BCG) with radius $\sim 1$ Mpc. We selected a cluster candidate if the density of galaxies of the cluster were larger than 3 times the mean density and had more than 10 galaxies in the red-sequence. We adopted as cluster radius the largest radius where the over density was detected.
In Figure 8 we can see the colour-magnitude diagrams for the six clusters of galaxies we have detected near SMG candidates 1, 2, 4, 5, 6, and 7, respectively. The $J-K_s$ colour of the BCG is an indicator of the redshift of the cluster (Tonini et al 2012). Only for candidate 3 we do not find a nearby cluster. Indeed this candidate is the only one that appears to be double in the VHS images of Figure 4, suggesting the presence of an intervening galaxy in the line of sight. We estimated the redshift of the clusters from the colours predicted by the red sequence fit (see Fig. 8) for each BCG. In Table 6 we present the main properties of all these clusters and group of galaxies, including reacheness and radius estimates. The VHS $J$-band images $10\times10$ arcmin$^2$ are shown in Fig. 9, blue circles denote clusters and are plotted with the cluster radii given in Table 6. The blue boxes mark the position of SMG candidates. Red circles denote the position of identified candidate galaxy members. Remarkablly candidates 1, 2 and 4 are located within clusters and candidate 5 is just at 1.8 arcmin from the centre of a cluster. The identification of clusters of galaxies within 3.5 arcmin of six out of 7 candidates further supports the suggestion that these objects are lensed SMGs.
{width="6.5cm"}
{width="6.5cm"}
{width="6.5cm"}
{width="16cm"}
{width="16cm"}
{width="16cm"}
----------------- ------------------ -------------------- -------------------- -------------------- -------------------- --------
ID $u$ $g$ $r$ $i$ $z$ Survey
6 VHSJ0319-4549 21.81 $\pm 0.01$ 21.161 $\pm 0.004$ 20.341 $\pm 0.003$ 20.083 $\pm 0.003$ 19.984 $\pm 0.006$ CFHTLS
----------------- ------------------ -------------------- -------------------- -------------------- -------------------- --------
$^1$ Estimate from the $J-K_s$ colour of the BCG and figure 3 in [@CS09; @TB12].
$^2$ Number of galaxies in the red-sequence of the cluster.
$^3$ The largest radius where the overdensity is detected in arcmin.
$^4$ Projected distance of the SMG to the geometric center of the cluster in arcmin.
$^5$ Overdensity factor from the mean denstity.
Conclusions
===========
We report results on a search for bright high redshift submillimetre galaxies , analogues of the lensed galaxy SMM J2135-0102, z=2.3, via a cross-correlation of the WISE and VHS databases performed over a high galactic latitude southern sky area of 6230 sq deg. We adopted as reference the near/mid-IR colours of this galaxy and searched for galaxies of similar colours with reported detections in all J, K$_s$, W1, W2, W3 and W4 bands. We find 7 galaxies brighter than K$_s$=18.2 matching the adopted colour criteria.
Using redshifted SEDs of the reference galaxy we determine the most likely redshift for each galaxy and find that our sample lie in the range z=1.6-3.2. From the best individual fits we estimate sub-mm and mm fluxes for each target and conclude that 5 out of the 7 galaxies may have fluxes above 20 mJy at 1.4 mm. The surface density that we determine for lensed SMGs ($\sim$ 1 galaxy per 890 sq. deg.) is about 10 times lower than that obtained by Weiss et al. (2013) for z=1.6-3.2 sub-mm galaxies with 1.4 mm fluxes above 20 mJy. The difference can be due to our selection criteria mostly selecting SMGs amplified by clusters of galaxies, as those amplified by intervening galaxies may have their infrared colours affected.
Using VHS and the J-K vs J cluster sequence method, we have identified potential clusters/groups of galaxies near 6 of these candidates. Photometric redshifts locate most of these nearby clusters in the range z=0.2-0.9.
We propose that our near/mid-IR selection procedure can identify good candidates to bright high redshift lensed SMGs. Follow-up sub-mm observations (already on-going with APEX) and mm-observations (with e.g. LMT and ALMA) will reveal if indeed the selected objects are bright lensed SMGs and will provide accurate spectral redshifts. If confirmed, these would be excellent targets to carry out a systematic determination of the properties in the optical, near-IR, mid-IR, millimetre and submillimetre of SMGs in the redshift range 1.6-3.2.
The near/mid IR colour selection used in this work can potentially be extended in redshift space, particularly to higher redshifts where the redshift distribution of SMGs displays a peak (e.g. Bèthermin et al. 2015). Higher redshift searches increasingly demand higher sensitivity near-IR data. Future large scale deep near-IR surveys, as planned for the Euclid mission (ESA), in combination with WISE may enable the detection of a large population of bright lensed SMGs at various redshifts.
Acknowledgements {#acknowledgements .unnumbered}
================
This work has been partially funded by projects “Participacion en el instrumento NISP y preparacion para la ciencia de EUCLID”, ESP2015-69020-C2-1-R (MINECO). H.D. acknowledges financial support from the Spanish Ministry of Economy and Competitiveness (MINECO) under the 2014 Ramón y Cajal program MINECO RYC-2014-15686
[100]{} Blain, A. W., 1996 MNRAS 283, 1340 Bothwell, M. S. et al. 2013, MNRAS, 429, 3047 Bèthermin, M. et al., 2015, A&A, 573, 113 Bussmann, R. S. et al. 2013, ApJ 779,25 Calanog, J. A., et al. 2014, ApJ, 797, 138 Cardelli, Jason A., Clayton, Geoffrey C. and Mathis, John S., 1989, ApJ 345, 245. Casey, C. M., Narayanan, D., & Cooray, A. 2014, PhR, 541, 45 Cañameras et al. 2015, A&A, 581, 105 Chapman, S. C., Blain, A. W., Smail, Ian & Ivison, R. J., 2005, ApJ 622, 772 Chen, C.-C., et al.,2016, ApJ,820, 82
Cohn, J.D. et al., 2007, MNRAS, 382, 1738 Condon,J.J.,1992,ARA&A,30,575 Dalton, G. B., Caldwell, M., Ward, A. K., et al. 2006 SPIE, 6269, 62690 Danielson, A. et al. 2011, MNRAS, 410, 1687 Dannerbauer, H., Lehnert, M. D., Lutz, D., et al. 2002, ApJ, 573,473
Dannerbauer, H. et al. 2008, ApJl, 673, 123 Dunlop, J. et al. 2004, MNRAS, 350, 769 Dye, S. et al.,2015 MNRAS 452, 2258 Emerson, J.P 2001, ASPC 232, 339E Emerson, J.P et al. 2004, SPIE.5,493, 401E Gladders, M.D., and Yee, H.K., 2000, AJ, 120, 2148 Gonzalez, Anthony H., Papovich, Casey, Bradac, Marusa, Jones, Christine, 2011, ApJ 720, 245 Greve, T. R. et al. 2005, MNRAS, 359, 1165 Harrington, K. C. et al, 2016, MNRAS 458, 4383 Hayward, C. C., Narayanan, D. ,Kereŝ, Dusân, Jonsson, Patrik, Hopkins, Philip F., Cox, T. J.& Hernquist, L. 2013, MNRAS 428, 2529 Hodge, J. A. et al. 2015, ApJL, 798, 18 Hodge, et al. 2016, ApJ, in press (astro-ph/1609 Hughes, D. H. et al., 1998, Nature, 394, 241 Ikarashi, S. et al. 2015, ApJ, 860, 133 Irwin, M. J., Lewis, J., Hodgkin, S., Bunclark, P., Evans, D., McMahon, R., Emerson, J. P., Stewart, M.& Beard, S., 2004, SPIE 5493, 411 Ivison R. J. et al., 2010, A&A, 518, L35 Ivison, R. et al. 2013, ApJ, 772, 137 Johansson, D. et al. 2011, A&A, 527, 117 Karim, A. et al. 2013, MNRAS, 432, 2
Lapi, A. et al., 2011, ApJ 742, 24L Lewis, J. R., Irwin, M. J. & Bunclark, P. S. 2010, ASP Conf. Ser., 434, 91 Ma, B., et al. 2015, ApJ, 814, 17 Ma, J., et al. 2015, ApJ, 812, 88 Magnelli, B. et al. 2012, A&A, 539, 155 Messias, H., et al. 2014, A&A, 568, 92 Michalowski, M. J.,Dunlop, J. S., Cirasuolo, M., Hjorth, J., Hayward, C. C.& Watson, D, 2012, A&A, 541, 85 Miettinen, O. et al. 2015, A&A, 577, 29 Narayanan, D. et al. 2015, Nature, 529, 496 Negrello, M., Perrotta, F., Gon$\acute{\rm{z}}$alez-Nuevo, J., Silva, L.,de Zotti, G., Granato, G.L., Baccigalupi, C., Danese, L.,2007, MNRAS, 277, 1557 Negrello, M. et al., 2010, Science, 330, 800 Perrotta, F., Baccigalupi, C., Bartelmann, M., De Zotti, G. & Granato, G. L., 2002 MNRAS 329, 445 Perrotta, F., Magliocchetti, M., Baccigalupi, C., Bartelmann, M., De Zotti, G., Granato, G. L., Silva, L. & Danese, L., 2003, MNRAS 338, 623 Polletta, M. et al. 2007, ApJ, 663, 81 Pope, A. et al. 2008, ApJ, 675, 1171 Schlegel, D. J., Finkbeiner, D. P. & Davis, M., 1998, ApJ, 500, 525 Simpson, J. M. et al. 2014, ApJ, 788, 125 Simpson, J. M. et al. 2015, ApJ, 799, 81 Smail, Ian, Ivison, R. J. & Blain, A. W., 1997, ApJ 490, L5 Smail, I., Ivison, R. J., Blain, A. W. & Kneib, J.-P., et al. 2002, MNRAS, 331, 495
Spilker, J. S et al., 2016, ApJ, 826,112 Strandet, M. L. et al. 2016, ApJ, 822, 80 Swinbank A. M. et al., 2008, MNRAS 391, 420 Swinbank A. M. et al., 2010, Nature, 464, 733 Swinbank, A. M., et al. 2015, ApJL, 806, 17 Tonini et al., 2012, ApJ, 759, 43 Vieira J. D. et al., 2010, ApJ, 719, 763 Vieira, J., et al. 2013, Nature, 495, 344 Younger, J. D. et al. 2007, ApJ, 671, 1531 Walter, F., et al. 2012, Nature, 486, 233 Weiss, A. et al., 2013, ApJ, 767, 88 Wright, E. L. et al., 2010, ApJ 140, 1868-1881
\[lastpage\]
[^1]: E-mail: [email protected]
[^2]: This paper makes use of data from catalogues VISTA and WISE
[^3]: <http://horus.roe.ac.uk/vsa/index.html>
[^4]: <http://wise2.ipac.caltech.edu/docs/release/allwise/>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present the discovery of a hot-Jupiter transiting the $V=9.23$mag main-sequence A-star KELT-17 (BD+14 1881). KELT-17b is a $1.31_{-0.29}^{+0.28}\,M_\mathrm{J}$, $1.645_{-0.055}^{+0.060}\,R_\mathrm{J}$ hot-Jupiter in a 3.08 day period orbit misaligned at $-115.9\pm4.1$ deg to the rotation axis of the star. The planet is confirmed via both the detection of the radial velocity orbit, and the Doppler tomographic detection of the shadow of the planet during two transits. The nature of the spin-orbit misaligned transit geometry allows us to place a constraint on the level of differential rotation in the host star; we find that KELT-17 is consistent with both rigid-body rotation and solar differential rotation rates ($\alpha < 0.30$ at $2\sigma$ significance). KELT-17 is only the fourth A-star with a confirmed transiting planet, and with a mass of $1.635_{-0.061}^{+0.066}\,M_\odot$, effective temperature of $7454\pm49$K, and projected rotational velocity $v\sin I_* = 44.2_{-1.3}^{+1.5}\,\mathrm{km\,s}^{-1}$; it is amongst the most massive, hottest, and most rapidly rotating of known planet hosts.'
author:
- 'George Zhou$^{1}$, Joseph E. Rodriguez$^{2}$, Karen A. Collins$^{2,3}$, Thomas Beatty$^{4,5}$, Thomas Oberst$^{6}$, Tyler M. Heintz$^{6}$, Keivan G. Stassun$^{2,3}$, David W. Latham$^{1}$, Rudolf B. Kuhn$^{7}$, Allyson Bieryla$^{1}$, Michael B. Lund$^{2}$, Jonathan Labadie-Bartz$^{8}$, Robert J. Siverd$^{9}$, Daniel J. Stevens$^{10}$, B. Scott Gaudi$^{10}$, Joshua Pepper$^{8}$, Lars A. Buchhave$^{11}$, Jason Eastman$^{1}$, Knicole Colón$^{12,13}$, Phillip Cargile$^{1}$, David James$^{14}$, Joao Gregorio$^{15}$, Phillip A. Reed$^{16}$, Eric L. N. Jensen$^{17}$, David H. Cohen$^{18}$, Kim K. McLeod$^{18}$, T.G. Tan$^{19}$, Roberto Zambelli$^{20}$, Daniel Bayliss$^{21}$, Joao Bento$^{22}$, Gilbert A. Esquerdo$^{1}$, Perry Berlind$^{1}$, Michael L. Calkins$^{1}$, Kirsten Blancato$^{18}$, Mark Manner$^{2}$, Camile Samulski$^{18}$, Christopher Stockdale$^{23}$, Peter Nelson$^{24}$, Denise Stephens$^{25}$, Ivan Curtis$^{26}$, John Kielkopf$^{27}$, Benjamin J. Fulton$^{28,31}$, D.L. DePoy$^{29}$, Jennifer L. Marshall$^{29}$, Richard Pogge$^{10}$, Andy Gould$^{10}$, Mark Trueblood$^{30}$, and Pat Trueblood$^{30}$'
bibliography:
- 'KELT-17b.bib'
title: 'KELT-17: A hot-Jupiter transiting an A-star in a misaligned orbit detected with Doppler tomography'
---
Introduction
============
The properties of planets orbiting high mass stars provide an important piece of the planet formation puzzle. The occurrence rate of giant planets increases with stellar mass [e.g. @Johnson:2007; @Johnson:2010; @Crepp:2011; @Jones:2014; @Jones:2016], at least until $\sim 2\, M_\odot$ [@Reffert:2015]. Observations of protoplanetary disks also show a correlation between the mass of the host star and the surface density and mass of the protoplanetary disk [e.g. @Muzerolle:2003; @Natta:2006], as well as the disk accretion rate [@Manara:2016]. As such, the conditions around young, massive stars are more conducive to the formation of giant planet embryos [e.g. @Liu:2016], and may even lead to more massive planets being formed [@Mordasini:2012].
Despite the apparent ease of giant planet formation around massive stars, only three transiting planets have been confirmed around A stars to date: WASP-33b [@Collier:2010b; @Johnson:2015], KOI-13b [@Szabo:2011; @Shporer:2011; @Johnson:2014], and HAT-P-57b [@Hartman:2015]. Confirming planets around these stars is difficult via traditional techniques: in addition to the low mass and radius ratios of these systems (therefore low radial velocity amplitudes and transit depths), main sequence A-stars have rapid rotation rates and few metal spectral lines, inhibiting precise radial velocity measurements that are typically necessary for the planet confirmation. One successful approach is to perform radial velocity searches for planets around ‘retired A-stars’ [e.g. @Johnson:2010; @Wittenmyer:2011] – giants and sub-giants with masses $>1.6\,M_\odot$ that have spun-down over their post-main-sequence evolution, and allow precise radial velocity measurements to be made. These surveys have revealed some intriguing trends, such as the apparent lack of high eccentricity warm-Jupiters around sub-giants [@Jones:2014].
Transiting planets offer an unique set of opportunities, such as the characterization of planet radius, orbital orientation, and atmospheric properties, that are not available to planets detected by radial velocities only. A sample of well characterized planets around massive stars is necessary to understand the mass-dependence of planet properties. The Kilodegree Extremely Little Telescope (KELT) [@Pepper:2007] is designed to target planets orbiting stars with brightnesses of $8 < V_\mathrm{mag} < 10$: systems around bright stars that are conducive to follow-up characterization. As discussed in @Bieryla:2015, a direct result of this KELT sample selection is that 55% of KELT-North targets are hotter than 6250K, with masses $\gtrsim 1.3\,M_\odot$ and median rotational velocities of $\gtrsim 20$kms$^{-1}$ [inferred from the Kepler sample of stellar rotational velocities in @Nielsen:2013]. A similar stellar sample will also be surveyed by the TESS full frame dataset [@Ricker:2014]. Strategies for confirming planets around massive stars from the KELT survey are therefore directly transferable to future planet candidates from TESS.
Transiting planets around rapidly rotating, high mass stars can be confirmed via Doppler tomography. During a transit, the planet occults parts of the rotating stellar disk, thereby distorting the observed spectral line profile of the star. For relatively slowly rotating host stars, this results in a net shift in the apparent velocity of the host star – the Rossiter-McLaughlin effect [@Rossiter:1924; @McLaughlin:1924]. In the cases where the rotational broadening of the star is significantly higher than other broadening factors, the shadow of the planet can be resolved as an intrusion in the rotationally broadened line profile of the star, yielding a Doppler tomographic detection of the planet. In addition to the three planets around A-stars confirmed via Doppler tomography, detections have been achieved for 9 more planets: WASP-3b [@Miller:2010], WASP-38b [@Brown:2012], CoRoT-11b [@Gandolfi:2012], HAT-P-2b and Kepler-25c [@Albrecht:2013], KOI-13b [@Johnson:2014], KOI-12b [@Bourrier:2015], KELT-7b and HAT-P-56b [@Bieryla:2015; @Zhou:2016]. The depth and width of the spectroscopic shadow of the planet is directly correlated with the planet-star radius ratio. In the cases where the depth agrees with that from the photometric transit, we can rule out blend scenarios often associated with transiting planet candidates. This is especially useful in eliminating the scenarios of background eclipsing binaries, where a Doppler tomographic observation will yield no planet detection. Subsequent out-of-transit radial velocities, at the kms$^{-1}$ level, are then taken to constrain the nature of the orbiting companion.
In this paper, we report the discovery of KELT-17b, a hot-Jupiter transiting a rapidly rotating $(v\sin I_* = 44\,\mathrm{km\,s}^{-1})$ A-star. KELT-17b was discovered in the equatorial field jointly surveyed by the KELT-North [@Pepper:2007] and KELT-South [@Pepper:2013] observatories. The discovery involves a series of photometric follow-up observations that confirmed and characterized the transit light curve, and spectroscopic monitoring that constrained the mass of the system. Finally, blend false positive scenarios were ruled out by two in-transit spectroscopic observations that confirmed the Doppler tomographic signal induced by the transiting planet.
Discovery and Follow-Up Observations
====================================
KELT-South and KELT-North
-------------------------
KELT-17, the first exoplanet host discovered through the combined observations of both the KELT-North and KELT-South telescopes, is located in KELT-South field 06 (KS06) and KELT-North field 14 (KN14) which are both centered on $\alpha =$ 07$^{h}$ 39$^{m}$ 36$^{s}$ $\delta =$ $+03\degr$ 00$\arcmin$ 00$\arcsec$ (J2000). At the time of identification, the post-processed KELT data set included 2092 images from KN14 taken between UT 2011 October 11 and UT 2013 March 26 and 2636 images from KS06 taken between UT2010 March 02 and 2013 May 10. The image reduction, light curve extraction, and candidate selection processes are described in @Siverd:2012 [@Kuhn:2016]. In brief, calibrated images are processed to light curves using a heavily modified version of the ISIS image subtraction package [@Alard:1998; @Alard:2000]. Extracted light curves are outlier-clipped, smoothed with a 90-day median window, and finally detrended with the Trend Filtering Algorithm [@Kovacs:2005]. This process was performed independently for the KELT-North and KELT-South data sets. Objects in common between the two fields were identified and given a new KELT-Joint field 06 (KJ06) designation. The candidate selection process was then run on these final combined light curves. KJ06C006046=KELT-17 emerged as a top candidate in the joint analysis of field KJ06. KELT-17 (TYC 807-903-1, 2MASS J08222820+1344071) is located at $\alpha =$ 08$^{h}$ 22$^{m}$ 28$\fs$21 $\delta =$ +13$\degr$ 44$\arcmin$ 07$\arcsec$2 (J2000). A list of the photometric and kinematics parameters for KELT-17 is shown in Table \[tbl:Host\_Lit\_Props\]. The box-fitting least squares (BLS) [@Kovacs:2002] periodicity algorithm was used to search for candidates in KJ06. Candidates are selected according to statistics produced with the VARTOOLS [@Hartman:2016] implementation of BLS, and from statistics calculated as per @Pont:2006 and @Burke:2006. Table \[tbl:BLS\_Selection\_Criteria\] shows the criteria and results for the KELT-17b candidate selection. The discovery light curves from both KELT-North and KELT-South are shown in Figure \[fig:KS\_discoverylc\].
{width="15cm"}
------------------ ------------------------------------------- ------------------------- --------- -------------------------------
Parameter Description KELT-17 Value Source Reference(s)
Names TYC 807-903-1
2MASS J08222820+1344071
BD+14 1881
$\alpha_{J2000}$ Right Ascension (RA) 08:22:28.21 Tycho-2 @Hog:2000
$\delta_{J2000}$ Declination (Dec) +13:44:07.2 Tycho-2 @Hog:2000
FUV Far UV flux 16.565 $\pm$ 0.034 GALEX @Bianchi:2011
NUV Near UV flux 13.261 $\pm$ 0.005 GALEX @Bianchi:2011
$u'$ 11.027 $\pm$ 0.001 SDSS @Abazajian:2009
B 9.553 $\pm$ 0.025 ASCC @Kharchenko:2001
B$_T$ Tycho B$_T$ magnitude 9.53 $\pm$ 0.02 Tycho-2 @Hog:2000
V$_T$ Tycho V$_T$ magnitude 9.23 $\pm$ 0.02 Tycho-2 @Hog:2000
V 9.286 $\pm$ 0.051 TASS @Droege:2006
$r'$ 9.223 CMC15 @Evans:2002
$I_C$ 8.948 $\pm$ 0.039 TASS @Droege:2006
J 2MASS magnitude 8.745 $\pm$ 0.027 2MASS @Cutri:2003 [@Skrutskie:2006]
H 2MASS magnitude 8.697 $\pm$ 0.042 2MASS @Cutri:2003 [@Skrutskie:2006]
K 2MASS magnitude 8.646 $\pm$ 0.018 2MASS @Cutri:2003 [@Skrutskie:2006]
*WISE1* WISE passband 8.616 $\pm$ 0.023 WISE @Cutri:2012
*WISE2* WISE passband 8.644 $\pm$ 0.02 WISE @Cutri:2012
*WISE3* WISE passband 8.630 $\pm$ 0.053 WISE @Cutri:2012
*WISE4* WISE passband 8.678 $\pm$ 0.384 WISE @Cutri:2012
$\mu_{\alpha}$ Proper Motion in RA (mas yr$^{-1}$) -22.9 $\pm$ 1.1 NOMAD @Zacharias:2004
$\mu_{\delta}$ Proper Motion in DEC (mas yr$^{-1}$) -0.7 $\pm$ 1.0 NOMAD @Zacharias:2004
U$^{*}$ Space motion ([kms$^{-1}$]{}) -25.6 $\pm$ 0.9 This work
V Space motion ([kms$^{-1}$]{}) 3.3 $\pm$ 0.9 This work
W Space motion ([kms$^{-1}$]{}) -0.1 $\pm$ 1.3 This work
Distance Distance (pc) 210 $\pm$ 10 This work
RV Absolute Radial Velocity ([kms$^{-1}$]{}) 28.0 $\pm$ 0.1 This work
------------------ ------------------------------------------- ------------------------- --------- -------------------------------
**<span style="font-variant:small-caps;">NOTES</span>**\
$^{*}$U is positive in the direction of the Galactic Center
\[tbl:BLS\_Selection\_Criteria\]
---------------------- ------------------------------------------------------------ -------------- --
Statistic Selection KELT-17b
Criteria /KJ06C006046
Signal detection SDE $>$ 7.0 10.56395
efficiency
Signal to pink-noise SPN $>$ 7.0 9.27154
Transit depth $\delta <$ 0.05 0.00433
$\chi^2$ ratio $\displaystyle\frac{\Delta\chi^2}{\Delta\chi^2_{-}} >$ 1.5 1.72
Duty cycle q $<$ 0.1 0.04
---------------------- ------------------------------------------------------------ -------------- --
: KELT discovery selection criterion
Photometric Follow-up {#sec:Follow-up_Photometry}
---------------------
To confirm the source, refine the transit depth, duration, period, and eliminate false positive scenarios, we obtained higher spatial resolution and precision photometric follow-up observations of KELT-17b in multiple filters. These datasets are uniformly reduced using AstroImageJ [AIJ, @Collins:2016]. These light curves are presented in Figure \[fig:All\_light curve\]. A description of each observatory is below. See Table \[tbl:detrending\_parameters\] for a list of the observations and their parameters that were used in the global fit.
![(Top) The individual KELT follow-up Network observations of KELT-17b. The best fit model for each light curve is shown in red. (Bottom) All follow-up transits combined and binned in 5 minute intervals to best represent the transit features. These binned data are shown for display only, and were not used in the analysis. The combined and binned models from each transit are represented by the red line.[]{data-label="fig:All_light curve"}](K17_Figure_4_Final.eps){width="1\linewidth" height="5in"}
![(Top) The individual KELT follow-up Network observations of KELT-17b. The best fit model for each light curve is shown in red. (Bottom) All follow-up transits combined and binned in 5 minute intervals to best represent the transit features. These binned data are shown for display only, and were not used in the analysis. The combined and binned models from each transit are represented by the red line.[]{data-label="fig:All_light curve"}](K17_Figure_6_Final.eps){width="1\linewidth"}
### CROW
An $I$ band transit was observed on UT 2015 March 05 at the Canela’s Robotic Observatory (CROW) with the 0.3m SCT12 telescope, remotely operated from Portalegre, Portugal. Observations were acquired with the ST10XME CCD camera, with a $30\arcmin$ $\times$ 20$\arcmin$ field of view and a $0\farcs86$ pixel scale.
### Kutztown
A full multi-color ($V$ and $I$) transit of KELT-17b was observed on UT 2015 March 12 at Kutztown University Observatory (KUO), located on the campus of Kutztown University in Kutztown, Pennsylvania. KUO’s main instrument is the 0.6m Ritchey-Chrétien optical telescope with a focal ratio of $f/8$. The imaging CCD (KAF-6303E) camera has an array of 3K$\times$2K ($9\,\mu$m) pixels and covers a field of view of $19\farcm5 \times 13\farcm0$.
### Swarthmore
The Peter van de Kamp Observatory (PvdK) at Swarthmore College (near Philadelphia) houses a 0.62-m Ritchey-Chretien reflector with a 4K$\times$4K pixel Apogee CCD. The telescope and camera together have a 26$\arcmin$ $\times$ 26$\arcmin$ field of view and a $0\farcs61$ pixel scale. PvdK observed KELT-17b on UT 2015 March 12 in the SDSS $z^{\prime}$ filter.
### Whitin
KELT-17b was observed in both $g^{\prime}$ and $i^{\prime}$ on UT 2015 March 12 at Wellesley College’s Whitin Observatory in Massachusetts. The telescope is a 0.6 m Boller and Chivens with a DFM focal reducer yielding an effective focal ratio of f/9.6. We used an Apogee U230 2K$\times$2K camera with a $0\farcs58$pixel$^{-1}$ scale and a $20\arcmin \times 20\arcmin$ field of view.
### WCO
One full transit of KELT-17b was observed from the Westminster College Observatory (WCO), PA, on UT 2015 November 4 in the $z'$ filter. The observations employed a 0.35m f/11 Celestron C14 Schmidt-Cassegrain telescope and SBIG STL-6303E CCD with a $\sim$ 3K$\times$2K array of 9 $\mu$m $\,$ pixels, yielding a $24\arcmin \times 16\arcmin$ field of view and $1\farcs4\, \mathrm{pixel}^{-1}$ image scale at $3\times3$ pixel binning. The stellar FWHM was seeing-limited with a typical value of $\sim 3.2\arcsec$.
### MVRC
Three full transits of KELT-17b were observed on UT 2016 February 26 ($g^{\prime}$ and $i^{\prime}$) and UT 2016 March 31 ($r^{\prime}$) using the Manner-Vanderbilt Ritchie-Chrétien (MVRC) telescope located at the Mt. Lemmon summit of Steward Observatory, AZ. The observations employed a 0.6 m f/8 RC Optical Systems Ritchie-Chrétien telescope and SBIG STX-16803 CCD with a 4K$\times$4K array of 9 $\mu$m pixels, yielding a $26\arcmin \times 26\arcmin$ field of view and $0\farcs39$ pixel$^{-1}$ image scale. The telescope was heavily defocused for all three observations resulting in a typical stellar FWHM of $\sim$ 17$\arcsec$.
### PEST
The PEST (Perth Exoplanet Survey Telescope) observatory is a backyard observatory owned and operated by Thiam–Guan (TG) Tan, located in Perth, Australia. It is equipped with a 0.3m Meade LX200 SCT $f/10$ telescope with focal reducer yielding $f/5$ and a SBIG ST-8XME CCD camera. The telescope and camera combine to have a 31$\arcmin$ $\times$ 21$\arcmin$ field of view and a $1\farcs2$ pixel scale. PEST observed KELT-17b on UT 2016 March 06 in the $B$ band.
Observatory Date (UT) Filter FOV Pixel Scale Exposure (s) Detrending parameters for global fit
------------- ------------------ -------------- ---------------------------------- ------------- -------------- -------------------------------------------------- -- --
CROW 2015 March 05 $I$ $30\arcmin$ $\times$ 20$\arcmin$ $0\farcs86$ 90 Meridian Flip, Airmass, BJD
KUO 2015 March 12 $I$ $19\farcm5\times13\arcmin$ $0\farcs76$ 30 Airmass, BJD
KUO 2015 March 12 $V$ $19\farcm5 \times 13\arcmin$ $0\farcs76$ 30 BJD, X Centroid, Sky Background
Pvdk 2015 March 12 $z^{\prime}$ 26$\arcmin$ $\times$ 26$\arcmin$ $0\farcs61$ 60 Airmass
Whitin 2015 March 12 $g^{\prime}$ $20\arcmin$ $\times$ 20$\arcmin$ $0\farcs58$ 48-80 FWHM, Total Counts, PSF Roundness
Whitin 2015 March 12 $i^{\prime}$ $20\arcmin$ $\times$ 20$\arcmin$ $0\farcs58$ 60-100 BJD, FWHM, Sky Background, Airmass, Total Counts
WCO 2015 November 04 $z^{\prime}$ $24\arcmin$ $\times$ 16$\arcmin$ $1\farcs4$ 180 BJD, Airmass, Total Counts
MVRC 2016 February 26 $g^{\prime}$ $26\farcm8 \times 26\farcm8$ $0\farcs39$ 30 Airmass, FWHM, X Centroid, Y Centroid
MVRC 2016 February 26 $i^{\prime}$ $26\farcm8 \times 26\farcm8$ $0\farcs39$ 60 X Centroid, Y Centroid
PEST 2016 March 06 $B$ 31$\arcmin$ $\times$ 21$\arcmin$ $1\farcs2$ 120 Target raw counts, FWHM, X Centroid
MVRC UT 2016 March 31 $r^{\prime}$ $26\farcm8 \times 26\farcm8$ $0\farcs39$ 30 Airmass, Total Counts, Sky Background
**<span style="font-variant:small-caps;">NOTES</span>**\
All the follow-up photometry presented in this paper is available in machine-readable form in the online journal.
Spectroscopic Follow-up {#sec:spec_fu}
-----------------------
Telescope/Instrument Date Range Type of Observation Resolution Wavelength Range ($\AA$) Mean S/N per res. element Epochs
---------------------- ------------------- -------------------------------- ----------------- -------------------------- --------------------------- -------- --
ANU 2.3m/WiFeS 02/2015 Low Resolution Spectral Typing 3000 3500–6000 135 1
ANU 2.3m/WiFeS 02/2015 Low Resolution Radial Velocity 7000 5200–7000 100 3
FWO 1.5m/TRES 04/2015 – 04/2016 High Resolution $\approx$44,000 3900–9100 $\sim$100 74
A series of spectroscopic follow-up observations were performed to characterize the KELT-17 system, they are summarized in Table \[tbl:spectroscopic\_parameters\].
In order to search for signs of stellar-mass companions of KELT-17, we performed low resolution, high signal-to-noise reconnaissance spectroscopic follow-up of KELT-17 using the Wide Field Spectrograph [WiFeS, @Dopita:2007] on the ANU 2.3m telescope at Siding Spring Observatory, Australia in February 2015. WiFeS is an integral field spectrograph employing $1\arcsec$ width slitlets to provide a field of view of $12''\times 38''$ when read out in the ‘stellar’ mode. Follow-up with WiFeS allowed us to obtain multi-epoch spectra for the target and all nearby stars within one KELT pixel, helping to eliminate astrophysical blend scenarios, such as nearby eclipsing binaries, that may mimic the signal of a transiting hot-Jupiter [@Bayliss:2013]. Stellar classification using the flux calibrated WiFeS spectrum provided an initial estimate for the stellar parameters of $T_\mathrm{eff} = 7200\pm200$K, $\log g_* = 4.0\pm0.4$ dex, $\mathrm{[Fe/H]} = -0.5 \pm 0.4$ dex. Three additional multi-epoch observations constrained the radial velocity variation of the target to be $<2\,\mathrm{km\,s}^{-1}$, indicating any orbiting companion responsible for the transit must be of substellar mass.
Following candidate vetting with WiFeS, in-depth spectroscopic characterization of KELT-17 was performed by the Tillinghast Reflector Echelle Spectrograph (TRES) on the 1.5m telescope at the Fred Lawrence Whipple Observatory, Mount Hopkins, Arizona, USA. TRES has a wavelength coverage of 3900–9100Å over 51 echelle orders, with a resolving power of $\lambda/\Delta \lambda \equiv R = 44000$. A total of 12 out-of-transit observations were taken to characterize the radial velocity orbital variations exhibited by KELT-17. The wavelength solutions are derived from Th-Ar hollow cathode lamp exposures that bracket each object spectrum. Each observation consists of three consecutive exposures, totalling $\sim 540$s in exposure time, combined to enable cosmic-ray removal.
The process of spectral extraction, reduction, and radial velocity analyses are similar to those described in @Buchave:2010. Absolute radial velocities are obtained by cross correlating the Mg b line region against a synthetic template spectrum generated using the @Kurucz:1992 atmosphere models. The Mg b velocities are used to determine the absolute velocity offsets presented in Tables \[tbl:Host\_Lit\_Props\] and \[tbl:KELT-17b\]. Precise relative radial velocities are obtained by cross correlating multiple echelle orders of each spectrum to synthetic spectral templates, and weight-averaging the resulting velocities. We adopt the relative radial velocities from the multi-order cross correlations for our radial velocity orbit analysis. The radial velocity orbit measured by TRES is shown in Figure \[fig:K11RV\], and individual radial velocity measurements are also presented in Table \[tab:RVs\].
In addition, we also observed spectroscopic transits of KELT-17b with TRES on 2016-02-23 and 2016-02-26 UT, gathering 33 and 29 sets of spectra, respectively. The exposures achieved a signal-to-noise ratio of 70–100 per resolution element over the Mg b lines, and were reduced as per @Buchave:2010. The in-transit series of spectra revealed the Doppler tomographic signal of the planet, described in Section \[sec:DT\]. Multi-order radial velocities were also derived for this dataset. These velocities clearly exhibit the Rossiter-McLaughlin effect [@Rossiter:1924; @McLaughlin:1924], are plotted in Figure \[fig:K11RV\]. In our global analysis with EXOFAST (Section \[sec:exofast\]), we model the Doppler tomographic signal, rather than the Rossiter-McLaughlin effect, to obtain the spin-orbit alignment of the system. The Doppler tomographic measurement, as presented in Section \[sec:DT\], provides an accurate measurement of the spin-orbit alignment of the system. The Rossiter-McLaughlin effect, however, is modelled with an approximatation to the measured velocities due to asymmetric cross-correlation functions induced by the shadow of the planet, and are subject to modelling assumptions [see discussions in e.g. @Boue:2013].
![TRES radial velocities of KELT-17. The TRES out-of-transit points are shown in black, in-transit points in grey. Only the out-of-transit radial velocities are used in the global fit. In-transit Doppler tomographic analysis, rather than radial velocities, were used to determine the spin-orbit alignment parameters (Section \[sec:exofast\]). (Top) Radial velocities shown as a function of time. (Middle) Velocities shown as a function of orbital phase. (Bottom) Zoomed in view of the in-transit velocities, with the predicted Rossiter-McLaughlin model [@Hirano:2011], with parameters determined from Doppler tomography, over-plotted. Note the in-transit velocities show a clear Rossiter-McLaughlin signal, and are largely consistent with the predicted model. Potential differences may be due to approximations in the Rossiter-McLaughlin model used to estimate the in-transit velocities.[]{data-label="fig:K11RV"}](RV_bjd.eps){width="1\linewidth"}
![TRES radial velocities of KELT-17. The TRES out-of-transit points are shown in black, in-transit points in grey. Only the out-of-transit radial velocities are used in the global fit. In-transit Doppler tomographic analysis, rather than radial velocities, were used to determine the spin-orbit alignment parameters (Section \[sec:exofast\]). (Top) Radial velocities shown as a function of time. (Middle) Velocities shown as a function of orbital phase. (Bottom) Zoomed in view of the in-transit velocities, with the predicted Rossiter-McLaughlin model [@Hirano:2011], with parameters determined from Doppler tomography, over-plotted. Note the in-transit velocities show a clear Rossiter-McLaughlin signal, and are largely consistent with the predicted model. Potential differences may be due to approximations in the Rossiter-McLaughlin model used to estimate the in-transit velocities.[]{data-label="fig:K11RV"}](RV_phase.eps){width="1\linewidth"}
![TRES radial velocities of KELT-17. The TRES out-of-transit points are shown in black, in-transit points in grey. Only the out-of-transit radial velocities are used in the global fit. In-transit Doppler tomographic analysis, rather than radial velocities, were used to determine the spin-orbit alignment parameters (Section \[sec:exofast\]). (Top) Radial velocities shown as a function of time. (Middle) Velocities shown as a function of orbital phase. (Bottom) Zoomed in view of the in-transit velocities, with the predicted Rossiter-McLaughlin model [@Hirano:2011], with parameters determined from Doppler tomography, over-plotted. Note the in-transit velocities show a clear Rossiter-McLaughlin signal, and are largely consistent with the predicted model. Potential differences may be due to approximations in the Rossiter-McLaughlin model used to estimate the in-transit velocities.[]{data-label="fig:K11RV"}](RV_RM.eps){width="1\linewidth"}
Analysis and Results
====================
UVW Space Motion
----------------
We calculate the UVW space motion of KELT-17 to better understand its place in the galactic population. The proper motions, absolute radial velocities, and resulting UVW values are laid out in Table \[tbl:Host\_Lit\_Props\]. To derive the UVW space motions, we derived an absolute radial velocity measurement of KELT-17, calculated by taking the TRES Mg b echelle order absolute velocity of the template frame, subtracting a relative offset of $\gamma = 0.073\,\mathrm{km\,s}^{-1}$ from the global analysis (Section \[sec:exofast\]), and shifting by $-0.61\,\mathrm{km\,s}^{-1}$ to the IAU absolute velocity reference frame, which is determined by our observations of a set of IAU radial velocity reference stars. Proper motion values are taken from NOMAD [@Zacharias:2004]. The distance estimate is derived from a spectral fit to the spectral energy distribution (Section \[sec:sed\]). We also adopt the local standard of rest from @2011MNRAS.412.1237C. The resulting U, V, W space motions are $-25.6 \pm 0.9$, $3.3 \pm 0.9$, and $-0.1 \pm 1.3\,\mathrm{km\,s}^{-1}$ respectively, giving a 99.4% probability that KELT-17 resides in the thin disk [@Bensby:2003].
Stellar Parameters from Spectra {#sec:stellar_params}
-------------------------------
The stellar atmospheric parameters were measured from each spectrum using the Stellar Parameter Classification (SPC) pipeline [@Buchave:2010]. The parameters effective temperature $T_\mathrm{eff}$, surface gravity [$\log g_*$]{}, metallicity \[m/H\], and projected rotational velocity $v\sin I_*$ are fitted for each TRES spectrum. SPC maximizes the cross correlation function peak of each spectrum, in the spectral order surrounding the Mg b lines, against a library of synthetic templates calculated using the @Kurucz:1992 atmosphere models. The resulting stellar parameters from the first round of fitting to all exposures were $T_\mathrm{eff} = 6975 \pm 50$K, $\log g_*=3.08 \pm 0.10$, $\mathrm{[m/H]}= -0.10 \pm 0.08$, $v \sin I_*=49.1 \pm 0.5\, \mathrm{km\,s}^{-1}$, the uncertainties describe the expected systematic errors and scatter between exposures. The initial spectroscopic stellar parameters for rapidly rotating stars are known to be unreliable. In particular, the surface gravity [$\log g_*$]{}is difficult to determine for hot and rapidly rotating stars, and an offset will lead to systematic errors in the other atmospheric parameters.
As such, we use the transit duration, which is directly related to the $a/R_\star$ parameter, to give a much better constraint on the stellar density $\rho_*$. Our global analysis (described in Section \[sec:exofast\]) simultaneously constrains the stellar parameters using the transit light curves and stellar isochrones, and yielded an updated [$\log g_*$]{}and \[Fe/H\]. We then re-ran SPC with the [$\log g_*$]{}fixed to that determined from the global analysis, and derived an updated set of stellar parameters of $T_\mathrm{eff}=7452 \pm 50$K, $\mathrm{[m/H]}=0.25 \pm 0.08$, $v\sin I_*=48.5 \pm 0.5$kms$^{-1}$. The derived $T_\mathrm{eff}$ agrees with that from the flux calibrated WiFeS low resolution spectrum (Section \[sec:spec\_fu\]). We note the SPC-derived \[m/H\] is slightly different from our final metallicity value quoted in Table \[tbl:KELT-17b\], since the metallicity, and [$\log g_*$]{}, are re-iterated through the global analysis, and are co-constrained by the transit-derived stellar density and the stellar isochrone models.
SED Analysis {#sec:sed}
------------
![The SED fit for KELT-17. The red points show the photometric magnitudes and adopted uncertainties. Horizontal error bars indicate the width of the photometric band passes. The blue points show the integrated model magnitudes from the best fit NextGen synthetic spectra [@Hauschildt:1999]. The parameters $T_\mathrm{eff}$, $\log g_*$, \[Fe/H\] are allowed free in the SED fit, while the reddening $A_V$ is limited to be less than the maximum from the @Schlegel:1998 dust maps.[]{data-label="fig:SED_figure"}](kelt17_sed_fin.ps){width="0.75\columnwidth"}
We use all available broadband photometry to construct the empirical spectral energy distribution (SED) of KELT-17 (listed in Table \[tbl:Host\_Lit\_Props\]), including GALEX near-UV fluxes [@Bianchi:2011], Sloan Digital Sky Survey release 7 [SDSS, @Abazajian:2009] $u'$ band, $B_T$ and $V_T$ Tycho-2 magnitudes, All-sky Compiled Catalogue-2.5 V3 $B$ band [ASCC, @Kharchenko:2001], The Amateur Sky Survey Mark IV [TASS, @Droege:2006] $V$ and $I_C$ bands, 2MASS [@Cutri:2003; @Skrutskie:2006] $J$, $H$, $K$ bands, and WISE [@Cutri:2012] magnitudes (Figure \[fig:SED\_figure\]). Only the $u'$ band magnitude is used from SDSS, as the other bands show signs of saturation. These wide band fluxes provide an independent check on the spectral classification of the host star. The SED is fitted against NextGen atmosphere models [@Hauschildt:1999], with maximum reddening limited to the local value from the @Schlegel:1998 dust maps. We adopt a minimum error bar of 0.03 mag if the reported error was smaller, except in the GALEX bands where a minimum error of 0.1 mag is adopted. We also set the WISE3 error to be much larger than reported in order to account for model uncertainties at $10\,\mu m$. We derive a $T_\mathrm{eff} = 7450 \pm 150$K, $\log g_* = 4.0 \pm 0.5$, $\mathrm{[Fe/H]} = 0.0 \pm 0.5$, $A_V = 0.02_{-0.02}^{+0.07}$, and inferred distance of $210\pm10$pc, with a reduced $\chi^2$ of 4.6 from the final fit. The derived stellar parameters are in full agreement with the final SPC stellar parameters in Section \[sec:stellar\_params\].
Evolutionary Analysis
---------------------
![Stellar evolutionary tracks for KELT-17. We match the $M_\star$ and \[Fe/H\] from the global analysis to evolutionary tracks from the YY isochrones [@Demarque:2004], plotted here in terms of $T_\mathrm{eff}$ and $\log g_*$, finding an age of 0.5–0.8Gyr for the system. The shaded region represents the $1\sigma$ model regime, and the blue markers note the ages (in Gyr) along the best fit isochrone. The $T_\mathrm{eff}$ and $\log g_*$ of KELT-17 is marked in red. []{data-label="fig:hrd"}](kelt17_hrd_fin.ps){width="0.75\columnwidth"}
To estimate the age of KELT-17, we match the system parameters to Yonsei-Yale (YY) evolutionary tracks [@Demarque:2004], shown in Figure \[fig:hrd\]. Following the procedure specified in @Siverd:2012 and subsequent KELT discovery papers, we adopt $M_\star = 1.635_{-0.061}^{+0.066}\,M_\odot$ and $\mathrm{[Fe/H]} = -0.018_{-0.072}^{+0.074}$ from the global fit (Section \[sec:exofast\]), and match these against the YY $T_\mathrm{eff}$ – $\log g_*$ isochrones, finding that KELT-17 is an A-star on the main-sequence with a relatively young age of 0.5–0.8Gyr.
Doppler tomographic analysis {#sec:DT}
----------------------------
During a transit, the planet blocks successive regions of the star, and imprints a ‘shadow’ on the observed spectral line profiles. For rapidly rotating stars, the line broadening profile can be derived via a least squares deconvolution of the observed spectrum against a weighted line list or an unbroadened spectral template [@Donati:1997; @Collier:2010b]. We follow the procedure set out in @Zhou:2016 to derive the broadening kernel for the set of TRES transit spectra. Each echelle order is blaze corrected by a flat lamp spectrum, and normalized by a polynomial fit to the continuum. We then stitch the spectra from every three echelle orders together, forming sections of the spectrum $\sim 200$Å long. A total of 34 echelle orders were used from each observation, spanning the spectral range of 3900–6250Å. For each spectral section, we generate a template using the SPECTRUM spectral synthesis program [@Gray:1994], with the ATLAS9 model atmospheres [@Castelli:2004]. The spectral template is generated using the measured [$T_{\rm eff}$]{}, [$\log g_*$]{}, and \[m/H\] values from SPC, without any rotational, macroturbulence, or instrumental broadening. The broadening kernel is derived from each spectral section via the least squares deconvolution between the observed spectrum and the template [as per @Donati:1997]. The global broadening profile of the exposure is then determined via the weighted average of the broadening kernel generated from each spectral section. Out-of-transit exposures provide an averaged broadening profile template from which the in-transit profiles are subtracted. The residuals and best fit models are shown in Figure \[fig:DT\]. The residuals are used in the EXOFAST global analysis in Section \[sec:exofast\] to help measure the spin-orbit alignment of the system, as well as co-constrain the planet transit parameters. In addition, we also derive a [$v \sin I_*$]{} of $44.5\pm0.2\,$kms$^{-1}$ and a macroturbulence broadening value of $5.10\pm0.47\,\mathrm{km\,s}^{-1}$ from the deconvolved broadening kernels using the fitting technique discussed in @Zhou:2016. We also tested performing the least squares deconvolution on the SPC template spectra (calculated using @Kurucz:1992 atmosphere models), finding no measurable difference in the resulting rotational profiles.
The detection of a Doppler tomographic signal eliminates blend scenarios that can mimic the transit signal of a planetary system. The depth and width of the Doppler tomographic signal is fully consistent with the photometric transit. In blend scenarios, the Doppler tomographic signal depth will be diluted, and the width of the signal will be wider than that of a planetary signal. In particular, if the photometric transit was caused by a background eclipsing binary blended with the target star, the Doppler tomographic signal would have been undetectable.
------------------------------------------------- -------------------------------------------------
{width="8cm"} {width="8cm"}
------------------------------------------------- -------------------------------------------------
Transit Timing Variation Analysis {#sec:TTV_analysis}
---------------------------------
To determine an independent ephemeris, we perform a linear fit through the mid-transit times determined for each follow-up photometric observation (listed in Table \[tbl:transitimes\]). This analysis gives T$_0$ (BJD-TDB) = 2457226.142194 $\pm$ 0.00033 and a period of 3.0801718 $\pm$ 0.0000053 days, with a $\chi^2$ of 19.94 and 9 degrees of freedom. Some outliers to this fit can be seen in Figure \[fig:TTV\]. However, the transit timing residuals are all within $2\sigma$ of a linear ephemeris, and do not show a coherent trend at levels larger than common systematic errors in transit timing [e.g. @Carter:2009]. We carefully ensured that all follow-up observations were correctly converted to BJD$_{\rm TBD}$ [@Eastman:2010]. These ephemerides are then used as priors for the EXOFAST global analysis, described in Section \[sec:exofast\].
------- ------------------------- ------------------ --------- -------------------- -----------
Epoch $T_{C}$ $\sigma_{T_{C}}$ O-C O-C Telescope
([$\rm {BJD_{TDB}}$]{}) (s) (s) ($\sigma_{T_{C}}$)
-45 2457087.536713 113 194.59 1.72 CROW
-43 2457093.695135 77 28.55 0.37 KUO
-43 2457093.695102 101 25.70 0.25 KUO
-43 2457093.692344 125 -212.59 -1.69 Pvdk
-43 2457093.694912 106 9.29 0.09 Whitin
-43 2457093.694187 130 -53.35 -0.41 Whitin
34 2457330.866145 76 -163.41 -2.13 WCO
71 2457444.834316 80 -6.78 -0.08 MVRC
71 2457444.835685 55 111.50 1.99 MVRC
74 2457454.078245 155 288.14 1.85 PEST
82 2457478.715457 53 -71.52 -1.34 MVRC
------- ------------------------- ------------------ --------- -------------------- -----------
: Transit times for KELT-17b.[]{data-label="tbl:transitimes"}
EXOFAST Global Fit {#sec:exofast}
------------------
To provide accurate system parameters and uncertainties for the KELT-17 system, we use a modified version of the EXOFAST exoplanet fitting package [@Eastman:2013], to perform a global fit of our follow-up photometric and spectroscopic observations. Here we provide an overview of the process with respect to KELT-17b; see @Siverd:2012 and @Kuhn:2016 for more detailed descriptions. To constrain $R_*$ and $M_*$, we adopt either the Torres relations [@Torres:2010] or the YY stellar evolution models [@Demarque:2004]. For both the Torres relations and YY Isochrones, we run a global fit constraining the eccentricity to zero. As a result of the high [$v \sin I_*$]{}, causing higher than typical errors on our radial velocity measurements from TRES, we do not attempt to constrain the eccentricity. Each follow-up photometric observation (with the best determined detrending parameters from AIJ), out-of-transit radial velocity measurements from TRES, and the Doppler tomographic observations from TRES are used as inputs for the final global fits. The results of both global fits are shown in Tables \[tbl:KELT-17b\] and \[tbl:KELT-17b\_part2\], and both fits are consistent with each other to within 1$\sigma$. The SPC determined [$T_{\rm eff}$]{}, \[m/H\], $v\sin I_*$, and line broadening due to instrumental resolution and macroturbulence $(v_\mathrm{broad})$ (and errors) from the TRES spectra, and associated $1\sigma$ uncertainties, were used as Gaussian priors during the fitting. We also adopt the period $P$ and transit epoch $T_0$ from the TTV analysis (Section \[sec:TTV\_analysis\]) as priors in our global fitting. Allowing for TTVs decouples the transit times from the planet’s orbit, adding these priors effectively encodes the information from the linear ephemeris into the global model, while still retaining the full flexibility for the transit times to vary. To simplify our interpretation, we adopt the YY circular results for the rest of this paper.
---------------------------------------------------------- ------------------------------------------------------------ --------------------------------------- --------------------------------------- -- --
Parameter Description (Units) **Adopted Value** Value
**(YY circular)** (Torres circular)
Stellar Parameters
$M_{*}$ Mass ([$\,M_\Sun$]{}) $1.635_{-0.061}^{+0.066}$ $1.515_{-0.071}^{+0.073}$
$R_{*}$ Radius ([$\,R_\Sun$]{}) $1.645_{-0.055}^{+0.060}$ $1.598_{-0.054}^{+0.058}$
$L_{*}$ Luminosity ([$\,L_\Sun$]{}) $7.51_{-0.55}^{+0.62}$ $7.07_{-0.51}^{+0.57}$
$\rho_*$ Density (cgs) $0.518_{-0.042}^{+0.045}$ $0.524_{-0.044}^{+0.046}$
$\log g_*$ Surface gravity (cgs) $4.220_{-0.024}^{+0.022}$ $4.211_{-0.025}^{+0.024}$
${\ensuremath{T_{\rm eff}}\xspace}$ Effective temperature (K) $7454\pm49$ $7451_{-50}^{+49}$
Age System Age (Gyr) 0.5–0.8
${\ensuremath{\left[{\rm Fe}/{\rm H}\right]}\xspace}$ Metallicity $-0.018_{-0.072}^{+0.074}$ $-0.274_{-0.072}^{+0.11}$
$v\sin{I_*}$ Rotational velocity $(\mathrm{m\,s}^{-1})$ $44200_{-1300}^{+1500}$ $44100_{-1300}^{+1500}$
$\lambda$ Projected spin-orbit alignment (degrees) $-115.9\pm4.1$ $-115.5_{-4.2}^{+4.1}$
$I_*$ $^a$ Line-of-sight stellar inclination (degrees) $94_{-10}^{+9}$
$\phi$ $^a$ True obliquity (degrees) $116\pm4$
$ v_\mathrm{broad} $ non-rotational line broadening $(\mathrm{m\,s}^{-1})$ $5100\pm470$ $5080\pm470$
Planet Parameters
$P$ Period (days) $3.0801716_{-0.0000052}^{+0.0000053}$ $3.0801718_{-0.0000038}^{+0.0000037}$
$a$ Semi-major axis (AU) $0.04881_{-0.00061}^{+0.00065}$ $0.04759\pm0.00075$
$M_{P}$ Mass ([$\,M_{\rm J}$]{}) $1.31_{-0.29}^{+0.28}$ $1.25\pm0.27$
$R_{P}$ Radius ([$\,R_{\rm J}$]{}) $1.525_{-0.060}^{+0.065}$ $1.478_{-0.058}^{+0.062}$
$\rho_{P}$ Density (cgs) $0.46_{-0.11}^{+0.12}$ $0.48_{-0.11}^{+0.12}$
$\log{g_{P}}$ Surface gravity (cgs) $3.144_{-0.11}^{+0.090}$ $3.149_{-0.11}^{+0.089}$
$T_{eq}$ Equilibrium temperature (K) $2087_{-33}^{+32}$ $2082_{-32}^{+34}$
$\Theta$ Safronov number $0.051\pm0.011$ $0.053\pm0.011$
${\langle F \rangle\xspace}$ Incident flux ([10$^9$ erg s$^{-1}$ cm$^{-2}$]{}) $4.31_{-0.26}^{+0.27}$ $4.26_{-0.25}^{+0.28}$
Radial Velocity Parameters
$T_C$ Time of inferior conjunction ([$\rm {BJD_{TDB}}$]{}) $2457287.74564\pm0.00030$ $2457287.74565\pm0.00021$
$K$ RV semi-amplitude $(\mathrm{m\,s}^{-1})$ $131_{-29}^{+28}$ $131\pm28$
$M_P\sin{i}$ Minimum mass ([$\,M_{\rm J}$]{}) $1.31_{-0.29}^{+0.28}$ $1.24\pm0.27$
$M_{P}/M_{*}$ Mass ratio $0.00077\pm0.00017$ $0.00079\pm0.00017$
$u$ RM linear limb darkening $0.5383_{-0.0020}^{+0.0028}$ $0.5437_{-0.0051}^{+0.014}$
$\gamma_{TRES}$ Offset for TRES relative velocities $(\mathrm{m\,s}^{-1})$ $73\pm24$ $74\pm24$
Linear Ephemeris
from Follow-up
Transits:
$P_{Trans}$ Period (days) $3.0801718\pm 0.0000053$ —
$T_0$ Linear ephemeris from transits ([$\rm {BJD_{TDB}}$]{}) $2457226.142194\pm0.00033$ —
---------------------------------------------------------- ------------------------------------------------------------ --------------------------------------- --------------------------------------- -- --
**<span style="font-variant:small-caps;">NOTES</span>**\
$^{a}$ From the independent differential rotation analysis described in Section \[sec:diffrot\].
------------------- ------------------------------------------ --------------------------------- --------------------------------- -- --
Parameter Description (Units) **Adopted Value** Value
**(YY circular)** (Torres circular)
Primary Transit
$R_{P}/R_{*}$ Radius of the planet in stellar radii $0.09526_{-0.00085}^{+0.00088}$ $0.09509\pm0.00086$
$a/R_*$ Semi-major axis in stellar radii $6.38\pm0.18$ $6.40\pm0.18$
$i$ Inclination (degrees) $84.87_{-0.43}^{+0.45}$ $84.93\pm0.45$
$b$ Impact parameter $0.570_{-0.035}^{+0.031}$ $0.566_{-0.035}^{+0.033}$
$\delta$ Transit depth $0.00907_{-0.00016}^{+0.00017}$ $0.00904\pm0.00016$
$T_{FWHM}$ FWHM duration (days) $0.12674_{-0.00067}^{+0.00068}$ $0.12667_{-0.00064}^{+0.00065}$
$\tau$ Ingress/egress duration (days) $0.0181_{-0.0011}^{+0.0012}$ $0.0179_{-0.0011}^{+0.0012}$
$T_{14}$ Total duration (days) $0.1448_{-0.0013}^{+0.0014}$ $0.1446_{-0.0013}^{+0.0014}$
$P_{T}$ A priori non-grazing transit probability $0.1418_{-0.0038}^{+0.0039}$ $0.1413_{-0.0038}^{+0.0041}$
$P_{T,G}$ A priori transit probability $0.1717_{-0.0048}^{+0.0050}$ $0.1710_{-0.0048}^{+0.0051}$
$u_{1B}$ Linear Limb-darkening $0.3713_{-0.0053}^{+0.0064}$ $0.3795_{-0.0098}^{+0.020}$
$u_{2B}$ Quadratic Limb-darkening $0.3509_{-0.0040}^{+0.0034}$ $0.3462_{-0.0100}^{+0.0062}$
$u_{1I}$ Linear Limb-darkening $0.1337_{-0.0032}^{+0.0047}$ $0.1504_{-0.0075}^{+0.023}$
$u_{2I}$ Quadratic Limb-darkening $0.3266_{-0.0027}^{+0.0024}$ $0.3122_{-0.019}^{+0.0056}$
$u_{1Sloang}$ Linear Limb-darkening $0.3418_{-0.0040}^{+0.0054}$ $0.3527_{-0.0087}^{+0.022}$
$u_{2Sloang}$ Quadratic Limb-darkening $0.3480_{-0.0033}^{+0.0024}$ $0.3408_{-0.012}^{+0.0055}$
$u_{1Sloani}$ Linear Limb-darkening $0.1511_{-0.0032}^{+0.0047}$ $0.1680_{-0.0083}^{+0.025}$
$u_{2Sloani}$ Quadratic Limb-darkening $0.3315\pm0.0025$ $0.3173_{-0.020}^{+0.0059}$
$u_{1Sloanr}$ Linear Limb-darkening $0.2179_{-0.0029}^{+0.0046}$ $0.2264_{-0.0096}^{+0.032}$
$u_{2Sloanr}$ Quadratic Limb-darkening $0.3501_{-0.0022}^{+0.0019}$ $0.3426_{-0.023}^{+0.0067}$
$u_{1Sloanz}$ Linear Limb-darkening $0.1023_{-0.0033}^{+0.0047}$ $0.1175_{-0.0058}^{+0.018}$
$u_{2Sloanz}$ Quadratic Limb-darkening $0.3194_{-0.0029}^{+0.0020}$ $0.3060_{-0.015}^{+0.0041}$
$u_{1V}$ Linear Limb-darkening $0.2795_{-0.0028}^{+0.0045}$ $0.2911_{-0.0091}^{+0.028}$
$u_{2V}$ Quadratic Limb-darkening $0.3467_{-0.0025}^{+0.0016}$ $0.3384_{-0.018}^{+0.0054}$
Secondary Eclipse
$T_{S}$ Time of eclipse ([$\rm {BJD_{TDB}}$]{}) $2457286.20555\pm0.00030$ $2457286.20556\pm0.00021$
------------------- ------------------------------------------ --------------------------------- --------------------------------- -- --
Since the follow-up light curves were obtained in multiple photometric bands, we can also search for signs of wavelength – transit depth dependencies. Large color-based transit depth dependencies can indicate the transit is actually caused by a stellar eclipsing binary system, while low level trends can reveal Rayleigh scatter signatures in the planetary atmospheres. The follow-up data includes observations performed in the $B$, $g'$, $V$, $R$, $i'$, and $z'$ bands. We allowed the transit model for each band to have independent $R_p/R_\star$ values, all other transit parameters are shared in the joint fitting. The limb darkening coefficients are fixed to those interpolated from @Claret:2011. We find no color-transit depth dependencies, with all derived $R_p/R_\star$ values agreeing to within $1\sigma$.
Constraining differential rotation via Doppler tomography {#sec:diffrot}
---------------------------------------------------------
Planets in strongly misaligned orbits can allow us to map the surface features on the host star. For example, spot-crossings during the transits of the polar orbit planet HAT-P-11b were used to construct a ‘butterfly-diagram’ for the spot evolution of the K-star [@Sanchis:2011]. Similarly, the Doppler tomographic shadow of a spin-orbit misaligned planetary transit can help map the projected spin velocity of the stellar surface underneath. In particular, this allows us to directly measure the stellar spin rate as a function of latitude [@Gaudi:2007; @Cegla:2016], thereby constraining the latitudinal differential rotation rate of the host star.
On the sun, differential rotation is thought to partially drive the solar dynamo [e.g. @Dikpati:1999], integral to the development of sun-spots and the 11-year solar activity cycle. Non-rigid rotation has been inferred for other stars by monitoring for long-term activity cycles in their spot modulated light curves [e.g. @Walkowicz:2012; @Reinhold:2013], via Doppler tomographic maps of spotted, active stars [e.g. @Donati:1997b; @Barnes:2000], or via modeling of the rotational broadening kernel [e.g. @Reiners:2002; @Reiners:2003; @Reiners:2003b]. Planetary transits also offer a method of directly detecting differential rotation for favorable systems.
Constraining differential rotation with planetary systems is also particularly interesting in the context of spin-orbit misaligned hot-Jupiter systems, like KELT-17b. A large fraction of hot-Jupiter systems are reported to be spin-orbit misaligned, with the G, K stars hosting a larger fraction of aligned systems than stars of earlier spectral types [e.g. @Winn:2010; @Albrecht:2012]. One idea is that internal gravity waves generated at the radiative envelope – convective core boundaries of hot stars can modify the apparent rotation of these stars at short timescales, in which case both radial and latitudinal differential rotation are expected [@Rogers:2012; @Rogers:2013b]. Direct measurements of host star differential rotation of a spin-orbit misaligned system is key to testing this idea.
We attempt to fit for any differential rotation in KELT-17 via an analysis that is independent, and simplified, from that of the EXOFAST fitting described in Section \[sec:exofast\]. Following @Cegla:2016, we model the stellar rotation at a given position, $v_\mathrm{stel}$ as $$v_\mathrm{stel} = x v_\mathrm{eq}\sin I_* ( 1 - \alpha y^2)\,$$ where $x$ and $y$ are the projected position coordinates on the stellar disk with respect to the stellar spin axis and equator, $v_\mathrm{eq}$ is the equatorial rotation speed, and $I_*$ is the line-of-sight inclination of the stellar spin-axis. The coefficient $\alpha$ describes the rate of differential rotation, where a rigid body has $\alpha=0$, while the solar differential rotation is described by $\alpha_\odot = 0.2$. The effect of various levels of differential rotation $\alpha$ and inclination $I_*$ for the KELT-17 system on the observed Doppler tomographic maps are illustrated in Figure \[fig:dfmodels\].
{width="14cm"}
We incorporate the differential rotation model in an independent global fit of the follow-up light curves and Doppler tomographic datasets, fitting for the differential rotation parameters $\alpha$ and $I_*$, transit parameters period $P$, transit time $T_0$, radius ratio $R_p/R_\star$, normalized semi-major axis $a/R_\star$, transit chord inclination $i$, projected spin-orbit angle $\lambda$, projected rotation velocity $v\sin I_*$, stellar parameters [$T_{\rm eff}$]{}, [$\log g_*$]{}, and first order light curve detrending coefficients for each of the instrumental trends set out in Table \[tbl:detrending\_parameters\]. The fitting procedure is largely described in @Zhou:2016, and is performed via a MCMC analysis with the *emcee* affine invariant ensemble sampler [@ForemanMackey:2012].
In @Zhou:2016, the tomographic signal of the planet is approximated by a Gaussian of width $v\sin I_* \times R_p/R_\star$, sufficient for the standard modeling of the planetary Doppler tomographic signal. However, a true model of the planetary shadow profile needs to account for asymmetries during ingress and egress, as well as the uneven limb darkening in the projected stellar surface under the planet. This is especially important when fitting for differential rotation, which relies on accurate centroids of the planetary shadow at each time step. Therefore, we model the planetary shadow via a numerical integration of the projected stellar surface underneath the planet, accounting for limb darkening, differential rotation, and instrumental broadening.
Figure \[fig:dfcorner\] shows the posterior distribution of selected parameters from our MCMC analysis. The differential rotation coefficient can be constrained to be $\alpha < 0.15$ at $1\sigma$ ($\alpha < 0.30$ at $2\sigma$), consistent with rigid body rotation, but also consistent with solar differential rotation. The line-of-sight inclination is constrained to $I_* = 94 _{-10}^{+9}$ deg. The line-of-sight inclination $I_*$ and the projected spin-orbit angle $\lambda$ can be combined to calculate the true spin-orbit angle $\phi$: $$\cos \phi = \cos I_* \cos i + \sin I_* \sin i \cos \lambda \, ,$$ giving a true spin-orbit angle of $\phi = 116\pm4\,^\circ$ for the system.
{width="17cm"}
@AmmlervonEiff:2012 found the fraction of differential rotators decreases with increasing $T_\mathrm{eff}$ and $v\sin I_*$ for A-F stars. KELT-17 is a late A-star with no evidence for differential rotation, consistent with this trend. In contrast, Doppler imaging of spots on rapid rotators from @Barnes:2005, @CollierCameron:2007, and models from @Kuker:2005 find the level of differential rotation to increase with temperature for late F to M stars. Direct detections of differential rotation via spin-orbit misaligned planets for a range of stars can be a way to provide clear benchmarks to re-examine these previous studies. Under the framework of @Rogers:2012, the lack of significant differential rotation suggests there is no evidence that the star is currently undergoing spin axis evolution.
Conclusions
===========
We report the discovery of KELT-17b, a hot-Jupiter around an A-star discovered by the joint KELT-North and KELT-South survey. KELT-17b is only the fourth hot-Jupiter found transiting an A-star, after WASP-33b, KOI-13b, and HAT-P-57b. The host star is also amongst the most rapidly rotating known transit-planet-hosting stars, with a $v\sin I_*$ of $44.2_{-1.3}^{+1.5}\,\mathrm{km\,s}^{-1}$, only WASP-33 [@Collier:2010b], KELT-1 [@Siverd:2012], KOI-12 [@Bourrier:2015], KOI-13 [@Szabo:2011], KELT-7 [@Bieryla:2015], and HAT-P-57 [@Hartman:2015] exhibit faster spin rates. KELT-17, with a mass of $1.635_{-0.061}^{+0.066}\, M_\odot$, is amongst the most massive (hottest) 3% (0.5%) of known planet hosts[^1]. Blend scenarios for KELT-17b are eliminated by the detection of the Doppler tomographic signal, from which we also measured a projected spin-orbit angle of $\lambda = -115.9\pm4.1^\circ$ for the system. With a mass of $1.31_{-0.29}^{+0.28}\,M_J$, and radius of $1.525_{-0.060}^{+0.065}\,R_J$, KELT-17b is inflated compared to standard evolution models. It receives an incident flux of $4\times 10^9\,\mathrm{erg\,s}^{-1}\mathrm{cm}^{-2}$, larger than the empirical threshold of $2\times10^8\,\mathrm{erg\,s}^{-1}\mathrm{cm}^{-2}$ where radius inflation is observed for the hot-Jupiter sample [@Demory:2011].
KELT-17b is one of 26 known transiting hot-Jupiters around a host star hotter than 6250K, of which 70% are spin-orbit misaligned ($|\lambda| > 10^\circ$). In fact, all four hot-Jupiters around A-stars are in severely misaligned orientations[^2]. The spin-orbit synchronization timescale for the KELT-17 system is $\sim 10^{11}$ yrs [using Equation 3, @Hansen:2012], so the current system misalignment is unlikely to have been modified by star-planet tidal interactions, and will be stable for the duration of the main-sequence lifetime of KELT-17. KELT-17b is super-synchronous: the host star has a maximum spin period of 1.7 days, while the planet orbital period is $\sim 3.08$ days, as with a number of other systems around F-A stars (CoRoT-3b, CoRoT-11b, HAT-P-56b, KELT-7b, KOI-13b, WASP-7b, WASP-8b, WASP-33b, WASP-38b). In contrast, no Kepler candidates, which are largely around cooler host stars, are found in super-synchronized orbits [@Walkowicz:2013]. While the synchronization timescale is long, the orbit circularization timescale should be only $10^7$ years, assuming $Q_\mathrm{planet} = 10^5$, and adopting the circularization timescale from @Goldreich:1966, so we expect the orbit of the planet to be circular in the present day.
The spin-orbit misaligned orientation of KELT-17b means the planet crosses a wide-range of stellar latitudes during the transit, which allowed us to constrain the latitudinal differential rotation of the star. As a result, we find KELT-17 to be consistent with both rigid body rotation and solar-levels of differential rotation ($\alpha<0.30$ at $2\sigma$). An equivalent technique was applied to the transits of HD 189733b [@Cegla:2016], a significantly more difficult case given the near-aligned geometry of the planet $(\lambda = -0.4\pm 0.2^\circ)$ and the low rotation rate of the star $(v\sin I_* = 3.25\pm0.02\,\mathrm{km\,s}^{-1})$. Nevertheless, they were able to rule out rigid-body rotation for the host star. Future Doppler tomographic follow-up of KELT-17 can further refine the differential rotation constraints on the star, and search for nodal precession of the planet’s orbit [e.g. @Johnson:2015].
The authors thank the referee for their insightful comments. K.K. McLeod acknowledges the Theodore Dunham, Jr. Grant of the Fund for Astronomical Research for the purchase of the SDSS filters used at Whitin Observatory. B.J.F. notes that this material is based upon work supported by the National Science Foundation Graduate Research Fellowship under grant No. 2014184874. Work by B.S.G. and D.J.S was partially supported by NSF CAREER Grant AST-1056524. Any opinion, findings, and conclusions or recommendations expressed in this material are those of the authors(s) and do not necessarily reflect the views of the National Science Foundation.
[^1]: NASA exoplanet archive <http://exoplanetarchive.ipac.caltech.edu/>
[^2]: Multiple $\lambda$ solutions are allowed for HAT-P-57b [@Hartman:2015]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In the paper we develop a framework for the alternative way of the study of a local geometry of almost cosymplectic manifolds with Kälerian leaves. The main idea is to apply the concept of a geometry and analysis of CR manifolds. Locally the almost cosymplectic manifold is modeled on the ’mixed’ space $\mathbb{R}\times\mathbb{C}^n$. There is given a complete local description of the underlying almost contact metric structure in the system of local, mixed - real, complex- coordinates. We also introduce a notion of a canonical Hermitian complex connection in the CR structure of a CR almost cosymplectic manifold. As an example we provide detailed descritpion of almost cosymplectic $(-1,\mu,0)$-spaces.'
author:
- Piotr Dacko
title: 'On CR (Cauchy-Riemann) almost cosymplectic manifolds'
---
Introduction
============
The paper is thought of as a preliminary to the subject of possible applications of the methods coming from the geometry of CR manifolds or complex geometry. The starting point is a notion of a CR manifold.
It is very common nowadays in the geometry of the almost contact metric manifold to impose conditions of a “CR integrability” of an almost contact metric structure. One of the first result concerning “CR integrable” almost contact metric structures is the theorem of S. Tanno - the characterization of a CR integrable contact metric manifold [@Tanno].
In the settings of almost cosymplectic manifolds CR geometry implicitly appeared in the time when Z. Olszak introduced quite naturally the notion of the almost cosymplectic manifolds with Kählerian leaves [@Olszak2]. Now it is clear that these manifolds are exactly CR integrable almost cosymplectic manifolds. And this statement is almost tautological from the point of view of the geometry of CR manifolds. However the main focus was on studying the Riemannian geometry using tensor calculus.
In the presented paper we propose an alternative way to study manifolds with Kählerian leaves, based on the CR geometry. We hope that this alternative allow us to solve some long-standing and yet unsolved problems. For example we do not know are there exist almost cosymplectic non-cosymplectic manifolds of a pointwise-constant $\varphi$-sectional curvature in dimensions $> 3$? Even in the case of manifold with Kählerian leaves the answer is unknown. The other benefit is that new classes of manifolds appear quite naturally on the base of the local description in the so-called CR charts (Sec. 3). The main point is that locally an almost cosymplectic manifold is modeled on $\mathbb{R}\times\mathbb{C}^n$ thus we have a ’mixed’ local coordinates and an almost contact metric structure $(\varphi,\xi,\eta,g)$ can be described using these ’mixed’ coordinates. As a working example we provide such description in very details for almost cosymplectic $(-1,\mu,0)$-spaces with $\mu=const$.
Preliminaries
=============
Almost cosymplectic manifolds
-----------------------------
A manifold $\mathcal{M}$ of an odd dimension, $\dim \mathcal M=2n+1\geqslant 3$ smooth, connected, being endowed with an almost contact metric structure $(\varphi,\xi,\eta,g)$, where $\varphi$ is a $(1,1)$-tensor field, $\xi$ a vector field, $\eta$ a 1-form and $g$ a Riemannian metric, and the following conditions are satisfied [@Blair] $$\label{structeq}
\begin{array}{l}
\varphi^{2} = -I+\eta\otimes\xi, \quad \eta(\xi)=1, \quad \eta(X)=g(X,\xi),\\[+4pt]
g(\varphi X,\varphi Y)=g(X,Y)-\eta(X)\eta(Y).
\end{array}$$ is called an almost contact metric manifold. If the forms $\eta$ and the fundamental skew-symmetric 2-form $\varPhi(X,Y)=g(\varphi X,Y)$ are both closed the almost contact metric manifold $\mathcal M$ is called almost cosymplectic [@GY; @Olszak1].
An almost contact metric manifold is called normal if [@Blair] $$\label{prelim-norm}
N_\varphi+2d\eta\otimes\xi=0$$ here $N_\varphi$ denotes the Nijenhuis torsion of $\varphi$ defined by $$N_\varphi(X,Y)=\varphi^2[X,Y]+[\varphi X,\varphi Y]-\varphi[\varphi X,Y]-\varphi[X,\varphi Y].$$
The normal almost cosymplectic manifold is called cosymplectic. From (\[prelim-norm\]) it follows that an almost cosymplectic manifold is cosymplectic if and only if the torsion $N_\varphi$ vanishes. In that case the tensor field $\varphi$ is integrable as a $G$-structure, i.e. there is a suitable atlas of local coordinates charts: the local coefficients of $\varphi$ on each chart are constant functions. Arbitrary cosymplectic manifold is locally a Riemannian product of a real line (an open interval) and a Kähler manifold. D. E. Blair proved [@Blair2] that an almost contact metric manifold is cosymplectic if and only if $$\nabla\varphi =0,$$ for the Levi-Civita connection $\nabla$. S. I. Goldberg and K. Yano studying harmonic forms proved the following theorem [@GY]: an almost cosymplectic manifold is cosymplectic if and only if $$\label{prelim-Rphi}
R(X,Y)\varphi Z = \varphi R(X,Y)Z,$$ $R$ the Riemann curvature operator $$R(X,Y) = \nabla_X\nabla_Y-\nabla_Y\nabla_X -\nabla_{[X,Y]}.$$ That theorem can be viewed as an analogue of a similar theorem for Kähler manifolds. Precisely for an almost cosymplectic manifold (\[prelim-Rphi\]) implies $\nabla\varphi=0$ and by the Blair result the manifold is cosymplectic.
An almost cosymplectic manifold $\mathcal M$ carries a canonical foliation $\mathcal M = \bigcup\limits_{p\in \mathcal{M}}\mathcal{N}_p$ which corresponds to a completely integrable distribution defined by $\eta=0$. A leaf $\mathcal{N}_p\subset\mathcal{M}$ can be in a natural way considered as an almost Kähler manifold $(\mathcal{N}_p, J, G)$. The almost Hermitian structure $(J,G)$ is given by $$\inc_*\circ J=\varphi\circ \inc_*,\quad\quad G=\inc^*g,$$ $\inc$ denotes the inclusion map $\inc:\mathcal{N}_p\subset \mathcal M$. If $\mathcal M$ is cosymplectic then all leaves are in fact Kähler manifolds. However the converse is not true. Z. Olszak introduced almost cosymplectic manifolds defined by imposing the following geometric condition: each leaf is a Kähler manifold. We have the following characterization [@Olszak2]: an almost cosymplectic manifold has Kählerian leaves if and only if $$\label{kaehleav}
(\nabla_{X}\varphi)Y = -g(\varphi AX,Y)\xi + \eta(Y)\varphi AX,$$ a $(1,1)$-tensor field $A$ is defined by $$\label{defA}
AX = - \nabla_X\xi.$$
If the curvature operator of the Levi-Civita connection $R(X,Y)Z$ of an almost cosymplectic manifold satisfies $$\label{prelim-Rxi}
\begin{array}{c}
R(X,Y)\xi = \eta(Y)PX-\eta(X)PY, \\[+6pt]
P=\kappa Id+\mu h +\nu A
×
\end{array}$$ $Id$ the identity tensor, $h = \dfrac{1}{2}\mathcal{L}_\xi \varphi$, $\mathcal{L}_\xi$ the Lie derivative, and $\kappa,\mu,\nu$ are functions such that for the 1-forms $d\kappa$, $d\mu$, $d\nu$ we have $$\label{prelim-dkappa}
d\kappa\wedge\eta=d\mu\wedge\eta=d\nu\wedge\eta=0,$$ then the manifold is called an almost cosymplectic $(\kappa,\mu,\nu)$-space [@DO2]. The particular case is when $\kappa$, $\mu$, $\nu$ are constants. Almost cosymplectic manifolds satisfying the condition (\[prelim-Rxi\]) with $\kappa=const$, $\mu=\nu=0$ were studied in [@Dacko]; and with $\kappa$, $\mu=const.$, $\nu=0$ in [@Endo1; @Endo2; @Endo3]. Manifolds with $\kappa=-1$, $\mu=const.$, $\nu=0$ are classified in [@DO3].
Given positive functions $\alpha$, $\beta$ on $\mathcal{M}$ an almost contact metric structure $(\varphi',\xi',\eta',g')$ defined by $$\varphi'=\varphi,\quad \xi'=\beta^{-1}\xi,\quad \eta'=\beta\eta,\quad g'=\alpha g+(\beta^2-\alpha)\eta\otimes\eta,$$ is called a D-conformal deformation of the structure $(\varphi,\xi,\eta,g)$. The structure $(\varphi',\xi',\eta',g')$ is by itself an almost cosymplectic if and only if $\alpha$ is a constant and the function $\beta$ satisfies $d\beta\wedge\eta=0$.
If $(\mathcal{M},\varphi,\xi,\eta,g)$ is an almost cosymplectic $(\kappa,\mu,\nu)$-space then its image by a D-conformal deformation (assuming an image is almost cosymplectic) is a $(\kappa',\mu',\nu')$-space where $$\kappa' = \dfrac{\kappa}{\beta^2},\quad \mu'=\dfrac{\mu}{\beta},\quad \nu'=\dfrac{\nu\beta-d\beta(\xi)}{\beta^2}.$$
So, roughly speaking we may say that the class of almost cosymplectic $(\kappa,\mu,\nu)$-spaces is closed with respect to the group of the inner D-conformal deformations. Deformations are inner in the sense that they preserve the class of almost cosymplectic manifolds.
For an almost cosymplectic $(\kappa,\mu,\nu)$-space the tensor field $A$ and the functions $\kappa$, $\mu$ and $\nu$ satisfy the relations $$\begin{array}[]{rcl}
A^2 &=& -\kappa(Id -\eta\otimes\xi), \\
\nabla_\xi A &=& \mu\,h+\nu\,A, \\
d\kappa(\xi) &=& 2\nu\kappa,
×
\end{array}$$ and these relations are fundamental in the process of the classification of $(\kappa,\mu,\nu)$-spaces. Eg. they imply that if $\kappa = const \neq 0$ then the function $\nu$ must vanish identically. Another consequence is that arbitrary $(\kappa,\mu,\nu)$-space is (generally speaking locally) D-conformal to a $(-1,\mu,0)$-space. Thus the almost cosymplectic $(-1,\mu,0)$-spaces are of particular importance and they are called the “model spaces”.
CR manifolds ([@BER; @Jacob; @Tai])
-----------------------------------
For complex vector fields $Z_1$, $Z_2$ a bracket $[Z_1,Z_2]$ is defined as a (unique) complex vector field such that $$[Z_1,Z_2]f = Z_1(Z_2 f)-Z_2(Z_1 f),$$ for arbitrary complex-valued smooth function $f$. A CR manifold is a pair $(\mathcal M, \mathcal H)$ where $\mathcal M$ is a smooth manifold and $\mathcal H$ is a $C^\infty$ complex subbundle of $TM\otimes\mathbb C$ such that: $$\mathcal H_p\cap\bar{\mathcal H}_p =0, \quad\quad p\in M,$$ and the set of sections $\Gamma(\mathcal H)$ of $\mathcal H$ is closed with respect to the bracket operation.
A $m=\dim_{\mathbb C}\mathcal H$ is called a CR dimension of a CR manifold $(\mathcal M,\mathcal H)$ and $n-2m$ a CR codimension.
The examples of CR manifolds are complex manifolds (CR codimension is $0$) and real hypersurfaces in $\mathbb C^n$ (CR codimension is $1$).
Now let $(\mathcal M,\mathcal H)$ be a CR manifold of a hypersurface type, i.e. CR-codim $=1$. For an imaginary non vanishing $1$-form $\tau$ annihilating $\mathcal H\oplus\bar{\mathcal H}$ (it is possible that $\tau$ is defined only locally), and a vector $z\in \mathcal H_p$ let $$\label{dlform}
z \mapsto L_p(z)=\tau_p([Z,\bar Z]_p),$$ where $Z\in \Gamma(\mathcal H)$ is an local extension of $z$. The function $z \mapsto L_p(z)$ is a real quadratic form on $\mathcal H_p$ called the Levi form. The form $L_p$ is defined up to a nonzero scalar for if $\tau$ is replaced by $\tau'=f\tau$, $f(p)\neq 0$ then (\[dlform\]) yields to $L'_p(z)=f(p) L_p(z)$. If $L_p$ is non-degenerate it is a pseudo-Hermitian form on $\mathcal H_p$.
A CR-manifold $(\mathcal M,\mathcal H)$ is said to be Levi flat if its Levi form vanishes everywhere.
CR structure of an almost cosymplectic manifold
-----------------------------------------------
Let $\mathcal{M}$ be an almost contact metric manifold and $\mathcal D$ a distribution $\mathcal{D}=Im \,\varphi$, therefore $\mathcal{D}$ is a field of hyperplanes on $\mathcal{M}$ $$\mathcal{D}: \mathcal{M}\ni p \mapsto \mathcal{D}_p=\lbrace x\in T_p\mathcal{M}: x = \varphi y\; \text{for}\; y\in T_p\mathcal{M}\rbrace.$$ Equivalently $\mathcal{D}$ can be defined as the kernel of the form $\eta$. We always have $\ker \eta= Im \,\varphi$ The complexification $\mathcal D\otimes \mathbb C$ splits into a direct sum $\mathcal D'\oplus \mathcal D''$ of $\sqrt{-1}$ and $-\sqrt{-1}$ eigenspaces of the (complexified) tensor field $\varphi$. For a section $Z\in\Gamma(\mathcal D')$ let $X=\dfrac{1}{2}(Z+\bar Z)$. The conditions $\varphi Z=\sqrt{-1}\hspace{1pt}Z$, $\eta(\varphi Z)=0$ together imply that $$\label{hsec}
Z=X-\sqrt{-1}~\varphi X, \quad\eta(X)=0.$$ Conversely given a real vector field $X$, $\eta(X)=0$ (\[hsec\]) defines a section of $\mathcal D'$.
Let suppose that the pair $(\mathcal{M},\mathcal D')$ is an almost cosymplectic CR manifold, i.e. $\mathcal{M}$ is an almost cosymplectic manifold and the complex distribution $\mathcal D'$ defined as above is formally involutive. In order to compute the Levi form $L$ we can take $\tau=-\sqrt{-1}\,\eta$ in (\[dlform\]) $$\label{lform}
L_p(z) = -\sqrt{-1}\ \eta([Z,\bar Z]_p)=2\,\eta([X,\varphi X]_p),$$ where $Z=X-\sqrt{-1}\ \varphi X$ and $Z_p=z$. Since $d\eta=0$ and $\eta(X)=\eta(\varphi X)=0$ it follows that $$L_p(z)=-4\,d\eta(X_p,\varphi X_p) = 0.$$ Therefore $L$ vanishes identically and $(\mathcal M, \mathcal D')$ is a Levi flat CR manifold (of the hypersurface type).
Intuitively: a Levi flat CR-manifold is foliated by a family of (real) hypersurfaces, each hypersurface admits a complex structure and this complex structure varies when passing from a hypersurface to another hypersurface according to a differentiability class of the manifold. From this of point view, having in mind the definition of almost cosymplectic manifold with Kählerian leaves, the next proposition is almost tautological
\[kleavcr\] The pair $(M,\mathcal D')$ is a (Levi flat) CR manifold if and only if $M$ has Kählerian leaves.
In other words an almost cosymplectic CR manifold is exactly the same concept as an almost cosymplectic manifold with Kählerian leaves.
Analytic almost cosymplectic CR manifolds. The canonical Hermitian structure.
=============================================================================
CR charts on an analytic almost cosymplectic CR manifold
--------------------------------------------------------
Suppose $\mathcal{M}$ is a real-analytic almost cosymplectic CR manifold (or a manifold with Kälerian leaves), and the tensor fields $\varphi$, $\xi$, $\eta$ and $g$ are assumed to be real-analytic. In a consequence the CR structure $\mathcal D'$ of $\mathcal{M}$ is real-analytic - there are local real-analytic sections spanning $\mathcal D'$. Now, according to the theorem of A. Andreotti and D.C. Hill ([@AndrHill]) there is a local embedding (real-analytic) $$f : \mathcal{M} \rightarrow \mathbb{C}^{n+1},$$ ($ \dim \mathcal{M} =2n+1$) such that $f(\mathcal M)$ is locally a real-analytic hypersurface. If $$\mathcal{M}\ni q \mapsto p=(z^1,\ldots,z^n, z^{n+1})=f(q) \in f(\mathcal M),$$ then there is a maximal complex subspace $\mathcal{H}_p$ in $T_p^{(1,0)}\mathbb{C}^{n+1}$ such that $$\Re (\mathcal{H}_p\oplus\, \overline{\mathcal{H}}_p)
= J(T_pf(\mathcal{M}))\cap T_pf(\mathcal{M}),$$ $J$ denotes the canonical complex structure of $\mathbb{C}^{n+1}$, and the complexification of the tangent map $f_*$ defines a $\mathbb{C}$-linear isomorphism between complex spaces $$:\quad \mathcal{D'}_q \xrightarrow{f^\mathbb{C}_*} \mathcal{H}_p.$$ If $\mathcal{U}_q\subset \mathcal{M}$ is sufficiently small then $f|_{\mathcal{U}_q}$ is a diffeomorphism onto its image $f(\mathcal{U}_q)$. Now let consider a real hypersurface in $\mathbb{C}^{n+1}$. That is a set of zeros $\mathcal{S}=r^{-1}(0)$ of a smooth real-valued function $r:\mathbb{C}^{n+1}\rightarrow \mathbb{R}$. We assume that $r$ is regular: $dr\neq 0$ for each point of $\mathcal{S}$. The space $\mathbb{C}^{n+1}$ endowed with its canonical flat Kähler metric becomes an Euclidean space $\mathbb{E}^{2n+2}$ so let $H$ be the second fundamental form of $\mathcal{S}$ treated as a hypersurface in $\mathbb{E}^{2n+2}$. Let define a real quadratic differential form on a complex subbundle $\mathcal{H}:p\mapsto \mathcal{H}_p$, $p\in \mathcal{S}$ ($\mathcal{H}_p$ is defined similarly as above for $f(\mathcal{M})=\mathcal{S}$) by the formula $$L(z) = H(X,X)+ H(JX,JX), \qquad z\in \mathcal{H}_p,\; z= X-\sqrt{-1}JX,\; X\in T_p\mathcal{S},$$ for a complex vector $z$ tangent to $\mathcal{S}$. The form $L$ is a Levi form of the real hypersurface $\mathcal{S}$ (cf. [@Tai]). Let consider two examples
1. $\mathcal{S}=\mathbb{S}^{2n+1}(r)$ a canonical sphere of the radius $r$; as the second fundamental form is non-degenerate and definite the Levi form is nondegenerate and definite: $\mathbb{S}^{2n+1}(r)$ is an example of a strictly pseudo-convex real hypersurface,
2. $\mathcal{S}= \lbrace p=(z^1,\ldots,z^{n+1})\in \mathbb{C}^{2n+1}: \Im\, z^{n+1} = 0\rbrace$; clearly $\mathcal{S}$ now is simply a hyperplane hence $H=0$ identically, in a consequence the Levi form vanishes; $\mathcal{S}$ is the simplest example of a Levi flat real hypersurface.
We need the following basic result: a Levi flat real-analytic hypersurface is locally biholomorphic to the hyperplane described in the example (2). So let assume that a Levi flat real hypersurface is passing through the origin $o$ of $\mathbb{C}^{n+1}$ then there is a small disk $D_o$ centered at $o$ and a biholomorphism $F: D_o \rightarrow D_o$ such that $$F(\mathcal{S}\cap D_o) = \lbrace \Im\, z^{n+1} = 0 \rbrace \cap D_o.$$ Now let consider the sequence of maps $$\mathcal{U}_q \xrightarrow{f} f(\mathcal{U}_q) \cap D_o \xrightarrow{F} \lbrace \Im\, z^{n+1} = 0 \rbrace \cap D_o,$$ the existence of $F$ follows from the fact that $f(\mathcal{U}_q)$ is a real hypersurface (for sufficiently small $\mathcal{U}_q$) as it is defined above, i.e. there is a regular real function $r$ such that $f(\mathcal{U}_q) \subset r^{-1}(0)$ and $f(\mathcal{U}_q)$ is Levi flat.
Let $(\mathcal{M}, \mathcal{D}')$ be a real-analytic almost cosymplectic CR manifold. Then each point $q$ of $\mathcal{M}$ admits a neighborhood $\mathcal{U}_q$ and a local diffeomorphism $f_q: \mathcal{U}_q \rightarrow (-a,a)\times D'$ where $(-a,a)$ is an open interval $a>0$ and $D'$ is a small disk in $\mathbb{C}^{n}$.
By the Andreotti, Hill theorem there is a local embedding $f$ such that $f(\mathcal{U}_q)$ is a Levi flat real-analytic hypersurface in $\mathbb{C}^{n+1}$. Now (we may assume $f(q) = o$) there is a local biholomorphism $F$ of an open disk $D_o$ which maps $f(\mathcal{U}_q) \cap D_o$ onto $\lbrace \Im\, z^{n+1}=0 \rbrace \cap D$. So an image of $\mathcal{U}_q$ by the composition $F\circ f$ is contained in the hyperplane $\Im\, z^{n+1}=0$. In the natural manner we identify $\Im\, z^{n+1}=0$ with $\mathbb{R}\times \mathbb{C}^{n}$ $$\lbrace \Im\, z^{n+1}=0\rbrace \ni p=(z^1,\ldots,z^{n+1}) \mapsto (t, z^1, z^2,\ldots, z^{n}), \quad t=\Re \,z^{n+1}.$$ The differential $(F\circ f)_*$ is non-degenerate at $q$. By the standard inverse function arguments we can assert that there is an open set of the form $(-a,a)\times D' \subset (F\circ f) (\mathcal{U}_q)$ with well-defined an inverse map $(-a,a)\times D' \rightarrow \mathcal{U}_q$. An image $\mathcal{U}'_q$ by this inverse map is the required neighborhood and we set $f_q = F\circ f|_{\mathcal{U}'_q}$.
The CR structure of $(-a,a)\times D'$ is defined by $T^{(1,0)}D'$ - the complex bundle of $(1,0)$ vector fields on $D'$ in natural way embedded into the complex tangent bundle of $(-a,a)\times D'$. The family $(\mathcal{U}_q, f_q)$, $q\in \mathcal{M}$ defines a very particular atlas on $\mathcal{M}$. So the manifold is covered by the coordinates charts of the form $$(t, z^1,z^2,\ldots,z^n): (-a,a)\times D'\rightarrow \mathcal{U}_q, \quad t\in \mathbb{R}, z^i\in \mathbb{C}.$$ The transition functions $$f_{qp} = f_q\circ f_p^{-1}: f_p(\mathcal{U}_p\cap\mathcal{U}_q) \rightarrow f_q(\mathcal{U}_p\cap\mathcal{U}_q)$$ (we verify this directly) are given as follows $$\begin{array}{rcl}
f_{qp} &:& (t,z^1,\ldots,z^n) \mapsto (t', z^{'1},\ldots, z^{'n}), \\[+4pt]
t' &=& t'(t), \\[+4pt]
z^{'1} &=& z^{'1}( t,z^1,\ldots, z^n), \\
\textellipsis \\
z^{'n} &=& z^{'n}(t,z^1,\ldots,z^n),
\end{array}$$ If $Z_1,\ldots,Z_n$ are sections of $\mathcal{D'}$ and such that in the local coordinates on $\mathcal{U}_q$, $Z_i$’s are ’base’ vector fields: $$Z_1 = \dfrac{\partial }{\partial z'^1},\; \ldots \;, Z_n= \dfrac{\partial}{\partial z'^n},$$ in the local coordinates on $\mathcal{U}_p$, $Z_i$’s have the decompositions: $$\begin{array}{l}
Z_1 = \sum_{i=1}^n f_1^i\dfrac{\partial}{\partial z^i}, \\
\textellipsis \\
Z_n = \sum_{i=1}^nf_n^i\dfrac{\partial}{\partial z^i}
×
\end{array}$$ $f_j^i$ are complex functions, therefore on $\mathcal{U}_q\cap\mathcal{U}_p$ $$\begin{array}{l}
\dfrac{\partial}{\partial z'^1} = \sum_{i=1}^n f_1^i\dfrac{\partial}{\partial z^i}, \\
\textellipsis\\
\dfrac{\partial }{\partial z'^n} = \sum_{i=1}^n f_n^i\dfrac{\partial}{\partial z^i}.
×
\end{array}$$ These identities imply $$\dfrac{\partial \bar z^j}{\partial z'^i}=0, \quad \dfrac{\partial t }{\partial z'^j} =0, \quad i,j =1,\ldots,n.$$ hence $\dfrac{\partial z^j}{\partial \bar z'^i} = 0$ and as $t$ is a real-valued function $\dfrac{\partial t}{\partial \bar z'^j}=0$.
A local description of an almost contact metric structure
---------------------------------------------------------
Once we have a clear idea how to understand a ’complex coordinate’ on an almost cosymplectic CR manifold we may describe the almost cosymplectic structure $(\varphi,\xi,\eta,g)$ in terms of these complex coordinates in the way similar as it is done in the complex geometry of Hermitian manifolds. Here we follow (with necessary changes) the monograph [@KobNomV2].
From now on we will consider [*complexified*]{} structure $(\varphi^{\mathbb{C}}, \xi^{\mathbb{C}}, \eta^{\mathbb{C}},g^{\mathbb{C}})$ however we use the same notation for both the structure and its complexification. It should be clear from the context when the structure is in fact the complexification. For example $\xi^{\mathbb{C}}$ is a complex vector field. i.e. a section of $T_{\mathbb{C}}\mathcal{M}=T\mathcal{M}\otimes\mathbb{C}$. The following convention is assumed: indices matching lowercase Latin letters $i,j,k,l,\dots$ run from 1 to $n$, while Latin capitals $A, B, C,\ldots$ run through $0,1,\ldots,n,\bar 0,\bar 1,\ldots,\bar n$, moreover $\bar 0 = 0$. We set $$\label{defZ}
Z_0 = \dfrac{\partial}{\partial t},\quad\quad
Z_1=\dfrac{\partial}{\partial z^1},\;\ldots\;, Z_n= \dfrac{\partial}{\partial z^n},$$ and $$Z_{\bar i} = \overline{Z_i } = \dfrac{\partial}{\partial \bar z^i}.$$
$$\label{cstruct}
\varphi Z_A =\varphi_A^B Z_B ,\quad\quad \xi = \dsum_A\xi^A Z_A, \quad\quad\eta_A=\eta(Z_A),\quad\quad
g_{AB}=g(Z_A,Z_B).$$
Note that $g_{\bar A \bar B}=\bar g_{AB}$. Since the vector fields of $Z_1,\ldots,Z_n$ span $\mathcal D'$ we have $$\label{fiza} \varphi Z_i =\sqrt{-1}Z_i,\quad \varphi Z_{\bar i} =-\sqrt{-1}Z_{\bar i}.$$ Similarly to a Hermitian metric on a complex manifold $$\label{isotr}
g_{i\,j}= g_{\bar i\, \bar j}=0,$$ and $(g_{i\,\bar j})$ is a $n\times n$ Hermitian matrix $\bar g_{i\, \bar j} = g_{j\,\bar i}$. The coefficients $b_A=g(Z_0,Z_A)$ satisfy $$\label{coisotr}
b_{00}=\bar b_{00},\quad b_{\bar i} = g_{0\bar i}=g_{\bar 0\bar i}=\bar b_i.$$ Summing up all above formulas we can write $$\label{metr}
ds^2= r\hspace{2pt}dt^2 +2\dsum_{i=1}^n \big(b_i\hspace{2pt}dt\hspace{1pt} dz^i +
b_{\bar i}\hspace{2pt} dt\hspace{1pt}d\bar z^i\big)
+ 2\dsum_{i,j=1}^n g_{i\,\bar j}\hspace{2pt}dz^i d\bar z^j,$$ here $r$ stands for $g_{00}$. All coefficients $\eta_A$ vanish except $\eta_0=\eta(Z_0)$ therefore $\eta=udt$. Without loss of the generality we may assume $u=1$ for $d\eta=du\wedge dt=0$. Since $\bar\xi=\xi$ ($\xi$ is real) and $\eta(\xi)=1$ it follows that $$\xi =Z_0 + \sum_{i=1}^n \left( a^i Z_i +a^{\bar i}Z_{\bar i} \right)\hspace{2pt},$$ and $\bar a^i=a^{\bar i}$. In virtue of $\varphi\xi=0$ the last formula implies $$\varphi Z_0 = -\sqrt{-1}\ \sum_{i=1}^n \left( a^i Z_i
-a^{\bar i}Z_{\bar i} \right)\hspace{2pt}.$$ We point out the following relations between $b_i$, $a^i$ and $r$ $$r=1+2\dsum_{i,j=1}^n a^i a^{\bar j}g_{i\,\bar j}, \quad\quad
b_i = -\dsum_{j=1}^n a^{\bar j} g_{i\,\bar j}\hspace{2pt},$$ they are consequences of $\eta(Z_A)=g(\xi,Z_A)$ and $g(\xi,\xi)=1$.
Let $( \mathcal{M},\varphi,\xi,\eta,g)$, $\dim \mathcal{M}=2n+1\geqslant 3$ be a real-analytic almost cosymplectic manifold with Kälerian leaves. Then
- there is an open covering $(\mathcal U_\iota)_{\iota\in I}$, such that for each $\mathcal U_\iota$ there is a diffeomorphism $$f_\iota : U_\iota\rightarrow (-a,a)\times D',$$ where $(-a,a)$ is an open interval, $a>0$ and $D'$ is a domain in $\mathbb C^n$;
- on the set $U_\iota$, the structure $(\varphi,\xi,\eta,g)$ is described by the following local expressions $$\label{eksi}
\begin{array}{l}
%\label{vphit}
\varphi\dfrac{\partial}{\partial t}=-\sqrt{-1}\ \sum_{i=1}^n \left( a^i\dfrac{\partial}{\partial z^i}
- a^{\bar i}\dfrac{\partial}{\partial z^{\bar i}} \right), \quad \\[+12pt]
\varphi \dfrac{\partial}{\partial z^i}=\sqrt{-1}\ \dfrac{\partial}{\partial z^i},\quad\quad
\varphi \dfrac{\partial}{\partial z^{\bar i}}= - \sqrt{-1}\ \dfrac{\partial}{\partial z^{\bar i }}, \quad i=1,\ldots,n,\\[+12pt]
\eta= dt, \quad \\[+8pt]
\xi = \dfrac{\partial}{\partial t} + \sum_{i=1}^n \left(a^i\dfrac{\partial}{\partial z^i}
+ a^{\bar i}\dfrac{\partial}{\partial z^{\bar i}} \right), \\[+12pt]
g= r\hspace{2pt}dt^2 +2\dsum_{i=1}^n \left( b_i\hspace{2pt}dt\hspace{1pt} dz^i +
b_{\bar i}\hspace{2pt} dt\hspace{1pt}dz^{\bar i}\right)
+ 2\dsum_{i,j=1}^n g_{i\,\bar j}\hspace{2pt}dz^i\, dz^{\bar j},
\end{array}$$
- the coefficients $a^i$, $a^{\bar i}$, $b_i$, $b_{\bar i}$, $r$, $g_{i\,\bar j}$ are related by $$r=1+2\dsum_{i,j=1}^n a^i a^{\bar j}g_{i\,\bar j}, \quad\
b_i= -\dsum_{j=1}^n a^{\bar j} g_{i\,\bar j}, \quad a^{\bar i} =\bar a^i, \quad b_{\bar i}= \bar b _i,$$ and $(g_{i\,\bar j})$ is $n\times n$ Hermitian matrix.
Hermitian structure of an almost cosymplectic CR manifold
---------------------------------------------------------
We start from an extension of the Levi-Civita connection to the complex connection in the complexified bundle $T_\mathbb{C}\mathcal{M}$. For a local section $Z$ of $T_\mathbb{C}\mathcal{M}$ $$Z = \sum_{i=1}^{2n+1}f^iX_i,$$ where $(X_1,\ldots,X_{2n+1})$ is a local frame of (real) vector fields on $\mathcal{M}$ and $f^i$ are complex valued functions we define $$\begin{array}{l}
Z\mapsto \nabla Z = \nabla (\sum_{i=1}^{2n+1}f^iX_i) = \\[+12pt]
\hspace{1.5cm}=\left(\sum_{i=1}^{2n+1} df^i\otimes X_i\right) +
\left(\sum_{i=1}^{2n+1} f^i\nabla X_i\right),
×
\end{array}$$ where $df = d(\Re f+\sqrt{-1}\Im f) =d\,(\Re f)+\sqrt{-1}d\,(\Im f)$. We have to of course verify that $\nabla Z$ is independent of the choice of a local real frame. This definition follows that $$\begin{array}{l}
\nabla g Z = dg\otimes Z+g\nabla Z, \quad \;\textnormal{for a complex-valued function}\; g\\[+6pt]
\nabla Z = \nabla (X+\sqrt{-1}Y) = \nabla X+\sqrt{-1}\nabla Y,
×
\end{array}$$ $X = \Re Z$, $Y = \Im Z$ are real and imaginary parts of $Z$. Now if $Z$ is a section of the CR structure $\mathcal{D'}$ then there is a real vector field $Y$ such that $$Z=Y-\sqrt{-1}\varphi Y,\quad \eta(Y)=0,$$ therefore $$\begin{array}{rcl}
\nabla_X Z &=& \nabla_X Y-\sqrt{-1}\nabla_X\varphi Y = \\[+4pt]
&=& \left(\left(\nabla_XY-\eta(\nabla_XY)\xi\right)\right) -\sqrt{-1}\varphi\nabla_XY
+\eta(\nabla_XY)\xi +\sqrt{-1}g(A\varphi X,Y)\xi.
×
\end{array}$$ Let denote $Y'= \nabla_XY-\eta(\nabla_XY)\xi$, $Y'$ is a tangential part of the derivative $\nabla_XY$, i.e. $\eta(Y')=0$, thus $$\left(\nabla_XY-\eta(\nabla_XY)\xi\right)-\sqrt{-1}\varphi\nabla_XY =
Y'-\sqrt{-1}\varphi Y',$$ is a $\mathcal{D'}$-component of $\nabla_X Z$. Let $$Z''= \text{a}\;\mathcal{D'}\text{-component of}\; \nabla_XZ,$$ then $$\nabla_XZ = Z''+\eta(\nabla_XY)\xi +\sqrt{-1}g(A\varphi X,Y)\xi.$$ Note that $\eta(\nabla_XY)\xi = g(AX,Y)\xi$ according to the definition of $A$ and $\eta(Y)=0$: $$\eta(\nabla_XY)\xi+\sqrt{-1}g(A\varphi X,Y)\xi =
\left(g(AX,Y)+\sqrt{-1}g(A\varphi X,Y)\right)\xi = g(X, A \bar Z)\xi,$$ $\bar Z$ is the complex conjugate of $Z$. Finally $$\nabla_XZ = Z'' + g(X,A\bar Z)\xi.$$
The map $$Z \mapsto \nabla'_XZ = \nabla_XZ-g(X,A\bar Z)\xi,$$ defines a complex connection in the CR structure $\mathcal{D'}$ - as a connection in a complex vector bundle. Moreover this connection is Hermitian with respect to a Hermitian metric $H$ on $\mathcal{D'}$ defined by $$H(Z,W) = g(Z, \bar W), \quad Z,W \in \Gamma(\mathcal{D'}).$$
To prove that $\nabla'$ is Hermitian with respect to $H$ we have to show that $$XH(Z_1,Z_2) = H(\nabla'_XZ_1, Z_2)+H(Z_1,\nabla'_X Z_2),$$ for arbitrary $X$ real vector field and arbitrary sections $Z_1$, $Z_2$ of $\mathcal{D'}$. From the definition: $$XH(Z_1,Z_2) = Xg(Z_1,\bar Z_2) = g(\nabla_X Z_1, \bar Z_2)+g(Z_1, \nabla_X \bar Z_2),$$ note that $$\begin{array}{l}
\nabla_X \bar Z_2 = \overline{\nabla_X Z_2} = \overline{\nabla'_X Z_2}+\overline{g(X,A \bar Z)\xi} = \overline{\nabla'_X Z_2}+ g(X,AZ)\xi, \\[+4pt]
g(\xi, \bar Z_2) = g(Z_1, \xi) = 0
×
\end{array}$$ hence $$g(\nabla_X Z_1, \bar Z_2)+g(Z_1, \nabla_X\bar Z_2)=
g(\nabla'_XZ_1, \bar Z_2) + g(Z_1, \overline{\nabla'_X Z_2})= H(\nabla'_X Z_1, Z_2)+
H(Z_1, \nabla'_X Z_2).$$
It is very interesting topic to study the geometry of the manifold from the point of view of that canonical Hermitian structure. However these problems deserve its own attention so we stop here - as a starting point for the further investigations.
Almost cosymplectic $(\kappa,\mu,\nu)$- spaces
==============================================
An almost contact metric structure of a model space $(-1,\mu,0)$, $\mu=const$ can be realized as a left-invariant structure on a Lie group $\mathcal{G}$. If $\mathcal{G}$ is simply connected (and connected as all manifolds considered here are assumed to be connected) then $\mathcal{G}$ is diffeomorphic to $\mathbb{R}^{2n+1}$ and there is a frame of left-invariant vector fields $(Z,X_1,\ldots,X_n,Y_1,\ldots,Y_n)$, the basis of the Lie algebra $\mathfrak{g}=\mathfrak{g}(\mu)$ of $\mathcal{G}$, such that $$\begin{array}{l}
\xi = Z, \quad \eta(Z) =1, \quad \eta(X_i)=\eta(Y_j)=0, \; i,j=1,\ldots,n ,\\[+4pt]
\varphi X_i = Y_i,\quad \varphi Y_i = - X_i, \; i=1,\ldots,n ,
×
\end{array}$$ and the frame is orthonormal. Let $(t,x^1,\ldots,x^n,y^1,\ldots,y^n)$ be a global chart on $\mathbb{R}^{2n+1}$ then the vector fields $X_i$’s, $Y_j$’s are described explicitly as follows (in all cases $\xi=\partial/\partial t$ and $i=1,\ldots,n$) [@DO3]:
1. $|\mu| < 2$, $\omega=\sqrt{1-\mu^2/4}$: $$\begin{array}{rcl}
X_i &=& (\cosh(\omega t)+\dfrac{\sinh(\omega t)}{\omega})\dfrac{\partial}{\partial x^i}
- \dfrac{\mu \sinh(\omega t)}{2\omega}\dfrac{\partial}{\partial y^i}, \\[+12pt]
Y_i &=& \dfrac{\mu \sinh(\omega t)}{2\omega}\dfrac{\partial}{\partial x^i}
+ (\cosh(\omega t)-\dfrac{\sinh(\omega t)}{\omega})\dfrac{\partial}{\partial y^i},
×
\end{array}$$
2. $|\mu| =2$ ($\omega=0$): $$X_i = (1+t)\dfrac{\partial}{\partial x^i} -\varepsilon\,t\,\dfrac{\partial}{\partial y^i}, \quad
Y_i = \varepsilon\,t\, \dfrac{\partial}{\partial x^i} + (1-t)\dfrac{\partial}{\partial y^i},\quad
\varepsilon = \mu/2 = \pm 1,$$
3. $|\mu|>2$, $\omega=\sqrt{-1+\mu^2/4}$: $$\begin{array}{rcl}
X_i &=& (\cos(\omega t)+ \dfrac{\sin(\omega t)}{\omega})\dfrac{\partial }{\partial x^i}
- \dfrac{\mu\sin(\omega t)}{2\omega}\dfrac{\partial}{\partial y^i},\\[+12pt]
Y_i &=& \dfrac{\mu\sin(\omega t)}{2\omega}\dfrac{\partial}{\partial x^i} +
(\cos(\omega t)-\dfrac{\sin(\omega t)}{\omega})\dfrac{\partial}{\partial y^i},
×
\end{array}$$
As we see the local descriptions are quite different depending on the value of $\mu$. Note that algebras $\mathfrak{g}(\pm 2)$ are ’limits’ (two-sided) $$\mathfrak{g}(\pm 2) = \lim\limits_{\mu \rightarrow \pm 2}\mathfrak{g}(\mu), \quad \mu\neq \pm 2.$$ Nevertheless the commutators of the vector fields can be described in a unique manner in all cases: $$\label{munu-comm}
\mathfrak{g}(\mu):\quad [\xi,X_i] = X_i-\dfrac{\mu}{2}Y_i, \quad [\xi,Y_i] = \dfrac{\mu}{2}X_i-Y_i,$$ the other commutators vanish identically.
On the base of the relations (\[munu-comm\]) here we will provide a different local representation. Let again $\mathcal{M} = \mathbb{R}^{2n+1}$, $\mathcal{M}\ni p = (t,x^1, \ldots,x_n,y^1,\ldots,y_n)$ and we set $$\label{munu-xiCR}
\begin{array}{l}
\xi = \dfrac{\partial}{\partial t} + \sum\limits_{i=1}^n(-x^i-\dfrac{\mu}{2}y^i)\dfrac{\partial}{\partial x^i}
+ \sum\limits_{i=1}^n(\dfrac{\mu}{2}x^i+y^i)\dfrac{\partial}{\partial y^i},\\
X_i =\dfrac{\partial}{\partial x^i},\qquad Y_i =\dfrac{\partial}{\partial y^i}, \quad i=1,\ldots,n
×
\end{array}$$ Direct computations show that such defined vector fields satisfy (\[munu-comm\]). Now, let identify $\mathbb{R}^{2n+1} \cong \mathbb R \times \mathbb{C}^n$ in the way that $$\label{munu-CRcoord}
\begin{array}{l}
p = (t,x^1,\ldots,x^n,y^1,\ldots,y^n) = (t, z^1,\ldots, z^n) \in \mathbb{R}\times \mathbb{C}^n, \\[+4pt]
z^1 = x^1+\sqrt{-1}\,y^1, \ldots,\; z^n = x^n+\sqrt{-1}\,y^n,
\end{array}$$ the vector fields $$\label{munu-Zframe}
\begin{array}{rcl}
Z_0 &=& \dfrac{\partial}{\partial t}, \\[+6pt]
Z_i &=& \dfrac{1}{2}(\dfrac{\partial}{\partial x^i}-\sqrt{-1}\;\dfrac{\partial}{\partial y^i})=
\dfrac{\partial}{\partial z^i},
\quad i=1,\ldots,n \\[+12pt]
Z_{\bar i} &=& \dfrac{1}{2}(\dfrac{\partial}{\partial x^i}+\sqrt{-1}\dfrac{\partial}{\partial y^i})=
\dfrac{\partial}{\partial \bar z^i},
\quad i =1,\ldots,n
×
\end{array}$$ form a frame (in our case global) of vector fields of the complexified tangent bundle $T_{\mathbb{C}}\mathcal{M}$. According to (\[munu-xiCR\]), (\[munu-CRcoord\]), (\[munu-Zframe\]) we have $$\begin{array}{l}
\xi =\dfrac{\partial}{\partial t}+ \sum\limits_{i=1}^n(-x^i-\dfrac{\mu}{2}y^i)\dfrac{\partial}{\partial x^i}
+ \sum\limits_{i=1}^n(\dfrac{\mu}{2}x^i+y^i)\dfrac{\partial}{\partial y^i} = \\
= \dfrac{\partial}{\partial t} +\dfrac{1}{2}\sum\limits_{i=1}^n(-(z^i+\bar z^i)-\dfrac{\mu}{2}\dfrac{z^i -\bar z^i}{\sqrt{-1}})(\dfrac{\partial}{\partial z^i}+\dfrac{\partial}{\partial \bar z^i}) + \\
+ \dfrac{1}{2}\sum\limits_{i=1}^n(\dfrac{\sqrt{-1}\mu}{2}(z^i+\bar z^i) + (z^i-\bar z^i))(\dfrac{\partial}{\partial z^i}-\dfrac{\partial}{\partial \bar z^i}).
\end{array}$$ Providing similar computations for the (complexified) tensor fields $\varphi$, $\eta$, and $g$ we obtain such expressions: $$\begin{array}{l}
\eta = dt, \\
\xi = \dfrac{\partial}{\partial t} +
\sum\limits_{i=1}^n(-\bar z^i +\dfrac{\sqrt{-1}\mu}{2}z^i)\dfrac{\partial }{\partial z^i} % eta
+\sum\limits_{i=1}^n (-z^i-\dfrac{\sqrt{-1}\mu}{2}\bar z^i)\dfrac{\partial}{\partial \bar z^i}, \\[+12pt]
\sqrt{-1}\;\varphi\;\dfrac{\partial}{\partial t} =\sum\limits_{i=1}^n (-\bar z^i+\dfrac{\sqrt{-1}\mu}{2}z^i)\dfrac{\partial}{\partial z^i} +\sum\limits_{i=1}^n(z^i+\dfrac{\sqrt{-1}\mu}{2}\bar z^i)\dfrac{\partial}{\partial \bar z^i},\\[+12pt]
\varphi \dfrac{\partial}{\partial z^i} = \sqrt{-1}\dfrac{\partial}{\partial z^i},\qquad
\varphi \dfrac{\partial}{\partial \bar z^i} = -\sqrt{-1}\dfrac{\partial}{\partial \bar z^i},\qquad i=1,\ldots,n ,\\[+12pt]
ds^2 = r\,dt^2 + 2\sum\limits_{i=1}^n (z^i+\dfrac{\sqrt{-1}\mu}{2}\bar z^i) dt dz^i
+ 2\sum\limits_{i=1}^n(\bar z^i-\dfrac{\sqrt{-1}\mu}{2}z^i) dt d\bar z^i + 2\sum\limits_{i=1}^n dz^i d\bar z^i,\\
r= 1+2\sum\limits_{i=1}^n|z^i+\dfrac{\sqrt{-1}\mu}{2}\bar z^i|^2.
\end{array}$$ Recently D. Perrone has classified Riemannian homogeneous simply connected almost cosymplectic three-folds under an assumption that there is a group of isometries acting transitively and leaving the form $\eta$ invariant [@Perrone].
[@Perrone] Let $(\mathcal{M},\varphi,\xi,\eta,g)$ be a be a simply connected homogeneous [^1] almost cosymplectic three-manifold. Then either $\mathcal{M}$ is a Lie group $G$ equipped with a left invariant almost cosymplectic structure, or a Riemannian product of type $\mathbb{R}\times \mathcal{N}$, where $\mathcal{N}$ is a simply connected Kähler surface of constant curvature.
From the list provided in [@Perrone] we are interested only in the case of non-cosymplectic three-manifolds and unimodular Lie groups as all Lie groups corresponding to the algebras $\mathfrak{g}(\mu)$ are unimodular:
- the universal covering $\tilde E(2)$ of the group of rigid motions of Euclidean 2-space, when $p > 0$,
- the group $E(1,1)$ of rigid motions of Minkowski 2-space when $p < 0$,
- the Heisenberg group $H^3$ when $p = 0$,
here $p$ is a metric invariant of the classification defined by $$p = \|\mathcal{L}_\xi h\|-2\|h\|^2.$$
Let $(\mathcal{M}^3,\varphi,\xi,\eta,g)$ be a simply connected 3-dimensional almost cosymplectic\
$(-1,\mu,0)$-space. Then $\mathcal{M}^3$ is a Lie group equipped with appropriate left invariant almost cosymplectic structure and
- $\mathcal{M}^3$ is a universal covering of the group of rigid motions of Euclidean 2-plane for $|\mu| > 2$,
- $\mathcal{M}^3$ is a group of rigid motions of Minkowski 2-plane for $|\mu| <2 $,
- $\mathcal{M}^3$ is a Heisenberg group for $|\mu| =2 $
[99]{}
A. Andreotti and D. C. Hill, [*Complex characteristic coordinates and tangential Cauchy-Riemann equations*]{}, Ann. Scuola Norm. Sup. Pisa (3) [**26**]{} (1972), 299–324.
M. S. Baouendi, P. Ebenfelt and L. P. Rothschild, [*Real submanifolds in a complex space and their mappings*]{}, Princeton Math. Series, 47. Princeton Univ. Press, Princeton, New Jersey, 1999.
D. E. Blair, [*The theory of quasi-Sasakian structures*]{}, J. Diff. Geom. [**1**]{}, 331–345
D. E. Blair, [*Riemannian geometry of contact and symplectic manifolds*]{}, Progress in Math. Vol. [**203**]{}, Birkhäuser, Boston, 2001.
P. Dacko, [*On almost cosymplectic manifolds with the structure vector field $\xi$ belonging to the $k$-nullity distribution*]{}, Balkan J. Geom. Appl. [**5**]{} (2000), no. 2, 47–60.
P. Dacko and Z. Olszak, [*On almost cosymplectic $(\kappa,\mu,\nu)$-spaces*]{}, PDEs, submanifolds and affine differential geometry, 211–220, Banach Center Publ., [**69**]{}, Polish Acad. Sci., Warsaw, 2005.
P. Dacko and Z. Olszak, [*On almost cosymplectic $(-1,\mu,0)$-spaces*]{}, Cent. Eur. J. Math. [**3**]{} (2005), no. 2, 318–330 (electronic).
H. Endo, [*On some properties of almost cosymplectic manifolds*]{}, An. Ştiin. Univ. Al. I. Cuza Iaşi. Mat. (N.S.) [**42**]{} (1996), no. 1, 79–94 (1997).
H. Endo, [*On some invariant submanifolds in certain almost cosymplectic manifolds*]{}, An. Ştiin. Univ. Al. I. Cuza Iaşi. Mat. (N.S.) [**43**]{} (1997), no. 2, 383–395 (1998).
H. Endo, [*Non-existence of almost cosymplectic manifolds satisfying a certain condition*]{}, Tensor (N.S.) [**63**]{} (2002), no. 3, 272–284.
S. I. Goldberg and K. Yano, [*Integrability of almost cosymplectic structures*]{}, Pacific J. Math. [**31**]{} (1969), 373–382.
H. Jacobowitz, [*An introduction to CR structures*]{}, Math. Surveys and Monographs, [**32**]{}, American Mathematical Society, Providence, Rhode Island, 1990.
M. Jurchescu, [*Variétés mixtes et cohomologie*]{}, Riv. Mat. Univ. Parma, [**10**]{} (1984), no. 4, 55–79.
S. Kobayashi and K. Nomizu, [*Foundations of differential geometry. Vol. II*]{}, Interscience Tracts in Pure and Applied Mathematics, No. 15 Vol. II Interscience Publishers John Wiley & Sons, Inc., New York-London-Sydney 1969
Z. Olszak, [*On almost cosymplectic manifolds*]{}, Kodai Math. J. [**4**]{} (1981), 239–250.
Z. Olszak, [*Almost cosymplectic manifolds with Kählerian leaves*]{}, Tensor N.S. [**46**]{} (1987), 117–124.
D. Perrone, [*Classification of homogeneous almost cosymplectic three-manifolds*]{}, Differential Geom. Appl. [**30**]{} (2012), no. 1, 49–58.
G. Taiani, [*Cauchy-Riemann (CR) manifolds*]{}, Pace University, Mathematics Department, New York, 1989.
S. Tanno, [*Variational problems on contact Riemannian manifolds*]{}, Trans. Amer. Math. Soc. [**314**]{} (1989), no. 1, 349–379.
[^1]: It is assumed that $\eta$ is invariant
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper introduces the Intentional Unintentional (IU) agent. This agent endows the deep deterministic policy gradients (DDPG) agent for continuous control with the ability to solve several tasks simultaneously. Learning to solve many tasks simultaneously has been a long-standing, core goal of artificial intelligence, inspired by infant development and motivated by the desire to build flexible robot manipulators capable of many diverse behaviours. We show that the IU agent not only learns to solve many tasks simultaneously but it also learns faster than agents that target a single task at-a-time. In some cases, where the single task DDPG method completely fails, the IU agent successfully solves the task. To demonstrate this, we build a playroom environment using the MuJoCo physics engine, and introduce a grounded formal language to automatically generate tasks.'
author:
- |
Serkan Cabi Sergio Gómez Colmenarejo Matthew W. Hoffman\
**Misha Denil Ziyu Wang Nando de Freitas**\
DeepMind\
`{cabi,sergomez,mwhoffman,mdenil,ziyu,nandodefreitas}@google.com`
bibliography:
- 'multihead.bib'
title: |
The Intentional Unintentional Agent:\
Learning to Solve Many Continuous Control Tasks Simultaneously
---
Introduction {#sec:intro}
============
Imagine a toddler in a playroom trying to bring two blocks together. While purposely focusing on this task, the infant is accomplishing many other goals incidentally both simpler and more complex: gazing, extending the arms, sitting, bending sideways, reaching, grasping, navigating around obstacles, dodging a looming object thrown by a sibling, sensing texture, sensing temperature, and so on. Over the first year, infants also display a wide range of spontaneous movements: kicks, stomps, sways, flaps, flails, rocks, rubs, nods, shakes, bounces, bangs, waves, wiggles and so on [@thelen:1979; @adolph:2015].
We hypothesize that a single stream of experience offers agents the opportunity to learn and perfect many policies both on purpose and incidentally, thus accelerating the acquisition of grounded knowledge.
To investigate this hypothesis, we propose a deep actor-critic architecture, trained with deterministic policy gradients [@silver:2014; @lillicrap:2016], for learning several policies concurrently. The architecture enables the agent to attend to one task on-policy, while unintentionally learning to solve many other tasks off-policy. Importantly, the policies learned unintentionally can be harnessed for intentional use even if those policies were never followed before.
More precisely, this [intentional-unintentional architecture]{}, shown in Figure \[fig:actor-critic-nets\], consists of two neural networks. The actor neural network has multiple-heads representing different policies with shared lower-level representations. The critic network represents several state-action value functions, sharing a common representation for the observations.
![The IU architecture. The actor network on the left consists of two shared MLP-tanh layers followed by non-shared MLP-tanh layers to produce the multivariate actions for each policy (4 policies in this diagram). The right hand side shows the critic network. The action vectors provided by the policies are fed into a non-shared MLP-tanh layer, which is then point-wise added to the ouput of a two layer MLP-tanh network applied to the observation. The resulting activations are processed by non-shared linear layers to produce the $Q$ values.[]{data-label="fig:actor-critic-nets"}](artwork/multihead-actor "fig:"){width="0.2\linewidth"} ![The IU architecture. The actor network on the left consists of two shared MLP-tanh layers followed by non-shared MLP-tanh layers to produce the multivariate actions for each policy (4 policies in this diagram). The right hand side shows the critic network. The action vectors provided by the policies are fed into a non-shared MLP-tanh layer, which is then point-wise added to the ouput of a two layer MLP-tanh network applied to the observation. The resulting activations are processed by non-shared linear layers to produce the $Q$ values.[]{data-label="fig:actor-critic-nets"}](artwork/multihead-critic "fig:"){width="0.27\linewidth"}
The architecture alone does not suffice for investigating our hypothesis. We also need a flexible way of generating many diverse tasks and a suitable environment. To this end, we introduce an automatic procedure to generate semantic goals for the agent. We also introduce a physical environment, with gravity, a ground, rigid objects, and a simple embodied agent. While some characteristics of the physical world change with experiences, the laws of physics and the body remain fixed to allow for transfer and continual learning.
While being specific, this study aims to address generality in artificial intelligence by designing agents capable of doing many things, to overcome the problem of sparse rewards associated with conventional reinforcement learning by generating and controlling a stream of reward functions. We design embodied agents situated in a physical environment resembling a playroom, not only as a consequence of being inspired by infant development, but also as a result of being interested in eventual transfer to the world of flexible manufacturing with robots.
Related work
------------
Thinking about an agent as immersed in a stream of multivariate rewards provides us with a powerful alternative to the conventional univariate reward reinforcement learning framework.
Recently, @Jaderberg:2017 consider Asynchronous Advantage Actor Critic (A3C) agents [@mnih:2016] immersed in a sensorimotor stream. The agents are designed to achieve an extrinsic scalar reward, but are endowed with *auxiliary control tasks* and *auxiliary predictive tasks*. The auxiliary control tasks (pixel changes and simple network features) are shown to enable the A3C agent to learn to achieve the scalar reward faster in domains where the action-space is discrete. This paper will show more dramatic gains for continuous action spaces.
We refer to the task whose behavior the agent follows during training as the [*intentional*]{} task, and to the remaining tasks as [*unintentional*]{}. Philosophically, in our work, the unintentional tasks are not thought of as playing a mere auxiliary role, but they can themselves become the intentional task. Our tasks are of a semantic nature, for example “[*move the red block east of the blue block*]{}”, and hence it is sensible to learn a library of such tasks for potential future re-use.
@sutton:2011 introduced the horde architecture to learn grounded knowledge from an unsupervised sensorimotor stream. While not focusing on the issue of representation, with the value functions being trained separately with different weights, the Horde provides much inspiration for this paper.
The concept of general value functions introduced in the horde architecture was further explored by @Schaul:2015, and has connections with research on options, successor representations and hierarchical RL [@Dietterich:1998; @Sutton:1998; @Kulkarni:2016]. At the time of writing this paper, @vanSeijen2017 introduced a linear decomposition of reward functions, related to [@russell2003], whereby several action-value functions are learned separately with Deep-Q-Networks (DQN) [@Mnih:2015].
Multivariate reward feedback appears in sequential multi-objective decision making [@Roijers:2013], in predictive decision making for 3D games [@Dosovitskiy:2017], and in agents that use auxiliary predictive tasks to reduce sample complexity [@li:2015; @lample:2016].
Our work is related to learning neural networks with a static or adaptive curricula [@Bengio:2009; @Zaremba:2014; @reed:2015], and to learning curricula for training neural networks using bandit techniques and Bayesian optimization [@Tsvetkov:2016; @Graves:2017].
In developmental psychology, there are many studies on incidental activity and its consequences on motor development, including fetal and neonatal movement [@sparling:1999], twitching during sleep [@blumberg:2013], stereotypies and flails in infants, and gross motor play [@adolph:2015].
Findings in developmental psychology have inspired the design or robotic systems in the field of developmental robotics [@lungarella:2003]. It has been also been argued that robots provide a platform for examining many of these findings. We believe that rich physical simulators are a viable alternative, provided that we keep advancing environments and task generation mechanisms as done in this paper. The importance of bodies and physical environments in the study of artificial intelligence has been championed by many, notably by @brooks:1991.
The automatic construction of grounded reward functions with formal languages has become a topic of great interest in recent months [@Littman2017a; @Yu2017; @Hermann2017; @Denil2017].
The Intentional Unintentional Agent
===================================
Policy gradient algorithms form a very popular, if not the most popular, class of continuous action reinforcement learning algorithms. The fundamental basis of many of these algorithms is the *policy gradient theorem* [@sutton:1999]. However, this approach necessitates the use of stochastic policies which can complicate the process of learning off-policy. More recently, a [deterministic policy gradient theorem]{} has been formulated by [@silver:2014] which removes this need. This approach was later extended in [@lillicrap:2016] to modern deep neural network actor-critic architectures, with *scalar rewards*.
In our setting, the agent perceives a *stream of rewards* $r_t^i$, indexed by $i$ at time $t$. To learn the actor neural network parameterized by $\theta$, we are interested in simultaneously maximizing the expected value of all tasks, that is $$\begin{aligned}
J(\theta)
&=
\E_{\rho^\beta}\Big[
\sum_i Q_\mu^i(\vs, \mu^i_\theta(\vs))
\Big],\end{aligned}$$ where $\mu_\theta^i$ is the actor’s policy associated with the $i$-th task, mapping a state vector $\vs$ to a continuous action vector $\va$, and $Q^i$ is the action-value critic associated with this task. The expectation above is taken with respect to $\rho^\beta$, the stationary distribution of some behavior policy $\beta(\va|\vs)$. Note that due to the fact that multiple policies are being learned at once we must necessarily be learning off-policy. The corresponding gradient for the actor is $$\begin{aligned}
\nabla_\theta J(\theta)
&
\approx
\E_{\rho^\beta}\Big[
\sum_i
\nabla_\theta \mu_\theta^i(\vs)\,
\nabla_{\va^i} Q_\mu^i(\vs, \va^i)\big|_{\va^i=\mu^i_\theta(\vs)}
\Big].\end{aligned}$$ The behaviour policy is effectively given by the intentional policies as we will detail shortly. Given an observation $\vs_t$, the behavior policy produces the action vector $\va_t$. In response, the environment returns a *reward vector* $\vr_t$, with one scalar component for each task, and the next state observation $\vs_{t+1}$. The tuple $(\vs_t,\va_t,\vr_t,\vs_{t+1})$ is stored in a replay buffer. Note that reward observations do not enter into the gradient estimate as they have instead been captured by the action-value function $Q_\mu$. However, since this quantity is never directly observed, we will instead replace this function with a parameterized critic $Q_w$ which must be trained by an appropriate policy evaluation mechanism. Here we update the critic in order to simultaneously minimize the temporal difference error of all tasks.
To update the critic and actor, we sample a mini-batch of tuples uniformly at random from the replay buffer and perform stochastic gradient descent with respect to both the actor and critic losses. Combining these updates we have, $$\begin{aligned}
\delta_j^i
&=
r_j^i + \gamma Q_{w'}^i(\vs_{j+1},
\mu^i_{\theta'}(\vs_{j+1}))
- Q_w^i(\vs_j,\va_j), \label{eq:delta}
\\
w
&\gets
w + \alpha_\text{critic}
\sum_j \sum_i \delta_j^i \,\nabla_w Q_w^i(\vs_j, \va_j),
\\
\theta
&\gets
\theta + \alpha_\text{actor}
\sum_j \sum_i
\nabla_\theta \mu_\theta^i(\vs_j)\,
\nabla_{\va^i} Q_w^i(\vs_j, \va^i)
\big|_{\va^i=\mu_{\theta}^i(\vs_j)}.\end{aligned}$$ Here $j$ represents the sampled set of indices for the mini-batch. Note that in equation (\[eq:delta\]), the same action sampled from the replay buffer, $\va_j$, is used to update every critic $Q^i$. We have also adopted target parameters $\theta'$ and $w'$ which stabilize learning of the critic, as in [@lillicrap:2016]. The target parameters are periodically updated to reflect the current values of the optimization parameters. In practice, rather than using fixed learning rates for the actor and critic we instead use dynamic learning rates, e.g. via the Adam optimizer [@kingma:2015].
We can now discuss the behavior policy used during training. In this work we use $$\begin{aligned}
\beta(\va|\vs) = \mu_\theta^i(\vs) + Z,\end{aligned}$$ where $Z$ is a random variable introduced for exploration; following [@lillicrap:2016] we utilize temporally correlated noise from an Ornstein-Uhlenbeck process. As before, the index $i$ denotes the intentional task that is being followed. We considered different variations on this process, for example following a single task, selecting a random task to follow at the beginning of each episode, and switching between tasks whenever the current task is successful. However, for clarity of presentation, we focus on having a single intentional policy solving one of the hardest tasks in the physical playroom domain. We will return to this point in Section \[sec:discussion\]
Experimental Setup {#sec:experiments}
==================
Our experiments are set in a virtual physical world in which the agent can interact with a variety of objects and obtain rewards by satisfying many procedurally-generated semantic relations between the objects. The interactions among the physical objects involve complex contact forces, which can pose significant challenges for control algorithms.
Crucially, the agent is physically embodied in this domain and its actions have consistent dynamics throughout the space. Embodiment, object commonality and consistent physics enable the agent to learn features that effectively generalize between different tasks. In this section we will first describe the physical environment in which the agent is situated and then detail the method by which we automatically generate tasks and grounded rewards.
The physical playroom domain
----------------------------
[r]{}[0.36]{}
We implemented a domain using the MuJoCo physics engine [@todorov:2012] which consists of a tabletop on which $N$ objects are placed. Each object may have different properties associated with it, for example color, size, friction, density, shape and so on, which allow the objects to be either partially or fully specified.
The agent is embodied as an actuated “fist” whose action space consists of 2-dimensional velocities. In other words, the agent can move its fist through the playroom domain by setting its immediate velocity. The agent is able to interact with other objects and affect their positions only indirectly by way of contact with these objects. Although the fist is only able to move in 2-dimensions the other objects can exhibit more complex behavior based on the speed and angle at which contacts occur.
The playroom consists of a 80 cm$^2$ square arena populated with $N$ objects, each of which is sized to have a diameter of 12 cm. Figure \[fig:playroom\] shows an example setting of this domain with $N=3$ cubes of different colors. In the figure we can also see the agent’s actuated fist, displayed as a white sphere. The fist can be thought of as a directly controllable object, and as a result we will be able to directly express relations between this object and other objects in the scene.
We have also added a border to the environment in order to reduce the occurrence of irrecoverable actions wherein objects are pushed into corners and consequently cannot be retrieved. The border allows the fist to move into the corners but not the blocks. Finally, we will also consider relations between objects and a given goal position. These goals are implemented as immovable goal objects for the purposes of allowing relational expressions; in particular this object is implemented as a *pad* sitting at ground level and hence does not physically interact with any object. An example of one such object is displayed in Figure \[fig:playroom\] as a white pad in the upper right corner.
Automatic reward generation with formal language
------------------------------------------------
Given the above domain we are now interested in generating rewards based on the properties of objects as well as relations between these objects. We will define a number of property functions $p:\calO\times\calS\to\{0,1\}$ where $\calS$ is the set of possible world states. Each property acts as an indicator over objects $o\in\calO$, taking value $1$ when the object satisfies a certain property. It is instructive to think of these properties as defining sets of objects in the scene. Often these properties will be independent of the state $s\in\calS$ in which case we can write them simply as $p(o)$. For example we might define $p_\texttt{blue}(o)$ to be a property of blue objects, which is independent of their location, velocity, etc.
Next we can introduce binary relations between objects which can be thought of as functions of the form $b:\calO\times\calO\times\calS\to\{0,1\}$. For example, we can introduce a “nearness” relation which holds when two objects are close to each other, that is when their center points are within some specified distance. By combining properties and relations we can write rewards of the form $$\begin{aligned}
r_\texttt{red\_near\_blue}(\vs)
&=
\sum_{o_1, o_2} p_\texttt{red}(o_1) \,p_\texttt{blue}(o_2)
\,b_\texttt{near}(o_1, o_2, \vs).\end{aligned}$$ where $o_1$ and $o_2$ represent object identifiers and we are summing over all pairs of objects such that the properties select a particular pair. The above reward represents the task “bring a red object near a blue object”. If the objects identified in the scene are uniquely identified by their properties this may also be more succinctly written as $$\begin{aligned}
r_\texttt{red\_near\_blue}(\vs)
&=
b_\texttt{near}(\texttt{red}, \texttt{blue}, \vs).\end{aligned}$$ where $\texttt{red}$ and $\texttt{blue}$ are the uniquely identified objects. While properties are frequently independent of the state, relations such as $\texttt{near}$ will depend on this state, specifically the positions of its two input objects $o_1$ and $o_2$.
We can generate many different rewards by logically combining these operations. In the experiments that follow we will make use of the following atoms:
1. color-based properties: $p_\texttt{red}$, $p_\texttt{blue}$, $p_\texttt{green}$;
2. properties identifying the fist $p_\texttt{fist}$ and the goal state $p_\texttt{goal}$;
3. near and far relations: $b_\texttt{near}$, $b_\texttt{far}$, parameterized by a distance parameter $\epsilon$ such that whenever the distance between two blocks is less than $\epsilon$ they are considered to be near;
4. directional relations: $b_\texttt{north}$, $b_\texttt{east}$, $b_\texttt{south}$, $b_\texttt{west}$ such that $b_\texttt{north}(a, b)$ is 1 whenever object $a$ is north of object $b$.
Finally, more complicated rewards can be expressed by and-ing these relations. A “gather to pad” task which collects blocks of the three different colors to a goal pad location can be written as $$\begin{aligned}
r_\texttt{gather\_to\_pad}(\vs)
&=
b_\texttt{near}(\texttt{red}, \texttt{pad}, \vs)\,
b_\texttt{near}(\texttt{blue}, \texttt{pad}, \vs)\,
b_\texttt{near}(\texttt{green}, \texttt{pad}, \vs).\end{aligned}$$ One can also simply gather blocks together or move them far apart from each other, without having to specify a pad or coordinates.
Results {#sec:results}
=======
![[**Left**]{}: Test performance for the task of gathering two blocks together in the playroom, using a varying number of additional tasks. [**Right**]{}: What happens when an extra block is added to the environment, causing significant physical interference. In both cases, the more tasks the IU agent solves simultaneously, the faster it learns the intentional task.[]{data-label="fig:two-blocks"}](figures/small-table-gathering-head-comparison "fig:"){width="49.00000%"} ![[**Left**]{}: Test performance for the task of gathering two blocks together in the playroom, using a varying number of additional tasks. [**Right**]{}: What happens when an extra block is added to the environment, causing significant physical interference. In both cases, the more tasks the IU agent solves simultaneously, the faster it learns the intentional task.[]{data-label="fig:two-blocks"}](figures/small-table-gathering-extra-block "fig:"){width="49.00000%"}
The agent observes the relative position of the fist actuator to each object in the scene. Actions are taken by moving the fist in 2-dimensions, which involves setting the velocities of the actuator. We followed the training protocol described in Appendix A of the DDPG paper of @lillicrap:2016. The actor and critic multi-layer perceptrons, shown in Figure 1, have 200 and 400 units in each layer respectively and standard hyperbolic tangent activations.
We first consider tasks involving two uncontrolled objects–––identified by the red and blue properties—as well as the fist actuator. Our tasks consist of 6 relations: north, south, east, west, near, and far, between each pair of objects including the fist. There are 3 distinct pairs of objects, which results in 18 total reward functions. We are interested in how the use of extra reward functions affects the ability of our agent to learn a single intentional policy.
Figure \[fig:two-blocks\] (left) illustrates the ability of our agent to learn to maximize the single intentional reward $b_\texttt{near}(\texttt{red},\texttt{blue}, \vs)$, when given access to additional signals. We consider three scenarios: no additional unintentional tasks (i.e. standard DDPG), all near and far tasks (6), and finally all 18 tasks. We see that by considering all 18 tasks simultaneously, the IU agent learns to control the system under a single intentional reward much more rapidly. In these plots the test-time reward is shown, averaged over 50 runs, where the error bounds show the min and max rewards for a given evaluation.
Figure \[fig:all-heads\] shows the test performance of the 18 policies learned simultaneously. Note that some tasks are easier than others, explaining the different average rewards to which each of the tasks converge. Those tasks which are more difficult require more time to obtain non-zero values and hence converge to an average return less than 1.0. All tasks, however, are able to improve over their initial baseline. Next we examine the ability of the agents to learn when an additional (green) cube is added to the playroom. The additional object results in 6 distinct pairs of objects and as a result increases the number of tasks to 36 ($6\times 6$). The results are shown in the right-most plot of Figure \[fig:two-blocks\]. We see that although learning is slower, the system is still able to learn to achieve the gathering task despite the distracting block, which can of course cause significant interference. Once again, the more tasks the agent solves the faster it learns.
[r]{}[0.55]{}
{width="46.00000%"}
In Figure \[fig:three-blocks\], we display the performance of the agent on tasks with three blocks. Here the intentional task involves placing all three blocks in the upper right corner pad. We display the average reward for this task when solving 1, 7, and 43 tasks. DDPG when following just a single task is incapable of succeeding at this task. However, the intentional unintentional agent succeeds when following 7 and 43 additional rewards. On the right panel of Figure \[fig:three-blocks\], we display the performance of the agent when trained on a larger task space, that is a ground with sides that are 50% larger in length. We observe that eventually, as the complexity of exploration increases, even the IU agents struggle. However by following 43 tasks the IU agent is still capable of gaining some reward, whereas DDPG completely fails.
Finally, in Figure \[fig:strip\] we display frames from example trajectories for various policies learned in the 3-block domain. Shown below these trajectories are illustrations of the object positions over time. The IU agent is able to learn reasonable action sequences which move the objects to the goal pad.
![IU agent executing three different policies. The top film-strip shows the agent solving the hardest task: gathering all three blocks into the upper right corner. The two strips below this show the agent moving a single block into the corner; red and blue respectively. Below these temporal illustrations is a figure illustrating a coordinate-centric view of each of these three tasks. Here the black line denotes the position of the fist as a function of time, and the colored lines represent the positions of each of the three colored blocks. We can see from the first plot that the agent moves each block to reach the goal, shown as an orange semicircle. For the second and third tasks, only the red and blue blocks are acted on respectively, although we can see in the second plot that the agent unintentionally moves the other two blocks slightly.[]{data-label="fig:strip"}](figures/strip_0 "fig:"){width="96.00000%"} ![IU agent executing three different policies. The top film-strip shows the agent solving the hardest task: gathering all three blocks into the upper right corner. The two strips below this show the agent moving a single block into the corner; red and blue respectively. Below these temporal illustrations is a figure illustrating a coordinate-centric view of each of these three tasks. Here the black line denotes the position of the fist as a function of time, and the colored lines represent the positions of each of the three colored blocks. We can see from the first plot that the agent moves each block to reach the goal, shown as an orange semicircle. For the second and third tasks, only the red and blue blocks are acted on respectively, although we can see in the second plot that the agent unintentionally moves the other two blocks slightly.[]{data-label="fig:strip"}](figures/strip_1 "fig:"){width="96.00000%"} ![IU agent executing three different policies. The top film-strip shows the agent solving the hardest task: gathering all three blocks into the upper right corner. The two strips below this show the agent moving a single block into the corner; red and blue respectively. Below these temporal illustrations is a figure illustrating a coordinate-centric view of each of these three tasks. Here the black line denotes the position of the fist as a function of time, and the colored lines represent the positions of each of the three colored blocks. We can see from the first plot that the agent moves each block to reach the goal, shown as an orange semicircle. For the second and third tasks, only the red and blue blocks are acted on respectively, although we can see in the second plot that the agent unintentionally moves the other two blocks slightly.[]{data-label="fig:strip"}](figures/strip_2 "fig:"){width="96.00000%"}
![IU agent executing three different policies. The top film-strip shows the agent solving the hardest task: gathering all three blocks into the upper right corner. The two strips below this show the agent moving a single block into the corner; red and blue respectively. Below these temporal illustrations is a figure illustrating a coordinate-centric view of each of these three tasks. Here the black line denotes the position of the fist as a function of time, and the colored lines represent the positions of each of the three colored blocks. We can see from the first plot that the agent moves each block to reach the goal, shown as an orange semicircle. For the second and third tasks, only the red and blue blocks are acted on respectively, although we can see in the second plot that the agent unintentionally moves the other two blocks slightly.[]{data-label="fig:strip"}](figures/coords_0 "fig:"){width="32.00000%"} ![IU agent executing three different policies. The top film-strip shows the agent solving the hardest task: gathering all three blocks into the upper right corner. The two strips below this show the agent moving a single block into the corner; red and blue respectively. Below these temporal illustrations is a figure illustrating a coordinate-centric view of each of these three tasks. Here the black line denotes the position of the fist as a function of time, and the colored lines represent the positions of each of the three colored blocks. We can see from the first plot that the agent moves each block to reach the goal, shown as an orange semicircle. For the second and third tasks, only the red and blue blocks are acted on respectively, although we can see in the second plot that the agent unintentionally moves the other two blocks slightly.[]{data-label="fig:strip"}](figures/coords_1 "fig:"){width="32.00000%"} ![IU agent executing three different policies. The top film-strip shows the agent solving the hardest task: gathering all three blocks into the upper right corner. The two strips below this show the agent moving a single block into the corner; red and blue respectively. Below these temporal illustrations is a figure illustrating a coordinate-centric view of each of these three tasks. Here the black line denotes the position of the fist as a function of time, and the colored lines represent the positions of each of the three colored blocks. We can see from the first plot that the agent moves each block to reach the goal, shown as an orange semicircle. For the second and third tasks, only the red and blue blocks are acted on respectively, although we can see in the second plot that the agent unintentionally moves the other two blocks slightly.[]{data-label="fig:strip"}](figures/coords_2 "fig:"){width="32.00000%"}
Discussion {#sec:discussion}
==========
Our experiments demonstrate that when acting according to the policy associated with one of the hardest tasks, we are able to learn all other tasks off-policy. The results for the playroom domain also showed that by increasing the number of tasks, all actors and critics learn faster. In fact, in some settings, learning with many goals was essential to solve hard many-body control tasks.
It is worth entertaining what happens when instead of following the policy associated with one of the hardest tasks, we follow a different behaviour strategy. For instance, what happens if we act according to a uniformly random mixture of actors? Is it advantageous to switch policies each time a policy succeeds in encountering a reward? Is it sound to choose behavior actors adaptively and in accordance with the critics’ values?
The above questions are related to the problem of learning curricula. We conducted experiments to explore each of these questions, but found that the naive strategy of choosing the one of the hardest tasks for the behaviour policy works best. There are reasons for this. The behaviour policy determines what information is written to replay memory. These memories in turn are used to update the remaining actor-critics off-policy by gradient descent. When acting according to policies solving simple tasks, the replay memory ends up consisting mostly of experiences associated with the simple tasks, and consequently, the IU agent fails to explore. Populating the memory with rich experiences is essential for learning to solve tasks that involve more exploration.
Our experimental setup was restricted in the sense that the tasks had a nested structure. That is, approaching a block is needed to bring two blocks together, which in turn is needed to bring three blocks together, and so on. Our conclusions regarding curricula might be different in the absence of this nesting of tasks. This should be studied in future work.
It is not difficult to make the hardest task sufficiently difficult, from an exploration perspective, that even the IU agent fails. For example, if we increase the size of the playground and ask the actuator to bring ten blocks together, all existing control agents are likely to fail. The solution to this hard exploration problem appears to be one of task decomposition, either through hierarchical reinforcement learning or an understanding of objects and relations.
In this paper we focused on tasks that are sufficiently hard that not even popular continuous control algorithms such as DDPG can solve them. For this reason we obviated other experimental factors, such as perception from pixels, articulated bodies with more degrees of freedom, and diverse sets of objects. Each of these constitute important challenges that should be addressed in future research.
This work also did not touch upon the topic of policy re-use. That is, once the various policies are learned how do we construct other controllers that can harness these policies to solve new tasks. A simple way to achieve this is to train agents that combine all the policies, either via weighted combinations or more sophisticated deep networks, to construct new policies. Related to this, our work did not address the problem of pruning irrelevant policies.
In regard to hierarchical RL agents, it would be worthwhile investigating architectures where intrinsic motivation provides the ultimate reward, and is used to guide the automatic online addition or pruning of actors in the IU agent architecture.
Conclusions {#sec:conclusions}
===========
We present in this work a novel architecture for learning several tasks at once. We also propose a flexible way of generating tasks in simulated physical environments. While these environments pose significant challenges because of complex contact forces among the objects, continuous action spaces, and difficult exploration, the body, objects and physical laws are shared among the tasks. In these domains, the more tasks, the faster the IU agents learn. In fact, they can learn to solve complex tasks where popular DDPG agents completely fail.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Crowd behaviour analytics focuses on behavioural characteristics of groups of people instead of individuals’ activities. This work considers *human queuing behaviour* which is a specific crowd behavior of groups. We design a plug-and-play system solution to the *queue detection problem* based on Wi-Fi/Bluetooth Low Energy (BLE) received signal strength indicators (RSSIs) captured by multiple signal sniffers. The goal of this work is to determine if a device is in the queue based on *only* RSSIs. The key idea is to extract features not only from individual device’s data but also mobility similarity between data from multiple devices and mobility correlation observed by multiple sniffers. Thus, we propose *single-device feature extraction*, *cross-device feature extraction*, and *cross-sniffer feature extraction* for model training and classification. We systematically conduct experiments with simulated queue movements to study the detection accuracy. Finally, we compare our signal-based approach against camera-based face detection approach in a real-world social event with a real human queue. The experimental results indicate that our approach can reach minimum accuracy of $77\%$ and it significantly outperforms the camera-based face detection because people block each other’s visibility whereas wireless signals can be detected without blocking.'
author:
- |
Fang-Jing Wu and Gürkan Solmaz\
NEC Laboratories Europe, Heidelberg, Germany\
[email protected], [email protected]
title: |
Are You in the Line?\
RSSI-based Queue Detection in Crowds
---
crowd behaviour analytics, cyber-physical systems, human mobility, internet of things, smart cities.
Introduction
============
Recently, *crowd behaviour analytics* has attracted much attention and has boosted many promising applications such as crowd detection and estimation for public safety [@Li2015_Wi-Counter][@2015Li_SenseFlow], social activity analytics [@Santani_2016:SocailNightActivities], and space syntax analytics for exploring new business opportunities [@Mashhadi2016-Wi-FiAnalytics]. These applications exploit Internet-of-Things (IoT) sensing technology with ambient sensors, Wi-Fi and Bluetooth Low Energy (BLE) sniffers, and built-in sensors of smartphones to capture human behaviour. Meanwhile, many research efforts have paid attention to individual human activity analytics such as transportation activity detection [@wu2014_UrbanMobilitySense] and daily activity recognition [@Riboni2016-DailyActivityRecognition]. However, compared to the human activity detection, crowd behaviour analytics focuses more on the behavioural characteristics of targeted groups of people instead of individual activities.
![An overview of queuing behaviour detection system.[]{data-label="Fig:SystemOverview"}](SystemOverview){width="\columnwidth"}
This work focuses on *human queuing behaviour* (“queuing behaviour” for short) which is a unique type of crowd behaviour created by groups of people rather than each individual. In reality, humans queue up for a specific purpose such as buying tickets, taking lifts in ski resorts, and waiting for taxis in which the starting point of a queue is known. Their movements observed by other people have unique patterns such as movements towards the same direction, small-scale and slow movements, and periodic movements. On the other hand, since Wi-Fi/BLE-enabled smart devices (e.g., smartphones and wearable devices) become more popular, the RSSIs in wireless packets from these devices carried by crowds provide insightful clues to capture these unique crowd mobility patterns. Thus, we consider wireless signal sniffers as crowd mobility observers which will capture broadcast packets (e.g, Wi-Fi probe request packets and BLE advertising packets) from mobile devices. This paper considers the BLE technology to implement a proof-of-concept prototype due to the privacy concern, where opt-in data collection is adopted [^1]. The system consists of multiple signal sniffers to capture queuing behaviour in crowds and extract feature patterns of queuing behaviour based on *only* RSSIs. [Fig. \[Fig:SystemOverview\]]{} shows an overview of our system which contains data sources, gateways, data analytics, and applications. These layers and the system is explained in detail in Section \[Sec:ProblemStatementAndSystemOverview\]. In this work, we deploy three sniffers to observe RSSI changes of crowds’ devices [^2]. One sniffer is deployed at the starting point of a potential queue such as a service counter, and another two are deployed at the left-hand side and the right-and side along the queue. The goal of our system is to determine if a person is in the queue based on the changes of RSSIs.
However, our system uses *only* RSSI information to extract queueing behaviour patterns which raises the following technical challenges.
- *Noisy and fluctuating RSSIs*: The RSSIs may vary dramatically over time, even in the case that people are staying static and closer to the sniffers.
- *Heterogeneity of antenna sensitivity*: The sensitivity of devices’ antenna are different from each other. Specifically, the RSSIs of same device observed by different sniffers may be very different from each other. Similarly, when multiple devices carried by the same person, the RSSIs of these devices observed by the same sniffer have different variations.
- *Cross-entity feature extraction*: Since queuing behaviour is created by groups of people, extracting feature patterns between multiple devices based on multiple observers (e.g., sniffers) is a new challenge.
To address the above technical issues, the key idea of our work is to extract feature patterns of crowd behaviour not only from each individual device’s data but also from cross-device data and cross-sniffer data. Specifically, we consider three types of feature extraction: (1) *single-device feature extraction*, (2) *cross-device feature extraction*, and (3) *cross-sniffer feature extraction*. The first type of features considers each individual device’s RSSI variation when the device’s owner makes movements along the queue. The second type of features considers the mobility similarity between multiple devices when these devices’ owners make movement together along the queue. The third type of features considers the mobility correlation observed by multiple sniffers when a particular device’s owner makes movements along the queue. Based on the three types of features extracted from individuals and heterogeneous entities, we use well-known classifiers implemented by Weka [@Weka] to verify the merit of these features. Note that our work mainly focuses on designing a holistic and plug-and-play solution to the queue detection problem using Internet-of-Things (IoT) sensing technology instead of designing classification algorithms. The proposed proof-of-concept prototype system can be extended to the Wi-Fi-based sniffing technology, where Wi-Fi probe request packets are captured for queue detection. We systematically conduct experiments with simulated queue movements to investigate how different parameters and setups affect detection accuracy. Finally, we compare our signal-based approach against camera-based face detection approach in a real-world team-building social event, where people queued up for taking cakes, biscuits, and drinks. The experimental results indicate that our approach can reach minimum accuracy of $77\%$ and it significantly outperforms the camera-based face detection approach since in the RSSI-based detection wireless signals are not blocked by the crowd themselves as in the case of camera-based detection.
Related Work
============
Recently, some research studies have paid attention to queue detection [@Barbagli2011_TrafficQueue] [@Lu2015_Taxi-PassengerQueue][@Zanin2003_image-BasedVehicleQueue][@Satzoda2012_Vision-basedVehicleQueue][@Li2014_QueueSense][@Li2016_QueueSenseTMC][@Wang2014_TrackingHumanQueues]. In [@Barbagli2011_TrafficQueue], an acoustic sensor network is deployed along road segments to monitor traffic queue. The work in [@Lu2015_Taxi-PassengerQueue] designs a system to detect queuing activities of taxis and passengers based on their GPS information. Research efforts in [@Zanin2003_image-BasedVehicleQueue][@Satzoda2012_Vision-basedVehicleQueue] consider video-based approaches to detect vehicle queues. In [@Li2016_QueueSenseTMC], a mobile application is designed to capture human behaviour and detect queueing behaviour, where built-in sensors (e.g., accelerometers, compasses, and Bluetooth) are exploited for capturing human mobility information. The work in [@Wang2014_TrackingHumanQueues] proposes an RSSI-based approach to detect three statuses of queuing behaviour including waiting period, service period, and leaving period based on a single Wi-Fi sniffer’s observations.
However, sensor-based approaches require well-deployed infrastructure, while GPS-based approaches work only in outdoor environments. Mobile sensing approaches require mobile applications to be pre-installed in people’s smartphones. Video-based approaches may compromise personal privacy and they are easily affected by light conditions and crowd blocking of visible area. The RSSI-based approach in [@Wang2014_TrackingHumanQueues] focuses on RSSI features from each individual device, while we consider not only single-device features but also cross-device features and cross-sniffer features since unique behavioural characteristics of human queues are created by groups of people as opposed to individuals.
Queuing Behaviour Detection {#Sec:ProblemStatementAndSystemOverview}
===========================
Queue Detection Problem
-----------------------
Assume that the starting point of a queue (e.g., service counter) is known. We consider three signal sniffers, denoted $\pi_1$, $\pi_2$, and $\pi_3$, where the BLE sniffing technology is used in this work due to privacy concern. One sniffer is deployed in the starting point and another two sniffers are symmetrically deployed along a straight line from the stating point towards the direction of a human queue. Suppose that each visitor carries a BLE transmitter in a designated environment (e.g., exhibitions). For a given time window $\Delta_k=[t_i, t_j)$, there are three time series data streams captured by the deployed three sniffers, denoted by $\Omega_1(\Delta_k)$, $\Omega_2(\Delta_k)$, and $\Omega_3(\Delta_k)$, where each $\Omega_i(\Delta_k)$ is a sequence of BLE advertising packets. Each BLE advertising packet contains the RSSI and the device ID.
Given three time series data streams, $\Omega_1(\Delta_k)$, $\Omega_2(\Delta_k)$, and $\Omega_3(\Delta_k)$ captured by $\pi_1$, $\pi_2$, and $\pi_3$ during a time window $\Delta_k$, for each device captured by the three sniffers, *queue detection problem* is to determine if the device is in the queue which is starting from $\pi_1$.
System Design {#SystemDesign}
-------------
[Fig. \[Fig:SystemOverview\]]{} shows our system design for queuing behaviour detection which is composed of four layers: *data sources*, *gateways*, *data analytics*, and *applications*. BLE sniffers serve as data sources which capture all types of BLE packets from users’ devices and report to the gateway layer. The crowd mobility sensing gateway registers the data types of interests and keeps updating on new data arrivals, where BLE advertising packets are in our interests. Meanwhile, the crowd mobility sensing gateway updates the crowd mobility database when there is new update on received BLE advertising packets. In addition to the crowd mobility database, there are three data analytical components: (1) *data preprocessing*, (2) *feature extraction*, and (3) *queue behavior model training and classification*. The data preprocessing performs data aggregation and low-pass filtering to eliminate noise form the collected RSSIs. The feature extraction identifies the mobility patterns from three aspects: single-device features, cross-device features, and cross-sniffer features. Single-device features are extracted from each individual device’s RSSIs. Cross-device features are extracted from multiple user devices’ RSSIs based on their mobility similarity. Cross-sniffer features are extracted from observations by multiple sniffers based on the correlation of their observations. Finally, the queue behavior model training and classification conducts off-line training for detecting the status of a device, where the trained classifier makes a binary decision on the device status (e.g., “in-queue” or “not-in-queue” statuses). Finally, the application layer is a knowledge consumer which requires results of queuing detection for improving user experience or services provided by operators.
Queuing Behaviour Data Analytics {#Sec:DataAnalytics}
================================
In this section, we first describe the methodology for collecting labeled data and explain the the proposed algorithm for feature extraction. Then, we describe the means for queuing behaviour classification.
Labeled Data Collection
-----------------------
[Fig. \[Fig:DataCollectionAndPreprocessing\]]{} (a) is our deployment for collecting labeled data, where human mobility behaviour can be categorized into three types. The first type of mobility behaviour is collected from in-queue devices which have movements periodically. The second type of mobility behaviour is collected from not-in-queue devices which take random walks. The third mobility behaviour is collected from not-in-queue devices which stay static at certain locations. We launch such data collection campaigns in an indoor office environment for collecting labeled data.
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![(a) Deployment for collecting labelled data. (b) Data preprocessing results.[]{data-label="Fig:DataCollectionAndPreprocessing"}](LabelledDataCollection "fig:"){width="0.53\columnwidth"} ![(a) Deployment for collecting labelled data. (b) Data preprocessing results.[]{data-label="Fig:DataCollectionAndPreprocessing"}](fig_features/Raw_RSSIs_exp1_node14 "fig:"){width="0.42\columnwidth"}
(a) (b)
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Data Preprocessing
------------------
Since RSSIs are noisy, we perform data aggregation and low pass filter to clean up noise in the collected raw data. First, in the data aggregation phase, we aggregate RSSI data streams from each device every $\lambda$ seconds. Then, we apply the dynamic exponential smoothing filter (DESF) to the aggregated RSSI data streams. We implement the DESF as follows. The $i$-th output sample is $O_i= \alpha \cdot O_{i-1}+(1-\alpha) \cdot I_{i}$ if $I_{i} < O_{i-1}$. Otherwise, $O_i= (1-\alpha) \cdot O_{i-1}+\alpha \cdot I_{i}$. Here, $I_{i}$ is the $i$-th input sample and $\alpha$ is a predefined parameter. [Fig. \[Fig:DataCollectionAndPreprocessing\]]{} (b) shows the data before and after data preprocessing, where the green one is raw data and the red one is the processed data after data aggregation and low-pass filtering.
Feature Extraction
------------------
We extract three types of features, (1) *single-device features*, (2) *cross-device features*, and (3) *cross-sniffer features*, from preprocessed data. The first type is mobility characteristic extracted from each individual device’s RSSIs. The second type is mobility similarity between crowds’ devices. The third type is the mobility correlation observed by multiple observers (e.g., sniffers). Our work extracts nine features in total from the three aspects. Below, we first make observations from the collected ground truth data to identify significant features of in-queue devices and then define the mechanisms to extract the features for the off-line training in the next step. The details of the proposed feature extraction mechanisms are explained in the following subsections.
------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------ -----------------------------------------------------------------
{width="0.29\linewidth"} {width="0.29\linewidth"} {width="0.29\linewidth"}
(a) (b) (c)
------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------ -----------------------------------------------------------------
C.1 Single-device feature extraction {#c.1-single-device-feature-extraction .unnumbered}
------------------------------------
Let $d_x(\Delta_k)$ denote the captured BLE packets during time window $\Delta_k$ for device $d_x$. For a given time window $\Delta_k=[t_i, t_j)$, for an observed device $d_x$, we define seven feature extraction functions to compute the seven single-device features based on only RSSIs as follows.
*Feature 1: Positive accumulated slopes*: [Fig. \[Fig:single-device-features\]]{} (a) shows the accumulated RSSI slopes of an in-queue device. As it can be seen, the accumulated RSSI slopes are all positive as the device moves closer to the starting point. Thus, we define the accumulated slope for a given time window $[t_i, t_j)$ as $f_1(d_x, \Omega_1(\Delta_k))=r_j-r_i$, where $r_i$ and $r_j$ are the RSSIs at the timestamps $t_i$ and $t_j$, respectively.
*Feature 2: Approaching-counter patterns*: In [Fig. \[Fig:single-device-features\]]{} (b), when the in-queue device is already very close to the starting point, the RSSIs dramatically increase even though the in-queue device makes only a little movement towards the starting point. Based on the experimental observation and the signal propagation theory, the function of distance to RSSI changes is not linear. We then define the following binary function to extract this nature feature pattern from RSSIs: $$f_2(d_x, \Omega_1(\Delta_k)) =
\begin{cases}
1 &\text{if $r_j-r_i > \tau_{f_2}$};\\
0 &\text{otherwise.}
\end{cases}$$ Here, $\tau_{f_2}$ is a pre-defined threshold. When there are dramatic changes on RSSIs during $\Delta_k$, the output value of the binary function is 1 since the device is approaching closer to the starting point. Otherwise, the output of the binary function is 0.
*Feature 3: Near-counter RSSIs*: [Fig. \[Fig:single-device-features\]]{} (c) shows that the RSSIs changes as these in-queue devices move closer to the starting point. As it can be seen, when devices stay near the staring point, the RSSIs are higher than devices far away from the starting point. We thus define the following binary function to extract this feature based on RSSIs.
$$f_3(d_x, \Omega_1(\Delta_k)) =
\begin{cases}
1 &\text{if $r_t > \tau_{f_3}, \forall t\in \Delta_k$};\\
0 &\text{otherwise.}
\end{cases}$$
Here, $\tau_{f_3}$ is a pre-defined threshold. When all of captured RSSIs during the time window $\Delta_k$, the output value of the binary function is 1 since the device is considered very close to the starting point. Otherwise, the output value of of the binary function is 0.
*Feature 4: RSSI stability observed by $\pi_1$*: The RSSI variances of in-queue devices are smaller compared to the RSSI variances of random walks since in-queue devices make movements slowly along the queue towards $\pi_1$. To track the RSSI changes, we consider $b$ backtracking time windows to extract this feature patterns. For each given time window $\Delta_k$, we backtrack $b$ time windows together with the current observations during $\Delta_k$ to compute RSSI variance as $$f_4(d_x, \Omega_1(\Delta_k))=Var(\Omega_1(\Delta_k), \Omega_1(\Delta_{k-1}), \ldots, \Omega_1(\Delta_{k-b})).$$ Here, $Var(\Omega_1(\Delta_k), \Omega_1(\Delta_{k-1}), \ldots, \Omega_1(\Delta_{k-b}))$ is the variance function of given sets of captured BLE packets’ RSSIs during these historical time windows.
*Feature 5: RSSI stability observed by $\pi_2$*: Similarly, we can extract RSSI stability observed by the sniffer $\pi_2$ as: $$f_5(d_x, \Omega_2(\Delta_k))=Var(\Omega_2(\Delta_k), \Omega_2(\Delta_{k-1}), \ldots, \Omega_2(\Delta_{k-b})).$$ Here, the size of the backtracking time windows is the same as in the feature 4.
*Feature 6: RSSI stability observed by $\pi_3$*: Similarly, we can extract RSSI stability observed by the sniffer $\pi_3$ as: $$f_6(d_x, \Omega_3(\Delta_k))=Var(\Omega_3(\Delta_k), \Omega_3(\Delta_{k-1}), \ldots, \Omega_3(\Delta_{k-b})).$$
*Feature 7: Longer stay duration observed by $\pi_1$*: In-queue devices have longer stay durations compared to not-in-queue devices which have wandering random walks. We thus define the following function to extract this feature from packets captured by the sniffer $\pi_1$. $$f_7(d_x, \Omega_1(\Delta_k))=t_l-t_f,$$ where $t_f$ is is the timestamp when the latest advertising packet received by the sniffer $\pi_1$ and $t_f$ is the timestamp when the first advertising packet is received by the sniffer $\pi_1$.
![Cross-device feature extraction.[]{data-label="Fig:MobilitySmilarity"}](MobilitySmilarity){width="0.8\columnwidth"}
C.2 Cross-device feature extraction {#c.2-cross-device-feature-extraction .unnumbered}
-----------------------------------
Since queuing behaviour is created by crowds and not by an individual, similar mobility behaviour of crowds creates similar RSSIs patterns to each other. Next, we first identify such kind of cross-device features based on experimental observations and define feature extraction function to speed up computation.
*Feature 8: Mobility similarity*: [Fig. \[Fig:MobilitySmilarity\]]{} shows the visualization of correlation matrix between all pair of in-queue devices’ RSSIs captured by the sniffer $\pi_1$. As it can be seen, there are high correlations between RSSIs of in-queue devices when they are queuing up for a certain purpose. Although the sensitivity of devices’ antennas are different from each other as shown in [Fig. \[Fig:single-device-features\]]{} (c), their RSSI patterns are similar to each other. Devices far away from the sniffer $\pi_1$ have higher correlation coefficients with others because of sufficient RSSI data samples. However, computing movement similarities between all pairs of devices incurs high computational complexity in the real world since there are many combinations of devices. Therefore, we design the following low-complexity feature extraction function to speed up the computation.
1. Sort devices based on the RSSI stability observed by $\pi_1$ (i.e., $f_4(d_x, \Omega_1(\Delta_k))$). The idea is to cross-check those devices with more stable RSSIs first because they are probably in the queue.
2. Let $\omega_x^{\pi_1}(\Delta_k)$ and $\omega_y^{\pi_1}(\Delta_k)$ denote the BLE packets captured by $\pi_1$ during time window $\Delta_k$ for devices $d_x$ and $d_y$, respectively. The movement similarity function between $\omega_x^{\pi_1}(\Delta_k)$ and $\omega_y^{\pi_1}(\Delta_k)$ is defined as $$\begin{aligned}
S(\omega_x^{\pi_1}(&\Delta_k), \omega_y^{\pi_1}(\Delta_k))\\
=Cor( &(\omega_x^{\pi_1}(\Delta_k), \omega_x^{\pi_1}(\Delta_{k-1}), \ldots, \omega_x^{\pi_1}(\Delta_{k-b})) \\
&(\omega_y^{\pi_1}(\Delta_k), \omega_y^{\pi_1}(\Delta_{k-1}), \ldots, \omega_y^{\pi_1}(\Delta_{k-b}))).
\end{aligned}$$ Here, $Cor(\cdot,\cdot)$ computes the correlation coefficient between two given sequences of RSSIs in BLE packets $(\omega_x^{\pi_1}(\Delta_k), \omega_x^{\pi_1}(\Delta_{k-1}), \ldots, \omega_x^{\pi_1}(\Delta_{k-b}))$ and BLE packets $(\omega_y^{\pi_1}(\Delta_k), \omega_y^{\pi_1}(\Delta_{k-1}), \ldots, \omega_y^{\pi_1}(\Delta_{k-b}))$. We backtrack $b$ time windows to compute the correlation coefficient between $d_x$ and $d_y$.
3. For a given $\omega_x^{\pi_1}(\Delta_k)$ and the sorted list in 1), we can compute movement similarity between $d_x$ and devices in the sorted list one by one until there are at least $m$ devices which have correlation coefficients greater than $\tau_{f_8}$. In this case, the $f_8(d_x, \Omega_1(\Delta_k)) = 1$. Otherwise, $f_8(d_x, \Omega_1(\Delta_k)) = 0$. Here, $m$ and $\tau_s$ are predefined thresholds.
Note that the above computation for this feature extraction can be terminated early if we can find $m$ devices which meet the above condition. In this case, we can speed up the computation instead of computing all pairs of combinations.
![The workflow of the queue behavior model training and classification.[]{data-label="Fig:Classification"}](Classification){width="0.9\columnwidth"}
![Implementation of sniffers and controllable experimental setup with a simulated queue.[]{data-label="Fig:Implementation"}](Implementation){width="\columnwidth"}
C.3 Cross-sniffer feature extraction {#c.3-cross-sniffer-feature-extraction .unnumbered}
------------------------------------
When an in-queue device makes a sequence of movements along the queue, there is a high correlation between the observations captured by the sniffers on both sides. Note that this feature exploits consensus between sniffers, while cross-device features focus on the mobility similarities between devices. Thus, we consider the cross-sniffer consensus to extract this feature as follows.
*Feature 9: Mobility correlation*: For a device $d_x$, the mobility correlation observed by the sniffer $\pi_2$ and $\pi_3$ during time window $\Delta_k$ is defined as follows. $$\begin{aligned}
f_8(d_x, \Omega_2(&\Delta_k), \Omega_3(\Delta_k))\\
=Cor(& (\omega_x^{\pi_2}(\Delta_k), \omega_x^{\pi_2}(\Delta_{k-1}), \ldots, \omega_x^{\pi_2}(\Delta_{k-b})), \\
& (\omega_x^{\pi_3}(\Delta_k), \omega_x^{\pi_3}(\Delta_{k-1}), \ldots, \omega_x^{\pi_3}(\Delta_{k-b}))).\end{aligned}$$ Here, we backtrack $b$ time windows to compute the correlation coefficient for a device $d_x$’s RSSIs captured by $\pi_2$ and $\pi_3$.
![Detection accuracy vs. the number of backtracking time windows.[]{data-label="Fig:ExpResults-backtrackingTW"}](Backtracking_accuracy){width="0.75\columnwidth"}
![Detection accuracy vs. the size of time window.[]{data-label="Fig:ExpResults-TWSize"}](TW_accuracy){width="0.75\columnwidth"}
Queue Behavior Classification
-----------------------------
[Fig. \[Fig:Classification\]]{} shows the working flow of the designed queue behavior model training and classification. After we extract the above features, these features are used for off-line model training. Here, we consider three existing classifiers: J48Tree, RandomForest, and Naive Bayes since all of these three classifiers work well for numerical and categorical data and incur lower computation overhead which can support real-time detection tasks in the future. After the model training phase, for each piece of the testing data, the classifier determines if the device corresponding to the testing data is in a queue.
Evaluation {#Sec:Evaluation}
==========
Implementation and Experimental Setup
-------------------------------------
We implement a packet sniffing program in Python for Raspberry Pi platforms. We use BLE sensing technology as an example for the prototype of our system. Practically, Wi-Fi sensing technology can be an alternative mechanism to capture human mobility through monitoring probe requests from smartphones. [Fig. \[Fig:Implementation\]]{} (a)-(b) is the software and hardware components of our sniffers. The [*packet monitor*]{} captures all of types of BLE packets. The [*packet filter and decoder*]{} parses only BLE advertising packets and discards other types of BLE packets. The [*data reporter*]{} updates the new packet arrivals to the crowd mobility sensing gateway and then to the crowd mobility database for performing queuing behaviour data analytics described in [Section \[Sec:DataAnalytics\]]{}. The designed software components are running as background processes on Raspberry Pi version 3 which has a built-in BLE module. The default parameters in our system are explained as follows. We consider $\lambda=30$ seconds for data aggregation and $\alpha=0.9$ for the low-pass filter in the data preprocessing component. Our system uses a fixed size of time window of 60 seconds. The number of backtracking time windows for extracting RSSI stability (in feature 4, feature 5, and feature 6) is $b=8$ time windows which leads to backtracking for 480 seconds. The pre-defined threshold for extracting the approaching-counter patterns (i.e., feature 2) is $\tau_{f_2}=5$ dbm. The threshold for extracting near-counter RSSIs (i.e., feature 3) is $\tau_{f_3}=-55$ dbm. Finally, we consider $m=3$ and $\tau_{f_8}=0.3$ for extracting nobility similarity (i.e., feature 8).
We conduct two types of experiments: (a) controllable experiments with simulated queuing behaviour and (b) real-world experiments with real human queuing behaviour in a team-building social event. In the controllable experiments, we compare the performance resulted from different classifiers. In the real-world experiments, we compare our signal-based approach against the camera-based face detection approach using OpenCV [@OpenCV].
First, we conduct controllable experiments with a simulated queue to verify our system, where [Fig. \[Fig:Implementation\]]{} (c) shows the experimental setup. Three sniffers are deployed in a big conference room. We deploy 7 BLE beacons on a straight-lined paper to simulate a queue. We pull the straight-lined paper from the starting point of the queue every 120 seconds to simulate human movements along the queue. We use three classifiers, J48Tree, RandomForest, and Naive Bayes, based on Weka [@Weka] implementation to evaluate the performance of the queuing behaviour classification. We change the number of backtracking time windows, the size of time window, and the number of sniffers to study how these parameters affect the performance of our system.
Then, moving towards a more realistic environment, the same setup is used in a real-world team building party, where 11 people gather together in an indoor conference room. In the beginning of the team building party, people queue up for taking cakes, biscuits, and drinks, where people casually talk with each other while waiting in the queue. We also deploy a camera in front of the cakes and record videos during the social event.
![Accuracy using different numbers of sniffers.[]{data-label="Fig:changeSnifferNumber"}](change_sniffer_number_accuracy_aglo){width="0.75\columnwidth"}
---------------------------------------------------------- ----------------------------------------------------------------
{width="0.4\linewidth"} {width="0.4\linewidth"}
(a) (b)
---------------------------------------------------------- ----------------------------------------------------------------
Experimental Results in a Simulated Queue
-----------------------------------------
First, we vary the number of backtracking time windows in the feature extraction component from 2 to 12. As shown in [Fig. \[Fig:ExpResults-backtrackingTW\]]{}, RandomForest provides the best accuracy compared to J48Tree and Naive Bayes. We can see that the accuracy resulted from all classifiers increases gradually as backtracking time window increases slightly and becomes stable. This is because a longer backtracking time window involves more mobility information to compute RSSI variances, mobility similarity, and mobility correlation. Then, we vary the size of time window from 60 to 360 seconds to extract features. [Fig. \[Fig:ExpResults-TWSize\]]{} shows the accuracy of queuing behaviour detection. We can see that the accuracy decreases as the size of time window increases from 60 to 240 seconds and then increases gradually when the size of time window varies from 240 to 360 seconds. This is because a longer time window considers sufficient RSSIs to compute RSSI variances, mobility similarity, and mobility correlation. However, a too long time window may incur larger RSSI variances because crowds make movements. In this case, RSSI variances cannot provide sufficient insightful information to differentiate in-queue and not-in-queue statuses. Next, we compare the detection accuracy using only 1 sniffer and 3 sniffers. [Fig. \[Fig:changeSnifferNumber\]]{} shows the experimental results. As it can be seen, multi-sniffer approaches improve the detection accuracy up to $7\%$. However, the accuracy provided by the Naive Bayes approach cannot be improved since some binary features are considered for model training and Naive Bayes approaches generally work well with numerical data which is used to estimate a distribution over continuous values.
Experimental Results with a Real Human Queue
--------------------------------------------
Finally, we compare our signal-based approach against the camera-based approach in a team-building social event in the real world. During the social event, we record the ground truth by human observations. The recorded video contains 21989 image frames. We use OpenCV face detection libraries to count people in the queue for each frame. [Fig. \[Fig:ExpResults-Camera\]]{} (a) shows the experimental results. We can see that the camera-based face detection approach provides lower accuracy since most of people in the queue are blocked by the people who are closer to the starting point. As shown in [Fig. \[Fig:ExpResults-Camera\]]{} (b), the number of detected faces does not change too much when the queue length becomes shorter. Generally, camera-based approaches are used for crowd mobility monitoring. However, it has limitations to detect queuing behaviour because of blocking visibility issues. We can conclude that signal-based approaches are more suitable for queuing behaviour detection especially for long queues without suffering from blocking issues.
Conclusion {#Sec:conclusion}
==========
This paper exploits RSSIs captured by multiple signal sniffers to classify if people are in a queue. We propose three types of feature patterns extracted from each individual’s device, cross-device mobility similarity, and cross-sniffer mobility correlation for classification model training. Our approach can be applied to both Wi-Fi and BLE sensing systems. The experimental results indicate that our approach can reach minimum accuracy of $77\%$ and it significantly outperforms the camera-based face detection approach since wireless signals are not blocked by crowds.
Acknowledgment
==============
This work has been partially funded by the European Union’s Horizon 2020 research and innovation programme within the project “Worldwide Interoperability for SEmantics IoT" under grant agreement Number 723156.
[10]{} \[1\][\#1]{} url@samestyle \[2\][\#2]{} \[2\][[l@\#1=l@\#1\#2]{}]{}
H. Li, E. C. Chan, X. Guo, J. Xiao, K. Wu, and L. M. Ni, “Wi-counter: Smartphone-based people counter using crowdsourced wi-fi signal data,” *Human-Machine Systems, IEEE Transactions on*, vol. 45, no. 4, pp. 442–452, 2015.
K. Li, C. Yuen, and S. Kanhere, “Senseflow: An experimental study of people tracking,” in *Proceedings of the 6th ACM Workshop on Real World Wireless Sensor Networks*, ser. RealWSN ’15.1em plus 0.5em minus 0.4emNew York, NY, USA: ACM, 2015, pp. 31–34. \[Online\]. Available: <http://doi.acm.org/10.1145/2820990.2820994>
D. Santani, J.-I. Biel, F. Labhart, J. Truong, S. Landolt, E. Kuntsche, and D. Gatica-Perez, “The night is young: Urban crowdsourcing of nightlife patterns,” in *Proceedings of the 2016 ACM International Joint Conference on Pervasive and Ubiquitous Computing*, ser. UbiComp ’16. 1em plus 0.5em minus 0.4emNew York, NY, USA: ACM, 2016, pp. 427–438. \[Online\]. Available: <http://doi.acm.org/10.1145/2971648.2971713>
A. Mashhadi, U. G. Acer, A. Boran, P. M. Scholl, C. Forlivesi, G. Vanderhulst, and F. Kawsar, “Exploring space syntax on entrepreneurial opportunities with wi-fi analytics,” in *Proceedings of the 2016 ACM International Joint Conference on Pervasive and Ubiquitous Computing*, ser. UbiComp ’16. 1em plus 0.5em minus 0.4emNew York, NY, USA: ACM, 2016, pp. 658–669. \[Online\]. Available: <http://doi.acm.org/10.1145/2971648.2971745>
F.-J. Wu and H. B. Lim, “[UrbanMobilitySense]{}: A user-centric participatory sensing system for transportation activity surveys,” *IEEE Sensors Journal*, vol. 14, no. 12, pp. 4165–4174, 2014.
D. Riboni, T. Sztyler, G. Civitarese, and H. Stuckenschmidt, “Unsupervised recognition of interleaved activities of daily living through ontological and probabilistic reasoning,” in *Proceedings of the 2016 ACM International Joint Conference on Pervasive and Ubiquitous Computing*, ser. UbiComp ’16.1em plus 0.5em minus 0.4emNew York, NY, USA: ACM, 2016, pp. 1–12. \[Online\]. Available: <http://doi.acm.org/10.1145/2971648.2971691>
“Weka,” http://www.cs.waikato.ac.nz/ml/weka/.
B. Barbagli, L. Bencini, I. Magrini, G. Manes, and A. Manes, “A traffic monitoring and queue detection system based on an acoustic sensor network,” *International Journal on Advances in Networks and Services*, vol. 4, no. 1-2, pp. 27–37, 2011.
Y. Lu, S. Xiang, and W. Wu, “Taxi queue, passenger queue or no queue?: A queue detection and analysis system using taxi state transition,” in *International Conference on Extending Database Technology (EDBT)*, 2015, pp. 23–27.
M. Zanin, S. Messelodi, and C. Modena, “An efficient vehicle queue detection system based on image processing,” in *International Conference on Image Analysis and Processing*, 2003, pp. 232–237.
R. K. Satzoda, S. Suchitra, T. Srikanthan, and J. Y. Chia, “Vision-based vehicle queue detection at traffic junctions,” in *[IEEE Conference on Industrial Electronics and Applications (ICIEA)]{}*, 2012, pp. 90–95.
Q. Li, Q. Han, X. Cheng, and L. Sun, “[QueueSense]{}: Collaborative recognition of queuing on mobile phones,” in *International Conference on Sensing, Communication, and Networking (SECON).*, 2014, pp. 230 – 238.
Q. Li, Q. Han, and L. Sun, “Collaborative recognition of queuing behavior on mobile phones,” *IEEE Trans. Mobile Computing*, vol. 15, no. 1, pp. 60–73, 2016.
Y. Wang, J. Yang, Y. Chen, H. Liu, M. Gruteser, and R. P. Martin, “Tracking human queues using single-point signal monitoring,” in *ACM Int’l Conf. on Mobile Systems, Applications, and Services*, 2014, pp. 42–54.
, “Opencv,” http://opencv.org/.
[^1]: For experimental purpose, we collected data only from a specific set of BLE beacons carried by participants.
[^2]: In general, more sniffers can provide richer information for detecting queuing behaviour, and we can deploy these sniffers strategically.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We discuss the structure of decoherence-free subsystems for a bosonic channel affected by collective depolarization. A single use of the channel is defined as a transmission of a pair of bosonic modes. Collective depolarization consists in a random linear U(2) transformation of the respective mode operators, which is assumed to be identical for $N$ consecutive uses of the channel. We derive a recursion formula that characterizes the dimensionality of available decoherence-free subsystems in such a setting.'
author:
- |
Jonathan L. Ball and Konrad Banaszek\
[*Clarendon Laboratory, University of Oxford\
Oxford OX1 3PU, United Kingdom\
e-mail: [email protected], [email protected]*]{}
title: |
Decoherence-Free Subspaces and Subsystems\
for a Collectively Depolarizing Bosonic Channel
---
Introduction
============
Decoherence induced by uncontrolled interactions with the environment is a major obstacle when implementing protocols for quantum information processing in real physical systems. The effects of such interactions can however be reduced by using particular quantum states that are robust against specific decoherence mechanisms. An interesting and physically relevant scenario of decoherence is when the physical system exhibits certain symmetries in interactions with the environment. Such symmetries imply the existence of whole subspaces that remain completely unaffected by decoherence and can therefore be used for faultless quantum information processing. This observation, which can be made for a number of interaction schemes from various perspectives [@PalmSuomPRS96; @DuanGuoPRL97; @ZanaRasePRL97; @LidaChuaPRL98; @KnilLaflPRL00], has led to the development of a general theory of decoherence-free subspaces and subsystems, reviewed recently in [@DFSReview].
The theory of decoherence-free subspaces and subsystems applies to a variety of physical systems. One such system is light travelling through an optical fiber, which usually exhibits random birefringence caused by fluctuating environmental conditions, such as temperature and mechanical strain. Mathematically, this system can be modelled as a bosonic communication channel. A single use of this channel is defined as a transmission of two bosonic modes corresponding to orthogonal polarizations. Birefringence is described as a random U(2) transformation between the operators of these two orthogonally polarized modes. In a realistic situation this transformation remains practically constant over many consecutive uses of the channel, for time intervals comparable at least with the round trip time in a fiber [@GisiRiboRMP02]. The presence of such a symmetry leads to the existence of non-trivial decoherence-free subsystems. Their structure has been discussed in our recent paper [@BallBanaXXX04]. In this contribution, we will concentrate on the mathematical details of the derivation briefly sketched in the previous paper, and will illustrate the presentation with diagrams that provide additional insights into the calculations.
The decoherence-free subsystems for a collectively depolarizing bosonic channel have recently attracted considerable attention in the context of quantum reference frames and various scenarios of quantum and classical communication over depolarizing channels [@BartRudoPRL03]. This has produced theoretical proposals for robust quantum key distribution [@WaltAbourPRL03; @BoilGottPRL04; @BoilLaflPRL04], alignment-free tests of Bell’s inequalities [@CabePRL03], and entanglement-enhanced classical communication [@BallDragPRA04], with some of these ideas implemented in proof-of-principle quantum optical experiments [@KwiaBergSCI00; @BourEiblPRL04; @BanaDragPRL04]. From the mathematical point of view, an interesting aspect of the collectively depolarizing bosonic channel is that in contrast to particle-based scenarios in which every elementary system (i.e. a particle) has a finite number of states, the elementary system in our case (i.e. a pair of bosonic modes) is infinitely dimensional and its transformation under the interaction with the environment is described by a reducible representation of the group U(2).
This paper is organized as follows. We start with a discussion of the polarization transformation for a single use of the bosonic channel in Sec. 2. We then analyze in Sec. 3 the emergence of decoherence-free subsystems for multiple uses of the channel. In Sec. 4 we derive a recursion formula for dimensions of decoherence-free subsystems. Finally, Sec. 5 concludes the paper.
Polarization transformation {#Sec:PolarizationTransformation}
===========================
We assume that the quantum system transmitted in a single use of the channel is composed of a pair of bosonic modes corresponding to orthogonal polarizations, with the respective creation operators labelled by $\hat{a}_{H}^\dagger$ and $\hat{a}_{V}^\dagger$. Random depolarization consists in a linear unitary transformation of the modes given by: $$\label{Eq:PolarizationTransformation}
\left( \begin{array}{c} \hat{a}_{H}^\dagger \\
\hat{a}_{V}^\dagger \end{array} \right)
\mapsto
{\bf\Omega}
\left( \begin{array}{c} \hat{a}_{H}^\dagger \\
\hat{a}_{V}^\dagger \end{array} \right)$$ where the matrix $${\bf\Omega} =
\left( \begin{array}{cc} \Omega_{HH} & \Omega_{HV} \\
\Omega_{VH} & \Omega_{VV}
\end{array} \right)$$ is an arbitrary element of the group U(2). It will be convenient to decompose ${\bf\Omega}$ into a product of a phase factor $e^{-i\alpha} \in \mbox{U(1)}$ and a special unitary matrix ${\bf\Omega}' \in \mbox{SU(2)}$: $$\label{Eq:OmegaDecomposition}
{\bf\Omega} = e^{-i\alpha} {\bf\Omega}'.$$ This decomposition is ambiguous, as both the factors can be multiplied by $-1$. However, the final results obtained with this decomposition will be free from this ambiguity, which makes the specific choice of the decomposition irrelevant. It is also worthwhile to note that in contrast to unitary transformations on the Hilbert space of quantum states, for which the overall phase factor is not physical, in the present case the overall phase factor $e^{-i\alpha}$ is physically meaningful, as it describes the phase of the fields which can in principle be measured with an external phase reference.
The two-mode Hilbert space ${\cal H}$ describing our system has a convenient orthonormal basis in the form of Fock states defined in general by $$|m_H n_V\rangle=
\frac{(\hat{a}_{H}^\dagger)^{m}(\hat{a}_{V}^\dagger)^{n}}{\sqrt{m!n!}}
|{\textrm{vac}}\rangle,
\qquad
m,n = 0,1,2,\ldots,$$ where $|{\textrm{vac}}\rangle$ is the vacuum state of the system. The polarization transformation defined in Eq. (\[Eq:PolarizationTransformation\]) does not change the total number of field excitations. It is therefore convenient to decompose the Hilbert space ${\cal H}$ into a direct sum $$\label{Eq:H=sumH(l)}
{\cal H} = \bigoplus_{l=0}^{\infty} {\cal H}^{(l)}$$ of finite-dimensional subspaces ${\cal H}^{(l)}$ that contain exactly $l$ excitations in both the modes. The subspace ${\cal H}^{(l)}$ has dimension $l+1$ and it is spanned by Fock states of the form: $$\label{Eq:H(l)=Span}
{\cal H}^{(l)} = \mbox{Span} \{ |m_{H} (l-m)_{V}\rangle \; | \; m=0,1,\ldots, l \}.$$ Under the polarization transformation given in Eq. (\[Eq:PolarizationTransformation\]), the state $|m_{H} (l-m)_{V}\rangle$ is transformed according to: $$\begin{aligned}
\lefteqn{|m_{H} (l-m)_{V}\rangle = \frac{(\hat{a}_{H}^\dagger)^{m}(\hat{a}_{V}^\dagger)^{l-m}}{\sqrt{m!(l-m)!}}
|{\textrm{vac}}\rangle}
& & \nonumber \\
\label{Eq:ml-m}
& \mapsto &
\frac{(\Omega_{HH}\hat{a}_{H}^\dagger+\Omega_{HV}\hat{a}_{V}^\dagger)^{m}
(\Omega_{VH}\hat{a}_{H}^\dagger+\Omega_{VV}\hat{a}_{V}^\dagger)^{l-m}}{\sqrt{m!(l-m)!}}
|{\textrm{vac}}\rangle \end{aligned}$$ If we now insert the decomposition of ${\bf\Omega}$ given in Eq. (\[Eq:OmegaDecomposition\]), this will produce an overall factor $e^{-il\alpha}$ times the same expression as in the second line of Eq. (\[Eq:ml-m\]), but with the elements of ${\bf\Omega}$ replaced by those of ${\bf\Omega}'$. It is easy to recognize in this expression the standard construction of irreducible representations of the group SU(2) using monomials [@Cornwell; @Greiner]. This yields the formula: $$|m_{H} (l-m)_{V}\rangle \mapsto e^{-il\alpha} \sum_{n=0}^{l}
D^{l/2}_{mn}({\bf\Omega}') |n_{H} (l-n)_{V}\rangle$$ where $D^{l/2}_{mn}({\bf\Omega}')$ are the elements of $(l+1)\times(l+1)$ matrices $\hat{D}^{l/2}({\bf\Omega}')$ which form the irreducible $(l+1)$-dimensional representation of the group SU(2). Thus the unitary transformation $\hat{U}({\bf\Omega})$ of an arbitrary quantum state in the Hilbert space ${\cal H}$ induced by the map given in Eq. (\[Eq:PolarizationTransformation\]) has the form: $$\label{Eq:UOmega}
\hat{U}({\bf\Omega}) = \bigoplus_{l=0}^{\infty} e^{-il\alpha} \hat{D}^{l/2}({\bf\Omega}').$$ Let us note that although the decomposition ${\bf\Omega}=e^{-i\alpha}{\bf\Omega}'$ is defined up to $-1$ multiplying both the factors, the product $e^{-il\alpha} \hat{D}^{l/2}({\bf\Omega}')$ does not depend on the specific choice of the decomposition. This follows from the fact that the elements of the matrix $\hat{D}^{l/2}({\bf\Omega}')$ are given by monomials of degree $l$ constructed from the elements of ${\bf\Omega}'$. Therefore each one of the substitutions $e^{-i\alpha} \rightarrow
- e^{-i\alpha}$ and ${\bf\Omega}' \rightarrow - {\bf\Omega}'$ will produce a factor $(-1)^{l}$, one multiplying $e^{-il\alpha}$ and another one multiplying $\hat{D}^{l/2}({\bf\Omega}')$, which will cancel each other. Consequently, the right hand side of Eq. (\[Eq:UOmega\]) is defined unambiguously as a function of ${\bf \Omega} \in \mbox{U(2)}$.
Collective depolarization {#Sec:CollectiveDepolatization}
=========================
We will now consider the scenario in which the polarization transformation ${\bf\Omega}$ is constant across $N$ uses of the channel. The entire Hilbert space in this case is given by an $N$-fold tensor product ${\cal H}^{\otimes N}$ of the two-mode space ${\cal H}$ analyzed in the previous section. The action of the collectively depolarizing channel on an arbitrary input quantum state $|\psi\rangle \in {\cal H}^{\otimes N}$ is given by: $$|\psi\rangle \mapsto
[\hat{U}({\bf\Omega})]^{\otimes N} |\psi\rangle.$$ Given the decomposition of $\hat{U}({\bf\Omega})$ derived in Eq. (\[Eq:UOmega\]), we can rewrite the tensor product $[\hat{U}({\bf\Omega})]^{\otimes N}$ as: $$\begin{aligned}
[\hat{U}({\bf\Omega})]^{\otimes N}
& = & \bigoplus_{l_1,\ldots,l_N \ge 0}
e^{-i(l_1+\ldots+l_N)\alpha}
\hat{D}^{l_1/2}({\bf\Omega}') \otimes
\ldots \otimes \hat{D}^{l_N/2}({\bf\Omega}')
\nonumber \\
\label{Eq:UotimesN}
& = & \bigoplus_{L=0}^{\infty} e^{-iL\alpha}
\bigoplus_{\substack{l_1,\ldots,l_N \ge 0\\
l_1+\ldots+l_N=L}}
\hat{D}^{l_1/2}({\bf\Omega}') \otimes
\ldots \otimes \hat{D}^{l_N/2}({\bf\Omega}') \end{aligned}$$ We see that the overall phase factor of the transformation ${\bf\Omega}$ enters the expression only with the total number of excitations $L=l_1+\ldots+l_N$ contained in all the $N$ uses of the channel. The $N$-fold tensor product of the SU(2) representations $\hat{D}^{l_1/2}({\bf\Omega}') \otimes
\ldots \otimes \hat{D}^{l_N/2}({\bf\Omega}')$ can in general be decomposed into direct sums by the iterative application of the formula [@BrinkSatchler]: $$\label{Eq:Dj1Dj2}
\hat{D}^{j_1}({\bf\Omega}') \otimes
\hat{D}^{j_2}({\bf\Omega}')
=
\hat{D}^{|j_1-j_2|}({\bf\Omega}') \oplus \hat{D}^{|j_1-j_2|+1}({\bf\Omega}')
\oplus \ldots \oplus \hat{D}^{j_1+j_2}({\bf\Omega}').$$ This formula allows one to convert all the tensor products into direct sums of irreducible representations, with any of the representations allowed to appear a number of times. Therefore we anticipate that the inner sum in Eq. (\[Eq:UotimesN\]) can be represented in the form: $$\label{Eq:DecompositionforfixedL}
\bigoplus_{\substack{l_1,\ldots,l_N \ge 0 \\
l_1+\ldots+l_N=L}}
\hat{D}^{l_1/2}({\bf\Omega}') \otimes
\ldots \otimes \hat{D}^{l_N/2}({\bf\Omega}')
= \bigoplus_{j} \hat{\mathbbm{1}}_{K^{j}_{NL}} \otimes \hat{D}^{j}({\bf\Omega}')$$ where $\hat{\mathbbm{1}}_K$ is the identity operator in a $K$-dimensional space $\mathbb{C}^{K}$ and the integer $K^{j}_{NL}$ tells us how many times the representation $j$ occurs in the decomposition of the $N$-fold product $[\hat{U}({\bf\Omega})]^{\otimes N}$ in the sector of the entire Hilbert space ${\cal H}^{\otimes N}$ that contains exactly $L$ excitations. The multipliticies $K^{j}_{NL}$ will define the capability of the system to protect quantum coherence against depolarization. By considering the parity of the indices $l_1,\ldots l_N$ when applying Eq. (\[Eq:Dj1Dj2\]) to multiple tensor products $\hat{D}^{l_1/2}({\bf\Omega}') \otimes
\ldots \otimes \hat{D}^{l_N/2}({\bf\Omega}')$, it is easy to observe that the representations which will appear in the decomposition in Eq. (\[Eq:DecompositionforfixedL\]) will be indexed with $j=0,1,\ldots, L/2$ for even $L$ and with $j=1/2,3/2,\ldots,L/2$ for odd $L$. Assuming for the time being that the multiplicities $K^{j}_{NL}$ are known, we thus arrive at the following decomposition of the polarization transformation in the entire Hilbert space ${\cal H}^{\otimes N}$: $$[\hat{U}({\bf\Omega})]^{\otimes N}
=
\bigoplus_{L=0}^{\infty} e^{-iL\alpha}
\bigoplus_{j=L/2-\lfloor L/2 \rfloor}^{L/2} \hat{\mathbbm{1}}_{K^{j}_{NL}} \otimes
\hat{D}^{j}({\bf \Omega}').$$
The above formula suggests the decomposition of the Hilbert space ${\cal H}^{\otimes N}$ into sectors with a fixed number of excitations ${\cal H}_{NL}$: $${\cal H}^{\otimes N}
=
\bigoplus_{L=0}^{\infty}
{\cal H}_{NL}$$ which in turn can be represented as isomorphic with the following structure: $$\label{Eq:HNLCC}
{\cal H}_{NL} \cong
\bigoplus_{j=L/2-\lfloor L/2 \rfloor}^{L/2}
\mathbb{C}^{K^{j}_{NL}} \otimes \mathbb{C}^{2j+1}.$$ The components of the state vector belonging to separate subspaces ${\cal H}_{NL}$ are multiplied by different phase factors $e^{-iL\alpha}$. If the parameters of the transformation ${\bf\Omega}$ are unknown, this implies the loss of quantum coherence between different sectors ${\cal H}_{NL}$. Within each sector, however, we have a number of subspaces for which the action of the collective depolarization operator is given by the identity $\hat{\mathbbm{1}}_{K^{j}_{NL}}$. This means that a quantum state which is encoded into a subspace isomorphic to one of the subspaces $\mathbb{C}^{K^{j}_{NL}}$ in the structure defined in Eq. (\[Eq:HNLCC\]) remains intact even if the transformation ${\bf\Omega}$ is completely unknown. We note that the coherence between the subspaces $\mathbb{C}^{K^{j}_{NL}}$ with different values of $j$ is destroyed as they are coupled to subsystems $\mathbb{C}^{2j+1}$ on which collective depolarization takes different forms. Consequently, the capacity of the system to protect quantum coherence is defined by the multiplicities ${K^{j}_{NL}}$, or to be precise, by the highest value over the permitted range of $j$.
Recursion formula {#Sec:RecursionFormula}
=================
We will now demonstrate that the multiplicities $K^{j}_{NL}$ are related via a simple recursion formula. As we have seen in the preceding section, when the total number $L$ of excitations is fixed, the overall phase of the polarization transformation ${\bf\Omega}$ is irrelevant. It is therefore sufficient to restrict ourselves to ${\bf\Omega} \in \mbox{SU(2)}$ and consider the decomposition of $[\hat{U}({\bf\Omega})]^{\otimes N}$ in the sector ${\cal H}_{NL}$ in the form: $$\label{Eq:UOmegaHNLExpansion}
\left. [\hat{U}({\bf\Omega})]^{\otimes N} \right|_{{\cal H}_{NL}}
=
\bigoplus_{j=L/2-\lfloor L/2 \rfloor}^{L/2} \hat{\mathbbm{1}}_{K^{j}_{NL}} \otimes
\hat{D}^{j}({\bf \Omega}).$$ The Hilbert space corresponding to $N$ uses of the channel has an obvious decomposition as ${\cal H}^{\otimes N} = {\cal H}^{\otimes (N-1)} \otimes {\cal H}$. If we now consider the sector of ${\cal H}^{\otimes N}$ containing exactly $L$ excitations, it can be constructed from the subspaces of ${\cal H}^{\otimes (N-1)}$ and ${\cal H}$ according to: $${\cal H}_{NL} = \bigoplus_{L'=0}^{L} {\cal H}_{N-1,L'}
\otimes {\cal H}^{(L-L')}$$ where ${\cal H}^{(L-L')}$ is a subspace of the two-mode Hilbert space introduced in Eq. (\[Eq:H=sumH(l)\]) that contains exactly $L-L'$ excitations of the field. This construction implies that the operator $\left. [\hat{U}({\bf\Omega})]^{\otimes N} \right|_{{\cal H}_{NL}}$ can be represented as: $$\left. [\hat{U}({\bf\Omega})]^{\otimes N} \right|_{{\cal H}_{NL}}
=
\bigoplus_{L'=0}^{L}
\left.[\hat{U}({\bf\Omega})]^{\otimes (N-1)} \right|_{{\cal H}_{N-1,L'}}
\otimes
\hat{D}^{(L-L')/2}({\bf \Omega})$$ We can now insert the decomposition of $\left.[\hat{U}({\bf\Omega})]^{\otimes (N-1)} \right|_{{\cal H}_{N-1,L'}}$ using an expansion analogous to Eq. (\[Eq:UOmegaHNLExpansion\]), assuming that the respective coefficients $K^{j'}_{N-1,L'}$ are known. This yields: $$\label{Eq:LostTheTrack}
\left. [\hat{U}({\bf\Omega})]^{\otimes N} \right|_{{\cal H}_{NL}}
=
\bigoplus_{L'=0}^{L}
\left(
\bigoplus_{j'=L'/2-\lfloor L'/2 \rfloor}^{L'/2} \hat{\mathbbm{1}}_{K^{j'}_{N-1,L'}} \otimes
\hat{D}^{j'}({\bf \Omega})
\right)
\otimes
\hat{D}^{(L-L')/2} ({\bf\Omega})$$ The tensor product $\hat{D}^{j'}({\bf \Omega}) \otimes \hat{D}^{(L-L')/2} ({\bf\Omega})$ appearing in the above expression can be decomposed using Eq. (\[Eq:Dj1Dj2\]) into a direct sum according to: $$\label{Eq:Dj'D(L-L')/2}
\hat{D}^{j'}({\bf \Omega}) \otimes \hat{D}^{(L-L')/2} ({\bf\Omega})
=
\bigoplus_{j=|j'-(L-L')/2|}^{j'+(L-L')/2} \hat{D}^{j} ({\bf\Omega})$$ Inserting Eq. (\[Eq:Dj’D(L-L’)/2\]) into Eq. (\[Eq:LostTheTrack\]) gives: $$\label{Eq:UOmegaHNLasN-1}
\left. [\hat{U}({\bf\Omega})]^{\otimes N} \right|_{{\cal H}_{NL}}
=
\bigoplus_{L'=0}^{L}
\bigoplus_{j'=L'/2-\lfloor L'/2 \rfloor}^{L'/2}
\bigoplus_{j=|j'-(L-L')/2|}^{j'+(L-L')/2}
\hat{\mathbbm{1}}_{K^{j'}_{N-1,L'}}
\otimes
\hat{D}^{j} ({\bf\Omega})$$ Comparing Eq. (\[Eq:UOmegaHNLExpansion\]) with the above expression allows us to relate the multiplicities $K^{j}_{NL}$, describing decoherence-free subsystems for $N$ uses of the channel, to the coefficients $K^{j'}_{N-1,L'}$ corresponding to the case when the number of channel uses is reduced by one. In order to find the explicit relation between $K^{j}_{NL}$ and $K^{j'}_{N-1,L'}$ we need to change the order of summations over $L'$, $j'$, and $j$ in Eq. (\[Eq:UOmegaHNLasN-1\]) such that the summation over $j$ is the outermost and its limits do not depend on other summation variables. Then the inner summations over $L'$ and $j'$ should yield a formula for $K^{j}_{NL}$.
The triple summation given in Eq. (\[Eq:UOmegaHNLasN-1\]) can be visualized by plotting in a three-dimensional space the set of points $(L',j',j)$ that are defined by the summation limits in Eq. (\[Eq:UOmegaHNLasN-1\]). The resulting grid is depicted in Fig. \[Fig:3DGrid\]. First, we note that the conditions $L'=0,1,\ldots,L$ and $j'= L'/2-\lfloor L'/2 \rfloor, \ldots, L'/2$ define a triangular grid for the pairs $(L',j')$. This two-dimensional grid is shown with the help of dark grey points in Fig. \[Fig:3DGrid\]. For every pair $(L',j')$ we have an allowed range of $j$s defined by the limits of the third sum in Eq. (\[Eq:UOmegaHNLasN-1\]). The upper limit for $j$ is always given by the plane specified by the equation $$\label{Eq:Maxj}
j = j'+(L-L')/2.$$ The lower limit for $j$ is specified by one of the two conditions: $$\label{Eq:Minj1}
j=(L-L')/2-j'$$ or $$\label{Eq:Minj2}
j=j'-(L-L')/2$$ whichever gives a higher value of $j$. It is easy to verify that Eq. (\[Eq:Minj1\]) is relevant when $j'\le (L-L')/2$, whereas Eq. (\[Eq:Minj2\]) gives the lower summation limit if $j'\ge (L-L')/2$. The three planes defined by Eqs. (\[Eq:Maxj\])–(\[Eq:Minj2\]) together with the fourth vertical plane specified by the condition $j'=L'/2$ form a tetrahedron which encloses the entire three-dimensional summation grid.
The structure depicted in Fig. \[Fig:3DGrid\] provides us with guidance on how to invert the order of summations in Eq. (\[Eq:UOmegaHNLasN-1\]), in order to make the summation limits of $j$ independent of other variables. Obviously, the summation over $j$ will run from $L/2-\lfloor L/2 \rfloor$ to $L/2$ in integral steps. In order to find the limits for $L'$ and $j'$ for a fixed $j$, we need to consider an intersection of the tetrahedron depicted in Fig. \[Fig:3DGrid\] with a horizontal plane corresponding to that value of $j$. This procedure is illustrated in Fig. \[Fig:Square\]. The intersection of the plane of constant $j$ with the planes specified in Eqs. (\[Eq:Maxj\])–(\[Eq:Minj2\]) gives three linear constraints on the values of $L'$ and $j'$ shown in Fig. \[Fig:Square\] with dashed lines. These constraints, together with $j'=L'/2$, define a rectangular region which, as it is easy to see, lies entirely within the triangular grid of pairs $(L',j')$. Because of the geometry of this region, it is convenient to define two new variables: $$\begin{aligned}
\mu & = & L'/2+j', \\
\nu & = & L'/2-j'.\end{aligned}$$ The limits for $\mu$ are given by Eqs. (\[Eq:Minj1\]) and (\[Eq:Minj2\]) and have the explicit form $L/2-j \le \mu \le L/2+j$, whereas the values of $\nu$ are bounded by $j'=L'/2$ and Eq. (\[Eq:Maxj\]) which combined together give $0 \le \nu \le L/2-j$. Both the variables $\mu$ and $\nu$ increase in unit steps, which follows from the form of the grid for the $(L',j')$ variables. Consequently, the summation over $L'$ and $j'$ for a specified $j$ can be given by a double sum over $\mu=L/2-j, L/2-j+1, \ldots , L/2+j$ and $\nu=0,1,\ldots,L/2-j$ with the old variables expressed as $L'=\mu+\nu$ and $j'=(\mu-\nu)/2$.
Using the new summation variables $\mu$ and $\nu$ gives us the following decomposition for $\left. [\hat{U}({\bf\Omega})]^{\otimes N} \right|_{{\cal H}_{NL}}$ with the inverted integration order: $$\label{Eq:Umunu}
\left. [\hat{U}({\bf\Omega})]^{\otimes N} \right|_{{\cal H}_{NL}}
=
\bigoplus_{j=L/2-\lfloor L/2 \rfloor}^{L/2}
\bigoplus_{\mu=L/2-j}^{L/2+j}
\bigoplus_{\nu=0}^{L/2-j}
\hat{\mathbbm{1}}_{K^{(\mu-\nu)/2}_{N-1,\mu+\nu}}
\otimes
\hat{D}^{j} ({\bf\Omega})$$ The two inner sums yield of course the identity operator acting in a larger space, defined by: $$\bigoplus_{\mu=L/2-j}^{L/2+j}
\bigoplus_{\nu=0}^{L/2-j}
\hat{\mathbbm{1}}_{K^{(\mu-\nu)/2}_{N-1,\mu+\nu}}
=
\hat{\mathbbm{1}}_{\sum_{\mu=L/2-j}^{L/2+j}
\sum_{\nu=0}^{L/2-j}
K^{(\mu-\nu)/2}_{N-1,\mu+\nu}}.$$ Inserting this equation into Eq. (\[Eq:Umunu\]) and comparing the resulting expression with Eq. (\[Eq:UOmegaHNLExpansion\]) gives a recursion formula for $K^{j}_{NL}$ in the form: $$\label{Eq:KRecursion}
K^{j}_{NL} =
\sum_{\mu=L/2-j}^{L/2+j}
\sum_{\nu=0}^{L/2-j}
K^{(\mu-\nu)/2}_{N-1,\mu+\nu}$$ which is the central result of this paper. In order to complete the recipe for calculating the multiplicities $K^{j}_{NL}$, we need to specify their values for $N=1$. This task however is straightforward. Let us recall that for $N=1$ we have ${\cal H}_{N=1,L}={\cal H}^{(L)}$ where the right-hand side has been defined in Eq. (\[Eq:H(l)=Span\]), and that for a polarization transformation ${\bf\Omega} \in \mbox{SU(2)}$ Eq. (\[Eq:UOmega\]) implies that $\left.\hat{U}({\bf\Omega})\right|_{{\cal H}_{N=1,L}} = \hat{D}^{L/2}({\bf\Omega})$. This means that for $N=1$: $$K^{j}_{N=1,L} = \delta_{L,2j}$$ which provides the initial condition to calculate $K^{j}_{NL}$ for an arbitrary $N$.
The recursion formula derived in Eq. (\[Eq:KRecursion\]) can be illustrated with a diagram shown in Fig. \[Fig:Recursion\] which provides a mnemonic recipe for carrying out calculations. In order to find $K^{j}_{NL}$ we need to add $K^{j'}_{N-1,L'}$ for all pairs of $(L',j')$ that lie within a rectangle which is rotated by $45^{\circ}$ with respect to the axes of the coordinate system, and has two vertices located at $(L,j)$ and $(L-2j,0)$, with the remaining two located on the line $j'=L'/2$. Fig. \[Fig:Recursion\] shows how to obtain this rectangle by starting from the point $(L,j)$ and going in two orthogonal directions that form $45^{\circ}$ with the axes, until reaching respectively the limits of $j'=L'/2$ and $j'=0$.
Conclusions {#Sec:Conclusions}
===========
We have derived a recursion formula for the dimensions $K^{j}_{NL}$ of decoherence-free subsystems in a bosonic channel experiencing collective depolarization described by the U(2) group. This model is relevant in quantum communication over single-mode optical fibers, for which collective depolarization is one of the dominant decoherence mechanisms. Although we have not been able to solve the recursion formula and obtain a closed analytical formula for $K^{j}_{NL}$, it can be easily implemented in numerical calculations.
The depolarization model considered here assumed that the phase relations between the consecutive uses of the channel are fixed. This requirement cannot be fulfilled if the communicating parties do not share a common phase reference to prepare and detect states. Then the phase factor $e^{-i\alpha}$ varies between the uses of the channel, and only the SU(2) transformation ${\bf\Omega}'$ remains constant. In order to implement decoherence-free encoding in such a scenario, the same fixed number of $l$ excitations must be transmitted in a sequence of channel uses. The structure of decoherence-free subsystems for $N$ uses of the channel is then given by a direct-sum decomposition of $[\hat{D}^{l/2}({\bf\Omega}')]^{\otimes N}$, discussed in [@HolbKribXXX04]. Alternatively, it is possible to devise schemes in which phases are self-referenced by employing multiport interferometers for state preparation and detection. Examples of such schemes have been described in [@WaltAbourPRL03; @BoilLaflPRL04]. Mathematically, this approach consists in introducing phase dependencies between multiple uses of the channel, which then give rise to non-trivial decoherence-free subsystems.
[*Acknowledgements.*]{} This work was supported by the UK Engineering and Physical Sciences Research Council and by Polish Committee for Scientific Research, Project No. PBZ KBN 043/P03/2001.
[20]{}
G. M. Palma, K. A. Suominen, and A. K. Ekert, Proc. R. Soc. London Ser. A [**452**]{}, 567 (1996).
L.-M. Duan and G.-C. Guo, Phys. Rev. Lett. [**79**]{}, 1953 (1997).
P. Zanardi and M. Rasetti, Phys. Rev. Lett. [**79**]{}, 3306 (1997).
D. A. Lidar, I. L. Chuang, and K. B. Whaley, Phys. Rev. Lett. [**81**]{}, 2594 (1998).
E. Knill, R. Laflamme, and L. Viola, Phys. Rev. Lett. [**84**]{}, 2525 (2000).
D. A. Lidar and K. B. Whaley, [*Decoherence-Free Subspaces and Subsystems*]{}, in Irreversible Quantum Dynamics, F. Benatti and R. Floreanini eds., Springer Lecture Notes in Physics vol. 622, Springer, Berlin 2003.
N. Gisin, G. Ribordy, W. Tittel, and H. Zbinden, Rev. Mod. Phys. [**74**]{}, 145 (2002).
J. L. Ball and K. Banaszek, arXiv.org e-print quant-ph/0410077.
S. D. Bartlett, T. Rudolph, and R. W. Spekkens, Phys. Rev. Lett. [**91**]{}, 027901 (2003).
Z. D. Walton, A. F. Abouraddy, A. V. Sergienko, B. E. A. Saleh, and M. C. Teich, Phys. Rev. Lett. [**91**]{}, 087901 (2003).
J.-C. Boileau, D. Gottesman, R. Laflamme, D. Poulin, and R. W. Spekkens, Phys. Rev. Lett. [**92**]{}, 017901 (2004).
J.-C. Boileau, R. Laflamme, M. Laforest, and C. R. Myers, Phys. Rev. Lett. [**93**]{}, 220501 (2004).
A. Cabello, Phys. Rev. Lett. [**91**]{}, 230403 (2003).
J. Ball, A. Dragan, and K. Banaszek, Phys. Rev. A [**69**]{}, 042324 (2004).
P. G. Kwiat, A. J. Berglund, J. B. Altepeter, and A. G. White, Science [**290**]{}, 498 (2000).
M. Bourennane, M. Eibl, S. Gaertner, C. Kurtsiefer, A. Cabello, and H. Weinfurter, Phys. Rev. Lett. [**92**]{}, 107901 (2004).
K. Banaszek, A. Dragan, W. Wasilewski, and C. Radzewicz, Phys. Rev. Lett. [**92**]{}, 257901 (2004).
J. F. Cornwell, [*Group Theory in Physics Vol. II*]{}, Academic Press, London 1984, pp. 444-445.
W. Greiner, [*Relativistic Quantum Mechanics: Wave Equations*]{}, 2nd ed., Springer, Berlin 1997, pp. 404-406.
D. M. Brink and G. R. Satchler, [*Angular Momentum*]{}, Clarendon, Oxford, 1963, Sec. 2.7.
J.A. Holbrook, D.W. Kribs, R. Laflamme, and D. Poulin, arXiv.org e-print math.OA/0402105.
| {
"pile_set_name": "ArXiv"
} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.